Nadella's Warning: Companies Using AI Are 'Paying Twice'

Satya Nadella has a blunt message for enterprises rushing to adopt large language models: you are paying for the privilege twice. The Microsoft CEO argues that companies spend heavily on token consumption, then unknowingly hand model providers a second form of payment in the shape of proprietary knowledge embedded in their prompts, feedback, and usage patterns.

Nadella's concern centers on how model makers often reserve the right to learn from customer interaction data. Every query an employee sends, every correction a team makes to a chatbot's output, can become training material that strengthens the very model a competitor might also be paying to use. His proposed countermeasure is distillation: enterprises should treat a frontier model's outputs as a resource to reverse-engineer cheaper, purpose-built models of their own, rather than remaining permanently dependent on outside providers who are simultaneously mining their data.

The warning lands at a moment when the economics of AI usage are shifting fast in the opposite direction, at least on paper. Research from Epoch AI shows per-token prices falling anywhere from 9x to 900x annually depending on the task, with the steepest drops occurring over the past year. Ramp's enterprise spending data corroborates the trend at a broader level, showing the average cost per million tokens across major providers dropping from roughly ten dollars to two dollars fifty within twelve months. Andreessen Horowitz has taken to calling this pattern "LLMflation."

Cheaper Tokens, Bigger Bills

That deflation hasn't necessarily translated into lower enterprise spending, because consumption is scaling even faster than prices are falling. Uber handed AI coding tools to 5,000 engineers in December and burned through its entire 2026 AI budget by April, even though 70% of committed code was AI-generated. Nvidia's Jensen Huang has framed heavy token usage as a feature rather than a bug, saying he'd be "deeply alarmed" if a $500,000 engineer's annual token bill came in under half their salary, and that Nvidia is targeting a $2 billion yearly token spend across its engineering organization.

Not every case of high usage reflects real productivity. A Disney employee reportedly queried Claude 460,000 times in nine days, part of a "tokenmaxxing" trend serious enough that Amazon shut down an internal AI leaderboard after staff started gaming rankings instead of producing business value. Gartner's survey of 350 executives at billion-dollar-revenue companies found that roughly 80% had cut headcount tied to AI deployment, with no measurable correlation to improved returns, suggesting that raw usage and genuine efficiency gains are not the same thing.

Against that backdrop, Nadella's data-ownership argument reads as much like a hedge against runaway infrastructure costs as a governance principle. The five largest US cloud and AI infrastructure providers, Microsoft included, have committed between $660 billion and $690 billion in capital expenditure for 2026, nearly double the prior year, according to Futurum Group. Anthropic has separately accused Chinese open-source model developers of funneling millions of prompts through Claude to accelerate their own training, prompting calls for tighter export controls, a dispute that underscores just how contested the boundary between "using" a model and "feeding" it has become.