One number
On September 13, podcast host Harry Stebbings cited data from Ramp on 20VC: among companies newly acquiring AI tools, Anthropic now captures 73% of new spending — a reversal from OpenAI's prior lead. The caveats matter. This is new spend, not installed base, and it comes from one company's reading of customer card flows, not a market audit. But it points the same way as another signal: enterprise buyers are paying for reliability before they pay for release cadence.
Nadella's argument
Microsoft CEO Satya Nadella gave the structural version this week in an All-In interview: the value of AI will not stay at the model layer. He is spending $175 billion on the "control layer" — the infrastructure, middleware and governance surfaces that let thousands of enterprises deploy AI safely and independently — rather than on being the smartest model. His advice: "use all models, independent of all models." He offers a concrete test: pull one model out and see whether your evaluations still hold. If they do not, you depend on something that may not belong to you.
Why now
Three forces are moving at once. Open-source models keep pressing prices down: Chamath Palihapitiya cited DeepSeek's latest model at $0.15–$0.60 per million output tokens, roughly 99% below frontier pricing of about $50. Enterprise buyers are waking up to single-vendor lock-in, the same dynamic that kept databases honest against PostgreSQL and MySQL. And the application layer now has a viable business model: memory systems, orchestration and middleware embedded in workflows are eating budgets that once went to model APIs.
The counter-reading: the same week, multiple outlets reported OpenAI is considering a funding round at a $1.2 trillion valuation, with an IPO possibly pushed to 2027. The gap between frontier labs' valuation narratives and corporate procurement behavior is widening. One side is a capital story; the other is a cash-flow story. The next few quarters of enterprise budget data will say which is closer to reality.
[1][2]