Last Week Ignite - 5.3.26
The Week AI Stopped Being About the Model
A founder I respect texted me on Monday morning with a screenshot from his Linear board. Twelve tickets, all assigned to a Codex agent, all moving. He hadn’t pushed a commit himself in two weeks. “Is this what we’ve been waiting for, or is this the part where I should be worried about my job?” he wrote.
I told him both, probably.
That same morning, OpenAI published Symphony, an open spec that turns Linear into a control plane for continuously running coding agents. The company said some internal teams had seen a 500 percent jump in landed pull requests in the first three weeks. By Wednesday, OpenAI and Microsoft had quietly rewritten their seven-year-old partnership so OpenAI can now ship products on any cloud. By Thursday, AWS announced OpenAI models were coming to Bedrock. By Friday, the Fed held rates with four dissenters voting for cuts, the most disagreement on the committee since October 1992, and the Q1 inflation print came in at 4.5 percent annualized.
If you take only one thing from the last seven days, take this. The frontier model layer, which a year ago looked like the moat, is becoming distribution. It is becoming infrastructure. It is becoming, in some cases, a commodity. The venture economy is reorganizing around that fact while we watch.
The capex print
Big technology companies reported earnings together this week, and the numbers told a single story.
Microsoft’s AI business now runs at a 37 billion dollar annual revenue rate, up 123 percent year over year. Azure grew 40 percent. Google’s cloud backlog reached 462 billion dollars, and the company raised its 2026 capital spending guide to 180 to 190 billion, with 2027 set to climb further. Meta lifted its capex range to 125 to 145 billion, citing component pricing and the cost of building future data centers. Amazon’s AWS hit a 15-quarter growth high at 28 percent. Free cash flow at Amazon dropped sharply because property and equipment purchases rose by 59.3 billion dollars, almost all of it AI-related.
Add it up across the four biggest spenders and you land somewhere between 650 and 700 billion dollars in 2026 spending on data centers, chips, power, and cooling. That is roughly the size of the entire United States defense budget.
The market reaction was telling. Microsoft and Meta both fell three to six percent after hours despite beating revenue. Public investors, for the first time in this cycle, are starting to ask whether the spending will pay back. They are not yet selling. They are starting to underwrite. That is a different posture, and it propagates downward into private markets.
The data center pipeline is also running into walls. On Wednesday, Compass Datacenters, owned by Brookfield, abandoned its 2,100-acre Prince William County project in Virginia after a March court ruling and a county vote against the rezoning. Sightline Climate estimates roughly half of the 16 gigawatts of US 2026 pipeline now faces delay or cancellation. Two-thirds of the new builds are heading to rural farm country, where transmission lines do not exist yet and local opposition often arrives before the steel does.
The bottleneck moved this week. It is no longer chips. It is power, land, and time.
What got funded, and why it matters
Six rounds last week tell a clearer story than any benchmark.
Hightouch raised 150 million dollars on April 29 to build an AI platform for marketers. Netomi raised 110 million dollars the next day for enterprise customer-service agents. JuliaHub raised 65 million dollars and shipped Dyad 3.0, an agentic platform for industrial digital twins. Scout AI raised 100 million dollars for a foundation model aimed at unmanned warfare. Aidoc raised 150 million dollars to expand its clinical AI deployment platform. And Ineffable Intelligence, the new lab from former DeepMind reinforcement learning lead David Silver, was reported to have closed a 1.1 billion dollar seed round at a 5.1 billion dollar valuation.
Five of those six are doing the same thing in different domains. They are putting AI inside a workflow that already had a budget owner, a system of record, and a cost of failure. Marketing operations. Customer service queues. Industrial design loops. Drone autonomy. Radiology departments.
The sixth, Ineffable, is a different category of bet. It is the third European AI lab in two months to raise a billion-dollar seed before shipping a product, alongside AMI Labs (Yann LeCun) and Recursive Superintelligence (which raised 500 million from GV and NVIDIA in mid-April). The thesis is that pure language model scaling has a ceiling and that whatever comes next is worth funding now, even at unicorn-plus prices, even without a roadmap. Sequoia, Lightspeed, NVIDIA, Google, DST, Index, and the UK Sovereign AI Fund all believe this enough to write checks.
The implication for early-stage investors is uncomfortable. The four or five labs taking pedigree-priced rounds are now sitting on cap tables that smaller firms cannot enter. The price of admission moved out of reach in a single quarter. Experienced early-stage investors are responding by funding the application and infrastructure layers that any of those bets will need to ride on. That market is wide open. The pedigree market is closed.
The True Anomaly round is in its own category. The space-defense company raised 650 million dollars on April 28 at a 2.2 billion valuation, four days after the Space Force named it among twelve primes for the Golden Dome interceptor procurement. The Golden Dome program is now estimated at 185 billion dollars, with 17.5 billion in the FY2027 request alone. True Anomaly’s previous round, less than twelve months earlier, valued the company at 260 million. Procurement visibility carried that eight-times step-up, with no help from the usual ARR-multiple math.
Defense procurement has become a real venture asset class this week, in the same way that healthcare AI did two years ago.
The partnership that broke
Three years ago, Microsoft’s deal with OpenAI looked permanent. Exclusive cloud rights. Exclusive IP license. A revenue share running both directions. An AGI trigger clause that gave Microsoft access to anything OpenAI built, until OpenAI declared itself something more than a normal company.
On Monday, all of that changed.
Under the new agreement, OpenAI can ship products on any cloud. Microsoft’s IP license remains in place through 2032 but is no longer exclusive. The Microsoft-to-OpenAI revenue share is gone. The OpenAI-to-Microsoft revenue share continues through 2030 but is capped. The AGI trigger clause is dead. By Tuesday, Andy Jassy at AWS had announced that OpenAI models were coming to Bedrock in the next few weeks. By the end of the week, AWS’s Matt Garman was publicly pitching that AWS could be a better partner to OpenAI than Microsoft had been.
The startup implication is the part that matters. For the last two years, a meaningful share of AI venture pitches arrived with some version of “we are built on OpenAI.” That sentence used to imply a cloud choice and a stack of integrations. Now it implies neither. Buyers will increasingly pick the cloud where their compliance, billing, and security already live, and pull the model from a menu. Selling access to a particular model has flattened as a wedge. What still works is multi-model routing, evaluation, observability, vertical data, and workflow ownership.
DeepSeek V4, released the week before, made the price-performance case more aggressive. The Pro version posts Codeforces and SWE-bench scores within shouting distance of frontier closed models, costs about a third as much per token, and ships under an MIT license. The US Commerce Department’s CAISI evaluation pegs the gap to closed frontier at roughly eight months. For any application where absolute frontier capability is not required, a 60 to 80 percent inference cost reduction is now available immediately.
The thin-wrapper thesis was already weakening. This week buried it.
The liquidity question
A different kind of news landed in private markets this week, quieter but worth tracking.
On Monday, Vinted closed an 880 million euro secondary at an 8 billion euro valuation, with EQT, Schroders Capital, and Teachers’ Venture Growth leading. Vinted is profitable, category-leading, and disciplined on unit economics. Secondary buyers love that profile. On Thursday, Lazard agreed to acquire Campbell Lutyens, a major secondary and GP advisory firm, for 575 million dollars plus up to 85 million in earn-outs. Avalyn Pharma priced an upsized IPO at the top of its range, raising 300 million dollars.
These three transactions together are a market-structure signal. The infrastructure of private liquidity, including secondaries, continuation funds, and tender offers, is being repriced upward. The SEC made this easier on April 16 by shortening the minimum issuer-led tender window from 20 business days to 10. This week, that change started showing up in client alerts and term sheets. Issuer-led tenders are about to become the dominant liquidity mechanism for top-tier private AI names, and the firms that intermediate them are positioning accordingly.
The takeaway for late-stage investors is that liquidity quality is opening, but not liquidity quantity. Vinted gets cleaner exits. Anthropic, reported by Bloomberg this week to be in talks with Google for as much as 40 billion dollars in additional cash and compute, gets cleaner pricing. The middle tier of “story stocks without exits” remains stranded.
Anthropic is worth a separate note. The company moved from 9 billion dollars in annualized revenue at the end of 2025 to roughly 30 billion by early April. The recent employee tender at a 350 billion dollar mark reportedly saw less selling than expected, while secondary buyers were quoted at 800 billion plus. That gap is unusual. Employees who sell early in a hot AI company are often the smartest indicator of where the price is going. When they refuse to sell at 350 billion, the market for the next tender will likely be higher.
What the macro is doing to the model
The Federal Reserve held rates at 3.50 to 3.75 percent on April 29, with eight votes for and four against. That is the largest dissent on the committee since October 1992. Two members wanted to cut. Two wanted to keep rates higher. Powell confirmed his last press conference as chair. He will remain on the Board.
The first-quarter GDP advance came in at 2.0 percent, up from 0.5 percent in the prior quarter. The PCE price index rose 4.5 percent annualized, with core PCE at 4.3 percent. Most of the inflation acceleration traces back to the Iran war fuel shock, which has held Brent crude near 105 dollars a barrel for two months and pushed US gasoline above four dollars a gallon. Spirit Airlines ceased operations on May 2, citing fuel costs. Forty countries are now reportedly considering nuclear power.
For venture, the macro implication is straightforward. The case for cheap capital re-emerging in 2026 is weaker than it looked in January. Private market valuation models that assumed multiple Fed cuts this year deserve a haircut. Founders pitching models that need rate relief to clear should expect skepticism. Companies whose customers are exposed to consumer discretionary spend should expect harder Q3 and Q4 pipelines.
What insulates a startup from this environment is pricing power, budget-owner buyers, and a workflow that costs more than the software does when it breaks. Marketing automation, customer service, clinical workflow, defense procurement, industrial simulation. The pattern in the funded rounds is not a coincidence.
What got weird
A few things this week worth treating as radar rather than thesis.
OpenAI’s Symphony released alongside reporting that its internal coding agents had cleared 5,000 production pull requests in three weeks of dogfooding. Symphony is open, which means by next quarter Linear, Jira, and GitHub-native versions will exist. The implication is that engineering organizations will spend 2026 redesigning themselves around stateful work pipelines that hand tasks to synthetic labor. The investable surface is in the seams between issue tracker and IDE, between agent and reviewer, between policy engine and compliance audit.
Anthropic’s Mythos Preview, gated under Project Glasswing to about 50 partners, continued generating cybersecurity ripples. The UK AI Safety Institute reported the model solving 73 percent of expert-level capture-the-flag challenges and chaining 32-step network attacks in three of ten attempts. Wordfence reported that AI-assisted vulnerability submissions grew from 16 percent to 66 percent of the total between November 2025 and April 2026. Berkshire Hathaway, Chubb, and Travelers all received approval this month to drop AI-related damages from corporate insurance policies. The next twelve to eighteen months are going to involve a sharp re-pricing of cybersecurity insurance, and a new procurement category for AI-vulnerability indemnity. SaaS contracts will start carrying language they have never carried before.
Google DeepMind launched an AI co-clinician research initiative on April 30. In blind evaluations on 98 realistic primary-care queries, the system recorded zero critical errors in 97 of them. In telemedical simulations across 140 assessment areas, it matched or exceeded primary-care physicians in 68. Expert physicians still performed better overall, especially on red flags and emergency triage. The framing is supervised teammate. That framing is what makes the procurement and reimbursement pathways begin to function. Healthcare AI is becoming a deployment category for growth-stage capital, with real workflow integration and real money behind it.
Figure announced its BotQ humanoid factory has hit 24-times its prior throughput, producing 55 humanoids per week, one per hour, with 80 percent first-pass yield. The company has shipped over 350 Figure 03 units. On the same day, China’s State Grid deployed 500 humanoid robots for high-voltage operations. Japan Airlines began piloting humanoid baggage handling at Haneda. The Beijing E-Town half marathon was won by a fully autonomous humanoid in 50 minutes 26 seconds. Production volume is no longer the constraint for humanoid robots. Demand validation is.
The cluster across these four signals is the same cluster. AI is moving from chat interface into workflow control. Whatever you build, build it with that fact in mind.
What experienced early-stage investors are doing
The pattern in this week’s rounds suggests a few questions worth asking founders right now.
What workflow state do you own, and what is the cost when that state goes wrong?
If AWS, Azure, or Google Cloud bundles a model, an orchestrator, and a compliance layer into the base offering, why does your buyer still need you?
How does your roadmap survive Symphony or Anthropic Skills becoming the default orchestration layer?
If DeepSeek dropped your inference cost by 60 percent next quarter, what part of your defensibility would still be intact?
For agent companies, can you describe your verification and audit story with the same depth as your generation story?
For security and trust startups, do you have a path into the consortium of companies with Mythos-class access, or is your TAM the long tail outside it?
The companies that have a clean answer to most of these are the ones being funded right now. The companies that do not are the ones whose seed rounds are taking three months instead of three weeks, and whose Series A pitches are bouncing.
For experienced late-stage investors, the watch list is concentrated. Anthropic and SpaceX are the two pre-IPO names where the price discovery is most active and the next ninety days will likely contain the cleanest entry windows. True Anomaly, Anduril, and Figure are the defense-and-physical-AI names where revenue visibility is starting to underwrite the valuation rather than narrative carrying it. Fervo Energy and X-energy are the first nuclear and geothermal names where public market issuance has cracked open. Vinted and similar profitable consumer marketplace assets are the secondaries to take seriously.
What to watch in May
A handful of dates over the next four weeks will tell us whether the patterns that landed this week harden into a regime, or get muddled by something new.
May 13. The EU Digital Omnibus trilogue meets again on the AI Act enforcement timeline. If August 2 holds as the binding date for high-risk and watermarking obligations, every European-exposed startup needs a compliance plan in hand by July.
The week of May 11. Senate confirmation vote for Kevin Warsh. The composition of the Federal Reserve through 2028 starts being decided here.
Mid-May. Recursive Superintelligence’s public launch. The first product release from the post-LLM lab cohort is the moment we find out whether the seed prices are real or theatrical.
Late May. SpaceX public S-1, with the roadshow targeted at the week of June 8. The cleanest test of whether the AI-frontier-tech IPO window opens or stays closed.
Through May. The first publicly accessible Mythos-class capability lands somewhere. Insurance coverage and SaaS contract language start changing the next week.
The infrastructure layer is more crowded and more expensive than ever. The model layer is becoming distribution. The application and workflow layers are wide open, but only for companies that own a state somewhere a budget owner cares about. Power, land, regulation, and procurement have moved from background conditions into active investment variables.
A reader who started Monday thinking AI was about chatbots and benchmarks should finish this week thinking it is about workflows, control planes, contracts, and kilowatts.
That is the regime. Build accordingly.