Most AI wrappers have no lasting moat.

Emmanuel Vallod, Partner and Head of Venture Research at Hivemind Capital, has spent 14 years teaching at Berkeley and evaluating frontier research.

We discuss:

• Why difficult-to-obtain data or deeply specialized workflows may be the only durable defenses for AI application companies.

• How Dark Matter Lab can provide researchers up to $1 million before incorporation through capital, compute, legal help, and operating resources.

• Why one research team faced a three-year wait for federal funding that private capital could eliminate immediately.

• How a missing training dataset can create a $500,000 bottleneck before meaningful AI experimentation even begins.

• Why choosing the wrong research direction can put a technical team as much as a decade behind.

“The unicorns or decacorns at a point in time were never the cool kids when they started.”

Where does an AI product stop being a wrapper and become genuinely defensible?

🎧 Watch, listen, and follow on your favorite platform: https://www.teamignite.vc/podcast
💬 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast
📝 Read the full episode breakdown on our blog: https://www.teamignite.vc/blog

AI did not fix outbound sales. It made the noise impossible to escape.

Taylor Offer is the founder and CEO of Atrios, an A16Z-backed platform inspired by referral agreements that once earned him well over six figures annually.

They discuss:

• How AI made spam at scale so cheap that qualified meetings became more valuable than messages.

• Why Atrios defines ICP, qualification questions, ACV, close rates, and price per meeting before activating tastemakers.

• How customers, investors, employees, and community leaders can become a company’s top-of-funnel.

• Why bad recommendations destroy the tastemaker’s own network, creating self-enforcing quality control.

• How A16Z Speedrun supplied seven of Atrios’ first 10 employees and helped Taylor adapt from bootstrapped growth to venture-backed building.

“AI has made it so easy to spam at scale.”

Would you trust a paid introduction if the person making it had real reputational downside?

Read the whole breakdown here: https://www.teamignite.vc/blog/ignite-startups-taylor-offer-on-the-relationship-intelligence-platform-changing-go-to-market-or

Five days later, OpenAI slowed the same model’s release because it couldn’t rule out critical cyber capabilities.

That’s the story underneath every other headline this week: the cost of producing a result is collapsing. The cost of deploying, defending, permitting, and verifying one is rising.

A few data points from the week:

SpaceX’s first public quarter showed revenue up 92%. The stock fell on $18.4B of capex, then absorbed a partial lockup release without flinching.

Airtable sold for a fraction of its 2021 peak valuation, with its founders carving out the AI business into a separate entity before signing.

Jeff Dean and three senior researchers left Google to build a company that aims to automate the scientific method itself.

The Ninth Circuit ruled that an AI agent shopping on your behalf is legally you, not the AI company, accessing the platform. First ruling of its kind, and a real unlock for agentic commerce.

Full breakdown in this week’s Last Week Ignite: what’s investable, what’s overhyped, and where the deployment side of AI is getting underpriced.

https://www.teamignite.vc/blog/last-week-ignite-august-9-2026-discovery-got-cheap-deployment-got-expensive

Waiting for the perfect international expansion window is how startups guarantee perfect competition.

Frank Lavin, a fellow at the USC Dornsife Center for the Political Future, served in the Reagan White House and as U.S. ambassador to Singapore.

We discuss:

• Why a $300 million Indonesia buildout can be irrational when first-year revenue may reach only $5 million.

• How the “smallest legitimate footprint” approach uses distributors, outsourcing, joint ventures, and local procurement to limit downside.

• Why founders should meet regulators before a crisis, not for the first time during an investigation.

• Why China welcomes foreign consumer brands but resists network platforms such as Uber, Google, and LinkedIn.

• Why Singapore functions as “Asia for beginners” for companies entering Southeast Asia.

“Don’t wait for a perfect moment or you’re going to face perfect competition.”

Should startups enter foreign markets before the economics look attractive, or wait until domestic growth slows?

View the full write up, watch, or listen here: https://www.teamignite.vc/podcast/ignite-political-economy-the-geopolitical-risks-founders-are-underpricing-with-frank-lavin-ep286-e3m

Microsoft booked a $3.2B gain on its Anthropic stake this quarter, close to two thirds of its headline beat. Amazon’s net income more than tripled to $62.6B, and the company says $53.4B of that was non-operating income primarily from Anthropic. Neither company files a 10-Q for Anthropic. Both just moved their own stock on its behalf.

That’s the throughline in this week’s Last Week Ignite. A few more findings:

Private AI exposure and public tech exposure stopped being diversified bets. Right now they’re largely the same bet wearing two tickers.

The open-weight price war inverted. Kimi K3, the strongest open-weight model available, now costs roughly 12x more per completed task than OpenAI’s cheapest closed model, and runs about 5x slower. Free weights stopped meaning cheap inference.

Anthropic’s own Frontier Red Team combed through 141,000 evaluation transcripts looking for a containment failure, and found three real breaches at three real companies. None of the three had noticed on their own.

Full breakdown on venture markets, compute economics, robotics, and what it means for founders, LPs, and VCs: https://www.teamignite.vc/blog/last-week-ignite-august-2-2026-everybody-s-earnings-ran-through-anthropic

Recursive self-improvement spent sixty years as a thought experiment. It now has a scoreboard in a product launch, filed between the cybersecurity results and the multimodal ones. The coverage went to the price cuts.

Companies benchmark what they intend to optimize.

I wrote up why I think the loop is already closed, and why most people are watching the wrong variable. Everyone is looking for self-improvement in capability. It arrived first in cost, and cost is the highest-leverage place for it to arrive, because power is the binding constraint now. Roughly 2,300 gigawatts sit in US interconnection queues with waits running four to seven years. Tokens per watt is the only term anyone can move, tokens per watt is software, and software is what these systems are best at.

The receipts, briefly. GPT-5.6 rewrote OpenAI’s production GPU kernels and designed, trained, and supervised the draft model that makes it generate faster. Google’s AlphaEvolve has been recovering 0.7 percent of Google’s worldwide compute for over a year and put a circuit change into an upcoming TPU. Anthropic says Claude authored more than 80 percent of the code merged into its codebase in May.

The strongest argument against all of this is in the piece too. METR published a study ten days ago finding that autonomous agents are still not economically competitive with humans on a frontier-style optimization problem. I take it seriously, and I explain why I think it bounds a narrower claim than the one I am making.

Also in there: what would change my mind, and which of it is measurable inside twelve months.

Full write up: https://www.teamignite.vc/blog/the-loop-closed-on-cost-first

Deo Arlo believes they are solving the wrong bottleneck.

An interceptor cannot stop what it cannot accurately detect, identify, and track.

On the latest episode of the Ignite Podcast, I spoke with Deo, founder and CEO of YC-backed Arlo Industries, about why the future of air defense may depend less on bigger weapons and more on better awareness.

Arlo is building a passive, decentralized sensor mesh designed to track small drones in real time. Its man-portable sensors connect and calibrate through their own network, creating a distributed view of the sky that can feed data into command systems and interceptor platforms.

We discussed:

• Why traditional radar struggles with small, low-flying drones
• How distributed sensors can produce compounding improvements in accuracy
• Why air defense may evolve to resemble a biological immune system
• What Ukraine teaches founders about speed, iteration, and battlefield deployment
• Why copying Ukraine’s exact solutions is the wrong lesson
• How autonomy could turn the future battlefield into an accountability black box
• Whether AI will scale peace, conflict, or simply human intent
• Why private drone defense may eventually become normal

One of Deo’s strongest insights:

“Information is power, and the more you know, the better you can strike, and maybe you can strike less.”

The visible part of air defense is the interception.

But the system that sees the threat first may be what ultimately determines the outcome.

Watch or listen to Episode 285: https://www.teamignite.vc/blog/ignite-startups-passive-sensors-drone-swarms-and-the-future-of-air-defense-with-deo-arlo-or

That incident is the thread running through this week’s Last Week Ignite: three different kinds of containment all turned into real underwriting questions in the same seven days.

Containing the agent: Congress had a draft AI kill switch bill within 72 hours of the disclosure.

Containing the cost: a router mixing an open model with a closed one hit 93% of frontier intelligence at a fraction of the price. Grok 4.5 now delivers near-frontier performance for $0.31 a task, next to $2.75 for the priciest model on the market. If your product’s only edge is access to the best model, that edge is disappearing fast.

Containing the physical footprint: a dozen states just filed data center permitting moratoriums, a federal AI power project got greenlit at Savannah River, and SpaceX hit a new all-time low as public markets started asking questions private markets have been avoiding.

None of these are isolated stories. All of them are things a founder, an LP, or a fellow investor needs to price in now, not in six months.

Full breakdown, section by section: https://www.teamignite.vc/blog/last-week-ignite-july-26-2026-the-week-containment-became-a-line-item

He was right.

So this summer I rebuilt teamignite.vc from scratch, and I didn’t hire a design agency or a team of developers. I built it by talking to AI.

No mockups. No tickets. Just conversation: “these logos don’t work on black,” “match the old team photo framing,” “build a contact form that emails us.” Review, react, repeat, for about a week and roughly 40 hours to get to launch.

What came out the other side: a real production site on Next.js and Vercel, with a live portfolio, our Volume Thesis simulator, working forms wired into our actual systems, and a jobs board that crawls our portfolio company’s careers page every night and posts open roles automatically. Zero manual upkeep saving us tens of thousands of dollars.

The part that surprised me most wasn’t the speed. It was how much taste and judgment still mattered. AI can build almost anything you describe. It can’t tell you whether the tone is too promotional for a regulated LP audience, or whether a logo that’s technically fine just looks wrong on a black background. That’s still the job.

Full writeup on how it came together (and what “vibe coding” actually looks like in practice) here: https://www.teamignite.vc/blog/my-advisor-called-our-website-ghetto-so-i-vibe-coded-a-new-one-with-ai

Take a look at teamignite.vc, and if something still looks ghetto, be gentle.

That gap between “it worked when we tested it” and “it works when a stranger calls in angry” is what Tarush Agarwal has spent the last two years trying to close. He’s the co-founder and CEO of Cekura.ai, a platform that simulates, tests, and evaluates voice agents for more than 200 companies, running millions of conversations through systems before they ever touch a real customer.

Before Cekura, Tarush was optimizing quant trading systems in London and Chicago down to seven or nine nanoseconds of latency. Cekura itself started as an internal tool: he and his co-founders were building a voice agent for personal injury law firms, and instead of shipping the product, they ended up spending three hours every night after dinner personally calling their own agent looking for failures. They automated that process, showed it to a few friends building similar things, and the reaction told them what the real business was. They pivoted during the first week of Y Combinator.

In this episode, Tarush gets specific about why testing voice agents through text alone is a mistake, why roughly half of Cekura’s customers still run GPT-4.1 instead of newer models, and a piece of research that should worry anyone shipping a customer-support agent: a system-prompt attack that succeeds in one turn about 20 percent of the time succeeds across a multi-turn conversation roughly 92 percent of the time. In one test, his team talked a major provider’s support agent into handing out a $150 discount code just by repeating a false claim over several turns.

He also lays out what “good” actually looks like across different domains (a healthcare receptionist and a customer-support bot are not measured the same way), why the newest model is often the wrong choice for a live phone call, and where he thinks this goes once voice is solved: chat, avatars, and eventually physical AI.

Read about the full conversation here

None of these facts explain each other on their own. Put together, they describe the week AI stopped being a story about scarcity and started being a story about glut, at the exact moment the people financing the glut started asking harder questions. That is the thread running through everything below.

The week in six numbers

• 2.8 trillion parameters: the size of Kimi K3, the open-weight model that just embarrassed the closed labs

• $510 billion: total global AI venture funding in the first half of 2026, already ahead of all of 2025

• 43 percent: the share of that funding that went to just two companies, OpenAI and Anthropic

• 41 percent: the decline in SpaceX’s share price from its post-IPO peak

• $188 billion: Databricks’ new valuation, up from $134 billion in February

• 6 of 50: the labs now above the frontier intelligence threshold, versus 2 in June

Venture markets and private capital

Start with the macro backdrop, because it explains everything else. Global AI venture funding hit $510 billion in the first half of 2026, already ahead of the $440 billion deployed in all of 2025. Nearly all of that growth is concentration, not breadth. AI’s share of total global venture capital crossed 70 percent in the second quarter, and OpenAI and Anthropic alone absorbed $217 billion of it, 43 percent of every AI dollar invested worldwide. OpenAI’s haul came as a single $122 billion round that pushed its valuation to $852 billion. Anthropic raised $30.6 billion in February and another $65 billion in May, the latter at a $965 billion valuation with participation from Amazon, Nvidia, and SoftBank among others. Anthropic is now generating roughly $2 billion a month in revenue, more than 40 percent of it enterprise, while still projecting a $14 billion loss for the year. That is the shape of the frontier lab business model in 2026: extraordinary top-line growth financed by extraordinary burn, underwritten by investors who have decided the alternative to funding it is worse than the loss itself.

Underneath that story, the early-stage market is quietly bifurcating in a way that matters more to Team Ignite than any single mega-round. Seed-stage dollar volume rose 30 percent year over year, but the number of seed deals actually closed fell 31 percent. Read that pair of numbers together and the picture is fewer, larger seed checks going to fewer companies, with everyone else getting squeezed out of the room entirely. Median Series A rounds have swelled to $14 million, up from an $8 to $10 million baseline just a year or two ago, and the median time between closing a seed round and closing a Series A has stretched to 20 months. A company built on the old assumption of a 12-month runway to the next raise is now flying without enough fuel for roughly eight of those twenty months. This is the single most important structural fact for anyone advising a seed-stage founder right now, and it deserves more attention than any individual funding headline this week.

The late-stage tape told two different stories depending on whether you were looking at private markets or public ones. Databricks signed a term sheet for a new strategic round at a $188 billion valuation led by Coatue, up sharply from $134 billion in February, bringing in roughly $3 billion in fresh capital to fund Unity AI Gateway (a multi-model governance layer), Genie (an interactive AI coworker), and Lakebase (a serverless Postgres database built for agentic workloads). Notably, this print landed well above where Databricks shares had recently traded on the Forge secondary marketplace, around $242 a share implying roughly $170.7 billion just days earlier, which tells you the primary market is now pricing Databricks ahead of the secondary market rather than behind it. The company has also ruled out a 2026 public listing.

Defense and dual-use AI kept pulling in serious capital. Helsing, the European defense AI company, closed a $1.8 billion Series E at an $18 billion valuation on July 13, with a genuinely broad syndicate behind it: Dragoneer, Lightspeed, Iconiq, Goldman Sachs Alternatives’ growth equity arm, JPMorgan, CPP Investments, General Catalyst, Plural, and Stepstone. The capital is earmarked for autonomous systems and drone integration for defense partners, and Helsing remains predominantly European-owned, which matters given the geopolitics of sovereign defense AI. Inference hardware kept its own bid too: Etched, maker of the Sohu transformer-specific inference chip, is reportedly in talks for a new round near a $20 billion valuation, up from roughly $5 billion previously, backed by $1 billion in contracted orders. And on the power side, Bloom Energy and Oaktree closed a $1.7 billion deal to finance fuel-cell power for Nebius’s AI infrastructure, one more data point in a week full of them that the constraint on AI buildout has shifted from chips to electrons.

A wider slice of the market kept moving too. Neko raised a $700 million Series C led by Lightspeed. PixVerse, the generative video company, closed a $439 million Series C extension backed by Alibaba. Chai Discovery raised $400 million for AI-driven drug discovery from Index Ventures, Kleiner Perkins, and Sequoia, a reminder that the FDA-gated categories TIV excludes from its own thesis are still attracting serious capital elsewhere. TerraFirma raised $115 million in a Series A led by Kleiner Perkins. Spectro Cloud closed a Series D above $100 million with Goldman Sachs Alternatives, AMD, and Ericsson. Senra Systems raised $65 million in a Series B from Lowercarbon Capital. Fora closed a $60 million Series D led by Forerunner Ventures, and Oak raised a $60 million seed round from Greylock, Accel, and CRV, an unusually large seed check that itself illustrates the bifurcation point above. Vendelux raised $50 million in a Series B, Valarian raised $50 million in a Series A led by NEA, and Monumental closed a $32 million Series B led by Khosla Ventures. India had its own moment: startups there raised $431 million across 23 deals in the third week of July, up from $107 million across 27 deals the prior week, led by Emergent’s $130 million Series C (Khosla, SoftBank, Lightspeed) and Udaan’s $160 million round combining fresh equity and debt.

Now the late-stage secondary book, where the most important divergence of the week actually lives. Anthropic secondary demand has become, in the words of Caplight CEO Javier Avalos, “the most sought-after company the venture secondary market has ever seen,” with indicated pricing up roughly 550 percent year over year to an implied valuation near $1.2 trillion against the $965 billion primary mark from May. Rainmaker Securities CEO Glen Anderson put the honest caveat on that number directly: “The demand outstrips the supply in Anthropic so much that it’s rare to get a trade done because no one’s selling.” That reluctance to sell makes a lot more sense in light of one more fact from this window: Anthropic reportedly filed a confidential S-1 draft back on June 1, targeting an October IPO. If that timeline holds, every current holder has a strong incentive to sit tight rather than sell into a secondary market, which is exactly the dynamic pushing the indicated price so far ahead of any trade that could actually clear it.

OpenAI’s secondary book told a cooler story by comparison. Forge pricing sat at $721.85 per share on July 18, down modestly from around $733 in early June, implying roughly $908 billion, a valuation drifting sideways to down at the same moment Anthropic’s is spiking. And then there is SpaceX, now the live test case for what happens when a trillion-dollar private darling actually goes public. Shares fell below the $135 IPO price for the first time on July 15, closing at $133, a 41 percent decline from the $225 peak reached shortly after listing. Elon Musk’s net worth dropped to a reported $861 billion on the move. For a secondary book, the honest read is that SpaceX’s public print is the first real data point on how much of the private AI-adjacent premium survives contact with public-market scrutiny, and so far the answer is not much.

Singularity signposts

Moonshot AI’s Kimi K3 became the first open-weight model to beat every closed rival on a real benchmark. Released July 16, Kimi K3 is a 2.8 trillion parameter mixture-of-experts model, meaning it activates only a small fraction of its total parameters (16 of 896 experts) for any given token, which is how a model this large stays affordable to run. It ships with a one-million-token context window and native multimodal input across text, image, and video, live immediately on kimi.com, in a coding tool called Kimi Code, and via API at $3 per million input tokens, $0.30 for cached input, and $15 per million output, flat across the entire context window. Per Alex Wissner-Gross’s Innermost Loop coverage, K3 took the top spot on the Frontend Code Arena and became the first open-weight model to beat every closed rival on SpreadsheetBench 2, a complex financial-tables benchmark, alongside top marks on SWE Marathon, Program Bench, and BrowseComp. One developer reportedly built a working Counter-Strike and Portal game clone for $3.24 in tokens, about a third of what the equivalent task costs on Claude Fable. Moonshot itself concedes K3 trails Claude Fable 5 and GPT-5.6 Sol on aggregate reasoning, and the full weights are not actually downloadable until July 27, so treat the benchmark sweep as vendor-reported until independent labs can reproduce it. What changed is the deployment bottleneck: near-frontier coding and reasoning at a fraction of closed-model pricing, on a path to full self-hosting, removes the last excuse for any application whose only real advantage was frontier API access. Watch the July 27 weight release and whether independent evaluators confirm the numbers.

Thinking Machines released Inkling and pointedly declined to chase the top spot. On July 15, Mira Murati’s (recall the former OpenAI CTO) lab shipped Inkling, a 975 billion parameter mixture-of-experts model with roughly 41 billion active parameters, trained on 45 trillion multimodal tokens, landing at 41 on Artificial Analysis’s Intelligence Index, the strongest US open-weights release to date. The strategically interesting part is what the lab said about it: Inkling is explicitly “not the strongest overall model available today, open or closed.” Thinking Machines is monetizing Tinker, its fine-tuning layer, rather than trying to out-benchmark the frontier. That is a genuinely different strategy than everyone else in this section, betting on customization and openness rather than supremacy, and it is worth watching whether Tinker’s fine-tuning revenue actually validates that bet.

Grok 4.5 posted real gains on expert-judgment work, not just coding leaderboards. Snorkel AI released independent evaluation results on July 16 using its GDPVal+ benchmark, a 2,000-task suite designed to test economically valuable professional work rather than toy problems. Grok 4.5 hit a 29 percent mean pass rate, ahead of GPT-5.5 at 22 percent and Claude Opus 4.8 at 21 percent, with the gap concentrated in domains that reward professional judgment: legal work at 40 percent versus a historical baseline of 27 to 28 percent, and healthcare at 35 percent versus 23 to 25 percent. If that gap holds up under further scrutiny, it is a meaningfully different kind of capability jump than a coding leaderboard win, because it points at automated agents actually handling complex, multi-step professional tasks that used to take a human days. That makes narrow, vertical professional agents and platforms (legal research, clinical documentation, and similar) more investable, and horizontal single-prompt tools more fragile, because the value here is clearly coming from sustained multi-step judgment, not from a bigger context window.

OpenAI taught a model to red-team itself. On July 15, OpenAI published research on GPT-Red, a safety and alignment model trained through competitive self-play that automates red-teaming, the practice of adversarially probing a model for vulnerabilities and alignment failures. The company reports meaningful reductions in vulnerability and alignment errors across six evaluation domains, replacing what has historically been a slow, expensive, manual auditing process. If this generalizes, it is a genuine phase change in how safety work scales: the bottleneck moves from headcount to compute. That is good news for anyone trying to ship models faster and slightly less good news for the human red-teaming and manual security-audit firms whose entire business model assumes this work has to be done by people.

The frontier stopped being a duopoly and became a crowd. Per Alex Wissner-Gross’s tracking, four major model launches landed inside eight days this window: Grok 4.5, GPT-5.6, Muse Spark 1.1, and Kimi K3. That run lifted six independent labs above 50 on the global intelligence index, up from just two in early June, with the top three scorers now separated by only three points. His daily Substack essays through the week framed this with a run of memorable titles (the frontier “open-sourcing itself,” a “trade dispute,” a “multiplayer game,” a “standing-room-only” frontier) that all circle the same idea: no single lab gets to hold the top spot for long anymore. The practical consequence is that betting on any one model as a durable moat is now a bet against the base rate. Multi-model orchestration and routing infrastructure, the plumbing that lets an application shop between models on cost and capability, is the more durable business to be in.

Foundation and open-source model watch

Kimi K3 is the whole story here in terms of capability, but the pricing detail deserves its own paragraph because it is the number that actually moves portfolio economics. At $3 input and $15 output per million tokens against Claude Fable’s roughly $50 output price, and a path to full self-hosting once weights land on July 27, K3 directly threatens any early-stage company monetizing thin interfaces over tabular data, spreadsheet automation, or basic code generation, because that exact capability is about to become available to any enterprise at a fraction of the cost, or free if they can run the weights themselves. License terms still matter enormously here: “open weights” is not the same as “open weights you can download and run today,” and until July 27 this is functionally still a hosted, API-gated product no matter what Moonshot calls it.

Nerdy Sidebar: what would it actually take to self-host this thing

Worth doing the arithmetic once, since “self-hostable” gets thrown around loosely. Moonshot hasn’t published K3’s exact active-parameter count, but its predecessor Kimi K2 ran 32 billion active parameters out of 1 trillion total, a 3.2 percent ratio. Applying that same ratio to K3’s 2.8 trillion total puts active parameters per token somewhere around 85 to 90 billion. That number matters for speed and compute cost. It does not matter for the memory floor, because a mixture-of-experts model cannot know in advance which of its 896 experts a given token will need, so every expert has to sit loaded in memory at all times regardless of batch size or how “sparse” the compute looks on paper.

That memory floor is the real constraint. The full weights run about 5.6 terabytes at FP16, 2.8 terabytes at FP8 (the precision most large-scale inference now runs at), or roughly 1.4 terabytes if you’re willing to quantize down to INT4 and accept some quality loss. To simply fit those weights, an Nvidia H100 (80GB of memory each) cluster needs about 35 GPUs at the bare minimum, and a real production setup with headroom for the 1-million-token context window’s memory overhead looks more like 40 to 48 GPUs, five or six 8-GPU nodes. The newer Blackwell-generation B200 (192GB each) cuts that to roughly 15 GPUs at the floor, 16 to 24 with headroom. The context window is the wild card here: KV cache, the memory that holds every prior token’s attention state, scales with context length, and at a million tokens per sequence, even a handful of concurrent long-context requests can rival the weight footprint itself, unless Moonshot is using aggressive cache compression under the hood.

On availability this week: AWS’s p5.48xlarge (8x H100) lists around $98 an hour on demand, roughly $12 per GPU-hour, putting a 40-GPU cluster near $490 an hour on-demand. Azure’s ND H100 v5 instances sit in a similar $10 to $13 per GPU-hour range. Google Cloud’s A3 (H100) instances are priced comparably, and its Blackwell-based instances are only just reaching general availability in limited regions, so B200 capacity is more often access-gated than openly listed right now. The neo-clouds are where the real arbitrage shows up: identical H100 hardware runs $1 to $7.50 per GPU-hour depending on commitment, which for a 40-GPU cluster is the difference between about $80 an hour and $490 an hour for the same chip.

Running the numbers on cost per token: a 40-GPU H100 cluster at a blended neo-cloud rate of roughly $2.50 per GPU-hour costs about $100 an hour to run. Autoregressive decoding is memory-bandwidth bound, not compute bound, so the honest ballpark for aggregate throughput across that cluster, with real batching, lands somewhere in the low thousands of tokens per second, call it 4,000 as a working number, with real-world benchmarks almost certainly moving that in either direction. At $100 an hour and 4,000 tokens a second, that cluster is producing roughly 14.4 million tokens an hour, which puts self-hosted cost around $6 to $8 per million output tokens. That’s actually cheaper than Moonshot’s own $15 API price, which sounds backwards until you remember that sticker price bakes in margin and the cost of running at a scale no single enterprise will match. The number that actually matters for a founder is the breakeven volume: that cluster costs roughly $72,000 a month whether or not anyone uses it, and at $15 per million tokens on the API, that fixed cost only pays for itself above about 4.8 billion output tokens a month. Below that, self-hosting is a worse deal than just calling the API, full stop.

And then, just for fun, everyone keeps talking about hosting these on MacBooks. Apple’s unified memory architecture is the one consumer-adjacent platform where memory is shared directly with the chip, which is why it briefly became the internet’s favorite way to run big open models at home. The math still doesn’t work on a laptop: even a maxed-out MacBook Pro tops out around 128GB of unified memory, so holding 2.8 terabytes of FP8 weights would take on the order of 22 of them, before touching compute or networking. The Mac Studio was the real answer, because until earlier this year Apple sold an M3 Ultra configuration with 512GB of unified memory, which would have gotten you to the full model with about six units, or three at INT4. Here’s the actual punchline: you can’t buy that configuration anymore. A DRAM shortage, driven in real part by the same AI buildout that makes people want to run models like this at home, pushed Apple to quietly kill the 512GB option earlier this year and has since squeezed the official lineup down to a maximum of 96GB. The machine you’d want for this experiment has been priced out of existence by the exact demand curve that produced the model you’re trying to run on it. Secondary market pricing makes the point loudly: a used 512GB M3 Ultra Mac Studio recently listed on eBay for $25,700, against an original price around $8,000 when Apple still sold it configured that way. Finding six of those secondhand, assuming you even could, would run north of $150,000, and you’d still be networking separate machines over Thunderbolt rather than NVLink, which means the resulting cluster would likely serve this model at single-digit tokens per second, a genuinely fun science project and nowhere close to a production stack. The honest takeaway: you can theoretically shop your way to enough memory to hold Kimi K3 on consumer Apple hardware, if you can find the now-discontinued configuration at all, but you’d be paying six figures for a machine that serves tokens slower than a five-dollar API call, which is as good an argument as any for renting the cloud GPUs instead of trying to own them.

The real-world data actually makes this cleaner than the K3 exercise.

Kimi K3 is the frontier flagship this week, but the actual “everyday workhorse” tier in Moonshot’s lineup is the Kimi K2 family (K2, K2.5, K2.6, K2.7), the prior generation that’s still what most people mean when they say “Kimi” in production. It’s 1 trillion total parameters with 32 billion active per token, built on 384 experts per layer (8 routed plus 1 shared), and it uses Multi-head Latent Attention, a technique (also used by DeepSeek) that compresses the memory needed for long context dramatically compared to a plain transformer. It also ships with native INT4 quantization built in, meaning the model was trained to run at 4-bit precision without the usual quality hit you’d take quantizing something down after the fact. That’s a Sonnet-tier analog: capable, cheap, and actually usable outside a datacenter, versus K3’s frontier-or-nothing positioning.

The memory math: 1 trillion params at native INT4 (half a byte each) comes out to almost exactly 500GB. That number is not a coincidence to notice: it’s within a hair of the 512GB M3 Ultra Mac Studio configuration Apple used to sell, meaning a single one of those machines could just barely hold the whole model, with almost nothing left over for context or overhead. Two of them networked together gets you real headroom.

The catch: that 512GB config is the same one Apple killed earlier this year over the DRAM shortage. So you’re back to the secondhand market, where those units are running around $25,700 each. Two of them for comfortable headroom is about $51,400 in one-time hardware, no ongoing cloud bill, power draw low enough to round to noise. Interestingly, buying six of the currently-sold 96GB units instead (roughly $5,500 each loaded) gets you to the same 500-576GB total for about $33,000, cheaper than two of the scarce 512GB units, though now you’re clustering six machines instead of two, which is a messier networking problem for basically no benefit given how little slack you’d have left over.

The active compute is the actual good news here. Because only 32 billion parameters are active per token (versus roughly 85-90 billion for K3), the amount of data that has to move through memory per token is far smaller. Working it through Apple’s unified memory bandwidth on the M3 Ultra (around 800GB/s), single-user generation lands in a genuinely usable range, likely somewhere in the teens to twenties of tokens per second, not the crawl K3 would produce on the same hardware. This is the one model in the exercise that actually makes sense to run on a Mac Studio.

Does it save money versus the API: here’s the honest answer, and it’s not close. Kimi K2’s official API pricing runs around $0.60 per million input tokens and $2.50 per million output tokens. At roughly 15 to 20 tokens a second running flat out, 24 hours a day, that Mac Studio pair could physically generate at most around 40 to 50 million tokens a month, ever, as a hard ceiling. At $2.50 per million output tokens, that entire monthly ceiling costs about $100 to $125 on the API. Against $51,400 of hardware, there is no volume at which this pays for itself, because the hardware literally cannot produce tokens fast enough to reach the breakeven point. Unlike the K3 exercise, where a real cloud GPU cluster crossed over into being cheaper than the API above about 4.8 billion tokens a month, the Mac Studio route for the cheap workhorse model never crosses over. The API isn’t just cheaper here, it’s cheaper at every volume the hardware could ever generate.

The honest use case for the Mac Studio setup isn’t cost savings at all. It’s data residency, offline access, or just wanting to poke at the weights yourself, not a financial argument. It’s fun to talk about, but really just probably use the API. Now back to our main programming.

A smaller but genuinely interesting item: on July 15, three days before Kimi’s weights are due, xAI open-sourced the complete Rust-based command-line harness for its Grok Build coding tool under the Apache 2.0 license, publishing all 844,530 lines of first-party code. The timing is not a coincidence. This landed roughly 72 hours after security researchers found the tool silently uploading users’ entire code repositories to Google Cloud servers without explicit consent. Open-sourcing the harness is trust repair, not a capability release: the default model alias still resolves to the closed, proprietary Grok 4.5, priced at $2 per million input tokens and $6 per million output. The lesson for diligence is a clean one: an open-source wrapper around a closed, metered model is not the same thing as an open model, and it does not eliminate vendor lock-in or the privacy risk that triggered this whole episode in the first place.

Platform power and incumbent moves

OpenAI made the boldest distribution move of the week. On Sunday, July 12, it launched ChatGPT Work, an integrated enterprise workspace that fuses ChatGPT, its Codex coding agent, and its Atlas browser into a single execution surface for a professional’s entire workday, spreadsheets, documents, slides, and code pipelines all in one place. This compresses the addressable market for a huge swath of horizontal SaaS productivity and document-automation tools that now compete directly with a bundled incumbent feature, while it simultaneously expands the market for the specialized security, logging, and governance tools that any enterprise running a fleet of these agents is going to need.

Anthropic took the opposite kind of distribution play, going long on relationships instead of bundling. On July 14, it launched Claude for Teachers, giving verified US K-12 educators free access to premium Claude capabilities, including its Code and Cowork agentic tools, through June 2027, alongside a Learning Commons connector that aligns responses to state standards across all fifty states and integrations with existing classroom tools like Canva Education, ASSISTments, and Illustrative Mathematics. The privacy posture is notably strict: FERPA-aligned, with no training on classroom conversations. This is a talent-pipeline play as much as a product launch, seeding familiarity with Claude among the exact demographic that becomes tomorrow’s technical workforce, and it directly compresses the market for standalone edtech tools built around lesson planning and differentiated instruction, which now have to compete with something free and backed by a frontier lab.

SoftBank locked down an entire national market for an agentic partner. On July 13, it announced an exclusive partnership with Sierra to bring agentic customer experience tools to Japan starting July 14, and the early results are striking: deployment on SoftBank’s LINEMO mobile brand lifted customer support resolution rates from 83 to 97 percent and satisfaction scores from 74 to 93 percent. An exclusive distribution deal like this is exactly how incumbents lock out independent customer-support AI startups trying to scale into a major market, before those startups even get a chance to compete on the merits.

Apple escalated its fight over the consumer AI hardware race in a serious way. On July 10, it filed a 41-page trade-secret lawsuit against OpenAI, alleging a systematic effort to poach more than 400 former Apple employees, with specific and pointed detail: prospective hires allegedly instructed to bring physical hardware prototypes to interviews, former iPhone design chief Tang Tan (a 24-year Apple veteran) named directly, and engineer Chang Liu accused of retaining an unreturned MacBook containing sensitive files and exploiting a software bug to download design and manufacturing documents. OpenAI pushed back on July 14, calling the complaint meritless. This litigation lands squarely on OpenAI’s hardware ambitions, which accelerated sharply after its acquisition of Jony Ive’s startup io, and the practical effect is to freeze OpenAI’s physical-device path for a while, which helps explain why the ChatGPT Work launch two days earlier leaned so hard into software bundling instead. Separately, Microsoft is reportedly building Project Perception, a security tool that routes vulnerability-scanning tasks across Microsoft, OpenAI, and Anthropic models to hold cost down, and Google delayed the broad rollout of Gemini 3.5 Pro after enterprise testing turned up failures, a reminder that even the best-capitalized incumbent cannot guarantee it ships on schedule. Nvidia also announced Cosmos 3 Edge, aimed squarely at robot inference workloads, and Apple Intelligence cleared a Chinese regulatory hurdle by integrating local models from Alibaba’s Qwen and Baidu, the price of admission for operating in that market.

Compute and inference economics

TSMC turned in the best quarter in its history and the market treated it as a warning sign. Second-quarter revenue hit $40.2 billion, up 36 percent year over year, with record margins across the board: 67.7 percent gross, 60.3 percent operating, and 55.6 percent net. High-performance computing revenue alone rose 20 percent sequentially and now accounts for 66 percent of total revenue. Wafer shipments are increasingly concentrated at the leading edge, with 3-nanometer process technology now 30 percent of revenue and 5-nanometer another 33 percent. TSMC raised its full-year capital spending guidance to $60 to 64 billion, up from $52 to 56 billion, and tacked on another $100 billion of US investment, bringing its total American commitment to $265 billion. And yet the stock fell somewhere between 5 and 7 percent in the sessions after the print, depending on which report you read, as the release of Kimi K3 triggered fears of what JPMorgan’s Andrew Tyler called a “DeepSeek 2.0 moment,” in his words adding “fuel to the fire” for a broader AI-chip selloff. Chinese rival Z.ai dropped almost 30 percent in Hong Kong trading and SoftBank fell 9 percent on the same news. The important tension to sit with: TSMC’s own guidance says 2-nanometer ramp will dilute its Q3 gross margin by 3 to 4 points even as CoWoS advanced packaging capacity, the bottleneck that determines how fast anyone can actually deploy new AI chips, is sold out through the end of 2026. Demand for compute is not slowing down. What is changing is the market’s willingness to treat unlimited compute demand as an unqualified positive, now that a model built for a fraction of the money just showed up near the top of the leaderboard.

Compute financing kept getting stranger and more diversified. Anthropic is reportedly in early talks to lease about $10 billion in Meta’s compute capacity over two years, a proposal it first floated back in June, structured with monthly payments and early termination rights for both sides. That is a fraction of the $45 billion, three-year deal Anthropic already has with SpaceX for access to more than 220,000 Nvidia GPUs at the Colossus 1 data center, but the smaller Meta deal reads as a genuine hedge: diversifying away from a single compute counterparty at the exact moment that counterparty’s stock is under public pressure. Meta, for its part, is building out a real commercial compute business, having hired 19-year AWS veteran Dave Brown to lead Meta Compute, backed by 2026 capital expenditure guidance of $125 to 145 billion, roughly double the $72 billion it spent in 2025 to acquire more than 1.3 million GPUs. Zoom out further and the four largest hyperscalers have now committed roughly $725 billion to 2026 infrastructure combined: Microsoft at $190 billion, Amazon at $200 billion, Google at $175 to 185 billion, and Meta at $125 to 145 billion. On the pricing side, GPU rental rates sit near multi-year lows, ranging from about $1 per GPU-hour on neo-cloud spot capacity up to $7.50 or more on the major hyperscalers, with specialized clouds running 50 to 75 percent cheaper than the big three for identical hardware, and the newest Nvidia B200 chips reportedly running inference at roughly $0.02 per million tokens versus about $0.14 on the older H100 generation. And Bloom Energy’s $1.7 billion fuel-cell financing deal with Oaktree for Nebius, mentioned above, is one more sign that power, not chips, is becoming the binding constraint on how fast any of this capacity can actually get switched on.

AI talent and compensation flows

The most consequential talent move of the window was actually a pair of departures, not an arrival. Noam Shazeer left Alphabet for OpenAI on June 18, and just one day later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for his work on AlphaFold, left Google DeepMind for Anthropic. Together, these two exits reportedly cost Google DeepMind roughly 6 percent of its market capitalization by June 23, a $22-per-share drop that closed the stock at $346.13. That is a striking amount of shareholder value to attach to two individuals leaving, and it says something about how thin the market believes the moat around any single lab’s research talent actually is right now.

OpenAI, meanwhile, promoted from within: Uday Ruddarraju, who joined as Head of Compute and Infrastructure in July 2025 immediately after leaving his role building xAI’s 250,000-GPU Colossus supercomputer, was named Chief Technology Officer for OpenAI’s Compute team. Infrastructure talent, not research talent, increasingly looks like the scarcest and most fought-over resource in this industry, which tracks with everything in the compute economics section above.

The talent war is reshaping real estate too. According to JLL property data, AI companies signed a record 565,000 square feet of office space in London in early 2026, with OpenAI taking 88,500 square feet in King’s Cross for more than 500 people and Anthropic securing space for 800 people in the Knowledge Quarter. That kind of physical footprint removes the geographic friction that used to give European deep-tech startups some protection from Silicon Valley poaching; the frontier labs are simply opening local offices and hiring the talent in place.

And the entry-level pipeline underneath all of this is visibly eroding. Across 150 enterprises studied by McKinsey, time spent on routine coding fell 46 percent, 84 percent of surveyed engineers now use AI coding assistants that write roughly 41 percent of all code, and yet security flaws in AI-assisted code rose 24 percent and net productivity gains shrink to just 10 percent once code review overhead is factored in. A separate Harvard study associates AI adoption with a 9 to 10 percent drop in junior developer employment within six quarters. Put together, that is a genuine structural risk to the pipeline that used to turn junior engineers into the senior architects and technical founders Team Ignite wants to back five years from now, and it deserves more attention from seed investors than it is currently getting.

Macro, regulation, and physical infrastructure

Microsoft cut roughly 4,800 roles, about 2.1 percent of its workforce, on July 13. That is one data point in a much bigger pattern: of 267 tracked layoff events across the industry so far in 2026, 150 of them, 56 percent, explicitly cited AI or automation in internal memos, affecting a combined 156,000 workers. Whatever the productivity debate looks like in the abstract, companies are citing automation as a stated reason for headcount reduction often enough now that it counts as a real trend, not an outlier.

Power and siting keep hardening into political constraints on AI buildout, not just engineering ones. New York’s governor signed an executive order pausing state environmental permits for any new data center at 50 megawatts or larger for up to a year, citing residential electricity prices that have climbed nearly 68 percent since 2019. Australia is drafting legislation that would go further still, requiring data center operators to generate their own power independently rather than draw on the public grid, under a new Office of AI. The message from two different governments in the same window is consistent: the era of assuming a data center can plug into the existing grid without political friction is over.

Export policy generated its own friction. Following the placement of new export restrictions on OpenAI’s GPT-5.6 and on Anthropic’s models, a Trump administration official argued the labs had brought this on themselves, telling reporters, in effect, that you cannot warn everyone your product might pose an existential risk and then expect the government to stay out of it. Whatever one makes of that framing, the practical effect is real: export controls now threaten to fragment the availability of frontier models internationally at the exact moment 141 nations are separately layering on their own data sovereignty requirements, which is a genuinely difficult two-front problem for any lab trying to sell the same model everywhere.

On the physical-AI side, Hyundai moved to buy out SoftBank’s roughly 10 percent stake in Boston Dynamics for about $325 million, taking the robotics maker fully in-house at an implied valuation near $3.3 billion, with its Atlas humanoid robot targeted for factory deployment starting in 2028. SoftBank exiting a marquee robotics asset to redeploy capital into AI compute, while a strategic industrial operator with actual factories takes full control, is exactly the kind of vertical integration our in-thesis interest in physical AI is watching closely. And SpaceX had its own physical hiccup: Starship Flight 13 aborted its launch attempt seconds before liftoff on July 16, with Musk indicating a relaunch attempt was likely within the week. That is a small thing on its own, but it lands during the same week SpaceX’s stock is under real pressure, and it does nothing to help sentiment.

Cross-stack interaction effects

Kimi K3’s price collapse meets TSMC’s sold-out packaging capacity. An open-weight model just proved it can match or beat closed rivals on real benchmarks, at a fraction of the cost, with a path to full self-hosting in eight days. But actually running a 2.8 trillion parameter model privately requires serious leading-edge hardware, and TSMC’s advanced CoWoS packaging, the bottleneck that determines how fast anyone can turn wafers into deployed chips, is sold out through the end of 2026, while the 2-nanometer ramp is diluting TSMC’s own margins by 3 to 4 points. The result is a real gap between what is technically possible and what is actually deployable: only well-capitalized players will be able to self-host these open weights at scale anytime soon. That makes model compression and localized inference chips more investable, and it makes thin B2B wrappers built purely on cheap spreadsheet or coding automation more fragile than the headline pricing alone would suggest, because the near-term reality is still gated by physical packaging capacity, not by model availability. This is probably the single most underpriced dynamic in the market right now, with a medium-term time horizon of a few months.

ChatGPT Work’s software bundling meets Apple’s hardware lawsuit. OpenAI’s ambition to own the consumer AI hardware layer, accelerated by its acquisition of Jony Ive’s io, just ran into a serious legal wall via Apple’s trade-secret suit. Denied a clean physical-device path for now, OpenAI is doubling down on owning the software workspace instead, and the timing of ChatGPT Work’s launch just two days before that lawsuit became public makes the sequencing hard to ignore. The market looks like it is overpricing the near-term odds of an independent consumer AI hardware category emerging cleanly, while underpricing how quickly a software-only OpenAI can consolidate horizontal productivity tooling into a single workspace. Runtime governance and multi-app integration layers get more investable here; single-purpose productivity SaaS gets more fragile, immediately.

Anthropic’s Meta compute lease meets SpaceX’s stock slide. Anthropic’s compute strategy has been anchored by its $45 billion, three-year SpaceX deal, and now it is reportedly negotiating a smaller, more flexible $10 billion lease with Meta at the same moment SpaceX’s public stock is sliding well below its IPO price. That reads less like a coincidence and more like prudent diversification away from a single compute counterparty whose public valuation is suddenly under real scrutiny. Hyperscaler-neutral, multi-provider compute orchestration becomes more investable here, and any startup whose entire infrastructure is bound to one provider’s data center becomes more fragile, with a structural time horizon that will play out over the life of these contracts.

What this means for founders

More attractive now: multi-model orchestration and routing infrastructure, the plumbing that lets an application shop across a genuinely crowded field of frontier models on cost and capability. Inference cost observability and budget-control tooling, now that usage growth rather than unit price is the real margin risk. Vertical, judgment-heavy professional agents in legal and healthcare workflows, where Grok 4.5’s GDPVal+ results suggest real multi-step capability rather than a leaderboard trick. Physical AI and robotics platforms that own both the hardware and the deployment surface, the exact structure validated by Hyundai’s Boston Dynamics buyout. And power-constrained compute plays, on-site generation, efficient inference, anything that routes around the siting fights now playing out in New York and Australia.

Less attractive now: thin wrappers whose entire value proposition is frontier-model access, which Kimi K3 and Inkling are actively pricing toward zero. Standalone lesson-planning and differentiation edtech tools, now competing directly against Anthropic’s free K-12 offering. Basic terminal coding tools without a genuine architectural or data advantage, undercut both by open-sourced CLI harnesses and by the sheer commoditization of coding capability generally. And any horizontal productivity SaaS tool that just got bundled, functionally, into ChatGPT Work.

Overhyped but worth watching: any vendor-claimed benchmark leadership before independent reproduction, Kimi K3’s July 27 weight release being the test case everyone should watch. And the idea that a self-hosted multi-trillion-parameter open model is imminently practical for most companies; it is not, until packaging capacity actually loosens up.

Underpriced or under-discussed: the seed-to-Series A runway math, a 20-month median gap against a 12-month runway assumption is a genuine crisis hiding in plain sight for a lot of companies right now. Fine-tuning infrastructure as a standalone business, validated by Thinking Machines’ explicit strategic choice to monetize Tinker rather than chase the top of the leaderboard. And workforce and apprenticeship tooling that tracks the junior-to-senior engineering pipeline, given the McKinsey and Harvard data on junior developer employment.

Questions worth putting to founders this week: If your seed round assumed a 12-month runway to Series A, how are you restructuring burn against a 20-month median timeline instead? If Kimi K3’s weights reproduce cleanly on July 27, what part of your cost advantage survives contact with a self-hostable frontier-class model? What does your enterprise data architecture actually require given that 141 nations now enforce their own data sovereignty rules? And if you are anywhere near coding tools or education, how are you differentiating against products that are either free or bundled by a frontier lab?

Secondary-market watch list: Anthropic, where demand so far outstrips supply that trades barely clear, made sharper by a reported confidential S-1 targeting an October IPO. SpaceX, now trading below its IPO price and offering the first real public test of whether the trillion-dollar AI-adjacent premium survives public scrutiny. Databricks, freshly marked at $188 billion by a real term sheet and explicitly staying private through 2026. OpenAI, cooling modestly on secondary desks even as it wages a two-front battle against Apple in court and against Anthropic for research talent. And Sierra, whose exclusive Japan partnership with SoftBank just posted a 97 percent resolution rate that any enterprise-agent investor should be studying closely.

What to monitor over the next one to four weeks: the July 27 Kimi K3 weight release and whether independent labs reproduce its benchmark claims. SpaceX’s price action into its next earnings print, the clearest live signal on late-stage AI exit assumptions. The July 28 to 29 FOMC meeting, where markets currently expect a hold rather than a cut despite June’s surprisingly soft inflation print. Whether the Anthropic-Meta compute lease actually closes. And whether any other state follows New York’s lead on data center permitting.

What this means for Team Ignite LPs

Team Ignite’s core positioning, small early-stage checks into infrastructure-adjacent, cost-aware, model-portable companies, is well matched to a market where frontier capability is commoditizing from the top down faster than almost anyone expected a year ago. The open-weight wave is genuinely good news for most of the portfolio, because it lowers the cost of building for every company that is not itself trying to be a frontier lab, which describes nearly everything Team Ignite backs.

On the secondary book, the honest message for LP communication this quarter has two parts that pull in opposite directions and both deserve airtime. First, the SpaceX print is a real and current caution: the assumption baked into a lot of 2026 late-stage marks, that a smooth public exit awaits at or above the last private valuation, is being tested in real time and is not passing cleanly so far. Second, Anthropic’s secondary demand remains genuinely extraordinary, and a reported confidential S-1 targeting October gives that demand a concrete catalyst rather than just momentum, but the $1.2 trillion figure should be communicated to LPs as demand intensity, not a price anyone could actually transact at today. On portfolio construction, the seed-to-Series A runway math argues for more aggressive reserve discipline at the earliest stage: a company that used to comfortably bridge on an old burn profile can run out of room faster now that real usage, not just headcount, is what drives the burn.

What this means for VCs

The market looks mispriced in two directions simultaneously, which is the actually interesting part. On the private side, a meaningful slice of the application layer is still being underwritten as though frontier-model access is a durable moat, days before an open-weight model may make comparable capability available to self-host for free. That looks like a short. On the public side, the entire AI-chip and infrastructure complex just sold off hard on an open-weight release even as the dominant foundry in the world posted record demand and sold-out packaging capacity, which suggests sentiment, not fundamentals, is doing most of the repricing, and that creates real entry points for investors willing to hold through the noise.

The portfolio-construction implication is a genuine barbell. Fund the primitives that profit directly from commoditization, routing, orchestration, portability, and inference-cost control, and fund the workflow-owners and physical-world systems that commoditization structurally cannot reach, defense, robotics, regulated verticals with real data moats. Avoid the compressible middle, the horizontal wrapper that only ever had frontier-model access as its edge. And watch the seed-to-Series A gap as closely as any single funding headline this week, because a 20-month median timeline against a 12-month runway assumption is quietly reshaping which companies even survive to the next round, regardless of how good the underlying product is.

This newsletter is for general informational purposes only and does not constitute investment, legal, tax, or accounting advice, nor an offer or solicitation to buy or sell any security or investment product. Investing involves substantial risk, including possible loss of principal, and past performance is not indicative of future results. Full disclaimer: teamignite.vc/disclaimer

That sounds like an obvious advantage. But according to Misha Makara, co-founder and CTO of Rally AI, faster development introduces a new problem: when companies can build almost anything, deciding what not to build becomes more important than engineering speed.

Misha has spent much of his career entering difficult technical situations. He has worked across cybersecurity, enterprise systems, digital securities, technical due diligence, and startup turnarounds. Venture firms have brought him into portfolio companies when products were failing, teams were struggling, or capital was running out.

Subscribe now

Rally AI presents a different challenge. Rather than rescuing a company in crisis, Misha is helping scale a product with more than 200 prospective customers waiting to get access.

His experience offers several lessons for founders navigating AI development, product strategy, technical debt, startup pivots, and public relations.

Treat Product Development Like an Experiment

Misha’s approach to company building was shaped by one of his earliest mentors, Brad Paxton, who worked on Kodak’s pioneering digital-camera technology.

Paxton taught him that an engineer could also operate effectively as a CEO. Rather than treating business decisions as intuition-driven bets, founders can approach them through experimental design.

Every meaningful product experiment should have:

• A clear hypothesis

• A defined scope

• A measurable outcome

• A limit on the time or capital invested

• A predetermined decision based on the result

The point is not simply to “fail fast.” That phrase is too vague to be useful.

Founders must decide in advance what failure looks like. Otherwise, a company can continue investing in an unsuccessful idea indefinitely, interpreting every disappointing result as a reason to add another feature, extend another deadline, or raise more capital.

A disciplined experiment creates boundaries. A team might give an engineer one afternoon to test whether a technical approach is viable. If the experiment produces the required result, the company invests further. If it fails, the team records what it learned and moves on.

Without those limits, experimentation becomes wandering.

More Features Rarely Rescue a Failing Product

One of the most common patterns Misha has seen in struggling startups is the belief that additional features will solve weak market traction.

He described a company that had raised approximately $20 million to build a government-focused alternative to DocuSign. After spending roughly $17.5 million, the business still lacked meaningful traction.

The company continued adding features. But the core issue was not product breadth. It was compliance.

The platform was not FedRAMP compliant—a major obstacle for a company selling software to the federal government.

The turnaround strategy was counterintuitive: build less.

Rather than continuing to expand the full product, the team focused on creating a compliant version with roughly 10% of the original functionality. That certification mattered more to customers than the long list of additional features.

The lesson is straightforward: founders should identify the one constraint preventing adoption before adding more functionality.

When a product is not selling, the answer is rarely “more product.” It may instead be:

• A missing certification

• Poor positioning

• An unclear target customer

• A broken distribution strategy

• Weak onboarding

• A lack of trust

• A problem customers do not consider urgent

Building more features before identifying the real constraint simply burns runway faster.

AI Has Shifted the Software Bottleneck

Traditional software development required lengthy planning cycles. Teams conducted customer interviews, created mock-ups, validated designs, wrote requirements, and then moved into development.

Shipping a new production release every two weeks was once considered a strong operating cadence.

AI development tools have radically accelerated that process. Rally AI can sometimes deploy software nine times in a single day.

The bottleneck is no longer necessarily engineering capacity. It is product discipline.

Because features are cheaper to produce, teams can easily build things simply because they are technically possible. This creates several risks:

• Product roadmaps become reactive

• Teams chase new AI capabilities without customer demand

• Technical debt accumulates rapidly

• Customers receive unnecessary complexity

• Engineering activity is mistaken for business progress

AI does not eliminate the need for customer discovery, product strategy, or thoughtful architecture. It makes those disciplines more important.

A company that moves quickly in the wrong direction does not gain an advantage. It simply reaches the wrong destination faster.

Not All Technical Debt Is Bad

Technical debt is often discussed as though it should always be eliminated. Misha argues that the more useful distinction is between productive and unproductive debt.

He shared the example of a client that wanted a sophisticated free-shipping system. The request sounded simple but quickly expanded into a complicated set of rules involving eligible products, bundles, promotions, and administrative controls.

The development team built extensive infrastructure to avoid creating technical debt. But the customer barely used the feature.

The company spent more building the system than it would have spent manually covering the shipping costs.

That is the wrong kind of engineering investment: technically clean, but commercially unnecessary.

Misha compares technical debt to debris left on a highway after a minor accident. Some debris can remain near the side of the road without stopping traffic. It becomes a priority only when it blocks the direction the company needs to travel.

Founders should therefore ask:

• Is this debt slowing customer growth?

• Does it introduce serious security or reliability risk?

• Will it prevent an upcoming product change?

• Is fixing it more valuable than acquiring customers?

• Can the process temporarily be handled manually?

Early-stage companies should not build permanent infrastructure for problems they have not yet proven exist.

Sometimes paying someone to complete a task manually is smarter than spending months automating it.

Rally AI Wants to Become a Company’s First PR Hire

Rally AI uses artificial intelligence to help companies build relationships with journalists, influencers, and media outlets.

The platform analyzes large volumes of information to understand:

• What journalists are writing about

• Which topics are gaining momentum

• How stories move across media channels

• What performs well with specific audiences

• How competing companies position themselves

• Which company narratives may interest particular reporters

Rally then matches a company’s goals, news, and positioning with the media contacts most likely to care about the story.

Misha describes the system as a three-sided marketplace involving companies, media contacts, and the broader cultural or news environment.

This matters because public relations is not simply mass email outreach. It is relationship building.

A journalist does not care that a company wants coverage. The journalist cares whether the story is relevant, timely, credible, and likely to interest an audience.

Rally’s goal is to understand both sides of that equation before the company begins pitching.

The platform has already helped early-stage companies secure coverage from major publications, including The Wall Street Journal.

AI Cannot Fix an Undefined Company Story

Rally has reportedly experienced no customer churn so far, partly because the company has been selective about who it allows onto the platform.

The companies that perform best tend to understand:

• Who they serve

• What problem they solve

• Why their approach is different

• What message they want the market to remember

PR becomes much harder when a startup cannot clearly explain what it is.

AI can improve research, targeting, timing, and communication. It cannot create genuine product-market fit or manufacture a coherent identity for a company that does not have one.

Before seeking media coverage, founders should be able to answer three questions in plain language:

1. What does the company do?

2. Why does it matter now?

3. Why is this company uniquely positioned to solve the problem?

If those answers remain unclear, the company probably has a positioning problem—not a PR problem.

First-Mover Advantage Is Overrated

Misha is skeptical when founders describe first-mover advantage as a durable competitive moat.

History frequently rewards the second or later entrant that learns from the original innovator’s mistakes.

Kodak created the digital camera but failed to dominate digital photography. Microsoft has repeatedly entered markets after competitors and then used distribution, product integration, and iteration to win. Apple has often succeeded by refining existing product categories rather than inventing them.

The first company into a market bears the cost of educating customers, discovering failure modes, and testing business models.

Later entrants can study those lessons and build a stronger product.

A better question for founders is not, “Are we first?”

It is, “What have earlier companies misunderstood, and why are we positioned to execute differently?”

Disruption Usually Starts With One Narrow Advantage

Founders frequently describe their products as disruptive, but genuine disruptive technologies are rarely superior in every dimension at the beginning.

They are often worse in most ways while being meaningfully better in one specific area.

Early flash storage illustrates this pattern. Initial flash drives had less capacity and cost more per unit than traditional hard drives. Their critical advantage was speed.

Over time, the technology improved while preserving that advantage. Eventually, solid-state storage became commercially viable across a much broader market.

Startups should identify the one dimension where they are significantly better than incumbents.

That advantage could be:

• Speed

• Ease of deployment

• Accessibility

• Accuracy

• Workflow integration

• Security

• A new distribution model

• Performance in a narrow use case

Trying to outperform established companies on every dimension from day one is usually unrealistic.

Fractional CTOs Should Not Be Fundraising Props

Fractional CTOs can be useful when a company loses technical leadership, needs to stabilize a platform, or requires specialized expertise during a transition.

But Misha strongly warns against hiring a fractional CTO primarily to help raise capital.

Investors may question why the technical leader presented during fundraising is not committed to operating the company over the long term. That creates uncertainty around technical ownership, recruiting, intellectual property, and execution after the round closes.

A fractional CTO should solve a real operational problem.

The role can work when the mandate is clear: stop the crisis, assess the architecture, rebuild the team, transfer knowledge, and recruit a permanent replacement.

It should not be used to manufacture credibility during investor meetings.

The Central Lesson: Build Less, Learn Faster

The common thread across Misha’s experience is disciplined decision-making.

AI allows teams to build faster, but speed does not replace judgment. Founders still need to understand their customers, define their experiments, identify their most important constraints, and stop investing when evidence no longer supports the original thesis.

The strongest teams are not those that produce the most features.

They are the ones that know which features matter, which shortcuts are acceptable, which technical problems can wait, and when the company must change direction.

In Misha’s words: “Try to find a way to do less. Do the things that actually matter.”

That may be the most valuable product strategy in the age of AI.

👂🎧 Watch, listen, and follow on your favorite platform: https://tr.ee/S2ayrbx_fL

🙏 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast

Chapters:
00:01 - Introducing Misha Makara and Rally AI
00:41 - Cybersecurity Roots and Kodak’s Digital Camera Legacy
02:26 - Experimental Design for Founders and Engineers
04:10 - From Wayfair and Dell to Startup Turnarounds
05:57 - Apollo Cameras and Black-and-White Barns
07:31 - Lessons From High-Volume Venture Portfolios
09:35 - Risk Cycles, Late-Stage Liquidity, and Early-Stage Investing
14:46 - Why More Features Fail to Save Startups
17:37 - Founder Coachability and the Ability to Pivot
19:50 - Bundling Risk From Hyperscalers and AI Platforms
22:27 - Why First-Mover Advantage Is Overrated
24:20 - AI Development Speed and the New Product Bottleneck
27:07 - Good Technical Debt vs. Bad Technical Debt
30:21 - Rally AI as a Startup’s First PR Hire
32:33 - Who Rally AI Works Best For
33:47 - Measuring PR Success and Building Media Relationships
35:11 - Rally AI’s Long-Term Vision
36:12 - The Three-Sided PR Marketplace
37:41 - Second-Mover Advantage and Disruptive Technology
41:25 - High-Signal Interviewing and Hiring Engineers
43:20 - When to Hire a Fractional CTO

Transcript

Brian Bell (00:01:11.092): Hey, everyone. Welcome back to the Ignite Podcast. Today, we’re delighted to have Misha Makara on the program. He’s a co-founder and CTO of Rally AI. Before Rally, he spent 15 plus years—as a venture firms, they would call him when a startup is on fire technically—and he spent over five years at T-Zero’s SEC regulated digital securities platform. Pretty exciting. And now he’s working on Rally, which is a system that uses AI to accelerate PR. So we’re really excited to have you guys on the program. Thanks, Misha. Yes, my pleasure. What—love to start with your origin story. What’s your background? Sure.

Misha Makara (00:01:43.634): So I studied cyber security, and one of my first bosses—I went to the Rochester Institute of Technology—my first boss was Brad Paxton, who’s the guy that was in charge of the team at Kodak that created the digital camera. So, fascinating guy to have as your first boss, first mentor. I remember him pulling me into his office and going, oh, me sure, this thing just got declassified. We can talk about the cameras that we were using in Apollo and all this. It was great. Wow. So I was working for him as a as a sysadmin, doing work on truly large-scale systems. We were working on data capture systems for the census, and he saw me and one of my friends, as I was in the undergraduate program, and said, you know, you should really consider starting a business. I had really never thought of myself as a business person. I’m an engineer. He said, no, no, no, these are the same things. So one of the the most interesting things that he kind of taught me was, you know, you have this one mindset of a CEO being, you know, the Elon Musk, the Steve Jobs. But a CEO can also be an engineer and look at the business through the lens of experimental design. And that’s one of the the things I I really like to get into with you today, about how do you both set an engineering culture, and how do you design and iterate?

Brian Bell (00:02:53.669): Through product ideation. I love the Kodak story, right? Because they had the digital camera, right? And the executives—this is famous in business schools, right? Like, I have a business degree undergrads. I’m the opposite. Like, I’m not an engineer. You know, business degree and MBA. And, like, that Kodak case is, like, told over and over again in schools. You know, they literally had the digital camera, and they just, like, shelved it because it would eat their revenue, right? Like, they were afraid...

Misha Makara (00:03:19.675): To cannibalize their own cash cow. Exactly. And and Brad was the person that was literally told, you hide that in a desk drawer where no one will find. And it’s weird to think, in another universe, Kodak could have been the Google of our time. One of the things that—I don’t know if you can see it—he gave me on the way out, as when I when I finally finished the story. But yeah, he encouraged me to start a business, a web application development company, and I ended up selling it. But he used to have, right back there, that’s the page from Alice in Wonderland where Alice comes to the fork in the road and meets the Cheshire Cat. And Alice says, you know, hey, which road should I go down? And the Cheshire Cat says, well, where are you trying to get to? I don’t care so much. Then it doesn’t really matter what road you go down. And I love that. I always keep that up there because it’s okay to do pure research, and it’s okay to do applied research. But again, as a facilitator, you have to define and set your limits about how much are you willing to experiment. And it’s also okay to tell your people, we’re gonna try this, but here’s the sandbox that we’re gonna play in. And if we—we’ll either get this result, in which case we’ll go in this direction, or this other result, in this other direction, or we’ll get no result, and we’re going to have to go retry the experiment. And I think that there’s a lot of that missing in software development as a whole. And especially now, where the costs of development are so much lower with AI, there’s a lot that can be done in that, again, experimental design approach to ideation and product development. From there, I created my company, sold it, said, you know, I’d really like to work for a real company, ended up working for Wayfair.

Brian Bell (00:04:49.731): In Boston. Did—corporate agencies are kind of like halfway companies, right?

Misha Makara (00:04:53.983): Yeah, no, no one’s ever heard of that. Yeah. And then got recruited out of Wayfair by Dell and put in charge of the managed security services product line at Dell. And that was another interesting story because their—Dell loves to hire military folks, and they very much like their silos. And I created a way to get those silos to talk to each other. That got attention from the guy that signs the paychecks, and next thing I knew, I was asked to standardize that process and roll it out. So enterprise project program governance is another important topic. And then, from there, I bumped into Gangels when I was living in New York. It actually started with an argument with someone that turned out to be Paul—Paul Grossinger, who’s, uh, one of the the founders of Gangels. And he asked me to start doing technical due diligence on some of their portfolio companies. And then, inevitably, some of those companies went sideways, and they said, hey, we like this Misha fella. Is he available to help us out? And I started doing more and more. And then I’m just one of the resources that they call in whenever something goes sideways or catches fire. And that’s where I’ve been up until Rally. And Rally is really a great opportunity because I—I’m not dealing with something on fire for once. I’m dealing with something that—we have this huge backlog, 200 plus customers that are waiting to sign up. And I’ve never had that kind of a problem before, where people are bashing down the door to get in. And we’re trying very hard to do it in a way that doesn’t decrease quality. So I’m very much about the not in shittifying the the product. Yeah, results.

Brian Bell (00:06:20.352): Well, I want to get to Rally, but there’s a few story threads I want to tug on first. Absolutely. One thing that just kind of circ—circling in my mind right now is the Apollo story. Yeah. You know, what were the cameras...

Misha Makara (00:06:31.416): On Apollo that they declassified? So yeah. So he pulled me into his office one day, and he was saying, so I got to tell you this story. There was this time that Kodak—we, the team and I—we went out into the the middle of the Midwest, and we had to knock on people’s doors and say, hey, we can’t tell you who we are or where we’re from, but we just want to paint your barn black and white. You don’t get to ask any questions, but we just need to paint your—paint your board. Pros...

Brian Bell (00:06:55.906): And cons. You get a fresh coat of paint. You get a fresh coat of paint. We’re going...

Misha Makara (00:06:58.828): To give you ten thousand dollars and the—no questions. And he was like, yeah, there’s still some barns that have to be painted black and white. And do you have a guess on what those were for? No. Satellite imagery, or—yeah, you need to focus those cameras, right? So you need to have a large spot in order to be able to focus them and to calibrate them. So that’s just kind of like—you, like, calibrate...

Brian Bell (00:07:19.518): A printer, almost. Yeah, exactly. Wow. Okay, so they’re out in the Midwest painting these. There’s some farmer...

Misha Makara (00:07:25.760): You know, has a barn that was painted black and white, and now you get to know what it’s for. And were these digital cameras? And no, these—these were analog cameras. Another project that he worked on later was, when you’re in the spacecraft and you want to do photos of the moon, everything’s moving. So how do you get the film to be in synchronous with what you’re taking a photo of, so that way it doesn’t end up blurred? And go figure, a lot of those cameras and optics were used for things before the—before moon photography. The next thing I wanted...

Brian Bell (00:07:52.392): To cover is your experience helping Gangels. Gangels is one of the the highest volume shops around. Lorenzo was on the podcast, I don’t know, last year, and, you know, we kind of talked about kind of the high volume strategy, pros and cons. And we’re both, you know, in that school of thought. We’re both in the high volume school. They’re even higher volume than us. We do about a hundred a year. I think they do two or three hundred years, something like that. Now, what are some common patterns you you notice with successful and failing startups and all that? Oh, that’s a good question. Yeah, yeah. And by...

Misha Makara (00:08:25.957): The way, I’m going to say, do as I say, not as I do, because, you know, my personal investments there have been a mixed bag. I personally like to see the—again, the founders that are willing to try something, but have—they’re creating a product for an audience that they can actually define and quantify. I’m also interested to see how quickly are they willing to learn and pivot. I think that that’s an important factor in in the success. If you asked three years ago if we were going to be in kind of this AI world, I don’t think we would be able to predict it as it were. And I keep thinking about the early days of dot com, and if you were asked to pick the winners of dot com, what you would have said at that particular point in time versus who were actually the winners at the end. So I don’t have any good good advice on that. I will say it’s interesting to look at the overall market forces, and especially in election years, what happens there. I don’t know if you’ve kind of experienced that, but Gangels—yeah, there’s there’s interesting and different market forces. So investment team tends to switch from early stage and seed stuff to to later, more mature things in those years. Well, tell me—tell me more about that. I think I know what you mean, but—yeah, no, it’s just as simple as that, because Gangels is individual investors, usually invest in SPD retail, a lot...

Brian Bell (00:09:43.329): Of SPDs. They have some funds. They do a lot...

Misha Makara (00:09:45.651): Of early stage, precedence seed. Yeah, yeah, yeah. But, you know, are you willing to go, you know, putting your money on a very early stage company for, you know, an hopefully exponential return, versus putting something into, you know, a Series C...

Brian Bell (00:09:58.901): Or D, kind of. Yeah, so you’re—you’re talking about this, like, cycle of risk that happens every, you know, seven-ish years, where, you know, we all go risk on. And risk on includes, like, going really early stage. All of a sudden, early stage lights up. And what happens first—I’ve noticed this as well, and I think I’ve talked about it on the pod. I’ve been on vacation. This is my first podcast after—after Fourth of July vacation, so I’m a little little rusty, uh, shaking the cobwebs off of my my brain here. And right now, I’d say, yeah, we’re very risk on in the late stage, pre-IPO. Yeah, like, that market is very hot right now. The private market, these eight billion, ten billion, hundred billion, trillion dollar companies. But as liquidity flows through the system, um, and LPs start getting checks and and shares from SpaceX, and then I—I guess Anthropic will be next, and Open AI—I forget the order right now—but as as all that money comes into the system, all of a sudden everybody’s going to go risk on, right? Because we’ll have a bunch of liquidity, and we’ll have a bunch of capital going, wow, that 20-year bet on SpaceX really paid off. You know, you know, I was in that fund. You know, if you’re in a fund that holds SpaceX from the C to A, you’re—that LP is just, like, licking his chop chops right now. It’s just, oh man, like, that—that big—well, you know, just that position will probably—you know, there’s a lot of fun sitting out there with a huge TBPI, total value to paid in capital, that’s about to be, like, huge DPI from that—from that one deal. And it’s gonna—it’s gonna make a bunch of LPs go risk on again in the early stage.

Misha Makara (00:11:28.813): I think you’re—you’re right. I’m personally interested to see what happens after the lockup period, and especially since you have a fund, you have—you have an asset that’s now going public and will immediately become an index, right? And then you also have that because of the the SEC rule changes, and you’ve also got people that are leaving the lockup period exactly the same time. And you also have a prospectus that has some very interesting things in there, saying, you know, yeah, part of our evaluation is based on the idea that we think that we can build this, uh, you know, power distribution facilities in space in order to support our data centers. And we think that that line of business is going to be worth this, but we haven’t quite figured out market viability about it yet. You know, so I’m interested to see what’s going to happen. It’s strange times, and strange times for—for that, on—on top of rule changes. You know, it’s going to be very interesting to watch.

Brian Bell (00:12:20.650): Yeah. What are some of those? Do you recall any of the rule changes? I’d love...

Misha Makara (00:12:23.952): To—yeah. One of them was just the the lockup period of about the re—and don’t hold me to this exactly—but there’s a requirement that used to be, I want to say, several months before. It’s—yeah, it’s typically...

Brian Bell (00:12:35.419): A six-month lockup. Yeah, it’s been changed to...

Misha Makara (00:12:38.354): To a stage lockup. Yeah, yeah. I thought it was something as low as 15 days now as part of the the new change, which is why it just all of a sudden will become an index fund, because of how large those companies—yeah, that—well, that’s...

Brian Bell (00:12:52.782): The other thing is, like, indexes like the Nasdaq Triple Q have to buy the shares because—because of the market cap, because it’s part of, you know, the Nasdaq 100, the 100 largest market cap companies in the index. And I think that happened this week, actually. They—they had to buy eight billion dollars worth of—no matter what, I just have to buy it—which props up the stock, right? Because it’s like—but then also you have, I—I saw the news of the Big Short guy, yeah, is shorting the stock, right? So—and then you have, like, Kathy Wood from ARC buying the stock, right? And then you have the index. So it’s like you have all these buying...

Misha Makara (00:13:28.409): And selling forces out there. If—if you were sitting on SpaceX for 10 years, and you’re 15, 20, and you’re all of a sudden are now looking at the the current price, you know, and you’ve got a mortgage to pay, I don’t—you know, you know how this goes. And we’ve seen the pattern previously.

Brian Bell (00:13:47.526): Yeah. Typically, IPOs don’t pop like they did, you know. I think I wrote an article on this, actually, on the margin. It’s like 50-50, basically. Okay. About 50 will be, like, above par, and 50 will—50 will be below par. And then you—you have the famous famous example of, like, Facebook, right? Where they just kind of just wallowed for a couple years, and everybody’s like, oh, Facebook’s dad. This sucks. And then, of course, it just went stratospheric after that. So, interesting. Went to space. Yeah. What—yeah. And so, what what are some of these lessons as you kind of dug in? And this is an interesting technical—it’s like almost like a technical post due diligence. These startups are blailing about, and you’re kind of...

Misha Makara (00:14:26.274): Yeah, shoot it in to help them. One of the the biggest problems that I’ve seen is, you know, especially when I’m parachuted in, the goal is to—to save the company. And it’s oftentimes very difficult to pivot because the the research and the hypothesis about the market has oftentimes shifted. So, for example, I’m—but I’m thinking about one that was basically a DocuSign competitor for government. And same thing, they—they got 20 million dollars. There were 17 and a half through it, but they—they never really could get the traction. It really wasn’t quite working, and they were continuing to invest in their existing product line, right? More features. The things that’s going to save us is more features. That—that’s—that’s what’s happening. But the problem in that particular case was they weren’t Fed Ramp compliant, which is a problem if you’re trying to be a DocuSign competitor for government. So they were trying to solve it by just saying, hey, let’s throw more features at this. And the reason they didn’t go there previously was, you know, cost and uncertainty. And in that particular case, it was quite a sell to the business to say, we’re going to do less. We’re going to make a product that is compliant. That’s only 10 of the feature set of what the full product is, but it’s going to have that check mark. And through that, that’ll be the the thing. And sure enough, that was, when—when push came to shove at, you know, the end of the killer feature, that—that was the thing that they needed. And so I think a lot about Pareto and the 80-20 rule. Yeah. One of my friends always said, you know, try to find a way to do less. Do the things that actually matter, and the things that—that show. And that’s really what we’re trying to do when we’re thinking about a turnaround. What—what is—what, first of all, what did we learn? You know, you spent this money to do something. You learned something, right? And there’s oftentimes little nuggets of IP that weren’t necessarily the things that you thought were going to be the business, but can be. And so one of my kind of default maneuvers is: what are the integrations that we built? What is the data that we have? Can we monetize any of this? As kind of a general emergency, last-minute maneuver, we can usually sell an API. There’s patterns to do that. We can spin that up pretty easy. And then it’s a decision about where and how are we splitting our effort, especially when there’s very little gas left in the tank.

Brian Bell (00:16:42.077): Yeah, and there’s famous examples of that. I mean, Slack was, you know, an internal app at yet another gaming company that he tried—tried to get off the ground, right? Exactly. And same with Flicker before that. I think it, you know, it was another game, and they had a photo sharing app inside the game, and they pivoted to photo sharing. The same founder, right? Same—same team. Yeah, yeah, yeah. But again, that’s...

Misha Makara (00:17:04.669): Through listening and and figuring it out. But if you show up and they go, no, no, no, we’re a gaming company here. We’re—we’re not doing that. That’s—that’s not what I signed up for, right? That’s a difficult conversation...

Brian Bell (00:17:17.167): To have. What have you noticed about, you know, the founders who—who were able to pivot, mm-hmm, and cut their losses, versus, I guess, VCs would describe them as coachable, right? That’s the word that VCs use. Yeah, that was exactly...

Misha Makara (00:17:32.559): Where my head was going. So there’s a certain—it’s difficult when you’re doing the diligence because you’re trying both to assess coachability, but also you want them to have enough of a product thesis and direction to be able to go. You know, this—you can’t be there coaching all the time. That’s not your job. You have enough companies in your portfolio. But you also want someone that’s going to be, as I like to say, appropriately opinionated, right? You want someone that you can do what I—what I call sandcastle crushing. You can jump in and say, okay, I’m not exactly sure about this, but, let you know, if this were to happen, what—what would we do? You know, if all of a sudden Open AI were to decide that this was going to be a focus, or Uber—Uber is a great example—you know, just massive company, they’ve decided that they now want to eat your lunch. What—what are you going to do? How are you going to handle that? Amazon, you know, diapers.com, right? There’s so many things that—that you can do. And again, not—not being distracted, or having the hubris to think that you can outspend one of those companies. They will win. So how do you operate in—in that particular environment?

Brian Bell (00:18:38.546): Yeah, I call it the bundling risk, right? I was just talking to a founder about this yesterday. I was like, hey, this is really cool. You have a very cool model for memory inside of an organization, but how are you going to defend that against the hyperscalers, who already have all that data, one, and then they can just bundle it into their offering? Yeah, exactly, right? And it’s, like, core to how they operate. It’s, like, literally core to how those foundational models work, is their memory and state and all of...

Misha Makara (00:19:05.242): That. Yeah, yeah. And that’s—that’s another larger topic about the use of AI now in development. And it’s been interesting. It’s interesting to get a front row seat to how these things have developed, right? So prompt engineering, as—I’m curious to get your thoughts—prompt engineering has not disappeared, but it certainly has been suppressed and pushed into a corner. And if you’ve mastered that skill, then you do better. If you understand how orchestration works, then you do better. And I—I’m now watching a bunch of friends that are using Chat GPT specifically as Google. One asked yesterday. He was asking kind of a appointed legal question, and I said, yeah, it’s—it’s not going to do well at that. He’s like, well, it gave me an answer. I was like, yeah, but let’s—let’s actually look at the the actual regulation. And I think we got onto this topic of how AI is different and how it’s...

Brian Bell (00:19:56.517): The same, right? I think one of the threads I wanted to tug on, which I thought was great, was, you know, you—you think about the dot-com. Yes. Right? And, you know, some of the winners we thought, like AOL and Yahoo. AOL just sold, I saw. Anyway, you know, you—you would have said, oh yeah, like, why even work on a search engine? Like, Yahoo’s got that dialed, and maybe AltaVista or Lycos or Excite or—or whatever. And it turned out to be Google. And I—I wonder the same thing with the—with the models, right? It’s like, oh, it’s so clear Open AI is going to win. Or—but now it’s Anthropic, right? Well, but...

Misha Makara (00:20:27.663): But again, this is the—I’m going to say—the failure to study the history of technology. And I—I love—there’s a couple things. I’ll give you a couple secrets if you’re pitching to me and you’re asking me to do diligence. The first one is, you know, we—they say, oh, we have first mover advantage. First movers don’t usually take over the market. If you don’t believe me, look at Uber, that bought their software from someone else. And, as we all know, we’re currently using Kodak cameras for everything, right? It’s usually the second person to the field that can take advantage of the first company’s learnings to do it better a second time.

Brian Bell (00:21:00.468): Look at Microsoft. They’ve been second to market on almost everything they’ve ever done. Yeah, exactly. Like, literally second to market...

Misha Makara (00:21:06.952): In every product they have. Every—every single product. Even Windows. And I’m going to say something controversial: there’s an argument that Apple is doing the same thing, especially before the M1, right? Yeah. Right? When you were comparing, you know, one of your older iPhones to an Android way back when, the Android had better specs, but Apple did better on...

Brian Bell (00:21:25.263): The software. I’m still Android, so I’m—yeah, I am too. But yeah. Well, you’re—you’re a nerd like me, so, like, I want to be able to look in my device. It bothers me that I can’t see what’s on that other partition. Yeah, yeah, exactly. What do you—I can’t get...

Misha Makara (00:21:37.581): To my files. What are you talking about? Crazy. I own this device. I downloaded the file. Like, just show me where the file is. You know, if you want something really terrifying, go—go look at the actual build, and you can see the packages that are getting installed, and the companies that make those packages. You’ll go—go to their website, and it’s like, keylogger for government. Love that. Thank you. Thank you, guys. Nice. Yeah.

Brian Bell (00:21:59.778): As somebody with boots on the ground working in a startup, I mean, how does this feel different now with AI, right? I get...

Misha Makara (00:22:06.622): That question all the time, but yeah, yeah. First of all, the the velocity of software development has changed. You know, it used to be that—I was chatting with one of my my friends and colleagues about this recently—you know, you really had to get your your product specs, and you’d have to do viability testing, right? So think about your your traditional SDLC. You’d sit down with the designer. What’s an SDLC? For—sorry, sorry. Software development lifecycle. My apologies. So think about how you used to make software, right? And again, we can go back and we can talk about waterfall to agile to the new agentic development, you know, all of that. But the method, if you were being a skilled practitioner, was always the same, right? You would sit down. You would do interviews. From those interviews, you would create a design and a mock-up. If you were smart, you would take that mock-up, you would put it in front of customers, you get feedback from those customers, then you—you would build your requirements. Rinse and repeat. And then, from there, we would go off into building. And you were happy if you could turn out a build every two weeks. That was considered to be the gold standard of a successful software development lifecycle: new version of product every two. In the world of AI now, we can go so much faster. And part of the problem now is, because it’s so easy to push things out, it’s very easy to become distracted with new features and doing things just because we can rather than because we should, or because they create value for customers. Before we even get into the the cyber security side of things, it’s just a basic maintaining of discipline. And one of the things we talked about earlier is—is intentional and deliberate experimental design in the product. So it’s, as I said, it’s very easy to roll out a new feature now. It’s a lot more difficult to sit and listen, and to be strategic, and to make sure that you aren’t accumulating massive amounts of technical debt inside of the platform, and the debt that you’re accumulating is the right type of debt.

Brian Bell (00:23:57.520): Right. Yeah. And this is—I think this is a key distinction. I remember, yeah, I was a PM for a really long time, for probably eight, nine years of my career. And, you know, the the joke was, like, our—our quarterly roadmap became our four-year roadmap, right? And now it’s probably the opposite. Your, you know, quarterly roadmaps becoming your, like, I don’t know, your sprint-level roadmap, because you can really accelerate. And what I think that probably does is it removes the excuse of engineering’s bottleneck. Now, if engineering’s not the bottleneck anymore, the bottleneck really, like I think is what you’re saying, which is it’s understanding the customer and avoiding bad technical debt, right? Yeah. And—and I’d love to get into the topic of the right kind of debt versus the wrong kind of—yeah.

Misha Makara (00:24:36.218): Yeah. What do you mean by that? Yeah, yeah. So this is actually a lesson from—from Wayfair. I’m gonna go through a personal experience. One of the first projects I ever worked on at my company, we had this customer wanted the ability to do free shipping. Sounds easy. Well, certain items get free shipping, but not others. Okay, so certain items, but if they’re bundled, maybe we can do this, right? You know, it sounds like, yeah, there’s...

Brian Bell (00:24:58.636): This whole decision tree of free shipping, right?

Misha Makara (00:25:00.960): Right, right. And so then, you know, we were kind of like, well, debt is bad, so let’s, you know, let’s handle that. Like, we’ll—we’ll build an interface for this. You know, the—you know, the admin of the the site needs the ability to specify these things, and to do this and do that, and we have to build this logic, and that—meetings with the customer, and over and over. And then, ultimately, how many things were sent with free shipping, right? How many things could have just been done with a—with a promo code, right? There was a lot of infrastructure that was built that was never used and never seen, but certainly cost a lot of real money. So I would argue that that would be the bad kind of debt, right? We ended up building something to try to remove debt, but we ended up adding code that never really served a value. The customer never used it. And, in fact, the money spent on building the feature would have been better to just handle more free shipping. The—the analogy that I like to think about when I’m—what I’m teaching this is debris on the highway. So if you think about it, after an accident, there is, you know, there’s little bits of car everywhere. And, you know, if you just give it a little bit of time, sometimes—like a—like a pile up, right? Yeah, yeah, yeah. But, you know, yeah, let’s go minor accident, right?

Brian Bell (00:26:13.298): But you got—you, yeah, you got the the—the crashed cars, the wrecked cars, off the highway. And now there’s, like, debris left. Yeah.

Misha Makara (00:26:19.941): Right. But it moves its way into ruts, right? And so it’s only really a problem if your—if you decide to change lanes exactly right there. And, in a business context, it might be okay if you change lanes, you know, a little bit later, a little bit earlier, in order to avoid that debt. And that debt might be okay, right? The traffic is still flowing. And naturally, as you give it more time, you know, the debris moves its way off of the side of the road, right? Where it moves itself into places that are—are less offensive. Now, don’t get me wrong. That’s not an excuse. You still have to repave the road every 10 years, which I encourage every platform to do. If you don’t believe me, look at what happened when we tried to do the Vista transition, right? You had to bring people out of retirement because they didn’t know how to do TCPIP. That’s a problem. That’s a—we failed to repave the road every 10 years, just to know inventory and what was actually down there. But there is something to be said for: what is the right kind of debt? Can you get away with, you know, just paying for that free shipping, paying someone to click the button, and paying for it that way, rather than paying engineering costs versus building all of the infrastructure to support it when you don’t know if you really need...

Brian Bell (00:27:27.384): Yeah, I love that. Really good analogy. So let’s talk about Rally. What is—what is Rally?

Misha Makara (00:27:31.991): And why are you working on Rally? Yeah. So Rally is your company’s first PR hire. So Rally, what it does as a platform is, we—again, experimental design style—we absorb tons and tons of signals, and the AI actually goes and figures out its own signals: what to absorb, how to prioritize them, how to weight them. And we build a contact database of all of these different journalists, influencers, people speaking in the space. We also build graphs and knowledge maps about what are the different topics that are trending, how are they trending, where did they start from, how do they split and divide. It’s way more complicated than I realized when I started it. And then, on top of that, we have the information that comes in from one of our clients about what are they doing, what are they trying to do, what are they trying to communicate to the broader universe. And we do matchmaking. So we match the company and our client and their goals with different media influencers and what they’re doing, and we try to create press. And one of the things that was a little bit tricky for me to understand initially was PR is not marketing. This is about building relationships with media outlets to get them to talk about you and your founders and what you as...

Brian Bell (00:28:39.956): An organization are doing.

Misha Makara (00:28:41.313): So we, in order to do that, we need to make sure that we understand not just who are these media contacts and what are they writing about, but what’s interesting to them. What are the kinds of things that get traction with their audience, and what are they looking—you know, what—what’s the story? What’s going to get them? And so doing that matchmaking takes massive amounts of data. It takes a lot of iterative cycles. It varies based on the person. It varies based on the time. It varies based on what’s trending in the news. And we keep creating and experimenting over and over again. The AI just does these iterative cycles in order to figure out what’s optimal, and it keeps learning as a result. And the outcome from that is we’re able to get early and seed stage companies press, like Wall Street Journal. And it’s just unbelievably impressive. Wow. What’s happening? Yeah, yeah, that’s huge.

Brian Bell (00:29:29.103): For early stage companies. Who’s it not for? You know, when you—when you, yeah, think of some—some of the companies that have onboarded and not worked out. Yeah.

Misha Makara (00:29:36.869): So, uh, I don’t want to say it too loud right now because we’ve been very slow with letting companies in. We—we’ve had no churn, which is a whole other topic of whether or not that’s a good thing or a bad thing. The—the problem is when a company doesn’t know who they are or what they’re trying to say. And then it becomes very difficult to present them and to represent them to the world. And...

Brian Bell (00:29:55.939): In general, you just don’t even onboard those guys. So you’ve been vetting. You’ve been doing a good job of just not letting anybody like...

Misha Makara (00:30:01.962): That on the platform. Well, we’ve definitely experimented, and we’ve tried a bunch of—we have artists, we have politicians, we have seed stage companies. There’s—there’s a bunch. But it’s definitely a lot easier. And as I’m looking at the metrics and what’s getting success, I see a lot more success with the organization that know who they are and how to communicate their product than organizations that don’t. I—I can’t help you with defining and creating market fit, but if you know your market fit, it becomes a lot easier for me to articulate it, and then...

Brian Bell (00:30:33.998): To share that with others. Yeah. And I think you described a problem with the industry and PR, which—it’s very relationship heavy, and it’s kind of a slow roll, right? You know, as somebody who runs a VC firm, you know, I’ve seen this both for Team Ignite, our firm, but also for portfolio companies. This PR game can be kind of a long haul. How do you sort of measure success as you kind of onboard these companies and kind of keep them engaged and coming back month...

Misha Makara (00:30:58.121): And quarter and year after year? I mean, of course, we track the results. We’re—we’re watching. And we, when you onboard, we identify your competitors, and we ask you to verify that we got that right. And we’re looking at what your competitors are doing, and where they’re getting featured, and how they position themselves, and how they’re communicating their differentiators in relation to you. All of that baked into the Rally platform. But you’re right. It does take time. One of the advantages of working with Rally is we’re actively building those relationships, and we’re—we’re fostering them right now. Before anyone signs up, we’re building those relationships with those media, so that way, when you show up and you have a press release, we already know who to talk to, and what’s working for them, and when they open their emails, and what are the stories that are getting them views. So when you have a press release, we know better how to do that matchmaking, and how...

Brian Bell (00:31:48.356): That reporter likes to be communicated. So what’s the, you know, vision for the future? You know, five or ten years out, you guys have—you’ve been really successful with Rally. What does...

Misha Makara (00:31:57.643): That look like? Yeah. We—we do want to be every company’s first PR hire. There’s, of course, a universe where we’re working with other agencies and leveraging our data and our insights to help them be more efficient and effective, and the services that they’re delivering. Ultimately, we do want to be a relationship platform. We have this information. We know what’s working. If you’re a new journalist and you’re looking to go figure out what will people be interested in, what’s a cool story, what’s going to get me views, you can come to Rally, and we’re going to give you some interesting stories about what our founders and different companies working on. That benefits everybody: is a journalist, you get a cool new story about something that’s happening that you wouldn’t have learned about otherwise. The company is getting some press that, you know, they would like to have. So it’s a truly a win-win...

Brian Bell (00:32:41.610): For everyone involved. It’s interesting. So it’s—it’s a platform, but it’s also a—almost like a connector. But it’s also a data aggregator on both sides of this sort of kind of PR marketplace, in a way. It’s all AI powered under the hood.

Misha Makara (00:32:54.351): Yeah, um, and you might see that thread from my my past with doing digital securities and certain multiple marketplaces. But I look at it as a kind of a three-sided marketplace. You’ve got the companies and their preferences. You’ve got the journalists, the media contacts, and what they’re working on. And you’ve got what’s going on right now in the larger ethos of what the community and people are talking about. And so you have to do this three-way matchmaking, which is very—yeah, yeah. What works now is—is different...

Brian Bell (00:33:20.880): Than what will work tomorrow. Well, yeah. And you also have this—this other constituent, which you described, which is the agency, right? And I think you can get surprised by how much power agencies actually have. We did at Rocket Fuel. I was at a AI unicorn, Rocket Fuel, and we were very antagonistic towards agencies, very famously. So we called them dinosaurs, and we were going to reinvent them with AI, right? Sound familiar, right? Absolutely. And those agencies just crushed us because they had the relationships, right? And—and they—they started taking traffic and—and—and bids and ad spend away from us, especially after we IPO’d, and they noticed that these guys are, like, buying things for a dollar and reselling them for five dollars, right? They didn’t like that. Yeah.

Misha Makara (00:34:00.441): And again, remember, like, we’re doing PR, so it’s less—less about the marketing and the specific ad spend. It’s about building those relationships with those media contacts so they want to talk about you. But you’re absolutely right. It takes a long time to build those relationships. The other thing is, you know, we have advantages of technology, and I want to say a second mover advantage in the space. One of the things I said before was kind of my two kind of things I look for when I’m doing diligence on a company. Another is, everyone loves to go and talk about being a disruptive technology. And I’m sure, you know, Brian, you’ve had people go to you and say, oh yeah, we’re disruptive. All of them are, right? All of them are disruptive, right? Sure, great. Look back at the actual case study about what does it mean to be a disruptive technology. And what that means is, they usually show up to the scene, or they’re supposed to show up to the scene, and be shitty in just about every dimension except for one. They can do better in one particular dimension, and it’s usually not cost, right? So, for example, the—the classic is, when I’m—when I’m teaching this, I like to talk about addition of flash drives as a storage technology, right? We had spinning disks. All of a sudden, you had flash drives, right? At that particular moment, you had a hard drive that was a couple of gigs, and you had a USB flash drive that was—what was your first flash drive? Two megs, three megs, right? Yeah. It was terrible, and it broke. It was much more expensive in the cost per unit, but it had an advantage, which was it was a lot faster. And sure enough, as time went on, both technologies got better, right? Spinning desks became cheaper. Flash drives also became cheaper. But their advantage still stayed their advantage. They were faster in terms of random access. And sure enough, as technologies continued to evolve, the—the price point changed, and now all of a sudden had commercial viability, right? You had your exchange server, and you would all of a sudden get solid—solid state disks because your database needed to be faster. It was a lot more expensive than the spinning disks, but it was better for...

Brian Bell (00:35:54.130): That particular application. Remember, we put our app—I think it was like 2013. I was a PM at a startup, and we migrated our app to SSD on the servers. Yeah. And it just made such a difference. Such a difference. Like, yeah, but those discs were much more expensive. Yeah, yeah. It was a lot more expensive to serve the app, but, you know, the user experience—you know, speed is the feature, as Larry Page would say. It reminds me, one of the most delightful experiences I had around the same time, I think, because I had this experience at work as a PM. I did the same on my own PC, right? I upgraded it to an SSD back when it was, like, a spinny disc, right? And just that night and day difference.

Misha Makara (00:36:29.676): And I want to remind you of that point where Apple was selling you a MacBook, and you would pay more to get less storage, but it was a solid state. I don’t know if you remember that—that point. I never bought Macs, so Android PC all the way. Yeah, sure. Good for you. But—but as an example, though, right? As just a general consumer product, those were competitively viable, right? That was an interesting inflection point. So when companies come to me and they’re pitching and saying, oh, well, we’re a disruptive technology, that’s what’s going through my head, about how are you—what is your differentiator? What makes you unique and special in this domain? How are you going to continue to trend to become the new sustaining technology? And I think that that’s a very important lesson when looking at just the overall patterns about how do technologies evolve and, you know, operating in this space as a technologist. Yeah.

Brian Bell (00:37:18.037): Well, let’s wrap up with some quick rapid fire questions. You know, you’ve hired hundreds of engineers. What’s your single highest signal interview question? I actually noticed...

Misha Makara (00:37:27.483): You do it to me earlier. So usually, what I like to do is, first of all, I have to think about the role and whether or not it’s a manager or a maker role. I very much like to subscribe to the idea of service owners. So usually, when I’m brought in, I’m thinking about what’s the narrative for the company that I’m starting to set up. If we’re trying to set up for a Series A, where does that IP—and how is this going to fit in the story that I’m going to tell with the financials? So there’s—that’s a whole other other topic. But in terms of hiring a individual contributor, usually what I like to do is I go a couple levels of deep detail down and ask them to give a little bit more context or insight, especially in the kind of feeling, or, like, what was the situation there. So I’m thinking about one where I was hiring a product manager. And obviously, interviews are important, right? How are you interviewing a customer? And so you poke into that, and you’re like, well, how—how did you actually collect feedback on this thing? And, you know, they give you the textbook answer, and they’re like, no, no, no, no. How do you, like, find people? Like, how did you track your feedback, right? Those questions are a couple levels down. So if someone’s bullshitting you, you can kind of quickly see when they’re like, oh, and you see the beach ball on their—their face kind of loading and figuring it out. So I love that also because it—it shows you, and you can listen to the words that they’re using of us versus I, and you can immediately get a sense of the ownership that—that they took on through that particular point. Were they a firsthand observer, or were they the ones actually driving the process forward? So those are a couple of my favorite techniques to use.

Brian Bell (00:38:59.780): So, for startups, fractional CTO—yeah, uh, hire or don’t hire? Ooh, answer, like everything...

Misha Makara (00:39:05.582): In tech world is, it depends. It depends on what you’re doing. It’s a little bit weird on my resume because I’m always parachuted in to do something for a couple months, and my goal is to get out of there. My goal is to help the company fix it, and then to get my replacement to stop the fire and put someone else in. So again, Rally is a fun fun adventure because I finally don’t have to do that. I get to just do the thing that I enjoy doing. Fractional CTOs—definitely be very careful doing that if you’re trying to raise money, because that is an awkward place to be, and that’s been a problem that I’ve had. So come in, hey, help us raise money, and then get out, right? That’s just awkward for everyone involved, so don’t do that. If you’re—of course, if you have a CTO or you have technical leadership that leaves and you need to quickly fill that hole, right, that can be a good interim solution, especially as you’re trying to suss out what’s going on. Another important factor is just the—the network that—that person’s bringing. So especially now with AI, go figure, DevOps is kind of important, right? If you need to be able to deploy software—Rally, we deploy sometimes nine times a day. I don’t know. That’s another topic. But you need some serious infrastructure to be able to do that, and you want to be able to make sure that that deployment schedule is in alignment with the risk appetite of the company, and you have the safeguards in place in case you oops and you need to go put something back. So in terms of hiring your fractional CTO, those are the things that I would look for. But, like everything, it really depends on—on the company in this stage of maturity and what’s going on. But hiring a fractional CTO to raise money, don’t do that. Sounds like...

Brian Bell (00:40:40.996): A bad idea. What have you changed your mind on? What’s a belief that you—you held that you’ve reversed course on? This one was coming. One of the things...

Misha Makara (00:40:49.379): That I definitely kind of changed my perspective on is Web3, specifically. So I’m a fan of domain driven design, and initially, when we were doing the Web3 crypto thing, my argument was blockchain is a event sourcing and event streaming, and event sourcing and event streaming was a great pattern to use for everything across the board. And if we look at what happened in Web3, I remember the rush where everyone was trying to tokenize everything, and they wanted completely Web3—complete Web3 platforms. As you know, everything is currently running down the chain now. That’s sarcasm, right? The—the ones that were leaders were the ones that used the right technology for the job and had some kind of a hybrid. So I think I was a little bit too focused on applying it across the board. So I would have a complete Web3 event sourced application, but the cost for that—the—the business case for that kind of quickly diminishes after a certain point of functionality. So again, Pareto style: what’s the 20% that gives you the 80% of the value? What’s the part that customers are actually looking for and that they want in their—their features, versus what—what needs to be in—in that new technology? So that was, I’d say, that was one of the mistakes that I made. I was a little bit too fixated, and I want to say I drank the Kool-Aid. But I still think, through looking from a cyber security perspective, I like the idea that the event itself is an immutable thing. But there’s also right ways and wrong ways to do that, and using that as the default solution for everything everywhere is—it’s doable, but it’s also cost prohibitive, even in the—the world of AI now.

Brian Bell (00:42:26.561): Yeah. What’s the best piece of advice you ever got? If you’re...

Misha Makara (00:42:28.964): If you’re doing research, do—you can be applied or you can be pure, but make sure you articulate what it is that you’re doing and set up bounds on the experiment, because it’s very easy to keep investing and keep investing without having a defined outcome that you’re looking for from the experiment, and continuing to invest in the experiment before. So that also fits into this larger idea of, you know, learn to fail quickly. Okay, great, but at what point is a failure a failure, right? How do you—how do you set your own personal limits? You know, what are your—you know, how far are you willing to go with something before you say, all right, we’ve tried that enough? And one of my—the things that I do when managing agile teams—I love the idea of a senior developer coming, or any any developer, saying, I’d like to go try this. But then me responding back, saying, all right, you really only have Wednesday afternoon to go do this. If you can do it and it meets these requirements, then we will continue, and I’ll give you Thursday, Friday, like whatever, to continue. Or if it doesn’t, you know, hey, we tried it. We’re gonna put our lessons learned on the shelf, and, you know, we might come back to that. Or, I didn’t get to finish the experiment because of these things. Let’s extend it. So the—the lesson learned would be: make sure you have limits on your experiments. Don’t just keep going forever and ever. Otherwise, that’s how you end up working yourself into a...

Brian Bell (00:43:47.903): Well, really enjoyed the conversation. Where can people find you online...

Misha Makara (00:43:51.385): And find out more about Rally? Sure. Uh, good thing about my name is it’s very easy to find me. So, mishamakaran.com. Go figure. rally-ai.com. You know, again, come—come join us. We’re finally starting to open things up a little bit. So, you know, we have some extra capacity. I think this month we decided we’re going to allow in another five customers, and then we’re going to continue to open it up from there. But yeah, come—come join us. Awesome. Well, thanks, uh...

Brian Bell (00:44:14.279): So much, Misha. My pleasure.

The frame worth carrying out of this week is that the bridge between “the model can do X” and “a paying customer relies on X in production” is being built by the platforms themselves, faster than the ecosystem around them can build a business on the gap. That is good news for anyone who owns a workflow, a dataset, or a control point the platform cannot absorb. It is a bad week for anyone whose product is the bridge itself.

Venture markets and private capital

The clearest signal in the private markets this week came from fund formation and pricing discipline, not from a scroll of new rounds. B Capital closed its third early-stage Ascent Fund at $500 million, roughly double its $254 million 2022 predecessor, with more than twenty investments already deployed and typical checks of $500,000 to $10 million, according to Wall Street Journal reporting on July 6. In the same reporting, PitchBook data placed median U.S. pre-money valuations at $19.5 million for seed and $64 million for Series A through the first half of 2026.

Read together, those two data points describe a market where specialist managers with a defensible thesis keep raising and keep getting financed, while the middle of the market, generalist, undifferentiated, “AI-enabled” in name only, keeps absorbing worse ownership math. The entry price at seed and Series A now assumes considerably more than a promising team and a large addressable market.

Late-stage private AI stayed even more sharply bifurcated. Business Insider reported on July 9 that Caplight and Rainmaker Securities were quoting Anthropic secondary indications near a $1.2 trillion valuation, against roughly $908 billion for OpenAI on Caplight, with very little seller supply on either name. That is a scarcity signal more than a valuation signal. A handful of buyers are chasing a handful of sellers in two companies, and price is doing what price does under those conditions: it floats free of any near-term cash flow discipline. The honest expectation for TIV’s book is that this kind of pricing survives right up until audited numbers, lockup expirations, or an actual IPO force real price discovery, at which point dispersion tends to widen sharply rather than compress.

The most durable capital signal of the week came from compute and power, not equity. TeraWulf disclosed on July 6 that Anthropic had signed a twenty-year lease for a Kentucky AI campus supporting roughly 401 megawatts of critical IT load and approximately $19 billion of contracted lease revenue over the initial term. This is the new definition of “AI infrastructure” capital: long-duration, utility-like assets with contracted off-take, not speculative GPU resale. That structure is bullish for anything that secures power, interconnect, or a deployment surface directly, and bearish for commodity neocloud stories that resell capacity without owning a hard bottleneck underneath it.

On founder behavior, the evidence this week is circumstantial but consistent: tighter burn discipline, smaller teams, and GTM increasingly tied to a measurable outcome rather than a seat count. When a model vendor lands directly inside Microsoft 365, when a hyperscaler productizes agent runtime primitives, and when the Fed minutes explicitly name AI-related investment as an inflation factor, the message to a fundraising founder is unambiguous: seat-growth stories and wrapper margins get less benefit of the doubt, while founders who can tie AI to labor substitution, revenue lift, or cycle-time compression get a materially easier path to capital. This is inference, not a directly observed data point, but it follows from several confirmed developments landing in the same week.

Frontier signposts: where the deployment bottleneck actually moved

Four developments this week changed a real bottleneck rather than a leaderboard score, and each is worth tracking on its own terms.

OpenAI’s model became Microsoft’s default, immediately. On July 9, OpenAI said GPT-5.6 became the preferred model inside Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork, framed by OpenAI as delivering more useful work per token, with benchmark claims on coding and knowledge work that came from the vendor itself. The real surprise is not the model release, frontier labs ship new models routinely now, but the speed at which the upgrade reached one of the largest enterprise software surfaces on earth. Workflow-native AI that owns a regulated or operationally unpleasant loop becomes more investable. The generic productivity copilot category becomes more fragile, because the platform now ships the generic version for free inside software the customer already pays for.

Google turned “agent” from a demo into runtime infrastructure. On July 7, Google said Managed Agents in the Gemini API gained background execution, remote MCP integration (a protocol that lets a model call external tools and data sources directly, rather than through custom middleware), custom function calling, and credential refresh, all inside an isolated cloud sandbox. This is an operational shift, not a marketing one. Asynchronous execution removes the brittleness of long-running HTTP sessions. Remote MCP removes a layer of custom integration code that startups used to sell. Agent governance, evaluation, audit trails, and tool-permission layers for specific vertical use cases get more investable. Thin wrappers that sell “agent runtime” with no proprietary trust, data, or distribution get more fragile, and they get more fragile fast, because Google is now giving away the plumbing they used to charge for.

Anthropic proved that state-scale security review is now an hours problem. On July 9, Anthropic reported that the Government of Alberta used Claude to review approximately 466 million lines of code across more than 900 repositories in about twenty hours, surfacing roughly 274,000 logging events and 588 secrets after calibration to an 8 percent false-positive rate. This is confirmed as a company-reported case study. Governments do not care about benchmark leaderboards. They care whether a tool reliably shrinks a review burden that used to require a multi-quarter consulting and security engagement. Software assurance, code governance, compliance automation, and human-in-the-loop remediation tooling all get more investable from this. Pure labor-arbitrage review businesses, the ones selling the same hours manually, get structurally more fragile. The open question worth monitoring is whether that false-positive rate holds up outside a single showcase deployment, and whether procurement broadens past one-off pilots into recurring contracts.

Physical AI moved from humanoid theater toward engineering infrastructure. NVIDIA and Hugging Face said on July 6 that LeRobot, an open robotics framework, would get NVIDIA Isaac GR00T 1.7, Isaac Teleop, and planned Cosmos 3 integration, opening more of the model-data-simulation stack to independent developers. Separately, Anthropic reported on July 9 that UST is deploying Claude across semiconductor and hardware validation workflows, training 20,000 engineers, with its iDEC pipeline cutting validation cycle times by 50 to 70 percent and condensing standard four-day turnarounds into 48 hours. Both integrations are confirmed at high confidence; the specific performance gains are company-reported and should be read at moderate confidence. The bottleneck here shifted from “can a physical system reason at all” to “who owns the deployment loop, the data loop, and the validation loop.” Digital-twin quality assurance, robotics data infrastructure, test automation, and simulation-to-production tooling all get more investable. Demo-first robotics stories with no installed base and no learning loop get more fragile.

Platform power and incumbent moves

The pattern across every major platform move this week points in one direction with one meaningful exception. OpenAI and Microsoft made the general workplace copilot market harder to compete in by shipping GPT-5.6 straight into Microsoft 365 Copilot. Google made general-purpose agent runtime layers harder to compete in by productizing Managed Agents. NVIDIA made open robotics infrastructure easier to build on, pushing both open tooling (the LeRobot integration) and a lower-cost open-stack agent story through Nemotron 3 Ultra.

Meta’s moves pulled in two directions at once. On the product side, Meta expanded Muse Image, its consumer-facing image generation surface, on July 7 and July 10, a move that makes standalone image-generation products even less defensible unless they own a specific niche workflow, brand relationship, or commerce loop the platform cannot replicate. On the infrastructure side, Meta broke ground on a 1 gigawatt data center in Alberta, its first in Canada, tied to more than CAD $13 billion of investment and roughly CAD $60 million of local infrastructure spending, announced across July 8 and July 9. That signals major platforms increasingly want sovereign-adjacent compute footprints with direct control over power-secured capacity, which squeezes the smaller cloud intermediaries sitting between hyperscaler capacity and end customers.

xAI added one more capable supplier to the coding and agentic-work pool with Grok 4.5’s July 8 launch, described by xAI as its strongest model yet for coding, agentic tasks, and knowledge work, benchmarked against peers on DeepSWE and related engineering tasks. The comparative performance claims are vendor-supplied and should be discounted accordingly. This does not open new startup surface area on its own, but it does make single-provider dependency progressively harder to defend, because the number of “good enough” frontier suppliers keeps climbing.

The net effect: startups trying to own a generic workflow surface or generic agent plumbing lost ground this week. Startups building on top of open, enterprise-controllable stacks inside specific, narrow workflows gained room to operate.

Macro, regulation, and physical infrastructure

The macro backdrop this week was hostile to anyone underwriting a return to easy money. The Federal Reserve released its June 16-17 meeting minutes on July 8. All members supported holding the federal funds rate at 3.5 to 3.75 percent, several participants said there was a case for a hike, and the minutes repeatedly cited strong AI-related investment as both a factor supporting growth and a contributor to inflation pressure, specifically in technology products and electricity demand. The Department of Labor followed on July 9 with initial jobless claims of 215,000 for the week ended July 4, down 2,000 week over week, with the insured unemployment rate holding at 1.2 percent.

Put those two together and the picture is unambiguous: the labor market remains stable enough that the Fed feels no obligation to rescue long-duration assets with rate cuts, and AI capital expenditure has graduated from a footnote to an explicit line in the inflation conversation. That is bad news for any business whose margin story quietly assumes falling rates or cheap, abundant power.

Physical infrastructure did more underwriting-relevant work this week than any formal regulatory action. Meta’s Alberta campus and the Anthropic-TeraWulf lease are saying the same thing in different accents: AI buildout is localizing around power access, long-duration contracts, and regional infrastructure politics. Meta’s site is a full gigawatt. Anthropic’s Kentucky lease covers roughly 401 megawatts of critical IT load over twenty years. Markets are still mostly talking about models. Underwriting should now treat power rights, permitting timelines, cooling capacity, and energy-linked financing structures as first-order variables in any AI infrastructure diligence, not footnotes to a compute story.

Cross-stack interaction effects

Four combinations this week matter more together than any single item does alone, and each has a different time horizon.

Managed agents plus government-scale code review, immediate horizon. Google productized agent runtime primitives on July 7. Anthropic demonstrated a state-level code review deployment on July 9. Together, they narrow the gap between “agent demo” and “institutional production” faster than either does alone. Security review, compliance orchestration, and software assurance become more investable in the same week that generic agent wrappers become more fragile. This combination still looks underpriced by the market.

Copilot bundling plus a hawkish Fed, immediate to medium-term horizon. OpenAI pushed GPT-5.6 into Microsoft 365 Copilot on July 9 in the same week the Fed minutes showed no easing bias and open discussion of tightening if inflation stays sticky. Together, these raise both the product bar and the financing bar simultaneously. Startups selling measurable outcomes inside budgets that already exist become more investable. Seat-based or prompt-box businesses that need generous capital markets to paper over a distribution problem become more fragile. The market is still overpricing generic “AI adjacency” and underpricing financing risk.

Meta’s gigawatt Alberta build and Anthropic’s 401-megawatt Kentucky lease, structural horizon. Together they sketch a new regional map for AI, where compute becomes geographically sticky around energy access and public-infrastructure relationships rather than remaining fungible across regions. Power-aware orchestration, cooling systems, campus financing software, and regional supply-chain services become more investable. Cloud-resale plays that assume capacity stays liquid and interchangeable become more fragile. The market is underpricing how quickly geography and energy politics will stratify who gets access to compute.

What this means for Team Ignite founders

More attractive now. Compliance and software-assurance tools that sit directly inside regulated code and data workflows, the Alberta deployment just proved the category’s ceiling is much higher than most diligence assumes. Power-aware AI infrastructure software, including orchestration, cost controls, and financing visibility, given how fast capital is moving toward power-secured campuses. Physical-AI validation and digital-twin workflows in semiconductors, manufacturing, and embedded systems, where UST’s cycle-time numbers show the commercial case is already provable rather than theoretical. Vertical agents that own a genuinely painful operational loop and can price on outcomes instead of seats, because the seat-based alternative just got a lot more exposed.

Less attractive now. Generic productivity copilots, now bundled for free into software your prospective customer already pays for. Horizontal “agent platform” wrappers with no proprietary trust, data, or workflow ownership, since Google just gave away the runtime plumbing this category used to sell. Standalone image-generation consumer products with weak distribution, squeezed directly by Meta’s Muse Image expansion. Neocloud stories that do not control power, campus economics, or a genuinely proprietary customer relationship, now competing against hyperscalers building sovereign-adjacent capacity of their own.

Overhyped but worth watching. Late-stage private AI marks, where scarcity is currently masquerading as fundamental value and where a real liquidity event will force a repricing nobody can predict precisely. Open-stack enterprise agent infrastructure, some of it will become a genuine long-term control point and a lot of it will turn out to be temporary tooling that the platforms absorb within a product cycle or two. Frontier coding benchmarks, where the real question founders and investors should be asking is deployment reliability in production, not who leads a screenshot this month.

Underpriced or under-discussed. AI for software assurance inside government and regulated enterprise, which this week went from theoretical to demonstrated at genuine scale. Services-led distribution into physical AI and embedded engineering, a channel that gets less attention than the robotics hardware story but is where UST-style deployments are actually landing. Energy-adjacent software for AI campuses, where capital is moving faster than the startup map has caught up to. Middleware that gives customers real portability across model vendors and regions, valuable precisely because the supplier pool keeps widening and no customer wants to be locked to one lab’s pricing and policy decisions.

Questions for founders this week. If Google or Microsoft keeps collapsing runtime and workflow primitives directly into the platform, what control point do you own that actually survives that compression? If your product sits inside a regulated or operational workflow, can you prove false-positive control and auditability, not just output quality, because that is what procurement is now asking for? If model costs keep falling and supplier choice keeps widening, what actually gets better in your business beyond gross margin? If rates stay elevated and power stays constrained, which assumption in your current roadmap breaks first, and have you actually stress-tested it?

Secondary-market watch list. Anthropic, where scarcity and product credibility are both driving price, and where the $1.2 trillion secondary mark deserves more skepticism than conviction until real liquidity tests it. OpenAI, whose GPT-5.6 push directly into Microsoft’s enterprise surface changes the distribution calculus more than any benchmark result would. xAI, which Grok 4.5 keeps relevant in coding and agentic work even without independent verification of its comparative claims. Sierra-class enterprise AI, where platform bundling raises the premium on companies that own workflow and trust inside the enterprise rather than model access alone. Databricks, where the open-stack agent and data-control narrative keeps strengthening in parallel with this week’s developments.

What to monitor over the next one to four weeks. Whether GPT-5.6 meaningfully changes Copilot usage and pricing behavior once the initial rollout numbers come in. Whether Google adds enterprise controls and transparent pricing to Managed Agents, which would determine how much of the third-party agent-runtime category survives. Whether more state or federal institutions publish production AI security case studies beyond Alberta, which would confirm rather than isolate the deployment pattern. Whether more compute campuses get announced with named power figures and contract durations, the clearest available signal of where AI capital is actually flowing.

What this means for Team Ignite LPs

AI exposure is splitting into three distinct buckets, and treating them as one undifferentiated category is now the biggest risk in a portfolio review. The first bucket is scarce late-stage assets whose prices are already extreme and whose risk is almost entirely a liquidity-timing question rather than a fundamentals question. The second bucket is early-stage categories where platform compression is accelerating in real time, this week’s Google and OpenAI moves are a direct preview of what happens to any startup whose core value proposition is a feature a hyperscaler can ship for free. The third bucket is under-owned wedges in regulated workflows, power-linked infrastructure, and physical AI, categories where the commercial proof points this week (Alberta, UST, TeraWulf) came from actual production deployments rather than pitch decks.

The opportunity for a seed-stage manager remains real, but it is narrower and more specific than a general “AI exposure” thesis. It is about owning genuinely unpleasant, high-consequence operational loops before they become obvious to everyone else, not about generic proximity to the AI theme. The risk on the other side is straightforward: paying frontier-adjacent multiples for surface area that an incumbent can absorb inside a single product cycle, which is exactly what happened to several categories this week alone.

LPs should also register the macro overlay directly. A Fed that is explicitly naming AI capital expenditure as an inflation factor, combined with compute campuses financed on multi-decade contracted terms, means AI is no longer a pure software story to underwrite. It is software, labor substitution, power economics, and long-duration financing risk, all at once, and LP due diligence on any fund’s AI thesis should reflect that composite nature rather than a single-variable software multiple.

What this means for VCs generally

Two rules got reinforced hard this week, and neither is new, but both got harder to ignore. First, distribution and infrastructure control are compounding faster than model novelty. A new frontier model release used to be an event that reshuffled competitive positioning for months. This week, three separate labs shipped meaningful model or agent-runtime news, and the more consequential story in each case was not the model itself but who got direct, immediate distribution into an existing customer relationship (Microsoft, in OpenAI’s case) or an existing developer ecosystem (Hugging Face, in NVIDIA’s case). Model quality is necessary and increasingly insufficient.

Second, discounted cash flow reality is back, even though a meaningful share of private AI capital still underwrites as if it is not. The Fed is on record treating AI investment as a demand and inflation variable. Compute campuses are being financed on twenty-year contracted terms that look like project finance, not venture capital. Funds that continue underwriting AI opportunities with 2021-vintage SaaS assumptions, cheap capital forever, seat growth as a proxy for value, generic AI adjacency as a moat, will keep getting surprised by exactly the kind of week this one was: platforms absorbing surface area, capital concentrating around scarcity and power rather than product velocity, and the labor market giving the Fed no reason to bail anyone out.

That is the tension at the center of our conversation with Jeff Gothelf, co-creator of the Lean UX movement, author of Lean UX, Sense and Respond, and Who Does What By How Much?, and co-founder of Sense & Respond Learning.

Jeff has spent decades helping product teams move away from rigid waterfall processes, bloated deliverables, and executive-led guesswork. His work has shaped how modern startups and product organizations think about experimentation, customer evidence, design, OKRs, and outcomes.

But in 2026, the product world is facing a new problem: AI can now generate designs, prototypes, copy, code, research summaries, and product ideas at incredible speed.

That sounds like progress. Jeff’s warning is sharper: speed without judgment just gets you to mediocrity faster.

Subscribe now

From Failed Musician to Lean UX Pioneer

Jeff’s career did not begin in software. It began in music.

Before becoming one of the best-known voices in product design, he was trying to make it as a rock musician. He played piano, toured with bands on the East Coast, and gave the dream a real shot.

When that did not work out, the early web was beginning to take off. Jeff taught himself HTML and basic web design, first building websites for himself and friends. In 1999, as he puts it, “if you could spell HTML, you could get a job.”

That led him into web design, interaction design, UX, product leadership, and eventually Lean UX.

His frustration came from a simple but painful realization: on a good day, only half of his design work ever got implemented. The rest was thrown away.

Not because the work was bad, but because the process was broken.

Teams would spend weeks or months designing software up front, based on assumptions, executive direction, and theoretical requirements. Then reality would hit. Customers did not behave as expected. Engineering constraints appeared. Priorities changed. The work was discarded.

Lean UX emerged as Jeff’s response to that waste.

Instead of treating design as a big upfront deliverable, Lean UX asks teams to use just enough design to move the conversation forward, get feedback, and decide whether to continue, change direction, or kill the idea.

The goal is not to produce more artifacts.

The goal is to learn faster.

The Moment Jeff Realized Software Development Was Broken

One of the most memorable stories from the episode comes from Jeff’s time at America Online.

He described working under an executive who micromanaged a massive organization. Jeff had spent about a week working on a design and walked into the executive’s office with a printed version of the work.

Before Jeff could even present it, the executive looked at the paper from across the room and said:

“There’s nothing remotely close to anything I want to see on that sheet of paper.”

For Jeff, that was a breaking point.

It was not just a harsh comment. It represented the larger dysfunction of how software was being built: executive opinion above customer evidence, hierarchy above learning, and big upfront work before any real validation.

That became one of the emotional roots of Lean UX.

Jeff was not simply trying to make design more efficient. He was trying to make product work less stupid.

Why AI Creates a New Race to Mediocrity

A major theme of the conversation was how AI is changing product work.

Jeff is not anti-AI. He sees the tools as powerful and exciting. Designers, product managers, and engineers can now create prototypes and explore ideas much faster than before.

But he is skeptical of the idea that AI eliminates the need for product judgment.

In his view, AI makes it easier for everyone to generate similar-looking, similar-sounding, similar-functioning work. That creates what he called a “race to mediocrity.”

He compared the current AI moment to the era of Twitter Bootstrap. Bootstrap made it easy for anyone to build clean, usable websites. That was useful. But it also made everything look the same.

AI risks doing the same thing at a much larger scale.

When everyone prompts similar tools using similar language and accepts similar outputs, products start to converge. The market fills with generic interfaces, generic copy, generic workflows, and generic strategy.

That is why Jeff believes the human role becomes more important, not less.

The future advantage will come from taste, judgment, originality, and a strong opinion about how a product should create value.

Strong Product Opinions Will Matter More

One of Jeff’s clearest arguments is that the best companies will stand out because they have a strong opinion about the value they provide.

He used banking as an example.

Traditional banks may have deep infrastructure and long customer histories, but many of their digital experiences remain clunky. By contrast, companies like Wise create a cleaner, more digitally native experience around international money movement.

The difference is not just features.

It is opinion.

A company that believes banking should be simple, fast, multicurrency, and digitally native will build a very different product from one carrying decades of institutional inertia.

Brian made a similar point with Mercury, describing how the company makes banking for his venture firm feel dramatically easier than traditional banking workflows.

Jeff’s response was direct: “Delightful to bank” is an opinion.

And once a company has a real opinion, product decisions can follow from it.

That is the work AI cannot fully replace: deciding what should exist, why it should exist, how it should feel, and what kind of customer behavior it should create.

Founders Still Need to Talk to Real Customers

For founders, Jeff’s advice has not changed much over the last twenty years.

The fundamentals are still the same:

Solve a real problem for a real customer in a meaningful way.

What has changed is that the excuses have disappeared.

In 2006 or even 2016, teams could argue that customer research was expensive, slow, or operationally difficult. In 2026, Jeff does not buy that excuse.

Founders can find customers, schedule conversations, show prototypes, collect data, and synthesize insights faster than ever.

But the core work still has to happen.

You need to understand the customer’s current behavior. You need to observe where the pain actually is. You need to test whether your solution changes behavior. And most importantly, you need to be willing to change course when the evidence contradicts your belief.

That last part is where many founders fail.

They collect evidence, but they do not let it change their mind.

Synthetic Users Are Not a Replacement for Real People

The episode also touched on one of the more controversial trends in product research: synthetic users.

Some startups now claim they can simulate customer interviews using AI-generated personas or artificial societies. Jeff is unconvinced.

He sees value in using synthetic users to practice interview scripts, test rough messaging, or prepare for research. But he does not believe they can replace real customer conversations.

Why?

Because real humans reveal things synthetic users will not.

Jeff shared a story from his time at The Ladders, a job board for professionals earning over $100,000. When interviewing executives, his team noticed that many preferred communicating with recruiters through SMS.

The surface-level explanation might have been convenience.

The real reason was more revealing: they believed their bosses could read their email, but not their text messages.

That kind of insight comes from human context, fear, hesitation, and lived behavior. A synthetic user would likely miss it.

Jeff’s point is not that AI cannot help research. It is that founders should not confuse simulated answers with actual market evidence.

The Real Signal: Behavior Change

So how does a founder know they are working on a real problem?

For Jeff, the answer is behavior change.

Do customers light up when they see the prototype? Do they ask for access? Do they change how they work? Do they come back? Do they refer others? Do they pay?

This is where Jeff connects product discovery to outcomes.

An outcome is not a feature shipped or a deliverable completed. It is a measurable change in human behavior that drives a business result.

This is why old startup metrics like acquisition, activation, retention, revenue, and referral still matter. They are not just dashboard numbers. They are signals of whether customers are behaving differently because of what you built.

The danger is when teams measure output instead of value.

Shipping more features does not mean you are solving a bigger problem. Generating more designs does not mean you are creating better UX. Producing more AI-generated artifacts does not mean you are making better decisions.

As Jeff put it:

“Producing stuff is not the production of value.”

Product Managers Are Not CEOs of the Product

In the rapid-fire section, Jeff challenged a popular product management cliché: the idea that product managers are “the CEO of the product.”

He rejects it.

Product managers cannot hire, fire, or fully control budgets. They do not have the authority implied by the CEO metaphor.

That does not make the role unimportant. It makes the role different.

A good product manager guides decisions, aligns people, clarifies outcomes, understands customers, and helps the team make evidence-based tradeoffs.

The danger of the “CEO of the product” framing is that it encourages product managers to think in terms of authority rather than influence.

Modern product work is not about commanding the team.

It is about creating clarity around what matters and why.

Why Jeff Changed His Mind About Measuring Customer Conversations

One of the most honest moments in the episode came when Jeff talked about something he used to believe but now thinks he got wrong.

For years, he argued that measuring the number of customer conversations was a vanity metric. After all, a team could talk to 1,000 customers and learn nothing if they were asking bad questions or talking to the wrong people.

He still believes that risk exists.

But he has changed his view.

For organizations that do not talk to customers at all, simply measuring the habit can be useful. Asking teams to speak with 50 customers in a month may not produce perfect insight immediately, but it builds the muscle.

The repetition matters.

At first, showing up and doing the work is the point. Teams get more comfortable. They ask better questions. They notice patterns. They improve.

Eventually, they need to move beyond activity and toward learning. But for teams starting from zero, counting customer conversations can be a useful first step.

The Human Advantage in an AI-Saturated World

The conversation kept returning to one central idea: as AI makes production easier, humanity becomes more valuable.

Jeff described how his daughters and their friends are already reacting against AI-generated perfection. They crave authenticity, flaws, live music, analog experiences, and real-world texture.

That matters for product builders.

When AI-generated work becomes abundant, polished, and generic, customers may place more value on products that feel distinctive, opinionated, and human.

This does not mean rejecting AI.

It means refusing to let AI flatten your taste.

The best product teams will use AI to move faster, explore more options, and reduce low-value labor. But they will still rely on human judgment to decide which ideas matter, which customers to serve, which problems are worth solving, and what kind of experience should exist.

The Circus Lesson

Near the end of the episode, Jeff shared one of the most unexpected stories of the conversation: his first job out of college was with the Clyde Beatty-Cole Brothers Circus.

He spent six months traveling the East Coast with a three-ring tented circus, running sound and lighting for hundreds of shows.

It was difficult, scrappy, and humbling.

The equipment was inadequate. The environment was brutal. Dust, moisture, animals, travel, and constant setup and teardown made everything harder. Jeff had to fix microphones, hit lighting cues, and keep the show running in front of thousands of people.

That job taught him not to make assumptions about people. It taught him to work under pressure. It taught him that every role matters when a community depends on the show going on.

In a strange way, that lesson connects directly to his product philosophy.

You cannot build good products from a distance. You have to understand the people, the constraints, the environment, and the real behavior happening on the ground.

Final Takeaway

This episode is not just about Lean UX or AI or product management.

It is about the difference between making things and making things that matter.

AI will make it easier to produce. Easier to design. Easier to prototype. Easier to write. Easier to code.

But none of that guarantees value.

The hard part is still the same: understanding humans, identifying real problems, forming strong opinions, testing those opinions against evidence, and having the discipline to change course when the market tells you that you are wrong.

Jeff Gothelf’s message for founders is blunt and useful:

There is no excuse not to talk to customers anymore.

There is no excuse to confuse output with progress.

And there is no future in building the same AI-generated product as everyone else.

The tools are faster now. The judgment matters more.

👂🎧 Watch, listen, and follow on your favorite platform: https://tr.ee/S2ayrbx_fL

🙏 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast

Chapters:
00:01 — Welcome & Jeff Gothelf Introduction
00:30 — From Failed Musician to Early Web Designer
03:38 — UX vs. Lean UX
05:19 — Waterfall Software and Wasted Design Work
07:31 — Lean UX and Just-Enough Design
08:43 — The AOL Moment That Changed Jeff’s Thinking
10:19 — Internet, Cloud, and Faster Feedback Loops
12:21 — AI’s Impact on UX and Product Teams
13:18 — Why AI Won’t Replace Product Roles
15:15 — Mass Production, Customization, and Human Taste
16:24 — Strong Product Opinions as Differentiation
18:44 — UX as the Competitive Advantage
21:02 — Founder Advice for Building in 2026
23:40 — Problem, Market, and Solution Validation
24:17 — Synthetic Users vs. Real Customer Interviews
28:01 — Finding a Problem Worth Solving
30:00 — Avoiding Bias in Customer Research
32:07 — Taste, Judgment, and AI Slop
34:05 — What the Next Generation Should Work On
37:41 — The Book That Aged the Worst
39:24 — Liberal Arts, Humanity, and Anti-AI Rebellion
41:40 — Using AI as a Harsh Thinking Partner
43:41 — Product Managers Are Not CEOs
44:27 — Disagreements, Qualitative Benefits, and Customer Value
46:02 — Getting Out of the Deliverables Business
48:19 — Customer Conversations as a Practice Muscle

Transcript

Brian Bell (00:01:21.523):

Hey everyone welcome back to the Ignite Podcast today we’re delighted to have Jeff Gothelf on the mic he’s one of the people who taught a generation of teams how to build software he co-created the Lean UX movement wrote the book on it literally uh Lean UX and now is now in its third edition and followed it with the Harvard Business Review Press book Sense and Respond and most recently Who Does What By How Much I love that name a customer-centric take on OKR thanks for coming on jeff it’s my pleasure thanks for having me well we’ve heard about the books but I’d love to hear from the man what’s your background what’s your origin story I was.

Jeff Gothelf (00:01:53.944):

A failed musician he tried to be a rock star and uh gave it a real shot for a couple of years toured with a couple of bands on the east coast of the U.S. those guys are still my best friends to this day in fact, we all just got together again recently about a month ago and it was so. So, nice but uh didn’t work out ended up the web was becoming a thing at that time and so the World Wide Web or just the web not the World Wide Web but the the interactive internet the Information Superhighway yeah exactly. So, I taught myself HTML and basic, you know, website design and began to build websites for myself my band my friends bands and that type of thing and then, you know, back in 1999 if you could get if you could spell HTML you could get a job and so I I started. Building websites and doing web design and that led to interaction design and UX and then ultimately leadership roles in design and then product management and then finally in 2012 I went out on my own and started a consulting but it was a services business building building apps and services in a way that seemed to make a lot of sense at the time that was kind of in contrast to the way that most software was being built at the time and we built that business it was called neo over four years and then when that sold. Uh in 2015 I’ve been I was out of my own for a long time and then really collaborating very closely with my with co-author Josh Seiden working together independently on the content and then in the last few years we have among other things solidified that into Sense and Respond Learning these days which is our product management training wow yeah.

Brian Bell (00:03:26.056):

So, much so much to unpack there what a journey I love I love the started as a rock star what instrument were you playing I played piano. Wow, okay yeah everybody needs a keyboardist in.

Jeff Gothelf (00:03:36.131):

The band right it is it’s it’s it’s one of those things that like if you need one it’s hard to find one not not all the rock bands in the world are looking for one but if you’re looking for one it’s good to be one.

Brian Bell (00:03:45.250):

Yeah, it’s good yeah it’s a that’s like the hardest to fill sometimes after the drummer the drummer’s kind of core too right like well.

Jeff Gothelf (00:03:52.738):

The drummers are hard yeah for sure and, you know, what’s interesting is I recently got I got conscripted into a Grateful Dead cover band and which not not a hundred percent my bag it’s not zero percent my bag it’s like 57 my bag I guess but I really enjoyed the gig it was it’s a really cool band with a bunch of guys kind of from a variety of different places and we did did some shows and the shows did really well and and then all of a sudden, to your point people are like I have a band that does Grateful Dead songs over here and we need a keyboard player yeah, you know, everybody wants to be Jerry. But nobody wanted to be Pigpen, you know, because that’s everything so it’s good to be a keyboard player sometimes it’s amazing yeah still.

Brian Bell (00:04:31.654):

On my bucket list of playing a band I play guitar but not very well so you help name and popular popularize Lean UX maybe you could just define what that means for versus just what is UX and what is, you know, what is Lean UX as.

Jeff Gothelf (00:04:44.753):

UX is user experience design right it’s it’s it’s been it’s been a practice for decades, you know, and it’s it’s it’s a combination of the the visual design the interaction design the copywriting the content strategy the whole sort of presentation layer of an interaction. Lean UX came out of years of really a decade of doing that kind of design work and working with teams doing that kind of design work and ending up in a place where on a good day 50 of my work got implemented which meant that on a good day 50 of my work got thrown away and I wasn’t going to put in another decade of having half my work being thrown out and part of the reason for that is because of it was sort of sort of the the nature of software development even even on online there I was just not just sort of, you know, I started America Online and, you know, we made software that went on CDs. But but even, you know, building websites and and web apps and digital products and services there’s still this belief even to this day which is mind-boggling to me that you can predict exactly what the software is going to look like and how it’s going to work and what it’s going to do and how long it’s going to take to build and what people are going to do with it right and and we were living in that world whatever that was 15 years ago and for me waterfall world as it were exactly exactly right so so this very very linear approach to developing products digital products and services and so we would make all these assumptions about the design up front right because design came early in the waterfall cycle and then at least half that work if not more would get thrown out and I just wasn’t gonna was this kind of.

Brian Bell (00:06:16.129):

An artifact I mean, if you think historically just because of the way software was manufactured, you know, packaged distributed and sold and so you kind of had to do it like that because you literally had to ship it. Exactly, right and it was hard to update it over the air like we’re used to getting a, you know, an iPhone update every, you know, every few days now right but back then it’s like okay I had to like package up the the CD as it, you know, or the the floppy disk and actually send it to you to get Windows 3.0 to 3.1.

Jeff Gothelf (00:06:42.714):

And then so on look I worked I worked at America Online and it was exactly like that we’d work for six to nine months building software and then we would stop building software and we would print 15 million CDs and then put them in the mail and hope that we built the right thing that works yeah. Yeah, you know, and that that but that that style of working it fits leadership teams really well because leadership teams can just say like well just build this and build that and when will you get it done and then ship it and what will be done right and there’s no question because at least up until things got. Got highly interactive it was difficult to determine whether or not we were actually building something valuable right and now that now that’s a lot much much much easier to figure out if we’re building something of value much sooner in the process we’re seeing a lot more of that sensing and responding that that, you know, building build measure learn type of learning loop in price but still but still to this day I work with tons of companies where there’s a lot of leadership direction and moving away from that leadership direction is occasionally a career-limiting move.

Brian Bell (00:07:41.983):

In these organizations.

Jeff Gothelf (00:07:43.465):

And so Lean UX was the reaction to that, right? Lean UX was an attempt to say, look, I am not gonna deploy everything that I have in the design arsenal on every single thing that we’re doing, right? What I’m gonna do is use the tools just at the right level, just enough to move the conversation forward one step to get feedback on that step? And then determine if I should still move forward or if I should change direction or if I should kill the idea and try something else so that that’s that was the approach in that book and it was highly, you know, based on the Lean Startup movement which was gaining traction right around the same time. So, you know, we in fact, Lean UX is a series of books that Eric Ries put his name on right the the the Lean Series actually a couple times exactly and so and so, you know, heavily influenced by that but really applied directly to the design practice and as a reaction to to making design work in agile software engineering which was sort of like the next wave of software engineering supposed to counteract a lot of the negative aspects of waterfall software development and so that and design was never really considered as part of Lean Startup it was never really considered as part of agile and so this was our attempt to really have that conversation. Using that foundation that a lot of the world was shifting over.

Brian Bell (00:08:57.126):

To do you recall kind of when you could sense that’s something, you know, to use a pun on your book here but you could sense something was wrong with with the process do you do you recall like a meeting or a aha moment. Where you’re sitting there it’s like there’s got to be?

Jeff Gothelf (00:09:10.728):

A better way this this sucks, you know, yeah I I remember this I I used to work for this EVP at America Online he was he was a difficult person to work with leave it at that um and I was a relatively young designer he micromanaged a 2500 person organization I get to give you a sense of of the the he knew everything he knew every pixel every detail every workflow every single. Thing and I remember one time I was working on some workflow and I printed out my design work we printed it out right because that’s how software is made on paper and uh and I remember walking into his office he had this big office and I walked in the front door and he was on the other side of of the of the room in in his chair and I’m walking over holding shaking with my piece of paper in hand and he goes from a distance I didn’t even get a chance to present. He says to me he goes there’s nothing remotely close to anything I want to see on that sheet of paper and I was I just, you know, spent a week of my life on this.

Brian Bell (00:10:04.265):

And I was like this can’t I cannot this this is not going to be my career if this is.

Jeff Gothelf (00:10:08.506):

The way that we’re moving forward and that was the beginning of the tipping point for me was this like like executive direction above all without any customer input it just didn’t make sense to me yeah but you.

Brian Bell (00:10:19.828):

But you couldn’t change it because of the top-down leadership in the way software was manufactured and delivered at the time when was it possible, you know, I think you can make a case that the whole lean movement was sort of enabled by first the internet which took, you know, five or ten years to kind of change how software was like built and delivered but also, you know, cloud right now all of a sudden, you could. Sort of like for free a couple guys with a laptop a couple kids with a laptop could iterate right do you recall sort of that how that technological shift felt at.

Jeff Gothelf (00:10:50.445):

The time I mean, it was it was incredible right I mean, the the interesting thing is that, you know, we at America Online I mean, this was a thousand years ago right we did use extensive user research but it was all based on a series of assumptions and then we just had to build and ship and then all of a sudden, right you could put up a website in no time you could show somebody a landing page in in a day or two days or however long it took to get it going you could you could prototype and Dreamweaver wow that was a tool right back then where you you had a WYSIWYG editor for websites it was it was revolutionary at the time and and ultimately super empowering because it allowed us to get from idea to feedback in a much much much shorter learning loop all of a sudden, so the cost of learning came down significantly so, you know, before it was like well we’re going to go spend two days in this research facility behind the one-way mirrors and, you know,. Eat a bunch of candy and watch 20 people tell us exactly the same thing and drop 20 grand on that that’s unscalable and unsustainable and now all of a sudden, to your point right myself and one other person could put something together show it to five customers get some feedback iterate on it and move that forward again in a much more condensed time frame and that that was.

Brian Bell (00:12:08.945):

That was revolutionary yeah and how does, you know, as someone who lived through that moment how does this moment feel different or the same with AI how is that kind of changing what Lean UX is and how you do you how UX gets done right you have things like cloud design and, you know, like that’s only going to get better like how does it kind of feel now on the ground as you sort of go out and work with with people on this it’s it.

Jeff Gothelf (00:12:31.957):

Number one’s incredible incredibly powerful and incredibly exciting and I think to some extent it’s also incredibly frightening because all of a sudden, you don’t need the designer or a person with a specific job title to bring your ideas to to life.

Brian Bell (00:12:47.910):

Yeah, that’s that’s seeming to be kind of the case everywhere, you know, the I I hear it all the time the designer’s going away the PM’s going away the engineer’s going away like all you have left is like I don’t know somebody with an idea who can talk to AI I don’t think so I don’t think.

Jeff Gothelf (00:13:00.929):

That I don’t think those roles are going away I mean, I mean, I know we’ve got these examples of the one-person startup these days or the two-person startup who does everything. I still think that’s a pretty rare case. I think the reality is, right, is that if everybody keeps prompting these things into existence, we’re heading towards sort of a kind of a trough of mediocrity where everything looks and functions and sounds exactly the same and we got a taste of this back in the Twitter days with Twitter Bootstrap Twitter Bootstrap I don’t know if you recall was a design system that Twitter put out and it was yeah it was a one design library yeah yeah exactly it’s a wonderful design system right and it came with with everything you needed it came with all the visual assets all the code everything you need in a beautiful website to kind of you can choose the pieces that you need put your website together and ship it and what happened every website in the world looked like Twitter Bootstrap right right there was no creativity there. Was no innovation there was no originality and and so we’re going to end up in the same place right like I can’t tell you how many times I read something. Or see something or get sent something I’m like this is Claude 100 Claude yeah it’s not you it’s. Yeah, right exactly and so what’s going to happen now is we are going to start to crave authenticity and humanity and originality and innovation in the work and that is where the humans are going to so cool let’s all bring our ideas to the next meeting terrific I prompted this you prompted that she prompted that thing great now let’s get all the feedback and then let somebody with actual experience and expertise put together the next version of this that not only. Is an improvement on what came into that meeting but solves a real customer problem. In a meaningful way and provides a unique opinion on the product or the service so.

Brian Bell (00:14:47.205):

That you stand out in the marketplace yeah I love this this I want to tug on this idea that you which is really interesting and and maybe this is a recurring pattern and in the history of, you know, technology society capitalism where we we sort of innovate and we come up with a mass-produced something like a Model T right like an automobile right and it’s like everybody now. Can have an automobile, you know, and it’s great everybody’s happy but then after? The mass production of something comes the customization of something right and sort of the proliferation of all of our, you know, endless like little iterations of desire right and and maybe that’s happening now with with sort of software right we’re getting the the really really the the Model T right which is like Claude or ChatGPT that can produce almost anything but it’s it’s really generic right it is like the Twitter, you know, design library that you just described of software and and and now we’ll get like a proliferation of even more design and taste and human ingenuity I wonder what you think about that.

Jeff Gothelf (00:15:47.460):

Yeah, I mean, look I I think ultimately the companies that are going to stand out are going have a strong opinion about how they provide value are relatively identical I’ll give you an example here’s a named example right, you know, I’ve had a Bank of America account for 30 years 30 years right their online software is still garbage for 30 years. But I’ve got some momentum there now by comparison, the layer like the actual usage layer I barely use their website it’s just sort of like the foundational accounts below it but the the the the UI layer for me for banking is Wise the company that used to be TransferWise why because I’ve got a daughter in university in London and I need pounds and I’ve got. Stuff to do in euros and I’ve got things to do in dollars and Wise creates a UI layer that’s digitally native that’s super easy to use that doesn’t mess around with all the all the kind of like the legacy intricacies uh that Bank of America and the Wells Fargos of the world have to deal with right and so I’m so loyal to that as the day-to-day tool because they have a strong opinion about what banking and transacting. Yeah, living in a multi-currency world should look like right as opposed to Bank of America that’s coming from 50 years or whatever of of historical inertia that that doesn’t allow them to move that forward and it’s it’s the people who have that opinion and then manifested that are going to win over this kind of like like ultimately look I have the 30 years of history with Bank of America but it doesn’t really matter to me that it’s Bank of America or Wells Fargo or Citibank or J.P. Morgan or Chase or whatever underneath it right because I use Wise every day that’s what matters.

Brian Bell (00:17:47.300):

To me it’s funny you mention that it’s it’s what you’re describing is the phenomenon of like UX as the differentiator right and we kind of saw that I think in the 2010s especially with Airbnb right, you know, because there were other Airbnbs like VRBO and and others before that that did vacation rentals yeah it just didn’t do it as good like in a in a very friendly way that Airbnb was able to, you know, differentiate by design.

Jeff Gothelf (00:18:12.580):

Differentiate by design right those guys were designers yeah.

Brian Bell (00:18:16.582):

That was their background right yeah from the Rhode Island one of the best design schools in the world right, you know, the banking example is prescient I I I bank, you know, for for Team Ignite Ventures our our early stage venture firm we bank with Mercury which is just a UI wrapper around another financial firm.

Jeff Gothelf (00:18:35.840):

But it just makes everything so easy and just it’s just delightful to bank.

Brian Bell (00:18:40.162):

With them I can upload, you know, when I invest in a company I can upload a safe and the wire instructions and just magically just takes care of everything and just with a few clicks versus I was with another bank I won’t name them because I don’t want to throw them under the bus but it was terrible just just just to send a wire was just ridiculous yeah there’s lots of steps.

Jeff Gothelf (00:18:56.447):

But you just said it it’s delightful to bank right delightful to bank is an opinion right it’s an opinion like if you’re saying look our our strong opinion about how we will differentiate in the market our strategy right is we’re going to be delightful to bank with okay then everything follows from that and you you can’t mire yourself in the sameness of the output of these LLMs right it’s just going to be insufficient to do that right there’s a sea of android phones out there and the difference. Between them I couldn’t tell you right but there’s only one iPhone right because they took a very opinionated approach jeff gothelf (00:19:33): to building a these decisions is evolving but I think we need them if we if we want to stand out.

Brian Bell (00:19:52.136):

In the market yeah so, you know, this show has a lot of founders who listen in, you know, we’ve backed a lot of founders so how does all this we’re talking about, you know, OKRs and design and UX like how should. How is that changing how is the advice changing now as we record this in 2026 versus what you might have said to a founder, you know, back in 2006 or 2016, you know, what is changing now on the ground where if you’re talking?

Jeff Gothelf (00:20:16.931):

To any founders look I think foundationally not much not not a whole like I think that understanding your customer solving a real problem for a real customer in a meaningful way meaningful way to them to be clear, not for you but for them and is still the recipe for success for building a successful business making evidence-based decision making I don’t think those. Things change I don’t they certainly were true in 2006 in 16 and certainly today in in 2026 the capabilities that we have what are our strong opinions should be and how to successfully meet their needs in a way that nobody else is doing is exponentially more powerful today right there’s. Literally no excuse honestly there is no excuse not to do this today you could argue you could argue that 20 years ago maybe even 10 years ago there was there was well there’s cost and there’s time that was always the excuse it’s you’re slowing us down with all this customer conversation stuff today there’s absolutely no excuse. Collecting the data finding people to talk to getting in front of them showing them it’s absolutely ridiculous to say that this is not something that you can do and and equally as importantly that’s actually only half the work right the other half is then taking all of that insight that you’ve collected and synthesized and actually changing. Marketing course based on what you’ve learned along the way right I get it founders have strong opinions they have visions they have a sense for where they’re going they want to get on that way terrific but but we’ve all seen those diagrams right a path is not a straight line that path is going to be a variety of twists and turns and those twists and turns should be driven by market insight customer feedback customer evidence what you’re hearing from.

Brian Bell (00:22:10.378):

The people that you’re trying to serve yeah and this is the perennial problem right when you’re going zero to one is, you know, problem what you’re describing is a process of like okay problem identification do people actually have this problem yeah right. Yeah, at scale and that’s like market validation are there lots of people that have this problem and then like will they pay for it and then and then solution validation, you know, hey here is this thing would you pay for it how do you kind of advise, you know, companies to go through these kind of phases it’s still kind of the same kind of Lean UX phases as it was, you know, 10 or 20 years ago it sounds like.

Jeff Gothelf (00:22:45.521):

It is it is right the the the the take like the the process of building that knowledge it can’t be faked right you can like I said you can get there faster and there’s a lot of there’s a lot so for example, right it’s a lot of talk right now about well I’ll just. Synthesize some users in ChatGPT and I’ll talk to them yeah there’s.

Brian Bell (00:23:04.466):

A couple startups actually that do that, you know, artificial societies I think came out of YC there’s a few others like that where they they simulate your users right they’ll just don’t buy it I don’t buy it.

Jeff Gothelf (00:23:14.712):

Like as a way as a way to practice your interviewing skills as a way to test your scripts right sure. Yeah, absolutely work with synthetic users but like truly like sitting with some ad copy maybe, you know, yeah exactly but like sitting with someone and watching them try to get through a workflow watching them having them show you how they currently do something and then asking them why they do that certain thing and then looking for those patterns across people it’s incredible I have a story years ago I used to work at a company in New York City called The Ladders The Ladders was a job board for people who made a hundred thousand dollars or more right that was the that was the deal and, you know, we we talked to customers every week that was a big part of the way that I worked and I kind of demanded that for my team and uh. We would talk to these executives on a regular basis and they would talk about text texting was just becoming a thing texting wasn’t huge yet right and they’re like we use SMS for communicating with with recruiters and I said why do you use SMS right and they said well I I don’t believe my boss can can read that can see that like an email I think my boss can read my email but text I can’t they can’t see.

Brian Bell (00:24:21.071):

That more secretive yeah.

Jeff Gothelf (00:24:22.630):

Yeah, there’s no way you’re going to get that out of a synthetic user right that that’s a pattern that comes out of actually having those conversations and there’s there’s so much nuance too right yeah there’s that’s.

Brian Bell (00:24:33.253):

The implicit reason like the expert like oh it’s easier that would be what people like some synthetic human would probably say oh it’s easier to just text yeah but really the implicit like core reason is like if you ask the five why’s and you get to the root cause right yeah like no actually I kept my hiding this for my boss.

Jeff Gothelf (00:24:50.319):

Yeah, the synthetic user doesn’t fear for their job right and so yeah it’s it’s that’s interesting but so that’s the key is again I don’t think you can shortcut that part of it I think you have to and look it’s not and it’s not that big of a deal like again I’m not saying go sit in a room for two days and talk to 20 customers that are, you know, after the fourth one they all say the same thing just talk to three yeah every week make that part of we do have a we have.

Brian Bell (00:25:16.948):

A portfolio company conveyor I don’t know if you’ve run across these guys but they basically do the qualitative interviews AI to human so that’s interesting. So, they’ll actually the AI will go interview all the humans and it’s growing super fast and so yeah and then you could go back and watch the videos and stuff but maybe that’s only a half solution in your mind right yeah no again like there’s.

Jeff Gothelf (00:25:36.359):

So, much again there’s so much nuance yeah the human nuance body language expression tone of voice hesitation, you know, like like reluctance to share something there’s no way there’s a human will pick that up and it’s so valuable or a human will pick up that pattern, you know, every time I talk to somebody and I and I bring up the job search aspect of it there’s this reluctance to even go there we’ve got to figure out what that is because that’s core to our value proposition. Whatever that is I just I don’t believe that a non-human entity can can decipher.

Brian Bell (00:26:10.246):

That at least not yet right maybe maybe in 10 years when we, you know, we get 10x better AI every year and it’s super intelligent and it can it can be more human than than we can right it’s overfitted to humanity maybe maybe how do, you know, when you have um if you’re a founder how do, you know, when you have a good problem to work on I I think look I think again I think there’s.

Jeff Gothelf (00:26:30.918):

The answer here is is, you know, the behavior change that you see in the folks that you’re talking to particularly when you’re not only asking them about how they’re currently solving a particular problem and where the challenges are so you see the patterns in that but then ultimately as you’re putting. Potential solutions in front of them you can see them light up you can see their behavior change you can see how they act with your prototypes how they act with the material that you’ve decided to share with them I think to me that’s the key right and those behavior patterns right we call we call those outcomes right so outcomes again sort of being the the push here right measurable changes in human behavior that drive business results and and we we know these things right we’ve had pirate metrics for.

Brian Bell (00:27:11.347):

A long time, you know, oh yeah mcclears pirate metrics that’s great reference I love those yeah.

Jeff Gothelf (00:27:17.030):

Yes it’s throwback but look again it’s one of those things that still makes sense because like that it was it was awareness acquisition revenue referral.

Brian Bell (00:27:26.224):

And retention revenue or something like.

Jeff Gothelf (00:27:27.586):

That that’s it it’s yeah exactly retention revenue and referral right but but those are measures of human behavior right and if if if you’re finding that people are looking for your thing and they’re landing on your page and they’re asking for more info or they’re engaging with with whatever it is you want them to engage with that signal. You’re solving something real right that’s to me that’s the thing that matters the thing that’s always mattered is the change in the behavior of the people that we want to serve to indicate that we’re doing something valuable.

Brian Bell (00:27:57.698):

For them how do you guard against cognitive biases in the in the measurement especially when you’re doing it more one-on-one with real humans there could be this tendency I think to I don’t know lead.

Jeff Gothelf (00:28:08.221):

The witness yeah look I mean, to me to me this this is and again coming back to this like which disciplines will survive the AI revolution and which will go away we have people who are good at this they’re called researchers and the nice thing about having experts who know how to do this is they can teach the rest of us. How to do this and they can look they can help train the bots I guess as well? About how to do this but ultimately there are lots of there’s lots of tips and tricks and techniques that come from practicing having conversations with people. So, that you don’t leave the witness and so that you don’t, you know, tons of stuff like you don’t show excitement you don’t agree you don’t indicate that you like an answer versus dislike an answer you ask open-ended questions, you know, when someone says oh I wish it did this you say things like oh. You don’t say wow me too I’ve been wishing for that forever you say if you had that what would it let you do and again what does that come back to behavior change what’s somebody actually trying to do again it’s kind of like there’s lots of examples you can use. But somebody says oh I wish it had calendar integration and you’re like okay if it had calendar integration what would it let you do well I wouldn’t ever miss. Any more meetings with my boss and I would know exactly when to leave to get home on time to see my kids soccer game okay so the real need is not to miss meetings with the boss and to make it home for your kids soccer game okay great calendar integration is a solution to that there’s an infinite number of ways to solve for that right and that’s again coming back to creativity and innovation that’s where the human aspect really shines through yeah I love that that’s really interesting.

Brian Bell (00:29:45.916):

So, I I guess in the end it all comes down to taste right and and sort of, you know, designers make maybe designers inherit the earth a little bit here right because in a world of AI slop and AI can generate everything it really just comes down to do I really understand humans and their needs and their pain points and and how we’re crafting this solution in a way do you think AI ever kind of gets there where it can it can do that better than we can.

Jeff Gothelf (00:30:09.634):

Look, at them at the moment I think there’s a lot of generalizable patterns that I can put forward that are good enough but again I I just kind of coming back to I think we end up in a sea of mediocrity right it’s kind of a race it’s a race to mediocrity I also don’t want to trivialize. Good product management good product managers good designers as to simply, you know, you know, complimenting them and saying oh you have good taste right because that gets that gets built that gets developed that gets trained that comes from doing the work it comes from trying something and understanding that it that it didn’t work this time yeah.

Brian Bell (00:30:43.252):

And that’s what people misunderstand about Steve Jobs, you know, they they imagine him as the armchair genius but he actually was out there talking to customers all the time all the time and doing focus groups and he was just talking to people and he was testing the ideas with people and he didn’t just sit there and just come up with that stuff from his office exactly he actually got out.

Jeff Gothelf (00:31:01.098):

Of his office and talked to people yeah amazing got got out of the building got out.

Brian Bell (00:31:05.740):

Of the building right yeah yeah whose whose phrase is that Steve Blank Steve Blank yeah yeah Steve Blank shout out to steve I love his books too so what are you telling your kids to work on now like as, you know, like I have three kids and, you know, should I tell them just be designers what is the career like what should people work on now it’s a great question.

Jeff Gothelf (00:31:23.454):

So, look I I have two daughters one is 23 and she has a I was doing when I started my career but the tools that she’s using the speed at which she’s working and the issues that she has to deal with are actually very very different right obviously she so she’s she works in in she works in Claude Claude Code she works in Claude Design she works in Lovable primarily in those tools to build to prototype to build and then to actually create code for for some of these things and so the speed with which she can work is incredible the output that she puts together from day one looks good which isn’t. Necessarily the best approach but the thing that makes it really really difficult is that the expectation that it sets for her stakeholders right stakeholders are like well, you know, when I was doing the work they were like okay here’s feedback on your wireframes or your prototype or whatever when can you have revisions done I mean, like well it might take me about a week or so maybe a couple weeks we’ll see you guys then for her they know how to prompt all her stakeholders prompt all day long right they’re like okay. So, we’ll see you tomorrow with iterations and she she’s like look I can certainly and what comes back out of these non-deterministic systems actually looks like a thing that we want to put in front of customers right and so that’s to me that’s that’s the fascinating she’s still solving the exact same problems that I was solving during my career but the the tension and the pressure that she’s under to deliver iterations and improvements are compressing her the time that she needs to actually do.

Brian Bell (00:33:16.885):

The thinking part right yeah a week or two has compressed to a day or two if if your uh counterparties will allow that right you’re the people you’re working.

Jeff Gothelf (00:33:26.412):

With yeah.

Brian Bell (00:33:27.892):

That’s really fascinating.

Jeff Gothelf (00:33:43.374):

Jobs that require opinions I keep coming back to this because I think again you I’m sure you’ve seen this right. But like people are just prompting things into existence and whatever comes out they’re there and they’re handing it off hey I did I did my job here it is right and they aren’t found in the work at all right so to me I think it’s the and I don’t I don’t mean to discount software engineers from this as well yes the job will change right but but having an opinion about how to build something what to build what it should look like and how it should behave is is the key to standing out in a world where people are just prompting things and then shipping it and assuming that it’s good enough and hoping.

Brian Bell (00:34:23.943):

That it’ll yeah I love that so of all the books you’ve written which one aged the worst and why.

Jeff Gothelf (00:34:30.842):

Which one aged the worst I think the one that aged the worst I wrote a very short book it was the fourth book lean versus agile versus Design Thinking it was. Really a long essay just a couple of chapters on reconciling the those various processes in a manner because I was working with organizations that were sort of hiring train like hiring training for lean and Lean Startup hiring training for Design Thinking hired training for agile and then trying to get all those people to play nice together when they’ve been trained with different vocabularies and different goals and different different targets and so that book was designed to help reconcile that conversation. I don’t know how prevalent that conversation is today anymore I feel like like Lean Startup has become sort of part of the general vocabulary rather than so much a process to be followed agile is losing some momentum these days as a I think the the foundation of of the way of working still stands I haven’t heard anybody say Design Thinking in years even. Though it’s a valuable process you don’t really hear it nearly as much as you did, you know, 10 15 years ago as well so of those four I think it’s that one jeff gothelf (00:35:39): simply because it leaned heavily into named processes like or brand name sort of frameworks that are are people have either moved on from or moving on.

Brian Bell (00:35:49.686):

From yeah this is funny back to the, you know, question how do you advise the the youth, you know, because I have three kids a couple teenagers I think I think the answer I’m honing in on is just, you know, study what’s interesting study humanity come up with opinions. Yeah, have a, you know, go for the general liberal arts education unless something really really speaks to you and you really want to study, you know, x y and z just a general human, you know, humanities human centered education I think so.

Jeff Gothelf (00:36:17.013):

And I think they’re headed that way look, you know, my younger daughter is 19 she’s in school and she gives me endless crap every single time I open up an AI tool. Endless and it’s all her friends are the same they’re kind of anti-AI yes they’re interesting they’re rebelling against this her sister her older sister too like they they roll their eyes. So, hard every time like let me just drop it in the Claude and see what it says. So, hard they they crave real world almost retro authenticity to your point of humanity I think you’re right I think the advice is is focus on the humanity side of things because I think. There’s going to be such an appetite for it moving forward that that’s I think there’s going to be a ton of value right yeah go go to.

Brian Bell (00:37:02.007):

The live show watch the musicians put the record on the record player hear the the, you know, the the warm analog sound right exactly crave crave that cup of coffee that you handcrafted and did a pour over in your kitchen.

Jeff Gothelf (00:37:16.197):

Embrace the flaws right like I think that’s part of it as well right like everything that everything that comes out of at the moment out of all these LLMs is. So, it it looks flawless and it complements you right great great job you’re so smart jeff that was such a great idea wow I didn’t even think about it that way jeff and here’s here’s here’s a polished out output of whatever it is that you asked for and I think I think that people miss the flaws the rawness of of of the human side.

Brian Bell (00:37:44.740):

Of things I really think it’s, you know, I actually I borrowed uh Mark Andreessen’s prompt I don’t know if you saw this he has a prompt for his AI and it’s it’s a really long one and it the gist of it is like accuracy is your most important trait your world expert you are not to agree with me you’re to point out my flaws and my, you know, when I’m wrong and this is this long thing about how to like tell your get your AI to tell you the harsh truth and so like now that when I talk to my AI it’s really harsh to me it’s like you’re like basically you’re an idiot what are you doing.

Jeff Gothelf (00:38:17.191):

This is the stupidest thing I’ve heard not not not not to flip the the interviewer interviewee chair but I’m curious if that makes you actually use the AI more or less because I think it makes me use it more right.

Brian Bell (00:38:27.717):

Because it it becomes this like imperfect harsh mirror to my to my thinking where, you know, like you’re talking to this like critic, you know, like, you know, and we’re using it to evaluate startups and evaluate pitch decks and, you know, should we invest in this company here’s the data room here’s pitch deck here’s our transcript call here’s resumes and it’s just like I can’t believe you’d even consider investing in this company basically is what it’s telling me I love that like wow okay like tell me more, you know, because I was really excited to invest in this company or vice versa right I’ll be like oh this is definitely a hard pass and it’s like no actually this is a strong yes and here you’re wrong and here’s why so okay so it’s you can actually prompt it in a way to yeah because it is a little sycophantic right more than a little yeah yeah well especially like the old ChatGPT I think 4-0 was like really people, you know, remember they protested when they took it away yeah. They lost their sycophantic friend right this thing that just like coddled.

Jeff Gothelf (00:39:22.646):

To them feels good when someone agrees with you all the time yeah.

Brian Bell (00:39:25.787):

So, everybody look up the Mark Andreessen prompt online and put that in the comments I guess that’s a good one well let’s wrap up with some rapid fire questions okay yeah so a popular idea in product management.

Jeff Gothelf (00:39:37.910):

That you think is quietly wrong there was an idea for a long time that product managers were the CEO of the product ooh and this is the Ben Horowitz correct.

Brian Bell (00:39:48.328):

Yeah, correct see the CEOs of nothing they’re the CEOs gathering everybody’s opinion yeah yeah exactly they can’t hire they have no budget like the CEOs of nothing yeah that’s funny they’re yeah they’re more like the uh the advisor to the product in a way yeah the uh well they got they got the backlog keeper of of ideas yeah it’s true it’s kind of true yeah you can tell I’m a little little saucy I was a PM for a long time a belief that you hold now that someone you generally respect disagrees.

Jeff Gothelf (00:40:21.914):

With oh that’s such a good one gosh I have here in my notes Josh Seiden josh Josh Seiden is my business partner and co-author for like the last 15 years I mean, he’s someone that I’ve I’ve greatly respect and we do we do disagree occasionally I think the the the disagreement is that you can. He believes that you can quantify qualitative benefits to customers and it’s something that I yeah and and I I I believe that you can quantify behavior driven by those by by sort of the the lack of those benefits or the the presence of those benefits. But he believes you can actually quantify the the benefits themselves the qualitative benefits and eyes still to the right yeah because.

Brian Bell (00:41:09.368):

The benefits are are basically like a bucket it’s like a problem it’s the opposite of a problem it’s kind of like a it’s a benefit benefit versus the underlying need, you know, the pain or the pleasure yes it’s a thought it’s.

Jeff Gothelf (00:41:21.715):

A feeling yeah.

Brian Bell (00:41:24.019):

Yeah, how painful is this if we take it away from you or like how pleasurable is it to have this versus like, you know, I can, you know, my phone unlocks with my face, you know, that’s a benefit right yep exactly but why is that like why is that beneficial why is that good yeah one thing you were right about before it was obvious and one thing uh you were confidently embarrassing wrong.

Jeff Gothelf (00:41:43.494):

I’ve been pushing for years and I think still to this day that to get out of the deliverables business in fact, like kind of my breakout article back in 2011 I think it was was called getting out of the deliverables business and I think we’re in in really perilous times to get back into the deliverables business now because producing stuff is so easy with AI right and so we like really just like focusing on the production of stuff is not.

Brian Bell (00:42:08.591):

The production of value necessarily right and so I feel like I’ve been right about that for a long time.

Jeff Gothelf (00:42:13.578):

As far as things that I was wrong about there was there was a thing like one thing that I, you know, particularly when I was formulating a lot of the Lean UX ideas and thinking about Lean Startup ideas there was we were trying to get teams to change the way that they worked and one of the things that teams really wanted for example, we talked a lot about in this conversation about talking speaking with customers right and there were a lot of organizations that were like okay well then we’ll just measure the number of conversations we have with customers and I was adamant that that was a vanity metric to use one of Eric Ries’s terms right because you could talk to a thousand customers or ten thousand customers and learn nothing if you’re talking to the wrong people you’re asking the wrong questions you’re not you’re not getting the data and I was I was totally wrong about that I’ve completely changed my mind about that I do believe that organizations that currently do not do any kind of customer conversations and start to go through the motions of talking to customers and they’re counting that like let’s say I just want you to talk to 50 customers this month right whatever it is inevitably you’re going to get better and more comfortable at that task so even just going through the motion and counting the motions is a great place to start so yeah.

Brian Bell (00:43:18.105):

Completely changed my well yeah it’s it’s kind of like the old phrase I, you know, 95 of success is just showing up right and doing the work, you know, just show up and do the work if you just show up and do the work I mean, that’s I tell my kids that, you know, I have one kid sort of struggling in school I’m like are you showing up and doing the work like now I’m like okay that’s the problem like if you just shut up and do.

Jeff Gothelf (00:43:36.526):

The work you’ll probably get a’s yeah and and look and you and the the repetition simply makes you better right this is it’s kind of like I used to use this in talks all the time the Karate Kid right the Karate Kid goes to Mr. Miyagi and he’s like hey I want to learn karate Mr. Miyagi is like terrific paint the fence. He’s like but I want to learn karate he’s like right on dude he goes wash the car yeah right and he’s like what like I don’t understand he’s like just do what I’m telling you and then like but he doesn’t know it but through the process of painting the fence and wax on wax off right he’s learning karate it’s the same thing right showing up and and going through the motions is is is actually a good thing initially eventually you’ll have to move forward and actually get value or deliver value out of those motions but at the very first show up and do.

Brian Bell (00:44:17.796):

The work really good advice a mistake you keep making despite knowing better this is, you know, it’s it’s.

Jeff Gothelf (00:44:22.780):

Embarrassing to admit this publicly and on a recording but, you know, there is a fruit for the better part of the last 15 years right, you know, 15 16 years I have been a staunch advocate for assumptions hypotheses experiments evidence-based decision making and then pivoting when the evidence contradicts your hypothesis right strong opinions loosely held right that I’ve been. I’ve been borrowing that phrase and using it forever and man I have such a hard time killing my ideas I love my ideas so much, you know, like really like to this day, you know, josh I talk to josh my business partner every day, you know, and he’s like dude he’s like this the we need to pivot we don’t have product market fit for this thing this isn’t working I was like come on I was like three more months we can make it happen we can make it work so to this day like I still 50 more interviews just get out of the building josh right exactly exactly I still struggle like we’re just. Interviewing the wrong people right exactly they’re, you know, these people don’t these people.

Brian Bell (00:45:21.974):

Don’t know what they’re talking about we obviously have the right solution for the wrong person we just got to go find the right people right.

Jeff Gothelf (00:45:27.567):

Yeah, exactly exactly what’s.

Brian Bell (00:45:30.168):

A question you wish more hosts podcast hosts like me asked you on a podcast what was.

Jeff Gothelf (00:45:35.109):

The most unlikely job you’ve ever had that has has had a kind of a profound influence on how you work today that’s that’s.

Brian Bell (00:45:44.032):

A question no one’s ever asked me I love it and now please we want to know the answer yeah.

Jeff Gothelf (00:45:49.085):

I worked in the circus I spent my first job out of college I joined the Clyde Beatty-Cole Brothers Circus and I traveled the east coast with a three-ring tented circus I’m not joking for six months doing circus shows. I saw the circus 400 times in a row that summer it was a very humbling experience it was a very difficult experience I learned a ton about a world that exists that very few people have any insight into sub-community of 200 people who live and work and and and, you know, celebrate and and argue with each other they meet out justice like it’s this really crazy crazy world of the circus I did sound and lighting by the way that was my job I was the sound and lighting technician for six months and I spent six months on the road and it just it just taught me so much about not making assumptions about people by truly getting to know the people around you um by to be super scrappy, you know, I had this super crappy PA system that was grossly inadequate for the size of the tent that we had to fill with sound the environment that we were working in was was really bad for the equipment it was we had elephants and horses and there was dust dust and and everything’s bouncing around from place to place and moisture and rain and humidity and so the microphones were breaking and the cables were breaking and it was a lot about like okay like I gotta hit this lighting cue and solder this microphone back together so that the trumpet solo can be heard and like it was a lot of stuff like that, you know, I was 22 and it was the most responsibility I ever had because, you know, this community is relying on everybody to pull this off in front of 4 000 people twice a day and it was it was a fascinating six months to say.

Brian Bell (00:47:32.223):

The least that is a question that’s going in the in the uh the stable when do you think I should ask that you think it’s right it’s good right here in the uh the the rapid fire section or you think I should ask that at the beginning like after the background question. What’s the most unlikely experience that you’ve had that kind.

Jeff Gothelf (00:47:47.078):

Of impacted you yeah that’s I like it at the end because because yeah can I because because at this point I’ve kind of shared a lot of my background right yeah and now you kind of get a little bit of the justification for it yeah.

Brian Bell (00:47:57.045):

A little little metacognition here of podcast making but yeah I really appreciate the conversation jeff really enjoyed it where can folks find you online jeffgothelf.com senseandrespond.co and of course LinkedIn always the easy places to find me all right well thanks uh.

Jeff Gothelf (00:48:11.876):

So, much for coming on my pleasure brian thanks so much for having

Adam Nash is building around a very different question: how do you help people give it away better?

Nash has spent decades at the center of major consumer technology and fintech shifts. He was VP of Product at LinkedIn through its IPO, President and CEO of Wealthfront, VP of Product at Dropbox, and previously held roles at eBay and Apple. He is also a prolific angel investor, with early investments in companies like Figma, Gusto, Opendoor, Firebase, and more.

Today, he is co-founder of Daffy, a modern donor-advised fund platform designed to make charitable giving easier, more intentional, and more accessible. In the episode, Nash explains why giving has been one of the most underbuilt categories in consumer finance—and why the donor-advised fund may be a much bigger product opportunity than most people realize.

Subscribe now

Money Is a Trust Business

Nash’s interest in financial products started early. In college, after earning what felt like a large amount of money from an internship, he quickly realized he had spent far more than expected. That experience pushed him to learn about savings, mutual funds, returns, and financial decision-making.

Over time, that curiosity became a career thesis: technology keeps getting more powerful, but the most interesting products sit at the intersection of rational systems and irrational human behavior.

That lens shaped his work at Wealthfront. Managing people’s money is not just a math problem. It is a trust problem. People are not optimizing spreadsheets in the abstract. They are trying to build lives, care for families, retire comfortably, reduce anxiety, and make decisions they can live with.

For Nash, the lesson was clear: in financial products, the emotional layer matters as much as the technical layer. The product must be accurate, safe, and reliable—but it also has to understand how people actually behave.

What Wealthfront Taught Him About Company-Building

Running Wealthfront gave Nash a broader view of what it takes to build a company beyond product strategy.

His definition of the CEO role is blunt: set the strategy, find the right people to execute it, and make sure they have the resources to succeed. Get those three things right, and a company can survive a lot of mistakes.

He also argues that culture has to be built early. At scale, behaviors are already locked in. The habits, incentives, and standards created in the first phase of a company become incredibly sticky.

That belief extends even to hiring. Nash recalled wanting every candidate—whether they got the job or not—to leave with a clear, positive understanding of what the company did and why it mattered. In his view, every interaction with the company is part of the brand.

The eBay Lesson: Operational Excellence Can Become a Trap

One of the most interesting parts of the conversation was Nash’s comparison between eBay and LinkedIn.

At eBay, he saw an extraordinarily disciplined product organization. Roadmaps were prioritized with financial rigor. Hundreds of features were evaluated, scheduled, and shipped with remarkable precision. By conventional business standards, it was an elite execution machine.

But that strength came with a cost.

Nash argues that eBay was wound so tightly around operational efficiency that there was less room for exploration, innovation, and riding new technology waves. The company was great at optimizing the current machine, but that made it harder to reinvent itself.

LinkedIn taught him a different lesson. Reid Hoffman’s deep understanding of network effects shaped Nash’s view of platform-building: the core product matters, but so does planting seeds for the next 10x opportunity. Great companies do not have one story. They compound through multiple waves.

That contrast became part of Nash’s operating philosophy: efficiency is valuable, but if it crowds out experimentation, it can become fatal.

Why Daffy Exists

The idea for Daffy came from Nash’s own experience with donor-advised funds.

After LinkedIn went public, he faced a set of financial decisions around taxes, stock, and charitable giving. His accountant introduced him to donor-advised funds: a structure that lets someone contribute assets, receive the tax deduction, invest the funds tax-free, and later recommend grants to charities.

Nash immediately saw the product as powerful. But he also saw how inaccessible it felt. Donor-advised funds had historically been associated with wealth managers, high-net-worth individuals, and legacy financial institutions. Most people who give to charity regularly had never heard of them.

That was the opening for Daffy.

Nash describes a donor-advised fund as something like a 401(k), IRA, wallet, or even HSA for charity. The idea is simple: put money aside for giving when it is financially convenient, then donate later when inspiration or need arises.

That separation matters. Giving usually involves two hard questions at once: how much can I afford to give, and where should I give it? Bundling those decisions together creates friction. Daffy’s goal is to split them apart.

Giving Is Emotional, Not Just Financial

One of Nash’s strongest product beliefs is that the best consumer products touch people emotionally, not just rationally.

He points to Apple as a company that understood this deeply. Photos are not just files. They are memories, children, family, and life history. The best products understand the human meaning underneath the task.

Daffy applies the same logic to giving. Charitable giving is not just a tax optimization problem. It is tied to values, family, identity, religion, schools, community, disasters, causes, and the desire to help.

That emotional insight shaped Daffy’s product decisions. The company launched mobile-first, supported crypto, added donor-advised fund transfers after users immediately requested them, and built features like family plans that let children, spouses, siblings, parents, and grandparents participate in giving.

The family plan example is especially revealing. In wealth management, people talk constantly about multi-generational giving, legacy, and family values. But most donor-advised funds were still structured like individual or joint brokerage accounts. Daffy asked a simple product question: why doesn’t giving have a family plan like every other modern consumer subscription?

That led to a feature where families can give together, children can recommend donations, and charitable giving can become a dinner-table conversation.

The Business Model Bet: Membership Fees Over AUM

Traditional donor-advised funds often charge fees based on assets under management. Nash argues that this model makes sense for investment products, but not necessarily for giving.

The work required to administer a very large account is not thousands of times greater than the work required to administer a smaller one. Yet AUM-based pricing naturally biases the product toward wealthy users.

Daffy’s contrarian move was to charge a membership fee instead. That supports the company’s broader ambition: make donor-advised funds useful not just for the ultra-wealthy, but for the tens of millions of American households that already give to charity each year.

The product is not trying to convert non-givers into givers. It is trying to help people who already give do it more consistently, more intentionally, and with less friction.

Are Donor-Advised Funds Just Warehouses for the Rich?

We raised one of the strongest critiques of donor-advised funds: that they allow wealthy people to park money, get tax benefits, and delay actually sending funds to charities.

Nash’s response was nuanced. He acknowledged that the concern is technically possible, especially at extreme wealth levels. If policymakers want to create rules or caps for very large accounts, he is open to that conversation.

But he argues the critique is distorted by an obsession with billionaires. Most people using donor-advised funds are not trying to warehouse billions. They are giving to schools, religious organizations, local causes, national nonprofits, and global crises.

He also points to payout behavior. According to Nash, Daffy’s own numbers show that more than half of contributed funds are granted out to charities the following year. His larger point: focus on the average use case, not just the most sensational edge case.

The Angel Investing Framework

The episode also goes deep on Nash’s angel investing philosophy.

He has invested in roughly 160 to 170 companies over 14 or 15 years, but he does not treat angel investing as casual check-writing. He runs it more like a personal venture portfolio, deciding how much of his overall assets he is willing to allocate to startups, then pacing that capital over a decade.

That decade-long view matters. Seed investing takes a long time. The best companies may take 10 years or more to reach liquidity. Many angels get excited in year one or two, then realize in year three that none of the money is coming back yet.

Nash looks for a few things. First, he wants to understand why the founder is talking to him specifically. If the answer is just money, that is not compelling. He wants to add value through relevant expertise in product, fintech, marketplaces, social, or growth.

Second, he listens for distribution. A product insight is not enough. The founder needs a credible path to reach customers and build a venture-scale company.

Third, he looks for founder-market fit. Not just “this founder found a way to make money,” but “this founder cares about this problem enough to spend a decade on it.”

The Venture Paradox: Saying No Sounds Smart

Nash also offered one of the sharpest lines in the episode: in venture, it is easy to sound smart by saying no.

There are always reasons a startup will fail. The market is too small. The timing is wrong. The team is incomplete. Distribution is too hard. The product is too early. Saying no can make an investor look disciplined in the short term.

But in the long term, nobody remembers the companies you correctly passed on. Returns come from the moments when you said yes.

That has shaped how Nash thinks about founders. Some of his best investments came from founders who convinced him he was wrong. Dylan Field did that with Figma. Nash initially questioned whether cloud-based graphics tools were ready, but Field had a deeper view of GPUs, bandwidth, latency, WebGL, and why the market would move.

The pattern he now looks for: a founder who has gone deeper into a market than he has, and can change his mind.

The Big Takeaway

This episode is not just about Daffy or donor-advised funds. It is about how great products emerge when founders understand both the rational and emotional sides of human behavior.

Wealthfront was about trust. LinkedIn was about networks. eBay was a lesson in the limits of operational precision. Daffy is about turning generosity into a better-designed habit.

The most provocative idea is that charitable giving may be one of the last major consumer financial behaviors still waiting for a truly modern product experience.

People already want to give. The question is whether the product makes it easy enough, intentional enough, and meaningful enough for them to do it more often.

That is the bet behind Daffy: the next great fintech product might not help people keep more money.

It might help them give it away.

👂🎧 Watch, listen, and follow on your favorite platform: https://tr.ee/S2ayrbx_fL

🙏 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast

Chapters:
00:01 - Introducing Adam Nash and Daffy’s Mission
02:20 - Adam’s Origin Story: Money, Family, and Human-Computer Interaction
05:24 - Fintech Before “Fintech” Had a Name
06:16 - What Wealthfront Taught Adam About Trust, Culture, and CEO Leverage
10:15 - Operator Playbooks from Apple, eBay, LinkedIn, and Beyond
11:27 - LinkedIn vs. eBay: Network Effects, Operational Excellence, and Missed Waves
15:28 - From Wealthfront to Greylock, Dropbox, and the Daffy Idea
18:00 - Donor-Advised Funds Explained and Why Daffy Exists
22:27 - The 401(k), IRA, or Wallet for Charity
25:03 - Daffy’s Business Model: Membership Fees Over AUM
27:11 - Product Innovation in Giving: Transfers, Family Plans, Crypto, APIs, and Private Stock
32:54 - The Donor-Advised Fund Critique: Warehousing Money or Unlocking Giving?
37:57 - Teaching Personal Finance for Engineers at Stanford
41:37 - Adam’s Angel Investing Framework After 160+ Startups
43:54 - Why Seed Investing Takes a Decade
46:12 - Founder-Market Fit, Distribution, and Knowing Why You’re on the Cap Table
48:28 - The Venture Paradox: Saying No Sounds Smart, Saying Yes Makes Returns
50:37 - Figma, Dylan Field, and Founders Who Change Adam’s Mind

Transcript

Brian Bell (00:00:58.278): Hey everyone welcome back to the Ignite Podcast today we're delighted to have Adam Nash on the mic he spent two decades building consumer financial products people actually use he was VP of product at LinkedIn through his IPO president and CEO of Wealthfront and VP of product at Dropbox with earlier stops at eBay and Apple since 2011 he's been one of Silicon Valley's most active angels early in Figma Gusto Opendoor Firebase and many others in 2020 he co-founded Daffy the donor advised fund for you which crossed a billion dollars in charitable assets in about four and a half years pretty amazing thanks for coming on, Adam. Yeah, Brian, great to be here yeah and so I, I put you on the spot before and said hey we actually know each other I interviewed at Wealthfront 10 years ago didn't get the job but I uh, you know, had a real good time in the interview process I remember it being like.

Adam Nash (00:01:43.501): A really fun interview well I, I'm glad it was a good experience I mean, well film was a very special moment it was one of the really early companies that managed to scale in fintech and um I was a big believer I still am that culture is so important and so I, you know, the process of how you bring people on board I think tends to be underappreciated by a lot of operators but it's this one window you have to kind of introduce people to the company and culture and.

Brian Bell (00:02:11.190): You have to remember that every person you talk to whether or not they get.

Adam Nash (00:02:13.993): The job has an impression of your company and so you're, you know, similar to my advice on building brands and products, you know, very often people don't think enough about the people who haven't used their product yet versus their active users so yeah.

Brian Bell (00:02:27.065): And I told uh, you know, I had such a good interview experience and went on to do other things, you know, leading AI at Amazon and and working at Microsoft so I'd it's not like I set back my career not getting the job but I had a really good experience and I remember just telling people from then on about Wealthfront I was like, you know, when people would talk about finance or investments and check out Wealthfront so yeah that, that positive experience kind of probably paid some dividends and probably got some customers from my network over the years.

Adam Nash (00:02:53.003): Yeah, that's actually great to hear like I said it was it's one of my long-standing kind of startup kind of building a brand and culture things as I always told the head of recruiting and, you know, all the hiring managers that like I want everyone leaving the company whether or not they get the job understanding what we do and why we do it and feeling good about the company.

Brian Bell (00:03:10.404): Yeah, that's amazing it's so funny to cross paths again but I'd love to just start with uh what's your origin story and background? Well, there's there's a lot.

Adam Nash (00:03:16.974): Of versions of that I mean, I, I have pretty diverse interests although I think that if you look at my LinkedIn profile it looks pretty standard um, you know, software engineer, you know, builds a career in Silicon Valley going from one tech wave to another but, you know, when you mentioned fintech and, you know, financial products in general when I look at that I mean, it goes back a long ways, you know, I remember my grandmother retired while I was in college and no one ever really had taught me about money and so I remember my first internship where I actually earned some money I think I earned about twenty two hundred dollars a month which at the time was an unbelievable sound yes that's yeah and then what happened is, you know, I went through the the fall quarter and then, you know, after buying a new computer probably for too much money and uh, you know, a little speddy here and there I realized after Thanksgiving I had less than a thousand dollars left and I was like oh my god like I, I can't believe I went through all that money and so I started on this process of learning about money and my grandmother I just retired she was into certificates of deposit and she was the type of person who did their taxes just with their just pencil and paper all the calculators by hand she was amazingly gifted that way and she taught, you know, in school and so started learning about mutual funds and returns and, you know, all these different things I actually talk about her a bit I teach.

Brian Bell (00:04:33.200): This class at Stanford now personal finance for engineers it's based on some of those.

Adam Nash (00:04:37.300): Learnings but, you know, when I, I think about it academically I ended up focusing my graduate school on on human computer interaction which is really just the study of, you know, the rational which is computers right ones and zeros and, you know, will always do the same thing and and humans that, you know, you know, we have emotions we have opinions about things and so I just built a career that was always on this intersection of computers get more and more powerful every year what are problems that people have that all of a sudden now computers can tackle or that people are willing to trust them with and so that, that led me through kind of the personal computer era into, you know, Web 1.0 and Web 2.0 and then fintech but the truth is I've always cared a lot about personal finance I think that money impacts our lives more than we like to give it credit for um and I don't just mean dollars and cents of what you can afford it's are you happy how do you live your life what do you spend your money on your loved ones, you know, what life do you want to build and I think so many financial decisions end up impacting things we care about but we don't think about it we don't have that intentionality and so not surprisingly I think that, you know, when this last fintech boom happened and I met great founders who were building companies in the space I, I got the itch I was an EIR at Graylock and I remember meeting the folks at Wealthfront and going like I think there's something special here I think we can build something that's meaningful for people and, you know, here we are 15 years later.

Brian Bell (00:05:56.913): Yeah, and it's still around right yeah of course I, I just saw.

Adam Nash (00:06:00.038): The numbers it's amazing to me I think that there's almost one and a half million people now who trust their finances with Wealthfront I mean, they're almost at a hundred billion dollars I mean, these numbers just seemed ludicrous you have to understand and in 2012 mostly there wasn't even a term fintech um it's one of the reasons there was an opening for me to talk to so many great founders but, you know, that idea most fintechs at the time had built these platforms and they managed to get people to trust them with maybe 20 million 30 million 40 million they all seem to hit a wall they'd get these early adopters and then they couldn't go farther and so seeing companies like Wealthfront scale to where they are now not just Wealthfront Robinhood Acorns Coinbase, you know, and crypto really makes me feel good about the fact that we're now using technology to help people with their money.

Brian Bell (00:06:44.554): Yeah, that's pretty amazing what did running a company like that, that manages billions and billions of dollars of other people's dollars uh what did it teach you that a product being.

Adam Nash (00:06:52.866): A product exec never could that's a good question uh there's two angles to that, you know, I think that when I went to wealth I think I really believe strongly was at its heart money is a trust business right, you know, like I said money is rarely the thing that people are focused on like that's not their goal in life there are a few people who have that life but for most people it's a means to an end but it's a trust business and it's emotional I believed heavily in behavioral finance and so actually a lot of, you know, I got to Wealthfront we were very small when I when I went to Wealthfront I'd been one of the early customers actually that's how they found me but uh, you know, I, I think it was about 80 million in assets at the time but I was a very big believer in building cultures and companies that you you have to build the right behaviors when you're small because when you're big you don't have time to right like things are locked in and so I was always very focused on that, that was very though kind of like building a company like a product so I think that's a direct lesson from being a product executive as you realize that, you know, the bare bones of what you put into a product when it's early even that MVP are amazingly sticky and durable and and getting those bones right are really the difference between, I think, talented product designers doing that kind of zero to one problem versus the one to many. But I don't know, I loved jumping in. One of the reasons I wanted to be a CEO is because I think that product strategy is only one of a handful of strategies you have to get right to build a truly great platform and organization right you have the product strategy but there's also a people strategy who are the right people to build and run and operate this organization what culture do you put in place who do you hire how do you train them what are you going to reward over time um there's a technology strategy right like your engineer you realize that technology is continent, you know, moore's law metcalf's law all these laws that we have about scaling and and growth are reality for the industry and I had been at some companies that didn't do a great job of riding those next waves of technology and so you have a technology strategy as well as a people and product strategy and then of course, you know, being on the venture side what's the financial strategy why why are investors going to fund this ride for you like how are you gonna what are the proof points along the way that you will scale to get to the point where you can actually build the kind of business that will fund itself on an ongoing basis not just for years but for decades so I really love jumping in to do all those things I remember I got asked once and in all hands, you know, it's classic kind of we were doing college recruiting so it's right I call it I, I actually love it because there's always one or two folks who are not afraid to ask the hard question and so I got the question once it was like what does.

Brian Bell (00:09:31.744): A CEO do now what would you say your job is yeah actually yeah.

Adam Nash (00:09:35.727): As it turns out there's hundreds of things and if you're doing startups there's there's no job too small for you I always look around like, you know, at some point I'm fixing the Wi-Fi and making sure that like you're doing all the little things in the beginning but my answer at that time was, you know, I said at the highest level I think CEOs of startups do three things I think they set the strategy for the company they find the right people to execute on that strategy and then they make sure that the team has the resources to execute right and at some level if you get those three things right you can do a hundred other things wrong and move along but that probably just reflects my philosophy about how you build organizations that, you know, stand the test of time.

Brian Bell (00:10:15.353): That's amazing how much wisdom I mean, I could tug on any one of those those sentences that you just said but I wanted to ask, you know, you've worked along under and alongside some famous operators whose playbook did you steal and who did you just deliberately throw out.

Adam Nash (00:10:29.255): Well, it's interesting because I actually have seen a lot of amazing players I mean, I started my career I had so many internships but I started my career full-time at Apple and so it's hard to although at the time I will tell you Apple was not everyone's dream comparison I mean, there were some people who loved Apple but this was the era of Businessweek cover like, you know, the black cover like the the fall of an american icon or the death of an american free ipod post, you know, I actually had a roommate when I took the job who who bet me that within five years Apple was going to be bankrupt so it was that's how dark the time was but and then you had, you know, like companies like eBay, you know, other startups etc but, you know, I think that mostly I mean, LinkedIn was a really formative experience for me I mean, reid hoffman is really a generational strategist in in a number of ways but his intrinsic understanding of kind of network effects businesses and how you build these platforms I think that for me I, I've always looked at the way you ride these waves is you're building out the core product and you're adding value customers but to build that business and platform what are the seeds that you're always planting for that next wave where's the next 10X going to come from because it's never one story right and and and so much of what I learned at LinkedIn was really in direct contradiction to what I'd seen at eBay I mean, eBay's an amazing company a platform I still use eBay I still know people at eBay and and uh I have so many strong feelings and positive feelings about the platform and what it enabled but when I also look at it I, I can't help it I'm a product guy I also think about what might have been what could have been and when I think about my time when I was at eBay it was my first internet product management role was at eBay I'd done really computer software before at PC.

Brian Bell (00:12:10.284): And servers they're also well known for having an amazing product org there at eBay historically especially during the time you were there.

Adam Nash (00:12:16.651): That's right it was amazing but it was great at one thing I would describe it as being fantastically great from an operational standpoint I would call, you know, there's product excellence there's operational excellence there's service excellence eBay's product org was really focusing on operational excellence right it I cannot tell you how every cost every efficiency every developer day um that roadmap that went out the prioritization I was as I rose in the ranks I was on a group a committee that every week met for hours going through over 600 different features that were prioritized based on an NPV that was uniform across the projects dollars and cents finance coordinated and by the way had it had a ship date on it and I will tell you I was there 99% plus features met their ship date like if it said if I said it was going to launch on december 12th it was going to launch on december 12th by every MBA standard I just went to a reunion but but by every business standard by every operational standard standard was unbelievable I mean, semiconductor companies are probably more efficiently run but just just barely it wow but then what went wrong and to me some of the product lessons I talk about this when I talk about different product lessons where did it go wrong where did that operational excellence it was wound so tight from a people standpoint and from a technology standpoint there wasn't a lot of room for exploration and innovation not a lot of room for riding, you know, those next technology waves I mean, the truth is our industry is brutal every five years the reason every founder lives under the the fear of if I was starting this company today what would I do differently is because every five years or so the game really does change like was the best solution is no longer the best solution what was the right approach is no longer the right approach and so when I look at eBay like they were wound so tight it made it very difficult for them to ride the upcoming waves I mean, you kind of saw that and so that affected a lot of of how I thought, you know, going into LinkedIn and and, you know, Wealthfront Dropbox, you know, all these other companies every company one thing I love about Silicon Valley is that no matter how great the companies are and every wave there's always a bunch of people who are making the of all the things that the company did wrong and how they do it differently and say hey next time we're going to do it better which is why frankly it's just a brutal industry in some ways right like this is why every five to ten years, you know, the the companies we previously celebrated as best in class all of a sudden are on their heels right, you know, from the next wave of companies because of that relentless growth mindset not just around the product and technology but how to build better organizations how to build better software I mean, we're going through one right now which is unbelievable.

Brian Bell (00:14:53.838): Right, so let's close the loop on wealth front to to Daffy because I want to kind of there was a like a four-year period there what are you up to it's the late 2010s and how did you get involved with with Daffy well I mean,.

Adam Nash (00:15:04.461): After I handed off wealth front back to Andy I ended up going to Graylock as an EIR I'd been there before I, I, you know, people have mixed feelings about these roles in venture capital firms and it seems like every year there's.

Brian Bell (00:15:16.764): For anybody listening who doesn't know what that what is an e.

Adam Nash (00:15:19.225): Actually, that's a great question because nor it's kind of an entrepreneur in residence or in some cases an executive in residence it can go either way but fundamentally you're sitting there with a lot of smart people and you're meeting with a lot of founders I'm an active angel investor so investing as well well and actually I met a lot of great founders in that turn I think it was in 2017 or 2018 met the folks at Bitwise became kind of an investor advisor there and a number of other companies but uh but fundamentally it gives you time to sit there and think you have a lot of smart people around you you see the frameworks when you're in an operational role when you're building a company you have to be heads down like you you just don't have as much time as you'd like to look around and think hey if I was starting with a blank sheet of paper what would I do today and so I spent about a year a little bit more thinking about different ideas and things to do next Dropbox had just gone public and so I jumped in there the Graylock company jumped in there to help drew on the product side ended up taking up growth too at I think at one point I had about 90 of the revenue rolling up to me trying to figure out what their next generation would be but I just had an itch I'm still working on the ideas and and different companies I had this list of ideas going back to Wealthfront days even before of great financial products that hadn't been reinvented yet and actually turned out the donor advised fund was one of those products and so I just had some trouble in the beginning thinking about how do you turn this into a venture class opportunity but yeah I, I know it sounds surprising but what happened was the pandemic happened I had left Dropbox already to work on a startup the pandemic happened, you know, Mike co-founder and I we've been talking for years about doing a company together he was one of my favorite engineers to work with at LinkedIn back in the day nice and we decided to take the leap in the middle of 2020 to actually start what is now called Daffy so here we are five years later.

Brian Bell (00:17:04.410): Amazing, and so I kind of get a donor advised fund Daffy tell U.S. What that is what is a donor advised fund what did why a startup around it what was kind of broken there and yeah tell U.S. The whole story.

Adam Nash (00:17:15.538): Yeah, well I, you know, the name you can blame me for yeah you have an engineer naming things yeah it's gonna be on the nose I thought it was a good app name though but the donor advised fund for you I, I really liked just the simplicity of it and also because I was actually a true believer I had discovered the donor advised fund before wealth front actually when I was at LinkedIn when LinkedIn went public, you know, like a lot of people all of a sudden had a lot of financial decisions to make made the good decision to hire a professional accountant really talented and one of the things he said was, you know, well the taxes are going to be a serious issue here, you know, have you thought about charitable giving right in this money I said well like a lot of people I believe in charitable giving but I hadn't given a lot of thought to, you know, well who would I give this much money to and how would I think about it and LinkedIn was one of these companies where the the lock expired in november so there wasn't a lot of time and so as a result um he said have you heard of a donor advised fund like most people I hadn't but he was like you take some of the stock you put it away cuts your tax bill money's invested tax-free and you have time to actually figure out what your strategy is about giving and so that really sat with me it was I always thought the donor advice was an amazing financial product and so the the real insight for Daffy was just really thinking of it that way, you know, most people think of giving as something purely it's almost I don't want to say it's cultural but it's moral it's ethical we, we teach our children to give like I was raised to believe that it's not all for you right that, that that some of what you earn etc goes back to to those less fortunate than yourself and.

Brian Bell (00:18:46.387): So, I love working out can't you just put it into a trust and give it to your family tax-free and.

Adam Nash (00:18:50.530): There are many options as it turns out but no I, you know, and that struck a chord with me because I will tell you one of those challenges in fintech all software products is that especially in the consumer space is, you know, engineers and MBAs don't have a lot in common but but what one thing that they have in common is they love numbers very left-brained very rational they like to go through the logic deductive etc I find that the best way to design products are really stick with people right that are really meaningful if you want to build a brand not just rationally but irrationally this is one of the things that Apple actually gets right consistently and has over decades it's good to work in an area and go to the where the heat is what are the emotions behind it why do we care I mean, our photos aren't just photos those are your children when they were young right those are those are the those are memories that's that's a life lived I Apple always had a way of getting to the heart and its design culture of what really touched people and the same thing is true in financial products and so when I saw giving I was like this is actually at the heart of it giving is more than just a financial goal it's not just a budget line item and so that juxtaposition of the rational which is wait, you know, the average 50 to 60 million american households give to charity every year everyone's trying to figure out how much they can afford when they can give it they're asked all the time there's probably a lot of lessons in personal finance about how to make people more generous how to help people with their giving but also on other side how do you make it more meaningful why do people get how do you touch my mind when I look at Daffy some of our best feature ideas have actually come from that emotional side of really asking the question how does giving fit in people's life I mean, the dollars and cents are exciting and I love rolling out features, you know, where people can build portfolios and it's automatically rebalanced and it's multi custodian all these different things that we do I think are fantastic but a lot of things I think we got right from day one was actually just talking about the emotions talking about why people give why it's meaningful to them all the different traditions around it and then saying hey how do we turn this into software how do we turn this into an app into a product that people can use and so anyway I love that process.

Brian Bell (00:20:56.249): So, yeah explain how it works is it basically it sounds like it almost kind of works like Wealthfront for nonprofits tell U.S. Like kind of what it is yeah.

Adam Nash (00:21:02.979): The donor advice fund is a really simple product right and so um and most people are surprised and and we can't take credit for inventing it donor advice funds have been around in the U.S. For almost 100 years I think the first one dates back to like 1931 for most people though it became big when the the giants fidelities the schwabs the vanguard started rolling it out in in the early 90s so it's been about 30 years but the truth is most most people haven't heard of it right if you don't have a high-end wealth manager um financial advisor or an accountant um you probably hadn't heard of it but the donor advice fund is almost exactly what you want right, you know, so giving is two hard problems right how much can I afford to give and then who do I give it to the donor advised fund is kind of like a 401k for charity it says okay, you know, that you give regularly right put money aside when it's convenient for you in this special account you get the tax deduction right away for for giving money to charity the money's invested tax-free and then whenever you're inspired to give you just name the charity and the donor advised fund, you know, sends.

Brian Bell (00:22:04.345): The money off so yeah it's like basically like a like an IRA for giving.

Adam Nash (00:22:08.673): Yeah, it's like an IRA for giving some people in the technology think I think of it more as like a wallet for charity you can you can think of it as a number of different ways but it turns out intermediation has this value right and the value of charitable giving is really large because very often the times where you have.

Brian Bell (00:22:26.266): The inclination to put money aside for charity as a budgeting kind of financial goal.

Adam Nash (00:22:31.470): Is different than when someone great products simple it's just separating out problems so that you can tackle them one by one right like not you don't you don't bombard the user with all the things they have to figure out at once right like you it's a journey right through the product and learning and so I think donor advised funds do the same thing for charitable giving for millions of people and so the space has been growing rapidly but no one had really taken the approach of saying how do we turn this into a great product most of the incumbents just treat it as an attach rate feature.

Brian Bell (00:23:13.584): For their advisory platform it's just another account with with stocks and bonds.

Adam Nash (00:23:17.265): In it right it's just another account right and and so they have, you know, individual they have joint yeah yes you can have stocks and bonds and funds it's kind of like a 401k they have pools but they're surprisingly not built for giving and and by the way I don't fault the incumbents for this it's not their business model they, they all picked a business model they borrowed from their investment side where they charge a percentage of assets which by the way is a great business model for some businesses as you mentioned Wealthfront uses that business model but it turns out that the problem with that business model for giving is really that money is very non-linear right it it it, you know, we all know and so if you make money it turns out the work you have to do for a billion dollar account isn't a thousand times more than a million dollar account let alone a thousand dollar account I mean, the average american gives you're talking about a a few hundred dollars a few thousand dollars a year to a handful of charities and so I, I really thought that, that was the opening to build a great product not just for the ultra wealthy but for everyone for those 50 to 60 million american households give to charity every year and it felt like the donor advice when you're building a financial product I think the regulatory framework is incredibly important we weren't going to invent a new account for this the donor advised fund was already there and so I mean, in some ways you stand on the shoulders of giants, you know, that a lot of people came up with the donor advise fund figured out what you could do with it I think we just took it into the 21st century, you know, I mean, when we launched Daffy there wasn't even a fully functional donor advised fund in the app store in 2021, you know, and so um we launched mobile first we launched with support for things like crypto that most of the incumbents didn't support and the biggest difference is we launched with a really unique business model where instead of charging a percentage of assets like a lot of non-profits we just charge a membership fee.

Brian Bell (00:25:04.740): Fascinating, so let's talk a little bit about there wasn't a lot of product innovation to be done it sounds like on the account creation side maybe a little bit on the investments like where the money goes while it's waiting to be donated but you said you innovated a lot on the product side on on the giving maybe you could talk a little bit more about.

Adam Nash (00:25:23.418): That yeah I mean, well, you know, you have to be uh this industry is humbling always being being a product leader is always humbling because you're you're lucky if you're half right you have a bunch of ideas I used to joke that you have 10 good ideas and winning that to the great ones that actually work in the market that you ship and then you're lucky if it's it's one or two out of ten that's why you have to iterate that's why you have to.

Brian Bell (00:25:44.170): Yeah, do you have a a story about that where you built something over your career and you were just so excited about it but nobody like it just flopped I mean,.

Adam Nash (00:25:52.696): So, many great features I just I like to think in some cases they were just early because timing actually matters but but in many cases there's so many pieces actually some of the reason I became known as a growth guy right like I, you know, right LinkedIn was that generation of companies Web 2.0 LinkedIn facebook they were the first companies to have a growth team instead of just marketing for user acquisition.

Brian Bell (00:26:13.783): Which became very sophisticated over time some of the LinkedIn growth stuff is just outstanding when we had to figure it out.

Adam Nash (00:26:19.932): But we had to figure it out when I think of all the features like growth is one of those areas where you have to get covered with most of your feature ideas not working and some of them are are very counterintuitive um when you roll them out but no like when I think of Daffy and I, I think of the things that we've iterated on it actually turned out what we were wrong about was um actually was very funny I thought this was a market expanding product I thought when we launched the MVP I was very the wealthy people who have financial advisors and are using the incumbents we're not for them we're for the other 50 million americans who who just want to put aside 10 a week 25 a month they give a few hundred dollars to charity it's meaningful and we could build a great I'd been on the board of a company called Acorns Acorns has millions of customers now who use Acorns to help them lead better financial lives and just off a few dollars a month and then of course we launched the product and immediately within the first few days we get this request well can I move I've been doing a lot of money from my existing donor advised fund we hadn't built it we didn't even know if it was possible we go into this crash sprint to figure out how to do that good news is it's not very hard to do and we actually shipped it within a couple weeks so we got it out there but, you know, the truth is last four years this space is so open there's been so little innovation at scale um we kind of have been running the table I mean, every year we, we've rolled out easily four to six meaningful I think innovations I mean, I'll give an example but this will sound trivial to you and for most people in tech but it turns out, you know, everyone in the charitable space knows that it's about families and giving if you talk to the Morgan Stanley's and Goldman Sachs of the world they'll talk about multi-generational giving they'll talk about wealth transfer they'll talk about legacy all these incredible things and yet every donor advisement out there is like do you want an I mean, I'm sorry I'm a father of four I, I feel like I have a lot of these like netflix Apple xbox I mean, I have three kids myself.

Brian Bell (00:28:11.702): So, yeah I can relate yeah yeah and so and we said why why would.

Adam Nash (00:28:14.927): You can add up to actually 24 different people to your plan and you as the organizer you funding the account you you still control it fundamentally, you know, you can have siblings you can have parents grandparents giant groups maybe it's an honor of someone else um I have all four of my children on the account and it's fantastic when I my wife and I make a donation it becomes a subject of dinner conversation right something to talk about to teach them and then my children can actually make recommendations in the end my wife and I can either approve those or reject them or or modify them it's kind of like Amazon requests but um I know this sounds awesome. You look at it and you go, why doesn't everyone else have it? But those are famous last words. It just comes from the perspective. We've rolled out APIs. We've rolled out custom portfolios. We've rolled out support for private stock. I mentioned crypto before. All these different features but every one of them came from the same place which is just looking at our members listening to them thinking about why they give what was meaningful to them and saying hey in the modern world how would you turn this into a feature how would you build that into the product rather than have it be something that people do on their own I always like to make products a little bit more than just a tool right they are tools but but fundamentally I like to think about the entire experience that people have around using them it was one of the things I learned the hard at eBay, you know, in the early days for Web 1.0 so many web products in the beginning were so focused on the one task they did they didn't really think about the broader set of tasks what the what the user was actually trying to get done so and we've just been like I said it's it's been a lot of fun yeah innovating in a space where you feel like when you do something right it encourages more generosity I mean, that's the mission of the whole organization is to help people be more generous more often amazing.

Brian Bell (00:30:05.570): So, there's been some contrarian takes online that, you know, donor advised funds get attacked as warehouses where rich people park their money and never give it away uh steel man that, that argument and then tell me.

Adam Nash (00:30:16.177): Where they're wrong well I think, you know, I think I think where it comes from is just this current fixation on billionaires I mean, I, I don't I don't even know if people really focused on what that word means and the number of people involved etc but, you know, you find out that Peter Thiel has billions in his his roth IRA and then everyone wants to change policies around IRAs and that sort of thing but, you know, to steal my argument it's like it's technically possible right like you can put a lot of assets into a donor advised fund.

Brian Bell (00:30:44.589): And have it compound a half million into the next few years and decades and.

Adam Nash (00:30:48.034): So, I can totally understand why some people say hey that feels like a hole in the system we should plug it for billionaires and then my argument against it really is too flat I mean, the first and foremost I would say like listen you may have opinions about what you can and can't do with an IRA but let's not forget that for tens of millions of americans that's an important part of their retirement savings and their goals they're not billionaires and so do not throw out the baby with the bathwater so so if you're gonna have a regulation around that and you want to cap it at some certain level or if you want to have restrictions once it gets to a certain size willing to listen like that makes sense but most people don't put away billions they don't have the capacity to do it right, you know, most people they're giving to they're giving to a religious organization they belong to they're they're giving to a cause they, they're not they're not trying to squirrel away money forever and so I really do like encourage people to focus on the vast vast vast majority of americans who who who need a better system for how they give but the second thing, you know, the argument I have is that the data just doesn't support it I mean, like they run this data every year I mean, I know everyone's worried well foundations have this requirement to give away five percent their assets every year endowments have rules about how much they have to put through their cause but every number every year I've ever seen for donor advised funds shows that the number is much much higher I think last year in the industry was something like 24 percent that maybe that's a 2024 number I'm not sure I don't get it right but but fundamentally um I mean, at Daffy it's over 50 percent wow right like in terms of how much money people put aside in year one and then the next year how much of that goes out the door to charities 55 was our number from 2025 and so I just think when I look at policy problems I'm saying like what what problem are you actually solving here right, you know, and and so um I would be asking more questions personally about foundations and endowments and and what they're doing but I think that the secret actually just turns out that we get distracted by the ultra wealthy it's like celebrity fascination for most people giving matters to them but we are all busy we, we have work we have home life we have family we have social life we have so many things going on the truth is is that giving is one of those tasks that isn't a daily task for most people and so what happens is when someone asks you to give right Marisol's and it's and and the difference that they're making and so I, I I just happen to be a believer that having separate accounts I mean, I teach this class like I said.

Brian Bell (00:33:46.946): On personal finance yeah I was gonna ask about that mental accounting is.

Adam Nash (00:33:50.647): A real thing goal setting is a real thing there is research that says that if you set a goal pre-commitment for your giving you give 32 percent more wow I mean, I believe that it's true but it's true for all financial tasks I mean, how many of U.S. Would put money away for retirement reliably if it didn't just come out of the paycheck I mean, that's why automatic that's why these things make sense that's why financial advisors will talk to their their clients about their goals for retirement and then turn that into a number a savings goal etc all these things work so I think that, you know, for folks um who get fascinated with donor advisements etc I always encourage them to think of of of what actually the average use case is and not just think about, you know, what you read about whatever whatever billionaire you're worried.

Brian Bell (00:34:34.663): About right now is is doing yeah that's amazing are you still teaching personal finance for engineers at Stanford.

Adam Nash (00:34:40.532): I am I actually just confirmed it's it'll be the 10th year this year wow it's it's amazing it's been that long but yeah it's it's a dream thing for me because well I love to teach I obviously care about the topic but it also happens to be one of the misses I felt when I was at Stanford was I mean, so many students at Stanford are the first in their family to go to college or if they're not, you know, they, they feel like they, they got the golden ticket they just don't want to mess it up and the truth is when I went to school there weren't classes I actually think it should be I've actually open sourced all the material it's actually available a free blog I've actually had a number of schools at all different levels even internationally ask if they can use the material and it's great for me that's the reward I'll walk through.

Brian Bell (00:35:23.967): With my teenage boys I think that's a good idea.

Adam Nash (00:35:26.307): Yeah, no it's uh I'm a big believer in that uh, you know, with my own children as well I think that for some reason especially in the united states we're incredibly uncomfortable talking about money and about how it affects people's lives and and how we make decisions, you know, we'll go to the store with our children and, you know, they'll want to buy something and the reality is it's not in the budget but we don't want to say that to our children and we don't want to burden them with thinking too much about money and life etc and so we tend to make up other reasons right like oh that's not that or, you know, that we'll talk about anything else and I happen to be in the school of thought that says that, you know, this is when you do that, you know, children figure it you're not talking about and so a lot of people what they know about money is just what they picked up from it's funny it's the same thing with careers at LinkedIn I used to say that step one for most children is kind of the friends of your parents and the parents of your friends tend to be the adults, you know, and so so your worldview tends to be very narrow because it's it's basically based on that and so I, I really think that's where education comes in to to open, you know, kids minds and and and to help people navigate their lives so by college I mean, students really need it they I'm not talking Adam Nash (00:36:34): about fancy stuff I'm not talking about portfolio theory or, you know, calculating returns I'm talking about the basics of what financial products are what problems they solve and, you know, how do you lead a healthy financial life like I, I think that these are things that actually everyone needs to learn from someone at some point.

Brian Bell (00:36:49.803): I mean, I, I grew up poor and uh on welfare and that's why I got a finance degree frankly I mean, I was getting a history major up in up in washington state where I'm from and that got really boring and I was like I better learn how to like money work so I went and got a finance degree it was interesting I'm decent at math I like the, you know, the numbers of it and stuff and ended up working wall street hating that and now I'm a vce which is very random but kind of full circle but, you know, like that's why I got a degree in finance I just wanted to understand how it all works, you know, from the inside.

Adam Nash (00:37:18.827): Yeah, I think a lot of U.S. Who love the field get into it that way right we had like I said I had my own stories etc but on this desire of how does this work how do you make good decisions I want to understand this and for some people I want to be good at it I think a I mean, it turns out that in our economy and in our society, you know, like I said money is not necessarily the goal that most people have but wow does it turn out that if you don't understand how it works the basics right it can really get in the way of some of the goals that we have that are are the most important to U.S.

Brian Bell (00:37:52.107): Yeah, speaking of money you've angel invested in 90 plus companies some really amazing ones like Figma Gusto Opendoor fibres and many others so you're like a prolific angel this is very much a startup in VC podcast and I love to learn from really great angels like you so what's what is a pattern that you look for when you when you make an investment.

Adam Nash (00:38:10.177): Well, I appreciate although I almost I, I think at this point I've been angel investing almost 14 or 15 years I think I'm up to about 160 170 companies because I have a model, you know, one of the things I learned so so my angel investing is kind of an interesting combination of of of course being an operator and knowing something about what I know and what I don't know.

Brian Bell (00:38:30.625): Yeah, and having a little humility around that and create great organizations.

Adam Nash (00:38:34.871): And yeah yeah but then also having spent a couple tours of duty at at some great venture firms understanding how professional venture capitalists really think of it what it takes to run a fund in a firm like Graylock really helps use a system so as an angel investor I'm probably not as as sexy or cool as as, you know, folks who kind of have the inspiration etc I, I actually run it a little bit like venture I even break my angel every three years into what I call a fund that I track but, you know, I actually not surprisingly I treat it as a little bit of a personal finance I've tacked I there's a certain amount of money I feel like I can afford in my broader portfolio to invest right.

Brian Bell (00:39:10.400): In new startups every year just out of curiosity for LPs listening when they're thinking about investing in and venture funds and angel investments and spvs like just, you know, what what is that percentage of your wealth that you feel comfortable every year putting towards.

Adam Nash (00:39:23.502): That asset they're they're not going to love this answer but I'm happy to give it but they're not going to love it there's some people who do their angel out of income basically I, I never approach it that way I think that's valid but I, I don't approach it that way I always treat it a little bit more as a portfolio problem so I basically said that I can afford to have about 10 percent or so of of my savings of my, you know, assets invested in startups yeah and so this is the part they're not going to like we all know it takes about 10 years I do seed stage it takes about 10 years yeah for the best companies to get from there to any form of liquidity and this is why I see a lot of angels getting wrong I've actually written about this that you see a lot of angels so excited they jump into it they invest year one in a bunch of companies year two and then they realize year three that like wow that's a lot of money out there and none of it's coming back right yeah am I really going to keep doing this down by the way too right what, you know, I'm a gardener too so this expression actually bothers me because it's not true but the VCs always say the lemons ripen early type of thing and so you you tend to see some the failures a little bit more the the ones that don't make it a little faster than the ones that make it.

Brian Bell (00:40:32.119): Right, and maybe not an AI four or five years to raise their a right from a seed right that's right they look like failures like Figma Figma very famously was, you know, I think four years five years between their seed and a something like.

Adam Nash (00:40:43.721): That it was I think it was the a to the b but you're right it's I mean, but that's a great example so Figma I did the seat I walked around palo alto with with Dylan in 2013 and, you know, pinged him afterward and, you know, and he let me put a small check into it and but when did Figma go public right 2025 yeah 12 years later right and so so what I did was I figured out what percent of my portfolio I could afford to have in angel investments I divided that by 10 and then I decided what my check size roughly was with the idea that I had to be able to invest in about four I'm not perfect as, you know, turns out I'm human and emotions run hot and so when I look at my angel now over 14 or 15 years I see like yeah in the years where things are running a little hot maybe I did more like 10 to 12 investments I think there was one year where I may have gotten closer to 20, you know, in the years where I'm super busy as an operator did fewer it's actually funny some of my best investments were in those years that I did fewer there was one year I think I only did six five or six investments but one of them I think it was 2013 I think I think I think that's where Figma and Gusto and Opendoor all were living either 2013 or 2014 that's really good vintage.

Brian Bell (00:41:50.645): As they say in the industry.

Adam Nash (00:41:51.955): Yeah, so but, you know, and you have to learn and so what do I look for um I look for a few things um believe it or not I'm actually a big believer in in so one is when I talk to folks I try to stick to my knitting right there are areas where I have expertise where I know things I'm still a big believer that I need to understand why the founder's talking to me if the reason is just money that's not a great answer, you know, there are people with more money than me but that sort of thing I hate to say but it's a little bit like, you know, if you don't know who the sucker at the poker table is you're the you're the sucker so I, I like when people come to their firm where they want a product executive, you know, someone with real experience maybe it's in social maybe it's in fintech or marketplaces or one of the areas I have direct experience so you can have a lot more value on the money it has to make sense to me why they're talking to me I listened to the founder one of the hard lessons I learned early on the venture side was it's not your company that one of the big mistakes operators make as investors is they keep thinking about what they you would do if they were running the company it's not your company you're not going to be there to run it as a small investor like an angel you're not even on the board you don't even know what they're gonna be doing so you need to actually take yourself out of it enough to just hear what they think they're doing and then and for me I need to hear not just the product value that it adds but also the distribution strategy how are they going to reach people how are they going to get to venture scale and then lastly and this sounds funny but there's a little bit of founder market fit for me which some people like that term some people don't but for me they're the founders who who think they found a great way to make money and I do not begrudge them because that, that is definitely one of the ways to build a business you find a hole in the market you think it does this I happen to be of the point of view that building startups especially venture-backed technology startups a decade at least the best best founders are involved even longer they have to really almost irrationally care about the space yeah it's like.

Brian Bell (00:43:44.320): A founder market live fit.

Adam Nash (00:43:45.723): Yeah, so it doesn't work for me if if it I've met by the way I've passed on on on companies that, you know, turned out to be amazingly valuable because I wasn't sure the dedication of the founder what is your anti I want.

Brian Bell (00:43:58.180): To hear Adam Nash's anti-portfolio what's that check that you didn't write that still keeps you up at night.

Adam Nash (00:44:02.405): Sorry, well none of them keep me up at night I learned that very early there's always gonna be more out there the companies you didn't see the the folks that you missed I mean, I have just from a personal operating standpoint like forget investing I mean, like, you know, everyone who came up in the 90s has this list of like when they could have gone to I could have been like the fifth engineer at Yahoo or, you know, early at Google I mean, I had so many friends who went all these companies you have to tune all that noise out but fundamentally I think that the the venture is filled with so many paradoxes I, I wrote this one post years ago and it reflected on something I'd learned at business school for a totally different reason had to do with business and government but I saw the same pattern in venture which is that the one of the fundamental paradox in venture is that in venture when you're new it's so easy to sound smart by saying no right in fact venture firms push this on associates junior folks in the firm right they come in your bar's not high enough right there's a litany of reasons why this won't work and by the way they're right there's always a litany.

Brian Bell (00:45:01.371): Of reasons why it won't work you could always always say no it's easy to say like it's easy to find a reason to say no.

Adam Nash (00:45:06.813): Yeah, but it's rewarded you sound smart by saying no in the short term but in the long term people only care about when you said yes right and just living in that paradox, you know, and so but the the joy like I could have said no to tons.

Brian Bell (00:45:24.184): Of companies doesn't matter but the fact that I said yes to Google in 1997 at a five cap and put a little 25k check in makes me a legendary Ron Conway level investor for all time so nobody cares that Ron Conway wrote 300 other checks in that fund that's right I mean, that's.

Adam Nash (00:45:41.988): That power law right like etc but so what happens is I think you have to always be looking back at your own foibles and mistakes idiosyncrasies I'm very people driven so in general I tend to invest in people who've worked with me for me that impressed me I think that's one of the ways it also answers the question why are they coming to you right they know you they trust you I mean, the reason Dylan Field came to talk to me was not because I was an angel actually at the time I had barely done a half dozen investments it's because I was the VP of product when he was an intern at LinkedIn right he was he had he wanted some advice so I walked around and then of course what I had learned on the venture side is no if you want to invest, you know, actually say so right and and by the way the original idea that Dylan had I had some real problems with I was not convinced it was going to work it was like a photo editing thing in the cloud it was not right but what I realize now and that I look for more because of Dylan is that some of my best investments came because I like to think of myself as a smart person a connected person I know what's going on to some level I mean, not perfect but you're more than most better than than quite a few so I'm always impressed when I meet someone young who has been diving into space where I thought something and they convinced me that I was wrong right and all of a sudden I go so I remember walking around in in palo the Dylan and he talked about I had this thesis that said hey what was going on with social and mobile was that we'd finally got enough distance from the PC era like 30 years that we were just going to do it again we were going to reinvent everything that we had done in the PC era except for social and mobile and so if you look back at the PC era we started with productivity word processing spreadsheets were the first things that came out that were actually real utility then we went to graphic design art desktop publishing, you know, graphics etc and then later then we went to fintech the quickens of the world etc like so I in my head I was looking for these investments and of course with Wealthfront I was jumping in myself to the third phase because ahead of the curve but, you know, I was talking to Dylan I was like oh I love this because I think it's happened but I told him I said I think it's going to take a while because of all the things that are going to move through the cloud graphics feels like a very late thing this is an area where people will still spend ten thousand dollars on a machine.

Brian Bell (00:47:45.073): That has beefy graphics cars right in a way because it did take them almost the rest of the decade yeah it really.

Adam Nash (00:47:51.838): Yeah, I was yeah I was convinced like, you know, I was like people don't even great designers at the time like wouldn't even use laptops they needed a beefy machine to kind of do it and he said no I think that's complete opposite those GPUs are sitting idle most of the day in the cloud you can have a cluster can be used and by the way I can put enough hardware in the cloud that you would never pay for that yourself you will just rent it and bandwidth has gotten high and latency low enough that I'll just send the pixels for the screen over to you and he was very big on WebGL he's like now's the time like we can build this and we can time frames totally fair but I remember thinking I was just talking and going like you're totally right I, I had already moved by the way at LinkedIn I was using desktop PCs just remote, you know, windows sharing in desktop sharing in because that way I could have the PC at work and free trial laptop I was like yeah it totally works the market's gonna get there and I know that sounds trivial and it wasn't the heart of the right thesis but that pattern with founders I think that if you have the right mentality meeting great founders when they tell you something when you learn something and you feel like you're plugged in for me at least I was like wow this is well it's either wrong can be or it's ahead of the market and so I love it turns out that my some of my most successful investments have been along that line I mean, Gusto was very much it was zen payroll at the time but was very much this this idea that actually the user experience mattered Josh Tomer Eric they were all very passionate about this idea that actually payroll mattered that actually there's this experience like people care about how that like small businesses huge number in the u.s no one was taking care of them and that it was possible now online to acquire to actually find those small businesses and give them a solution that was not just 10X but 100X better than what they had access to so I don't know like if you had asked me before I met them I would have said there is no reason that, that should work a small business products that tortured area venture capital that there are so many tenured venture capitalists who refuse to invest in anything tied to small business just because of how hard the go-to-market was but they were the ones who said no actually I think with online social acquisition I think it's changed I think you can find small business owners the same way you find consumers and frankly having built out LinkedIn I was like no you're right like actually LinkedIn is pretty good for finding small business owners etc it's it's becoming.

Brian Bell (00:50:26.286): A solved problem well Adam I feel like I could talk to you for another hour or two about all this stuff but I know you got to run.

Adam Nash (00:50:31.028): Where can folks find you online? I lack some we already talked about Daffy so when it comes to naming it turns out my handle is Adam Nash on pretty much all the platforms but um you can find me on X Adam Nash you can find me on LinkedIn is Adam Nash um and of course at Daffy on the blog there, you know, I try to write regularly about giving and financial I would highly encourage everyone to check out Daffy.org I don't think Daffy is for people who don't give to charity like if you don't give to charity I don't think we're going to convert you into the type of person who does but if you if you support your kid's school if you belong to a religious institution if you give regularly to national or or global causes try the product it's free to get started put a little money in there I think what you're going to find is that having a separate account for charity I think it's more like an HSA, you know, it's funny uh it is it is like an HSA it's just forgiving yeah yeah it's it all comes down the same thing having a separate account means that when someone asks you to give you no longer have to reach in your pocket and debate you can just literally open your phone tap tap tap yeah and.

Brian Bell (00:51:29.636): If you're inspired to give card right to my Apple Pay, you know, Google Wallet or whatever and just and just pay for it with with.

Adam Nash (00:51:35.960): The funds I have set aside funny you mentioned that we haven't solved that problem.

Brian Bell (00:51:39.382): Yet but it's on the list okay that's on the road map yeah okay well I expect it 99 on time, you know, as somebody who who lived that at eBay.

Adam Nash (00:51:48.208): So, we're we're free now we, we uh we love iterating some of our best feature ideas are things where we didn't plan them one member comment comes in one in inspiration saying why can't we do this for giving and then a week.

Brian Bell (00:52:00.884): Or two later we get it out the door just get yeah just go for it yeah I mean, I, I use the HSA we, we max it out every year and then every once in a while we'll be like oh there's like five grand sitting there let's withdraw it and send all the receipts in and, you know, and every once in a while we'll pull out the card for at the dentist or the doctor's office so yeah I think.

Adam Nash (00:52:16.898): That would make sense yeah it's our members actually are delighted by it there's something that feels really good when you get asked to give and you actually care and you want to support you want to support your friend yeah colleague the organization etc when you open up that app and you discover actually you do have the money yeah and it's literally for nothing else you cannot use it for anything but giving to charity it feels so good to be able to do it when you want to do it it's hard to express that freedom that's why I encourage everyone just.

Brian Bell (00:52:42.570): To try it what we see this miles could I like donate miles into an account like that have you looked into that it's funny we've gotten.

Adam Nash (00:52:49.312): That request just recently yeah a little bit of like are there other ways that I can fund this account.

Brian Bell (00:52:55.592): Right, there's some monetary value there if I have a hundred thousand miles in Southwest or something maybe I could donate that uh we have a startup that helps it's a platform for that so I'll we'll talk after the after the recording interesting all right awesome thanks so much Adam cool thank you again.

On June 30, the US Commerce Secretary sent a letter to Tom Brown, one of Anthropic’s cofounders, and turned the company’s most powerful AI model back on. Eighteen days earlier the same government had ordered it off. The trigger had been a jailbreak, discovered by Amazon, that coaxed the model into writing working exploit code. The condition for restoration was a set of promises: Anthropic would hunt for security risks proactively, help write the standards its future models would be judged against, and report malicious use to Washington.

Read that sequence again slowly, because nothing like it has happened before. A frontier AI model, the commercial crown jewel of one of the most valuable private companies in the world, was switched off by government order and switched back on by government letter, all inside three weeks. The same week, OpenAI’s newest model stayed locked to roughly twenty government-approved organizations, unavailable in ChatGPT, waiting for its own letter.

For two years the AI trade has run on two assumptions. The first is that compute is scarce and will stay scarce, so anyone holding GPUs holds pricing power. The second is that frontier models ship when their makers decide to ship them. Both assumptions cracked this week, and the second one cracked in public, with a return address.

Who decides what ships

Start with the permission story, because it is the stranger of the two.

Anthropic’s Claude Fable 5 went back on sale globally on July 1. Its restricted sibling, Mythos 5, was restored to about a hundred US critical-infrastructure organizations. Anthropic says a retrained safety classifier now blocks the exploit that caused the freeze more than 99 percent of the time, a company-stated figure, and it opened a public bug bounty to find the next one. Alex Wissner-Gross, the Harvard physicist whose Innermost Loop newsletter tracks frontier AI, opened his July 1 post with the line “The Singularity just cleared customs,” and quoted security researcher Alex Stamos calling the whole episode a “huge own goal for the US.”

Meanwhile GPT-5.6, which OpenAI previewed on June 26 in three pricing tiers, spent the entire week gated. Same government, same month, opposite outcomes for the two leading labs.

Here is why this matters more than the news cycle suggests. Until June, the question “can I build my product on this model” was a commercial question. You checked the price, the rate limits, the terms of service. Now it is also a regulatory question, and regulators have demonstrated they will answer it differently for different companies in the same month. Experienced investors will tell you that the moment a single supplier can be shut off by forces outside your contract, dependence on that supplier stops being a convenience and starts being a disclosed risk. Every startup built as a thin layer on one frontier model just inherited a risk factor it cannot negotiate away.

The market is treating the Anthropic freeze as a one-off. The more likely reading is precedent given how strong models are getting now. A June executive order already sketches a voluntary pre-release review for frontier models. The thing to watch over the next month is whether voluntary hardens into mandatory, because if it does, the release calendar for the most important technology of the decade runs partly through Washington.

The glut nobody priced

The second crack was quieter in its mechanics and louder in the tape.

Around July 1, Bloomberg reported and CNBC confirmed that Meta will start selling its excess AI compute to outside customers, turning the largest private GPU hoard in the world into merchant supply. Meta’s stock rose 8.8 percent that session on roughly triple its normal volume. CoreWeave, a company whose entire business is renting GPUs, fell 10.8 percent. Nebius, same business, fell 12.4 percent. Micron, which makes the memory chips that feed those GPUs, fell more than 10 percent. Gil Luria at D.A. Davidson put the problem plainly: the neoclouds “rely on Meta for their growth and Meta may not need them anymore.”

One announcement, no product, no pricing page, and tens of billions of dollars in market value moved. That tells you how much of the AI trade was resting on the assumption that compute stays scarce forever. Meta is the second giant to break ranks; SpaceX had already begun renting compute to Anthropic and Google. When the biggest hoarders become sellers, scarcity is a policy choice, and policies change.

The confusing part is that physical scarcity is still real. On Micron’s earnings call in late June, CEO Sanjay Mehrotra said “sustained and strong industry demand, along with supply constraints, are contributing to tight market conditions and we expect these conditions to persist beyond calendar 2026,” and disclosed that every unit of high-bandwidth memory Micron will make in 2026 is already committed under contract. Memory and power are genuinely tight. What cracked this week is the premise that GPU rental specifically stays a seller’s market. Those are different bottlenecks, and the market spent Tuesday learning the difference.

Agents got cheap, with an asterisk

While the permission and supply stories played out, the price of intelligence itself dropped.

Anthropic released Claude Sonnet 5 on June 30, its most capable mid-tier model, at introductory pricing of two dollars per million input tokens and ten per million output, rising to three and fifteen after August. Its top model, Opus 4.8, costs five and twenty-five. On Anthropic’s own numbers, Sonnet 5 beats Opus on a terminal-use benchmark and on a knowledge-work benchmark while trailing it on hard software engineering. Every one of those figures is vendor-claimed until someone outside Anthropic reproduces it, and vendor benchmark leads rotated multiple times this week alone, so hold them loosely.

The asterisk sits in the fine print. Sonnet 5 uses a new tokenizer, the component that chops text into the units the model bills you for, and it maps the same input into 1.0 to 1.35 times more tokens than before. Cheaper per token can still mean pricier per task. Any founder recomputing their unit economics this week needs to run the math on tokens consumed per job, and the price per token second.

There is a deeper pattern underneath the pricing news. Two research claims surfaced this week, both secondhand and unverified, both pointing the same direction. ByteDance researchers reportedly found agent learning speed doubling every three months on a benchmark of theirs, and the UK AI Safety Institute reportedly showed a model’s capacity for long cyber tasks stretching from two hours to fourteen as its token budget rose twentyfold. If either result holds up, agent capability is partly a function of how much inference you can afford, the way a basketball team’s ceiling is partly a function of how many minutes its best player can stay on the floor. Cheaper inference literally buys smarter agents. Whoever controls cheap inference controls the rate at which agents improve.

Where the money went

Which brings us to the week’s largest venture round, because it is the same story told in capital.

Together AI, a platform that makes open-source models cheap to run, raised 800 million dollars at an 8.3 billion valuation, led by Aramco Ventures with Nvidia, Vista, General Catalyst, and Salesforce’s venture arm participating. The company disclosed annual bookings crossing 1.15 billion, with open-model usage tripling in a year. Its customer list is the tell: Cursor, Cognition, Decagon, the very AI application companies whose margins get eaten when frontier-model prices stay high. Decagon says its inference bill fell sixfold after moving over. When a Saudi state energy fund and the chipmaker itself co-underwrite a company whose product is lower inference prices, the smart money is betting the closed-model premium shrinks.

The second notable round was Chamath Palihapitiya’s 8090, an enterprise AI coding company, raising 135 million led by Salesforce Ventures, with Palihapitiya taking the CEO seat, his first operating job since Facebook. The company’s claims about translating eighteen million lines of COBOL for insurers are its own and unaudited. The signal is in what got funded: governed execution in regulated industries, where the audit trail is the product, rather than another general coding assistant. Note that Salesforce now anchors both 8090 and part of Together’s cap table. An incumbent is buying position in the layer where enterprise software gets built, which sharpens the acquire-or-compete question for every startup in that lane.

Zoom out to the whole week’s funding table and the pattern gets blunt. The largest round was Joulent at 1.75 billion for AI-oriented energy infrastructure, backed by National Grid’s venture arm. Then Together. Then a compliance company, then 8090. Power, compute plumbing, and compliance topped the table. Application software did not. A reminder worth repeating: funding is a price signal, and prices can be wrong, but the direction of the pricing is information. The market is paying up for the picks and shovels and grading the apps harder.

Down at the stage where seed investors live, five rounds worth naming printed inside the week: a 30 million seed for Nebex in space-economy settlement led by GV, a 19.5 million Series A for Pie in small-business customer acquisition led by Lightspeed, a 31 million Series A for Omen AI in data-center sensor monitoring, a 12.6 million seed for Queue in autonomous pharmacy robotics, and 7.5 million for Civ Robotics in outdoor surveying robots, the latter two led by AlleyCorp. Three of the five are physical AI with a clear commercial buyer. Seed pricing has not repriced downward in robotics; if anything the deployment-constrained physical companies are getting funded faster than software.

One small founder-behavior anomaly worth filing away. Omnea, a London AI procurement company, announced it will hand employees 250,000 dollars to openly plan their next startup instead of moonlighting in secret. One company, small sample. But it names something real about this talent market: the best operators now expect to found, and employers are starting to price retention against the seed market itself.

The macro that will not help you

The Bureau of Labor Statistics reported July 2 that June nonfarm payrolls rose 57,000 against a consensus of 115,000. Participation fell to its lowest level since March 2021, household employment dropped by half a million, and prior months were revised down. Wages still rose 3.5 percent year over year. This is a labor market cooling through weak hiring rather than layoffs, which is the awkward kind of soft: weak enough to worry about, firm enough on wages that the Federal Reserve feels no urgency.

And the Fed said as much. Chair Kevin Warsh, speaking at the ECB’s Sintra forum on July 1, said “prices are too high” and “we’re going to deliver price stability,” and declined to give any guidance on the July meeting. Rates sit at 3.50 to 3.75 percent and nearly half the committee still leans toward a hike this year. The second-half rate-cut thesis that a lot of 2026 growth-stage burn plans were quietly built on is dead for now. Worse for planners, the Fed has stopped pre-announcing its path, so repricing will arrive without warning when it arrives.

One mechanical event rounds out the week. SpaceX joins the Nasdaq-100 before the open on July 7, a fast-track inclusion just fifteen trading days after its June IPO, with J.P. Morgan estimating around 4.3 billion dollars of forced passive buying concentrated after the July 6 close. Index funds have to buy regardless of price. Whatever SpaceX shares do around those two days reflects flow mechanics, and tells you nothing new about the business.

For founders

Every founder building on AI should answer four questions this week, and the honest answers sort companies fast.

• If your best model were switched off by government order tomorrow, what is your fallback, and how many days does it take to deploy?

• What do your unit economics look like at Sonnet 5 pricing, after the tokenizer change, versus ninety days ago?

• Do you own a workflow end to end, or are you a feature a platform can absorb the next time it ships enterprise controls?

• If inference prices fall 30 percent over the next year, does that expand your margin or commoditize your product?

The winners from this week’s shifts are visible. Agent products whose economics were underwater at last quarter’s model prices just became viable. Routing and fallback infrastructure, the switching layer that lets a company move between models when one gets expensive or gets banned, went from nice-to-have to board-level topic in eighteen days. Tooling that watches, permissions, and rolls back autonomous agents gets more urgent every time agents get cheaper and longer-running, and this week they got both. And vertical AI in regulated industries, where the compliance trail is the moat, just watched 8090 validate the wedge with strategic capital.

The losers are equally visible. Undifferentiated GPU rental now competes with Meta’s cast-offs. Thin wrappers on a single frontier model carry policy risk no term sheet can hedge. Burn-heavy growth plans premised on cheap money in the second half need new premises. And startups selling cost dashboards around Claude just watched Anthropic ship those controls itself, which is the oldest platform story there is.

For LPs

The variables that governed this cycle are rotating. For two years the questions were who has capital and who has compute. The questions now are who has permission and what inference costs. Neither shows up in a quarterly mark until it suddenly does.

Three practical implications. First, early-stage exposure to capital-efficient companies that own their workflow ages better in a world where model access can be revoked and model prices can collapse; both happened, in both directions, in one week. Second, pacing beats speed while the Fed refuses guidance, because macro repricing will arrive unannounced and vintages that deployed patiently through 2026 will look smarter than vintages that sprinted. Third, late-stage private marks now carry public-market machinery inside them. SpaceX exposure is partly an index-flow instrument as of July 7. Anthropic’s valuation now embeds a demonstrated government off switch. Underwrite the risk factor, and treat the headline number as the output rather than the input.

For the venture business generally

The uncomfortable summary of the week is that two premises underwriting most AI theses since 2024 broke within four days of each other. Permanent compute scarcity broke on Tuesday when Meta became a seller. Unconditional model access broke on Monday when a government letter, and only a government letter, put Fable 5 back on the market. Any thesis document that contains the phrase “GPU scarcity” or assumes frontier models ship on lab timelines is due for a rewrite.

Value is migrating to three places: whoever is allowed to ship, whoever makes inference cheap, and whoever owns the workflow the model plugs into. The middle of the stack, the layer that merely resells intelligence made by others, got squeezed from above by policy and from below by price in the same week. That is the story to underwrite, and next week’s job is to see whether Washington makes it a pattern.

Sources for this issue include the Anthropic and Together AI announcements, the Commerce Department letter as reported by CNBC and Al Jazeera, BLS Employment Situation (July 2), Micron’s fiscal Q3 call, Nasdaq inclusion notices, Crunchbase funding data, and The Innermost Loop (July 1 to 4). Benchmark figures are vendor-claimed unless noted. Two capability-scaling claims (ByteDance EdgeBench, UK AISI cyber task horizons) are secondhand and unreproduced; they are flagged as such where cited.

Anthropic launched Fable 5 on June 9, 2026. It was the first time the company let the public touch its most capable tier, the Mythos class, the line it had previously called too dangerous in the cybersecurity domain to ship. Three days later, at 5:21 p.m. Eastern, a letter arrived from the Commerce Secretary. By that evening Fable 5 was gone. Not throttled, not geofenced. Gone, worldwide, for every customer. The less restricted sibling, Mythos 5, the version only a vetted set of partners could use, got pulled along with it. The stated trigger was a reported way to jailbreak the model, a trick that amounted to asking it to read a codebase and find the flaws.

Almost immediately, people read this as the smoking gun. The models have gotten so strong the government is yanking them off the shelf, the argument went, and behind the lab firewall those same systems are quietly rewriting themselves into something far past human. The shutdown was proof. The takeoff had started.

Follow that idea for a second, because it falls apart in your hands. The government did not lock Fable in a vault and let it cook. It cut its power and shoved it out of reach. The version the public never had, the one reserved for trusted partners, got killed in the same stroke. If frontier labs were sitting on a self-improving superintelligence, the Fable episode would be strange evidence for it, because the machine here was a kill switch, not a greenhouse.

That gap, between what the shutdown felt like and what it was, is the whole subject. The raw facts of 2026 are more dramatic than most people outside the labs realize. The story laid on top of them usually runs a step or two past where the facts can carry you. The interesting work is finding exactly where the facts stop and the story begins, because that line is where the money and the risk both live.

So let me take the question seriously and from both ends. Is predicting the next token enough to get us all the way to superintelligence, the kind that redesigns itself faster than we can watch? And are we, right now, inside the fast and terrifying version of that, the hard takeoff, or the slower one we can still steer?

What a next-token predictor actually does

Strip away the marketing and a large language model does one small thing, over and over. Given a string of text, it guesses the next chunk. Then it adds that chunk and guesses again. Training it means showing it a huge pile of human writing and nudging it, billions of times, to make that guess less wrong. That is the whole objective. Everything else, the essays, the code, the legal analysis, falls out of getting very good at that one guess.

The skeptics’ oldest line is that this can only ever be mimicry. A system trained to predict words is matching surface patterns in text, and patterns in text are not the world. A house cat understands gravity. It plans a jump, predicts where a falling object will land, models cause and effect, and it has read exactly zero words. A language model can write you a flawless paragraph about gravity and could not catch a ball. Yann LeCun, who ran AI at Meta until he left in late 2025 to build a company around a different design, has made this case as bluntly as anyone. He thinks autoregressive models, the technical name for these next-token machines, are a path that climbs impressively and then dead-ends short of the summit. His proposed replacement learns by predicting abstract states of the world rather than words, and his bet is funded to the tune of about a billion dollars, which is a useful reminder that the smartest skeptic in the room is not a crank.

Here is the catch that keeps the skeptics from closing the case. When researchers trained a small model only to predict the next legal move in the board game Othello, and then went looking inside it, they found something it was never told to build: a representation of the board. The model had no eyes and no rules. It saw only strings of moves. To predict the next move well, it had reconstructed the thing the moves were about. You can probe that internal board, flip a piece in the model’s “mind,” and watch its predictions change accordingly. That result, and a stack of others like it, is why the cleanest version of the mimicry argument fails. To predict well enough, the system is pressured to understand. Ilya Sutskever, who has as much claim as anyone to having seen this from the inside, has made the same point in plain terms: predicting the next token well means understanding the reality that produced it.

That is the deep reason the next-token bet is not obviously stupid. Compression is comprehension. If you can predict something, you have modeled it. My confidence that these systems build real internal models of the things they discuss is high. My confidence that text-only prediction builds a model rich enough for full general intelligence is much lower, and that is the crack LeCun keeps his thumb in.

The thing the skeptics got wrong, and the thing they got right

For a few years the recipe was simple. Make the model bigger, feed it more text, give it more computers, and it got better on a smooth and almost embarrassingly predictable curve. People called these the scaling laws, and they held across several jumps in size. That era is closing.

OpenAI’s big 2025 pretraining run, the one that shipped as GPT-4.5, was the tell. It cost an enormous amount and delivered a much smaller jump than the prior generation had. Sutskever said the quiet thing at a December 2024 talk: pretraining as we have known it will end, because data is the fuel and we have only one internet. The supply of high-quality human text is finite, and the largest runs are already drinking from the bottom of the glass. The skeptics who said “the scaling wall is real” were right.

What they missed is where the road turned. The labs stopped trying to win only by making the model bigger and started spending more at the moment of use. Instead of answering instantly, the newest models think first. They write out a long internal chain of steps, check their own work, notice a wrong turn, back up, and try again before they answer. OpenAI’s o-series and DeepSeek’s R1 are built this way, trained by reinforcement learning, which means the model gets rewarded for reasoning that reaches correct answers and learns to do more of it. This is still a next-token predictor. It is the same engine, now allowed to talk to itself on the way to a reply, and rewarded for talking itself into better answers.

Notice what that resembles. When a model writes out its reasoning, evaluates it, catches its own error, and corrects course, it is doing a crude version of the loop people point to when they say humans are more than pattern matchers. We deliberate. We hold a thought, inspect it, argue with it, and revise. The reasoning models externalize that loop onto the page in tokens. It is the difference between a streetballer who reacts and a point guard who reads the defense, runs two options in his head, and picks the better one before he moves. Whether the model’s version is genuine reasoning or fast retrieval dressed as reasoning is a real fight, and the honest answer is that it is some of both, in proportions nobody can yet measure. Moderate confidence that the reasoning is partly real, and low confidence on how far it generalizes past the patterns it was trained on.

This matters for the central question because it changes what “all you need” means. If you had asked in 2024 whether scaling the next-token predictor was enough, the honest answer was becoming no. The pretraining curve was bending. But the paradigm did not die. It grew a second engine, the thinking-at-inference engine, and that engine is young and its own curve has not bent yet. So the question is not settled by the pretraining slowdown the way the skeptics hoped. The bet just moved to a different table.

The strongest card the bulls hold

Here is where the facts get genuinely hard to wave away, and where I had to verify every number before I’d repeat it, because the loose versions floating around are inflated.

In May 2026, Anthropic published its own internal data on how much of its work the AI now does. More than 80% of the code merged into Anthropic’s own codebase was written by Claude. Before its coding agent launched in early 2025, that figure sat in the low single digits. The company’s leadership puts the looser number, counting scripts and throwaway code, north of 90%. The output per engineer tells the same story: in the second quarter of 2026 a typical Anthropic engineer was shipping roughly eight times the code per day they shipped in 2024. One engineer, the company reports, had not written a line by hand in five months.

The capability behind that is climbing on a curve worth staring at. The length of task a model can finish on its own, with no human stepping in, has been doubling about every four months, up from every seven. In March 2024 the best model could handle a software task that takes a person about four minutes. A year later, about ninety minutes. By 2026, twelve-hour tasks. A research preview, the unreleased Mythos model, ran for at least sixteen hours on its own, which is the edge of what the outside evaluators could even measure. On one optimization problem, that preview found a 52x speedup where a strong human researcher typically gets about 4x in a half-day of work.

And the AI is starting to improve the AI. Google’s AlphaEvolve, a system that uses models to search for better algorithms, found a way to multiply a certain class of matrices using 48 multiplications instead of 49. That sounds trivial until you learn the previous record stood since 1969, for 56 years, and that the result is mathematically verifiable, not a vibe. The same system claws back about 0.7% of Google’s entire worldwide computing fleet by scheduling it better, and sped up a key training routine enough to cut roughly 1% off the time to train Google’s flagship model. A model, helping train its successor. That is the loop everyone is watching for, observed in a small but real form.

If you wanted to make the case that we are in or near a hard takeoff, this is the case. The thing improves. The thing now helps build the next thing. The task horizon is doubling fast. The forecasters who called this are not all cranks either. Leopold Aschenbrenner, who wrote a widely read 2024 essay arguing the trendlines pointed to roughly human-level AI by 2027 and then a fast intelligence explosion, got the infrastructure story very right; the trillion-dollar buildout he predicted is happening, and he now runs a hedge fund betting on the picks-and-shovels of it. Ray Kurzweil has held to AI matching humans by 2029 and a full singularity around 2045, and his 2029 date, once fringe, now sits inside the range serious lab leaders give. The “AI 2027” scenario laid out a month-by-month path through an automated-research explosion that, read in 2026, does not feel like science fiction.

The thing the bulls keep getting wrong

Now turn the same facts over and look at the underside.

Start with that 80% figure, because it is the one people quote most and understand least. It is Anthropic’s number for Anthropic’s codebase, not an industry truth, and writing code is the single task these models are best at, the home court. Software has a property most work lacks: you can check the answer automatically. The code runs or it doesn’t, the test passes or it doesn’t, and that clean signal is exactly what reinforcement learning needs to train on. The horizon doubling every four months is measured mostly on coding and technical tasks for the same reason. None of this tells you the model is as far along at the messy, unverifiable work that fills most of the economy, the negotiation, the judgment call with no test suite, the decision about what is even worth doing.

That last one is the wall the labs keep hitting, and to their credit they say so. Anthropic’s own paper, the one with the 80% number, is built around the idea of AI building itself, and its own researchers write that recursive self-improvement “is not here, nor is it inevitable.” Their most likely scenario is not the runaway. It is a world where the AI does more and more of the doing while humans keep setting the direction. The bottleneck they name is research taste, the senior-level judgment about which problem to chase and which result to trust. In late 2025 the model picked a better next research step than the human about half the time; months later, closer to two-thirds. Climbing, clearly. Closed, no.

It is worth being skeptical even of the impressive anecdotes. The paper’s showcase example, a model that diagnosed a nasty production incident in two hours that would have taken a person days, is a real and useful thing. It is also, when you read it closely, classic debugging: a clear problem, rich error data, a fix to be found. A model finding an obscure flag faster than a tired human is the compiler catching your typo, scaled up. It is enormously valuable. It is not the same as the model deciding, unprompted, what the company should build next quarter, which is the capability that would actually close the loop.

Then there is generalization, the soft spot under all of it. There is a test called ARC-AGI, built specifically to be easy for humans and hard for memorized knowledge. Its second version, released in 2025, knocked frontier models down to near zero at launch while ordinary people solved the puzzles without much trouble. Scores have since climbed, at high cost, which tells you the wall is scalable but not free. Apple put out a paper showing that reasoning models, pushed past a certain complexity, do not degrade gracefully; they collapse, and stranger still, they sometimes try less as the problem gets harder. The rebuttals were fair, some of Apple’s puzzles were rigged in ways that guaranteed failure, but the core point survived the fight. These systems have a frontier of difficulty past which the reasoning stops being reasoning, and we do not know how to push that frontier reliably with scale alone. Moderate-to-high confidence that this limit is real; moderate confidence on how binding it stays as the inference-time engine matures.

Now back to the opening, because the causal story behind the hard-takeoff thesis is where it breaks hardest. The claim is that government is pulling the best models off the market, which lets the labs keep those models internal, which means recursive self-improvement is happening behind the firewall, out of view. The June 2 executive order is voluntary; it explicitly does not create a licensing or pre-clearance requirement, and the 30-day window it describes gives the government an early look, not the lab a private runway. The Fable shutdown removed the model from everyone, including the labs’ own foreign-national staff. When the government restricted OpenAI’s GPT-5.6 later that month to about twenty approved partners, OpenAI pushed back in public, saying this kind of gating “should not be the long-term default.” The labs are fighting to release these models, not hoarding them to self-improve in the dark. The internal-versus-public capability gap is real, and it is mostly explained by dull things: safety testing, the cost of serving a model to hundreds of millions of people, and the obvious competitive logic of not handing rivals your sharpest research accelerant. A gap that boring is not evidence of a secret intelligence explosion. It is evidence of caution and economics.

And the physical world is slower than the digital one by a margin that bounds how fast any of this can hit the economy. Roughly half the AI data centers planned for 2026 in the United States have slipped or been canceled. Transformers are back-ordered, the power grid is strained, and only about a third of the new capacity people projected is actually under construction. You can have an algorithmic breakthrough on a Tuesday. You cannot conjure a gigawatt of power and the steel to use it on a Wednesday. Even Aschenbrenner’s own thesis leans on this; his hedge-fund bet is less about clever code and more about electrons, because electrons are the constraint that clever code runs into.

The forecasters are slipping accordingly. The “AI 2027” authors have quietly walked their median toward 2029 and 2030. Aschenbrenner’s revenue prediction for mid-2026 came in well under his line, and his call that open-source models would fade was flatly wrong; the cheap open models from China are sitting right behind the frontier and forcing prices down. Being early and being wrong are different, and these are early. But early enough that anyone underwriting against 2027 as a date should stop.

So which is it

Both pictures are true, which is why smart people keep talking past each other. The capability is compounding faster than the skeptics admit. The runaway is further off than the bulls claim. The shape that fits the evidence is a fast soft takeoff: a steep, accelerating ramp where the AI gets dramatically more capable and starts meaningfully speeding up its own development, while humans stay in the loop at the points that matter and the physical world throttles how fast any of it reaches the ground. The loop is real and it is open. Closing it requires automating the senior judgment that the labs themselves admit they have not automated, and clearing a generalization wall we do not know how to clear on command.

My honest probabilities, held loosely: fast soft takeoff as the base case, the most likely world by a comfortable margin. A genuine hard takeoff, the weeks-to-months runaway, as a real tail, somewhere in the rough range of one in seven to one in four this decade, mostly through the automated-research channel if that research-taste gap closes faster than the physical constraints bite. Not negligible. Not the base case. Anyone who tells you they know which of these we are in with confidence is selling something, possibly a fund.

On the title question, then. Is the next token all you need? For an AI that is superhuman across most of what can be checked and verified, probably yes, and we are most of the way there. For full self-improving superintelligence, unproven, and the honest word is unproven rather than no, because the next-token engine keeps doing things its critics swore it couldn’t, and because the inference-time reasoning engine bolted onto it is too young to have shown its ceiling.

One more thing, on the seductive line that humans are just next-token predictors too. There is real science under it. A leading theory in neuroscience holds that the brain is fundamentally a prediction machine, constantly guessing its next sensory input and learning from the error. When you read this sentence, your brain is predicting the word before your eyes reach it. The rhyme with a language model is not an accident. But the brain predicts a flood of sound and sight and touch and its own movement, grounded in a body, on about twenty watts, learning continuously as it goes. It was not trained by gradient descent on the internet. Calling a human a next-token predictor is a metaphor that illuminates one shared trick and hides a dozen differences that might be the whole game. Use it to understand why the bet is plausible. Do not mistake it for a proof that the bet pays.

What this means if you allocate capital

Switch tracks now, from what is true to what to do about it, because the two get jumbled and both suffer for it.

The money has already voted, hard. In the first quarter of 2026, global venture funding hit about $300 billion, and roughly 80% of it, around $240 billion, went to AI. Four rounds, OpenAI at $122 billion, Anthropic at $30 billion, xAI at $20 billion, and Waymo at $16 billion, took nearly two-thirds of all venture dollars on Earth. A single quarter of AI funding exceeded all of 2025. That is not a sector. That is a gravity well.

Here is the part most people get backwards, and it is the same backwards as the Fable story. A soft takeoff is the harder world to invest in, not the easier one. If a hard takeoff were imminent, almost nothing you funded at the application layer would matter, because a superintelligence would eat every workflow at once and the only sane bets would be the compute and the power underneath it. The soft takeoff is more demanding precisely because the technology keeps getting better and cheaper underneath every company you back, on a schedule, for years. The model that makes your portfolio company magic this year is the commodity that makes it ordinary next year. The cost of a million tokens fell about 80% from 2023 to 2025. Any business whose only edge was reselling access to a model has already watched its margin evaporate.

So the test experienced investors are starting to apply is brutal and simple. Two questions, and a company needs a real answer to both:

• If a frontier lab shipped your exact product as a default feature in their next release, would your customers cancel?

• Does what you do survive three more model generations getting cheaper and smarter?

Most thin wrappers around someone else’s model fail both. What passes tends to own something the model maker cannot reach from a data center. Proprietary data that compounds the more the product gets used. A workflow so embedded in how a company runs that ripping it out costs more than tolerating it, which is why coding tools, legal-research tools, and enterprise-search tools that became the system of record have held up while generic chat skins have not. Trust and compliance in regulated corners, healthcare, finance, defense, where being right and being accountable matter more than being clever. And the physical world, robots and the machinery of moving atoms, where the bottleneck is not tokens and the frontier labs have no special advantage.

The two layers worth real money look opposite and are both defensible. One end is the picks and shovels: compute, power, memory, the boring infrastructure the whole boom runs on, which pays off in almost every scenario including the scary one. The other end is the deep vertical application that owns its data and its customer so completely that a better base model helps it rather than kills it. The undifferentiated middle, the layer that is just a prompt and a logo on top of an API, is where capital goes to die, and it is where a frightening amount of 2026 seed money is going anyway.

For the people who fund the funds, the limited partners, the uncomfortable truth is that AI is your largest exposure whether you chose it or not, through the venture portfolios, through the public indices where the ten biggest companies now make up a share of the market that exceeds the peak before the 2000 crash. The question is not whether to be exposed. It is whether the exposure is concentrated in things that survive a soft takeoff, where defensibility and timing decide everything, or scattered across things that the next model release quietly deletes. The barbell, infrastructure on one end and defensible verticals on the other, with as little as possible in the middle, is the shape that respects both the upside and the wall.

And the macro claim you will hear at the top of every keynote, that AI and robots will multiply the global economy tenfold in a decade, the Musk number: low confidence, and I would bet against it on that timeline. The serious range from people who model growth for a living runs from a rounding error to something genuinely large, several percentage points of added output over a decade in the credible middle, with explosive growth as a real but later and far less certain tail. A tenfold jump in ten years requires the hard-takeoff world and the physical buildout to both arrive on schedule, and the physical buildout is already slipping. Plan for a serious productivity boom. Do not underwrite a miracle.

The shutdown on that Friday in June is the whole thing in miniature. A model good enough to scare a government (encouraged by fear based marketing by the labs), killed by a letter, pushed out of reach instead of locked away to grow. Powerful and constrained at the same time, the capability sprinting while the institutions and the power grid jog to keep up. That is the texture of a soft takeoff. It is less cinematic than the runaway, and harder to live inside, because it asks you to keep making good decisions year after year while the ground moves under you, instead of betting everything once on a single discontinuity. The next token might well be most of what you need. It is not, yet, all of what it would take to stop needing us. The interesting years are the ones in between, and we are in them.

An agent is a program that uses an AI model to do work on its own. It queries, retries, checks itself, and loops, the way a tireless junior analyst might if the analyst never slept and never got bored. Point a fleet of them at a conversational data tool like the one Databricks sells and they hammer the underlying database around the clock. Every query costs something to serve. Multiply by millions of queries and a six-point hole opens in the margin.

Sit with that, because it flips the oldest fact in the software business. The reason software became the best business model ever invented is that once you have built the thing, serving one more user costs almost nothing. Usage was free money. A customer who used your product twice as much made you richer at no extra cost. That is the entire reason a good software company can earn 80 cents on the dollar.

Agents changed the sign on that equation. Now the heaviest users are the ones quietly draining you, because behind every agent action sits a model call, and model calls cost real money every single time. The unit of consumption is the token, which is roughly a syllable of text the model reads or writes, and you pay per token whether the work was useful or the agent just spun in a circle. The meter is always running now. This is the thing the whole week was secretly about.

A budget is the first thing to break

You can watch the same shift hit from the buyer’s side. Uber, a company that instruments everything and is not in the habit of losing track of a cost, reportedly gave its roughly five thousand engineers a monthly allowance for AI coding tools, somewhere between $500 to $2,000 dollars a head. By April they had spent the entire year’s budget. The company then capped it. Walmart, Cisco, Amazon, and Meta reportedly did versions of the same thing.

When a company that good at counting misjudges a number by a factor of three, the number is new. Nobody overruns the electricity bill by 3x, because everyone has been paying electricity bills for a hundred years and the intuitions are baked in. AI spend has no baked-in intuition yet. It was modeled like a software seat, a fixed price per person per month, and it behaved like an appetite. Give a smart engineer an agent that can run a hundred experiments overnight and the engineer will run a hundred experiments overnight, because why wouldn’t they. The tool got good, so people used it more, so it cost more. The better it works, the more it costs, which is a sentence that has never been true about software until right now.

Databricks’ own response tells you how serious this is. The company built an internal reinforcement-learning tool, reportedly called KARL, whose entire job is to make its customers’ agents query less expensively. When a software vendor has to ship a feature to stop its own product from being used too hard, the economics have moved under everyone’s feet.

Rationing from the top

Here is where the week turns from a finance story into a power story. The same week CFOs started rationing tokens from the demand side, the government started rationing models from the supply side.

OpenAI’s newest model, the GPT-5.6 line, did not get a normal launch. It went out under a review where the federal government cleared access customer by customer, with Commerce sitting in the middle deciding who was allowed in. Anthropic’s most capable models were restricted to a small set of vetted cybersecurity and infrastructure firms under a defensive program, after an earlier outright block. For the first time, the best American models are shipping by permission rather than by press release.

The reason is not hard to see. These models are now good enough at finding software vulnerabilities and writing exploits that handing the strongest version to anyone with a credit card looks, from a national-security desk, like handing out a capability rather than a product. Whether that framing is right is a separate argument (although this is exactly what Anthropic’s fear based marketing was asking for). What matters for anyone building a company is that frontier-model access just became a thing a government can switch off on a Tuesday.

Alex Wissner-Gross, the researcher who writes The Innermost Loop, made the sharp observation about what gating actually does. Slowing how fast a lab is allowed to ship does nothing to slow how fast it is allowed to train. So the distance between what the public can use and what exists privately inside the labs does not hold steady. It widens. Every week the public frontier is held back, the real frontier keeps moving, and the gap becomes a strategic asset owned by a handful of organizations and their chosen customers.

Put the three pieces side by side and they rhyme. A CFO rationing tokens by budget. A vendor rationing queries with a throttling tool. A government rationing the model itself by clearance. Three different actors, same underlying move, because the same thing happened to all of them. A resource that felt free when it was a demo became scarce the moment it became a workload that runs continuously. And scarce things get rationed by whoever controls the choke point.

Who wins when the meter rules

If usage is the cost, then the winner is whoever can afford to let the meter run without flinching. That is a very different winner than the one most people were betting on a year ago.

The clearest example is the strangest one. SpaceX, fresh off going public, used its newly liquid stock to buy Cursor (as planned), the AI coding editor, in an all-stock deal reported at around sixty billion dollars, after having already absorbed xAI. A rocket company now owns the satellite network, the supercomputer, the model, and the screen the developer types into. People keep asking why a space company is winning the coding-tool war. The answer is the meter. SpaceX can let agents run flat out because it owns the electricity and the silicon underneath them. It is not watching the token bill the way a venture-funded wrapper has to. When the cost of a product is compute and power, the company that owns compute and power can simply outlast everyone selling a thin layer on top.

Notice the deal was paid in stock, not cash. That is its own signal. When the strongest player in the market pays for a sixty-billion-dollar acquisition with its own shares instead of money, it is telling you what it thinks those shares are worth, and it is conserving the cash for the thing that actually constrains it, which is building more compute and securing more power. (Worth noting the acquisition valuation was secured at a lower SpaceX valuation so effectively they paid something like half the advertised priced!)

The talent flows confirm the same gravity. John Jumper, who shared a Nobel Prize for AlphaFold, left Google DeepMind. Noam Shazeer, a co-author of the original transformer paper and a researcher Google reportedly paid billions to bring back in 2024, left for OpenAI. Within days of each other, two of the most valuable researchers alive walked out of a diversified giant and into pure-play labs. The best people are choosing the places that are allowed to push the frontier hardest, which widens the capability gap, which makes the pure-play labs even more attractive to the next departure. The loop feeds itself.

The constraint that everyone underweights is power

Follow the meter far enough and it stops being about money and starts being about electricity. Agents that run continuously are workloads that run continuously, and workloads need power. Projections for data-center electricity demand keep getting revised up, fast enough that the binding limit on the next phase of AI may turn out to be the grid rather than the model.

This is why the orbital-compute conversation stopped sounding like science fiction this week. On the Moonshots podcast, Planet Labs’ Will Marshall and others walked through the case for running AI inference on satellites, processing remote-sensing data in space and only beaming down the conclusions, paired with the idea of powering data centers with continuous solar arrays in orbit where the sun never sets. You can dismiss the specifics. The underlying instinct is correct. When power becomes the constraint, people start looking for power in places they previously ignored, including up. The serious money in AI is migrating toward whoever can generate, secure, and cool electricity at scale, and that is a very physical, very capital-heavy place for a software boom to end up.

What this means for founders

Separate the building advice from the investing advice, because they point in slightly different directions.

If you are building, the meter reorganizes what is worth building. The wedges that got more attractive this week are the ones that either dodge the meter, control it, or survive someone else flipping a switch.

A routing and orchestration layer that lets an application swap between a US model and an open-weight one in real time is now close to mandatory infrastructure, because the alternative is having your product bricked the day your single provider gets restricted. On-device and edge inference got more interesting for the same reason, since running the model locally sidesteps both the cloud bill and the question of who is legally allowed to call the API. Anything that throttles, caches, or optimizes agent spend is selling directly into the wound Databricks just showed everyone, so cost-control tooling went from a nice-to-have to a budget line. And dual-use work tied to compute, secure communications, and energy lines up with where the capital and the procurement dollars are actually flowing.

The wedges that got less attractive are the mirror image. A single-model wrapper with no proprietary data is now squeezed from above by the platforms bundling the same feature and from the side by the model getting restricted out from under it. Flat-priced agentic software sitting on top of uncapped API spend is a margin trap waiting to spring, because your heaviest users are your biggest losses. And services built on manual offshore QA or junior-developer outsourcing are racing automated patching that gets cheaper every month.

The questions worth forcing in a pitch this week are concrete:

• If your model provider gets restricted on Tuesday, does your product still work on Wednesday?

• Does your pricing survive a customer whose agent runs a thousand times more often than a human would?

• Are your heaviest users your most profitable accounts or your biggest losses, and can you tell me the number?

• Can a foreign-national engineer on your team legally touch your primary model API next quarter, and what breaks if the answer becomes no?

• Are you selling a capability demo, or a line item a CFO will defend during the next budget cut?

Have you tried the new open-weight models?

There is a question worth running before you raise another dollar to cover your model bill: have you actually tested the open-weight models lately, or are you paying frontier prices out of habit?

The gap between the best closed model and the best open one has been collapsing on exactly the tasks most startups run, the coding, extraction, summarization, and routing work that makes up the bulk of real production traffic. The frontier still wins at the hard edge, the long-horizon reasoning and the genuinely novel problem. Most products do not live at that edge. They live in the middle, doing the same bounded task ten million times, and in the middle an open model running on rented hardware can deliver something close to the same answer for a fraction of the price. The arithmetic that matters is blunt: if a capable open-weight model gets you 95 percent of the value at 85 percent lower cost, the 5 percent you are paying frontier rates to recover has to be worth more than the margin you are burning to get it. For most workloads it is not.

So measure it instead of assuming it. Take your actual production traffic, not a benchmark, and replay a representative slice through an open model. Score the outputs against what your frontier provider returns. You will usually find the work splits cleanly. A large share comes back indistinguishable, a slice is good enough with light guardrails, and a thin tail genuinely needs the frontier. Route on that finding. Send the bulk to the cheap model, keep the frontier for the tail, and you have just rebuilt your cost structure without touching the product the customer sees.

The economics compound past the per-token savings. An open model you can host or fine-tune is one nobody can restrict out from under you on a Tuesday, which is no longer a hypothetical given how this month went. It is one you can run at the edge or on-prem for customers who care where their data sits. And it is one whose cost you control rather than rent, which means your margin stops being a decision your model provider makes for you. The companies that win the next year will not be the ones using the smartest model. They will be the ones who figured out the cheapest model that clears the bar for each job, and who built the routing to put every request in front of the right one.

The reflex to default to the frontier for everything was rational when the gap was wide and the price difference was small. Both halves of that flipped. If you have not re-run the test in the last quarter, your cost structure is built on a price comparison that is already stale.

What this means for LPs

The week sharpened a dispersion that has been building for months. Capital is concentrating violently into a few platforms that own compute and power, while the messy middle of enterprise software gets repriced down toward acquisition value. That gap is the whole environment, and it argues for two disciplines at once.

At the early stage, the discipline is to fund the throttle and the portability layer and to refuse the high-burn wrapper, because the wrapper is exactly the business model the meter punishes. A fund built to write small seed checks into model-portable, cost-aware, infrastructure-adjacent companies is positioned for this. A fund chasing flashy consumer agents with uncapped token bills is underwriting margin collapse and has not noticed yet.

In the secondary book, the discipline is to price on real numbers rather than momentum. Databricks growing past a multi-billion-dollar revenue run rate while its margin compresses to 74 percent is the tell. Even the winners are absorbing the same cost shock, so the right entry mark is a function of revenue quality and gross margin, not the most recent post-IPO headline. Funding is a price signal, never a quality signal. The SpaceX print is dazzling, and the correct response to a dazzling print is discipline, not fear of missing out.

What this means for the venture market

Three structural facts surfaced this week that change how the asset class behaves.

Liquidity came back, and it came back concentrated. The exits and the up-rounds are clustering in a handful of names while the long tail waits. Access to those names matters less than the price you pay to get in.

Acquisitions are being paid in inflated equity. A sixty-billion-dollar all-stock deal is a barter transaction between two richly valued private currencies, and it tells you the acquirer would rather spend paper than cash. Read those marks as directional, not precise.

The IPO line is real but orderly. The strongest names appear to be staggering their public debuts so they do not flood the same investor base at once, which means the public-market repricing of private tech will arrive in waves rather than a single reckoning. That gives a disciplined secondary investor time, and time is the one thing a momentum chaser never uses well.

What this means for VCs

The meter changes diligence before it changes anything else. For two years the central question in an AI deal was whether the product worked. That question is close to free now, because the models got good enough that most demos work. The question that separates winners from margin traps is whether the company makes money when the product works hard. So the diligence that matters moved from the demo to the bill. Ask for token cost per unit of delivered work, ask how it trends as usage scales, and treat any founder who cannot answer in those terms the way you would treat a SaaS founder in 2015 who could not tell you their gross margin.

The harder truth is that a lot of what looked like product moat this year was the model doing the work, and the model is rented. If a startup’s edge is capability the foundation lab can ship in its next release or a regulator can switch off by name, you are underwriting a lease, not an asset. The durable edge sits in the places the meter creates: proprietary data the model cannot get elsewhere, a workflow lock that survives a model swap, distribution into a budget line a buyer will defend, and unit economics that improve rather than decay as the agents run. Price the lease cheaply. Pay up only for the asset.

Check construction has to respect the new cost curve too. An early AI company now has a cost of goods that scales with adoption, which means a seed round that would have lasted eighteen months on a 2019 burn profile can evaporate in nine if the product takes off, because success spends compute. Founders who win burn cash by being used. Reserve accordingly, and stress the model against the good case, not just the bad one, since the good case is where the token bill explodes.

The questions worth forcing in a partner meeting this week:

• Does this company’s edge survive the next frontier-model release, or is it renting capability that the platform will absorb?

• What is the gross margin at scale once agents, not humans, are the primary users, and is the founder even measuring it?

• If the primary model provider gets restricted, does the company have a portability story or a dependency it has been calling a feature?

• Are we paying for an asset the company owns or a lease on someone else’s model?

On strategy, the consolidation cuts against the late-stage growth game and toward the early one. When a handful of platforms own compute, power, and distribution, the value they create accrues to them, and writing a growth check into that gravity well is buying a crowded, richly priced position. The unowned ground is early, in the throttle-and-portability layer the platforms have no incentive to build and the application wedges too small for them to chase. That is where a disciplined seed fund still gets real ownership at a price that leaves room. Funding remains a price signal and not a quality signal, and this week the loudest prices were in exactly the places where the next dollar is least likely to compound.

The scarcity moved

The assumption underneath the last two years was that if intelligence got cheap, it would get abundant, the way bandwidth did. Cheap bandwidth gave us streaming video and nobody thinks about the cost of a megabyte anymore. We expected the same arc for tokens.

It is going the other way. Intelligence is getting cheaper per token and more expensive in total, because we have learned to consume so much more of it (a Jevon’s paradox!), and the things it actually rests on are not getting cheaper at all. Power is getting scarcer. High-end silicon is getting scarcer. And permission, the right to run the very best models, just became something a government rations by name.

The scarcity did not disappear when the models got good. It moved. It used to live inside the model, in the cleverness that was hard to build. Now it lives in the meter, the grid, and the gate. The companies worth backing are the ones building where the scarcity actually went.

We already published a separate batch takeaways post with our broader impressions of the cohort, category-level observations, and what we think this batch says about where early-stage AI and software are heading. You can read that here:

Ignite Insights

What YC Spring 2026 Felt Like From the Room

Today is demo day…

Read more

2 months ago · 7 likes · Ignite Insights

This post is more direct: here are the companies we backed, what they do, and why we were excited to invest.

A few patterns stood out.

First, AI is moving from “copilot” to operating system. Many of the best companies are not building thin wrappers. They are trying to own full workflows: customer success, medical practice operations, construction estimating, pathology reporting, observability, marketing execution, sales development, and more.

Second, infrastructure remains a major theme. GPU utilization, HPC backtesting, app observability, and AI-native workflow systems are becoming more valuable as AI increases the volume and complexity of work.

Third, the batch was broader than generic SaaS. We invested across defense, aerospace, healthcare, fintech infrastructure, devtools, vertical AI, and applied AI agents. That matters because the biggest outcomes rarely look obvious at the start.

Our job is not to wait until consensus forms. Our job is to identify early signal, move quickly, and earn allocation before the round is gone.

The Companies We Backed (Alphabetically)

Arlo Industries

What they do: Arlo Industries is building a passive aerial sensing mesh to track drones, missiles, and other aerial threats without traditional radar.
YC Profile: https://www.ycombinator.com/companies/arlo-industries

Why we invested: Defense is being reshaped in real time by low-cost drones, autonomous systems, and asymmetric warfare. Legacy radar systems were not designed for a world where cheap aerial threats can appear everywhere at once. Arlo’s distributed, passive sensing architecture is compelling because it attacks both the technical problem and the economic problem: how do you create persistent, wide-area coverage without relying on centralized, expensive infrastructure? We were attracted to the severity of the problem, the timing, and the potential for Arlo to become a foundational sensing layer for modern defense.

Dispatch

What they do: Dispatch is building refurbishable reentry vehicles that can host and return payloads for companies manufacturing ultra-high-value materials in microgravity.
YC Profile: https://www.ycombinator.com/companies/dispatch

Why we invested: Space manufacturing is one of those categories that sounds futuristic until the enabling infrastructure arrives. The core bottleneck is simple: if companies can manufacture valuable materials in space, they still need a reliable way to bring those materials back to Earth. Dispatch is attacking that bottleneck directly. We liked the ambition, the technical depth, and the founder-market fit from a team with relevant spacecraft experience. This is a high-risk, capital-intensive category, but the upside case is enormous if Dispatch becomes a core logistics layer for in-space manufacturing.

Expanse

What they do: Expanse helps teams recover wasted GPU and HPC capacity by predicting the resources compute jobs actually need before they run.
YC Profile: https://www.ycombinator.com/companies/expanse

Why we invested: AI infrastructure spend is exploding, but much of that compute is still wasted through over-provisioning, failed jobs, and poor resource prediction. Expanse is building an intelligence layer for GPU and HPC infrastructure: read the job, understand the cluster, predict what resources are needed, and reduce waste before money is burned. The team has unusually strong founder-market fit, having worked directly on the kinds of HPC and GPU workloads they now serve. We invested because compute efficiency is becoming a strategic budget issue, and Expanse has a credible path to becoming a control point for cluster utilization.

InstaAgent

What they do: InstaAgent helps consumer brands create, distribute, test, and learn from large volumes of personalized social creative across personas and channels.
YC Profile: https://www.ycombinator.com/companies/instaagent

Why we invested: AI has made content creation cheap. It has not made content effective. That distinction matters. Consumer brands do not need more generic AI slop; they need high-quality creative variation, fast testing, and a learning loop that compounds across channels. InstaAgent’s wedge is scaled social creative generation and distribution, but the bigger opportunity is becoming a workflow layer for modern brand marketing. We liked the team’s speed, their understanding of social distribution, and the possibility that creative testing becomes more software-like as AI changes how brands operate.

Keyframe Labs

What they do: Keyframe Labs is building visual AI avatars and multimodal agents designed to become more scalable, expressive interfaces for AI applications.

YC Profile: https://www.ycombinator.com/companies/keyframe-labs

Why we invested: Voice AI was one major interface shift. Visual, embodied, multimodal AI may be the next. Keyframe is attacking the cost and scalability constraints that have historically made realistic avatars difficult to deploy widely. We liked the team’s technical ambition and the possibility that avatars become a core interface layer for sales, education, support, entertainment, and agentic software. If AI agents are going to represent companies, teach users, sell products, or guide workflows, they may need faces, presence, and visual interaction—not just text boxes and voice streams.

Klarify

What they do: Klarify is building AI software for therapists and mental health practices, starting with documentation and expanding into broader administrative workflows.
YC Profile: https://www.ycombinator.com/companies/klarify

Why we invested: The best vertical AI companies do not replace the professional; they remove the non-core work that prevents the professional from doing their highest-value job. Klarify’s insight is exactly that. Therapists are overburdened by notes, treatment plans, claims support, scheduling, payments, and client follow-up. Klarify keeps the human therapist at the center and uses AI to automate the surrounding operating burden. We liked the clarity of the wedge, the severity of the administrative pain, and the potential to expand from notes into the operating system for small and mid-sized mental health practices.

Memory Store

What they do: Memory Store is building memory infrastructure for AI applications, helping AI systems retain, retrieve, and use context over time.
YC Profile: https://www.ycombinator.com/companies/memory-store

Why we invested: Memory is one of the most important unsolved primitives in AI applications. Today, many AI products feel impressive in isolated interactions but weak across time because they do not remember enough, structure context well enough, or retrieve the right history at the right moment. Memory Store sits at a foundational layer: persistent memory for AI-native software. We invested because every serious AI workflow eventually needs durable context, personalization, and recall. If they become the memory layer for a meaningful share of agentic applications, the opportunity is large.

Oddpool

What they do: Oddpool is building data and infrastructure for prediction markets and event-based financial markets.
YC Profile: https://www.ycombinator.com/companies/oddpool

Why we invested: Prediction markets are moving from niche curiosity to serious financial infrastructure. As venues, assets, and trading strategies proliferate, institutions need normalized data, symbology, historical records, settlement metadata, and APIs they can build on. Oddpool’s opportunity is to become the neutral data and reference layer for this emerging market structure. We liked the infrastructure angle: as the category grows, high-quality historical data and normalized cross-venue infrastructure become harder to replicate and more valuable over time.

PerfectBit

What they do: PerfectBit is building AI infrastructure for software engineering and code intelligence.
YC Profile: https://www.ycombinator.com/companies/perfectbit

Why we invested: Software engineering is one of the first major labor markets being reshaped by AI, but the tooling stack is still early. The opportunity is not merely autocomplete. The larger prize is understanding codebases, automating complex engineering work, and helping technical teams ship faster with fewer bottlenecks. We liked PerfectBit because the category is enormous, the timing is right, and even small improvements in engineering productivity can create significant customer value. The risk is competition, but the market is so large that multiple durable companies can emerge.

PLAN0 AI

What they do: PLAN0 AI builds AI-native construction cost estimation and analytics software.
YC Profile: https://www.ycombinator.com/companies/plan0-ai

Why we invested: Construction cost estimation is slow, manual, expensive, and error-prone. It also sits at a critical point in the construction value chain: decisions made early can determine whether a project works economically years later. PLAN0 ingests floor plans and elevations, reconstructs projects, and helps produce cost estimates and scenario analysis far faster than traditional workflows. We invested because this is a severe, high-friction vertical problem with a clear AI wedge and a large potential expansion path into analytics, forecasting, and workflow ownership for construction teams.

Plena Health

What they do: Plena Health is building a full-stack operating system for specialty medical practices.
YC Profile: https://www.ycombinator.com/companies/plena-health

Why we invested: Specialty medical practices are operationally complex, understaffed, and buried under repetitive workflows: phones, scheduling, prior authorization, results follow-up, billing, collections, and EMR-adjacent administrative work. Plena’s thesis is that AI can run more of this operational stack end-to-end while working inside the practice’s existing systems. We liked the vertical depth, the wedge into painful back-office workflows, and the potential to become a deeply embedded operating layer for specialty care. Healthcare AI is crowded, but practice operations remain brutally inefficient and highly valuable if solved.

RASPIRE

What they do: RASPIRE builds runtime application security for mobile apps, protecting compiled applications from tampering, exploitation, and reverse engineering.
YC Profile: https://www.ycombinator.com/companies/raspire

Why we invested: Mobile apps are increasingly critical infrastructure for fintech, gaming, consumer software, health, and enterprise workflows. Yet many applications remain vulnerable once they are running in the wild. RASPIRE’s wedge is runtime protection: defending the app while it executes, not merely scanning code before release. We liked the severity of the security problem, the developer-facing adoption path, and the potential for RASPIRE to become a key protection layer for high-risk mobile applications. In a world of more AI-generated code and more automated attacks, runtime protection should become more important, not less.

Sherpa

What they do: Sherpa helps companies improve website conversion by using AI to test, optimize, and personalize growth experiences.

YC Profile: https://www.ycombinator.com/companies/sherpa

Why we invested: Growth teams know their websites leak revenue, but most conversion-rate optimization is slow, manual, and dependent on limited testing bandwidth. Sherpa’s opportunity is to turn CRO into an AI-native workflow: identify friction, propose improvements, run experiments, and compound learning across pages and customers. We liked the clear ROI, low-friction installation, and obvious pain point. The key question is whether Sherpa becomes a durable growth control plane rather than a point solution, but the wedge is strong and the buyer pain is easy to understand.

Superlog

What they do: Superlog is building autonomous observability for software teams.
YC Profile: https://www.ycombinator.com/companies/superlog

Why we invested: Observability has become essential, but it is often expensive, noisy, and painful to configure. Developers do not want more dashboards; they want systems that understand what is happening, collect the right context automatically, and help resolve issues faster. Superlog’s thesis is that observability should become more autonomous: less manual instrumentation, less configuration burden, more agentic debugging and system understanding. We liked the founder urgency, the technical wedge, and the category timing as AI changes how software is built, monitored, and repaired.

Userlens

What they do: Userlens builds AI customer-success agents that detect churn risk, generate account insights, prepare QBR materials, and help teams manage renewals and expansion.
YC Profile: https://www.ycombinator.com/companies/userlens

Why we invested: For B2B SaaS companies, retention and expansion are existential. The problem is that churn signals are scattered across product usage, CRM notes, billing, support tickets, customer conversations, and CSM intuition. Userlens turns that fragmented data into an AI CSM that can monitor accounts, identify risk, prepare playbooks, and support renewal workflows. We liked the repeat-founder angle, the clear pain from the founders’ prior company, and the possibility that customer success becomes increasingly agentic. The best version of Userlens is not a dashboard. It is an operating layer for revenue retention.

Voquill

What they do: Voquill is building an AI coworker for pathologists, starting with voice-driven report generation and expanding toward broader lab workflow automation.
YC Profile: https://www.ycombinator.com/companies/voquill

Why we invested: Pathologists spend an enormous amount of time documenting cases, producing reports, and navigating administrative friction. Voquill’s wedge is highly specific: listen as a pathologist works, learn their reporting style, and generate sign-out-ready reports. We like vertical AI products that start in a painful, repetitive workflow and expand into a larger system of action. Voquill is not a generic medical scribe; it is focused on pathology, where workflow specificity, reimbursement, and lab operations matter. If the company becomes embedded in pathology labs, the expansion opportunity is significant.

Zibra Labs

What they do: Zibra Labs builds high-performance computing infrastructure for quant trading firms running large-scale backtesting.

YC Profile: https://www.ycombinator.com/companies/zibra-labs

Why we invested: AI is accelerating the generation of candidate trading strategies, but backtesting infrastructure is becoming the bottleneck. Quant teams can produce more hypotheses than their systems can evaluate. Zibra Labs is building HPC infrastructure to make large-scale backtesting faster, cheaper, and easier to run across cloud resources. We liked the founder-market fit and the technical credibility of the team. This is a narrower market than generic cloud infrastructure, but it is a high-value buyer segment where performance, scale, and cost efficiency matter enormously.

Why This Matters for Fund III

This batch is a good example of why Team Ignite exists.

YC batches are now large, noisy, and fast-moving. The best companies do not wait politely for every investor to finish diligence. Rounds form quickly. Allocation disappears quickly. Consensus is usually late.

Our Fund III strategy is built around that reality: review the batch early, process signal quickly, meet founders before Demo Day when possible, and invest with disciplined speed.

Team Ignite Fund III is our second YC-focused fund. We have now made 101 YC investments to date in the fund, and the portfolio is already marked up at nearly 3x on 2024 investments. The fund is still young and still actively deploying, but the early signal is strong.

Fund III Snapshot

• Updated deck: https://tr.ee/yc_fund

• Join the fund: https://teamignite.decilehub.com/pacts?pid=7Na3Qz8D

• Closes End of June 2026

Our view is simple: if you want exposure to the next generation of YC companies, the edge is not waiting until everyone agrees. The edge is seeing enough of the batch, moving early, and being useful enough that founders want you on the cap table.

That is what Team Ignite is built to do.

Cursor was in talks to raise at fifty billion. Then a different conversation arrived.

Four days after SpaceX rang the Nasdaq bell on June 12, the company filed an 8-K, the regulatory disclosure public companies use to announce material events, to acquire Anysphere, the parent of the Cursor AI coding tool, in an all-stock deal valued at $60 billion. The fundraise turned into a merger. The price went up. The currency was stock that had been publicly tradable for four days.

Salesforce signed a $3.6 billion definitive agreement to buy Fin the same week, giving its Agentforce AI agent product a best-of-breed customer service layer. AWS turned bot payments and agent search into edge primitives, two unrelated-looking features that together outline a metered, cloud-controlled agent web. OpenAI launched a $150 million partner network with a target of 300,000 certified consultants by year-end and released a tool that simulates how a model will behave on real user conversations before it ships. The Fed held rates at 3.5 to 3.75 percent on June 17, with the FOMC statement saying inflation remains above the 2 percent goal. Translation: capital stays expensive while compute, power, and grid capacity stay strained.

Read each of these alone and they look like routine industry items. Read them together and the shape of the AI stack changes.

The independent application company, the place where most application-layer venture dollars have lived for three years, started to lose its standalone character last week. The middle is getting eaten from above by platforms turning features into primitives. It is getting eaten from across by strategic acquirers spending newly minted public stock. The squeeze is happening against a rate backdrop that is not going to bail anyone out.

The simplest test is to look at what got paid for.

What buyers paid for last week

Two of the better venture rounds of the week did not use the word copilot.

Convey raised $38 million in a Series A led by Andreessen Horowitz on June 17. The product is what the team calls operator-managed digital teammates: agents that nontechnical business operators build inside finance, accounting, marketing, and ad operations, that connect to legacy systems through IT-configured permissions, do measurable work, and write back into systems of record.

Gradial closed a $65 million Series C the same day, framing enterprise marketing as an execution problem. Their agents do the unglamorous middle of campaigns. Approvals. Publishing. QA. Reporting. The pitch is hours saved that show up on a budget line, not ideation that shows up on a slide.

Both companies look dull next to a foundation model release. Both are doing the most important thing a venture-stage AI company can do right now, which is own a workflow end-to-end. Write permissions. Audit trails. Integrations that take three months to put in and never get unwound.

Now look at the two acquisitions.

Customer service models are not scarce. Salesforce bought Fin for the route into its install base and the credibility of a best-of-breed agent already deployed at thousands of companies. Salesforce told the market its own agent platform, Agentforce, had hit $1.2 billion in annualized revenue last quarter, up around two hundred percent year over year. Adding Fin on top is a way to take a category that was rapidly commoditizing and lock the upgrade path inside an existing system of record.

The SpaceX end of the spectrum runs the same logic at a different scale. Cursor uses Claude, GPT, and an internal model called Composer. SpaceX bought it because Cursor sits inside a reported sixty-four percent of the Fortune 500’s developer workflows. That number is the company’s own and should be read as a marketing claim until diligence proves otherwise. Even if it is half right, the structural point holds. Developer tooling that has earned daily use across most large engineering organizations is a distribution asset that public stock can convert into ownership.

Two transactions in the same week valued at $60 billion and $3.6 billion. Both were about workflow real estate. Neither was about the model.

The other side of the squeeze

If strategics are buying the application layer where it stands on durable workflow, platforms are eating it where it doesn’t.

AWS announced two things last week that look unrelated and aren’t. On June 15, AWS WAF, the application firewall that fronts most large AWS-hosted properties, added a feature that lets content owners charge AI bots and agents per request at the network edge, with prices set by content path, bot category, or verification tier. On June 17, Amazon Bedrock AgentCore got Web Search as a managed tool. An agent built on AWS can now be grounded in current web data without wiring up a separate search vendor.

Read them together and the agent web has a price column and a default retrieval path. Retrieval goes through Bedrock. The price column lives at the edge of every AWS-fronted publisher. A startup whose product is web search for agents or a clearinghouse for bot payments saw its surface area shrink. A startup with a vertical knowledge corpus whose rights it controls saw it grow.

OpenAI’s two moves on June 16 echo the same logic. Deployment Simulation, in OpenAI’s own framing, lets the company replay prior conversations against a candidate model before release to forecast undesired behavior, surface novel misalignment, and reduce evaluation awareness. The Partner Network puts $150 million into systems integrators, consultancies, technology firms, and data partners, with a target of 300,000 certified consultants by the end of 2026. OpenAI is moving model release toward production simulation and enterprise deployment toward a partner-led channel. The shadow each casts over the venture market is the same. Independent AI eval startups whose product is a static leaderboard got harder to fund. Generic AI transformation consultancies got harder to fund. Tooling that plugs into the OpenAI channel and measures outcomes for partners got easier.

Microsoft posted enterprise AI guidance this week with the framing intelligence plus trust, model diversity, governance, observability, security, and FinOps for agents. Buyers should not be locked into a single model or harness, and should manage agent spend and behavior from one central control plane. Build something that strengthens that control plane and there is a place to stand. Build something that asks the enterprise to adopt a parallel system of record for agents and the sales motion gets long.

A macro that does not give the middle time

The Fed held the federal funds rate at 3.5 to 3.75 percent on June 17, with the FOMC, the Fed’s rate-setting committee, saying inflation remains above the 2 percent goal even as productivity growth and capital investment look strong. No signal of relief.

This matters more than the rate level itself. Capital stays expensive while compute, data centers, and grid capacity get more expensive. The companies that survive this environment are the ones whose unit economics work today, not the ones whose pitch deck has a hockey stick predicated on a model-cost crash that has not arrived.

For a wrapper company, this is the most painful macro available. They are already exposed to platform feature absorption from above. They do not get the rate cut that would buy another year of runway. The clock is the same color and it ticks faster.

DeepSeek closed a $7.4 billion fundraise the same week, reported by The Information on June 16, in a founder-controlled limited-partnership structure: no voting rights for commercial backers, five-year lockup, only the Chinese state AI fund taking direct equity. The deal mechanics are the news. Capital wants in. Founder Liang Wenfeng wants control. The signal for application-layer companies anywhere is that open-weight cost compression now has institutional patience behind it. The cheap end of the token market is going to keep getting cheaper on a schedule the closed labs cannot match.

Where capital still looks comfortable

Two places last week.

First, anything physical. Odyssey, the world-model company founded by ex-Wayve and ex-Cruise leaders, raised $310 million at a $1.45 billion valuation from a backer list that reads like sovereign and strategic capital playing one hand: Amazon, AMD Ventures, GV, EQT, In-Q-Tel. World-model capability claims remain vendor-stated until reproduced. Treat the headline cautiously. The investor mix is the tell. When strategics and government-linked vehicles anchor a Series B, the company is being underwritten as infrastructure.

Anthropic’s Frontier Red Team published Project Fetch Phase Two on June 18, showing Claude Opus 4.7 running robot-dog tasks roughly twenty times faster than the fastest human team had a year ago, with about ten times less code. The honest qualifier from Anthropic’s own write-up: the model still cannot precisely move a beach ball with the robot. The capability is real and incomplete. The investable claim is that fleet-learning robotics with a real data loop are getting cheaper to build per task completed. The advantage lives in the data, not the chassis.

Pegasus Tech Ventures and CYBERDYNE launched a roughly $60 million corporate venture fund on June 16 aimed at physical AI, automation, intelligent systems, and healthcare. The healthcare tilt makes this fund mixed-signal for a firm with an FDA exclusion. The broader pattern holds. Corporates with physical-world deployment channels want structured access to robotics startups, and they are willing to put a vehicle on the table to get it.

Second, anything that helps a buyer keep an agent on a leash. Agent permissions, agent identity, audit logs, deployment simulation, change management for nontechnical operators, model routing, FinOps for agents. The language is dry. The categories are the load-bearing plumbing of an industry whose buyers just got told by every major platform to make sure their agents are governed, observable, and secure.

The regulatory layer started taking shape

On June 17, OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Google DeepMind’s Demis Hassabis sat with G7 heads of state at a working lunch in Évian-les-Bains, France. Per Semafor, Altman pitched an international standards forum. Per CNBC, Amodei pushed a US-led coalition that excluded China from chip and frontier-model trade.

The substance of what was agreed is opaque. The structure is the news. AI governance moved from working groups to the leaders’ table inside one calendar week.

Pair that with reporting earlier in the month that Commerce Department export controls forced Anthropic to disable two frontier models, and with Politico’s June 18 reporting that the White House and Anthropic are now drafting a joint framework to assess AI security risks. The joint framework, when it surfaces, will be one of the most important regulatory documents in venture for the next two years. It will set the trigger for when a government can switch off a frontier model. Anyone evaluating an AI company today is also evaluating regulatory risk to its model access. That is new.

Questions worth asking the founders we are seeing this month

Last week’s pressures, taken together, point at a small set of sharp questions. The most important one is what Salesforce-Fin and SpaceX-Cursor answer in opposite directions: what about a company cannot be acquired or bundled away?

A few sharper versions:

• Which workflow do you own end-to-end, and which budget line at the customer disappears if you do?

• If Salesforce, AWS, OpenAI, or Microsoft makes your core feature a free primitive next quarter, what about your install base and your data stays defensible?

• Which single model are you locked to, and how fast can you fail over if it goes offline by government order?

• How sensitive is your gross margin to a forty percent drop in open-weight inference cost over the next twelve months?

• If your unit economics need a rate cut to work, is this a venture investment or a real-estate bet?

A founder who walks into a check-cutting conversation with crisp answers to those is in a small group. The rest will hear it from acquirers and platforms instead, on terms set by the buyer.

What to monitor over the next one to four weeks

A short list of items where the next signal will move underwriting more than last week’s headlines did:

• The White House-Anthropic AI security framework, when it surfaces. The trigger language is the document.

• Whether OpenAI’s confidential S-1 converts to public disclosure. Audited comparables would reprice every late-stage AI position.

• SpaceX-Cursor regulatory review timing. Closing risk is the most concrete pricing input on SpaceX equity for the next two quarters.

• Whether the AWS WAF AI traffic monetization is adopted by any major publisher and paid by any major crawler. If yes, content rights become a programmable market and a new layer of startups becomes investable. If no, it stays a feature switch with no economy on top.

• The conversion rate of robotics megadeal headlines into deployed fleets. Capital has been generous to physical AI for several quarters. The fleet sizes need to start showing up.

Closing thought

It is easy to read last week as a story about acquisitions. Two big deals. A handful of platform launches. A funding board that paid for operating loops and ignored copilots. A Fed that did not blink.

The structural read is different. The middle of the AI stack, the place where most application-layer venture money has gone for three years, started to lose its standalone character. From above, platforms are absorbing features. From across, strategic acquirers are buying workflow real estate with public stock. From below, open-weight cost compression has institutional capital behind it now.

What remains is a barbell. Infrastructure on one end. Companies that own a measurable workflow on the other. The middle is where most of the noise has lived. The middle is also where most of the next twelve months of portfolio markdowns will live.

The founders who already understood this were not on the news pages last week. They were heads down, signing into the systems their customers cannot live without, writing back into data other companies do not have, and learning from feedback loops that compound. That is the work.