Last Week Ignite 6.28.2026
The Week the Meter Started Running
Special Note: Today is the last day to join our marked up Fund III. Review the Deck and Join Here today if you want in.
Databricks makes money the way most software companies do. It charges more to serve a customer than the customer costs to serve, and for years that gap was wide and getting wider. Then this week the company reported that its gross margin, the slice of revenue left after paying to run the service, had slipped from somewhere above 80 percent to 74 percent. There was no price war. There was no botched quarter. The reason margins fell is that its customers’ AI agents would not stop asking questions.
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.