The most interesting thing about Satya Nadella's blog post this week is who wrote it. The argument that AI labs are quietly extracting value from the companies that use them has been circulating for a while, pushed by venture capitalists like Jason Calacanis and by Palantir's Alex Karp. It is a familiar complaint from people positioned to benefit from making it. It is a different thing coming from the chief executive of Microsoft, a company that sells AI, buys AI, and owns a large piece of OpenAI.
Nadella's claim is that enterprises are paying 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 writes. "The better you want the model to perform, the more of that knowledge you have to feed it." His sharper point concerns what he calls exhaust: the prompts employees write, the tools agents call, and above all the corrections people make when a model gets something wrong. Every correction, he argues, is distilled into institutional know-how of a kind a competitor could never simply buy. Companies are teaching the models the nuances of their own businesses, and they are doing it without recognising it as a transaction.
His proposed remedy has a certain symmetry to it. If model makers claim fair use rights to train on the public internet, he argues, they cannot coherently turn around and forbid their customers from distilling their models in return. He calls the status quo ironic, which is a diplomatic word for hypocritical. It is worth noting that Anthropic accused Chinese open source developers in February of doing precisely that to Claude, sending millions of prompts to harvest its outputs, and asked Washington to tighten export controls in response. Nadella is proposing that enterprises be allowed to do legally what the labs have described as theft.
This argument is landing in a market that is already flinching at the invoice. KPMG surveyed more than 2,000 senior executives across 20 countries and found 29 percent struggling to understand how operating costs scale as enterprise AI deployments grow. Nearly half said they were looking to re-phase deployments where costs outweigh expected value. The Register, discussing the findings, reached for the image of a drug dealer who hooks you first and raises prices after, which is unkind but captures the structural complaint: usage-based billing means the bill grows precisely as the dependency does.
The numbers behind that anxiety are real. Research from the Ramp AI Index puts spending at the most AI-committed businesses at roughly $7,500 per employee per month. One organisation reportedly spent $500 million on Claude usage in a single month, an extreme outlier but an instructive one. Some of this is self-inflicted. Meta factors AI usage into performance reviews; engineers run several agents in parallel because they are expected to produce more than ever. When a company mandates consumption and pays per token, it should not be surprised by the invoice.
Microsoft's own position is uncomfortable enough to be worth stating. Its stock is down roughly 23 percent year on year, driven by enormous AI capital spending and an inability to monetise Copilot at the scale that spending implies. Last week it cut 4,800 jobs, a little over 2 percent of its workforce, with the gaming division absorbing the worst of it, and it is not slowing AI spending at all. A CEO arguing that AI buyers are being quietly fleeced is also a CEO whose company is spending tens of billions to be the seller.
Put the pieces together and a coherent picture emerges. Costs are becoming visible, value is proving harder to demonstrate than promised, and buyers are looking for exits. Some are re-phasing deployments. Some are moving workloads to cheaper open-weight models. And now the industry's most powerful platform vendor is telling them the money was never the whole price. What none of these responses do is answer the underlying question of whether the productivity gains justify the spend. That question is still owed an answer, and the bills keep arriving while everyone waits for one.