Microsoft’s CEO wants companies to keep their models swappable. The spending data says enterprise buyers gave up that option two years ago, and picked a different layer to get stuck in.
Satya Nadella told CNN’s Fareed Zakaria on Sunday that companies relying wholly on the proprietary AI labs will not survive as firms. He wants enterprises to retain their prompts and metadata, keep memory and coding tools separate from any single model, and preserve the ability to switch. Enterprise buyers have moved the opposite direction on every measurable axis. Purchased AI use cases climbed from 53% to 76% in one year, open-weight models fell from 19% of enterprise usage to 11%, and three vendors now hold 88% of enterprise LLM spend. Microsoft sells the infrastructure that fixes the problem Nadella describes.
Read the coverage of Nadella’s warning and you get one of two stories. Either the Microsoft CEO has broken ranks with the AI labs his company funds, or he is running a sales pitch with a scary hat on. Both readings skip the more useful question, which is whether enterprise buyers can still take the advice. The spending data says most of them cannot, and the reason has nothing to do with which model they picked.
What Happened
On Sunday, July 26, Nadella appeared on CNN’s “Fareed Zakaria GPS” and extended a warning he first published two weeks earlier. Asked what counts as sharing too much with a model provider, he said companies should be wary of everything they hand over, from data to prompts.
His prescription was specific. He wants a setup where every model call leaves the metadata in the customer’s hands, so the company can eventually train its own weights or its own open model. Without that control, he said, a firm has outsourced its thinking and will not remain a firm.
He also named the target. Nadella wants companies to stop depending on the coding tools the labs ship with their models, the layer the industry calls a harness. Anthropic’s Claude Code and OpenAI’s Codex are the two he was pointing at. Keep the harness, the context and the memory separate from the model, he argued, and any single model can disappear without taking the business down with it.
Microsoft has invested in both Anthropic and OpenAI. Its cloud division sells the gateway and orchestration layer he is recommending. TechCrunch flagged the conflict and then agreed with him anyway, which is close to the correct read.
The Backstory
Nadella’s July 13 essay, posted to X under the title “The Reverse Information Paradox,” carries the actual argument. He borrows from economist Kenneth Arrow, who observed that a buyer cannot value information until they have seen it, at which point they no longer need to pay for it. Nadella flips the direction. In AI, the seller learns about the buyer with every transaction.
“You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful,” he wrote. The better you want the model to perform, the more you have to feed it.
He gave the resulting asset a name. Token capital: the knowledge created in the traffic between employees, applications and AI systems. Over enough time, he argues, it matters as much as intellectual property or human capital. The line he closed on was that in consuming intelligence you are creating intelligence, and what you create should belong to you.
The 2023 version of this fear came from venture investors worried about startups. Jason Calacanis warned Y Combinator founders that accepting OpenAI credits meant showing OpenAI the roadmap. Palantir’s Alex Karp made a louder version of the same point. Nadella has now aimed it at the Fortune 500, and he did it two days before Microsoft reports fiscal fourth quarter earnings on July 29, with Azure guided to 39% to 40% growth and third quarter capital spending already at $31.9 billion.
The Plan
Strip the framing away and Nadella is asking enterprises to do four things: keep their interaction logs, run more than one model, put an independent gateway between their applications and any lab, and own the harness rather than renting it.
Microsoft benefits from all four. Every one of them pushes spend toward Azure’s orchestration and governance products and away from a direct relationship between the customer and a frontier lab. That does not make the analysis wrong. It does mean the recommendation and the product catalog line up too neatly to take at face value.
Worth remembering what Microsoft’s own model has always been. Lock-in through switching costs sits alongside razor-and-blade and subscription in any honest description of how Microsoft makes money. Windows, Office and Azure each work because leaving costs more than staying. Nadella is not opposed to lock-in. He is arguing about which layer should hold it.
The Business Model Angle
Menlo Ventures surveyed 495 US enterprise AI decision-makers in November and published the results in December. The numbers describe a buyer moving away from control on every axis Nadella cares about.
In 2024, enterprises built 47% of their AI solutions in house. By the end of 2025 that number fell to 24%, with 76% of use cases purchased ready-made. Open-weight models, the foundation of Nadella’s “train your own” scenario, dropped from 19% of enterprise usage to 11%. Chinese open models, which dominate token volume among startups and indie developers, account for 1% of enterprise LLM API usage. Anthropic, OpenAI and Google together hold 88% of the market. Prompt design remains the dominant customization technique, and fine-tuning is still a frontier-team activity.
The spending split is the clearest signal. Of the $18 billion enterprises put into AI infrastructure in 2025, foundation model APIs took $12.5 billion. Model training infrastructure took $4 billion, most of it inside the labs themselves. The layer that handles storage, retrieval and orchestration, the one that would let a company move between models without rewriting anything, took $1.5 billion.
Buyers are spending more than eight dollars on model access for every dollar on the machinery that would make model access replaceable. That is the whole argument in one ratio.
The mechanism behind it shows up in the coding numbers. Anthropic’s share of enterprise coding spend went from 42% to 54% in six months while OpenAI sat at 21%, and Menlo attributes the move to Claude Code. Coding reached $4 billion in 2025, 55% of all departmental AI spend and the largest single category in the application layer. Enterprises did not choose Anthropic’s weights. They chose a harness, and the model came attached.
That is where the switching cost lives. Token prices for frontier-class access have collapsed since 2023, and the collapse changed almost nothing about vendor behavior, because the friction was never the price. The friction is the accumulated evals, the prompt libraries, the tool definitions and the corrections that make a deployment work. Nadella named that asset correctly. He is late on when it stopped being portable.
We mapped the margin migration toward compute, cloud and orchestration earlier this year. This is the buyer-side half of the same story, and it explains the mechanism: orchestration captures margin because the customer’s accumulated context lives there, not in the weights.
The Risk
The strongest case against Nadella is that his advice is unavailable to the people he is addressing. Retaining metadata “so you could train your own weights” is a live option for a few hundred companies on earth. For a mid-market insurer or a regional bank, it is a slide, not a plan. The achievable version is narrower: negotiate log portability, standardize on open protocols, keep evals in your own repository. None of that requires owning a model, and none of it appears in the headline.
The second problem is that Nadella’s fix relocates the dependency rather than removing it. An AI gateway that sits between your applications and every model becomes the thing you cannot leave. Ask any company that consolidated onto a single cloud in 2015 for the reason they consolidated. Convenience, and the promise of never being locked in again.
The case for him is straightforward and worth stating. Moonshot’s Kimi K3 pricing showed what happens once a vendor wins default status and starts charging for it. OpenAI’s leaked financials show a business that has to raise effective prices eventually. Buyers who kept optionality will pay less than buyers who did not, and the gap will show up in 2027 renewals rather than this quarter’s invoice.
Quick Questions
Is Nadella wrong about the risk? No. Labs do learn from customer usage, and the incentive to move up into a customer’s category is real. His diagnosis holds. His timing assumes enterprises still have the flexibility to act on it.
Why does switching stay rare when prices keep falling? Because price was never the barrier. Rebuilding evals, prompts, tool schemas and agent scaffolding around a new model costs engineering months, and the finance team never sees that number on an invoice.
What is a harness? The scaffolding around a model that turns it into a working tool: the agent loop, the file access, the terminal, the memory. Claude Code and Codex are harnesses. They are also where the lab captures the customer.
Does any of this apply to a small business? The four-point plan does not. Two habits do: keep your prompt and eval work in your own repository rather than a vendor’s console, and prefer tools that let you point at a different model without a rewrite.
The Business Model Analyst Take
Nadella has described the right asset and the wrong clock. Token capital is real, and every enterprise generating it should think harder about who ends up owning it. The problem is that his audience already traded the option he wants them to protect, and they traded it for something they valued more, which was speed. Menlo found AI deals convert to production at 47% against 25% for traditional software. Buyers were not careless. They were fast, and speed and optionality are the two things you cannot buy at once.
Watch the layer, not the logo. Whoever holds your evals, your memory and your agent loop holds your renewal, and that has never been the same company as whoever holds the weights. Microsoft understands this better than anyone in the industry, which is precisely why its CEO is telling you to keep the model at arm’s length while offering to hold everything else. The enterprises that come out of this decade with pricing power will be the ones that treated the harness as a build decision in 2026 instead of a procurement one.
Microsoft reports on July 29. Azure’s growth rate will tell you how the pitch is landing.
