Simon Willison published a pointed analysis arguing that Anthropic and OpenAI have found product-market fit. His evidence isn’t user satisfaction scores — it’s billing data and infrastructure contracts.
The most direct signal: in April 2026, both companies restructured enterprise pricing to charge at API rates rather than offering per-seat discounts. Willison calculates that he personally would have consumed approximately $2,180 worth of tokens in a month on what had been a $200/month plan. The pricing change wasn’t arbitrary; it reflects actual usage patterns that made the old structure unsustainable for the providers. When engineers use a tool constantly enough that the underlying compute cost vastly exceeds the subscription fee, the subscription model breaks.
That usage pattern shows up elsewhere. Uber exhausted its entire 2026 AI coding budget in roughly four months. This isn’t a story about AI being expensive; it’s a story about demand exceeding forecasts by a factor that planning cycles didn’t anticipate. Engineers found the tools valuable enough to use them at high volume, budgets followed, and then ran out. That’s a distinct kind of signal from “engineers like using AI coding tools” — it’s closer to “engineers used them as fast as they physically could until told to stop.”
On the infrastructure side, a recent SEC filing disclosed a $1.25 billion per month compute agreement between Anthropic and xAI — Anthropic renting excess capacity from xAI’s Colossus cluster near Memphis, locked in through 2029. Anthropic’s projected Q2 2026 revenue is $10.9 billion, with what may be a first profitable quarter. OpenAI has filed a confidential IPO registration with the SEC. These numbers tell a different story than the one you’d construct from benchmarks and model releases alone.
Willison’s framing — product-market fit — is deliberately precise. The claim is not that AI has been broadly validated as useful, or that consumer adoption is accelerating. The specific claim is that a particular use case (enterprise coding agents, sitting inside developer workflows) has achieved the kind of fit where customers consume the product aggressively, budget for it formally, sign multi-year infrastructure agreements around it, and are willing to pay at full price when discounts disappear. The fit is with the use case, not with AI in general.
That narrowing is worth taking seriously. The companies’ own hiring patterns support it: Willison tallied 703 open positions at OpenAI (32.6% enterprise-focused) and 390 at Anthropic (26.9% enterprise-oriented). A meaningful share of both companies’ go-to-market motion is now pointed at large organizations deploying coding agents at scale, not at the developer hobbyist or consumer markets that drove early growth.
The interesting implication is what this means for the next wave of AI product development. If the revenue case for enterprise coding agents is now established, you’d expect both companies to go deeper on the same loop: better tool use, longer context, more reliable autonomous execution, tighter integration with existing software delivery infrastructure. The capabilities that matter commercially aren’t necessarily the ones that generate benchmark leaderboard attention.
The budget shock that Uber and others experienced is a forcing function in both directions: organizations are now seriously modeling AI compute costs as a line item, and providers have demonstrated they can price accordingly. That’s a more mature market dynamic than the year prior, when token prices were dropping and usage was essentially free at enterprise scale.