By Stuart Kerr, Technology Correspondent, LiveAIWire
Satya Nadella just told enterprises they are paying for AI twice, and only realising it once. In a personal blog post published on 12 July, the Microsoft chief executive argued that every company using a proprietary AI model is handing over something more valuable than the subscription fee: the accumulated knowledge of how that company actually works. Nadella calls it the Reverse Information Paradox, and it lands at a genuinely awkward moment for a man whose company has poured tens of billions of dollars into OpenAI and holds a major stake in Anthropic too.
His argument, stripped of the branding, is simple. Every prompt a business writes, every correction it makes when a model gets something wrong, every tool call an AI agent runs inside a company’s systems teaches the model something about that company that no competitor could otherwise buy. Feed the model enough of that information, Nadella argues, and a business ends up paying twice: once in subscription fees, and again in the institutional knowledge it hands over just to make the AI useful. As he put it, you “pay for intelligence twice.” The more useful you want the AI to be, the more of your own institutional knowledge you have to give it.
What This Means for Anyone Using AI Right Now
If your business uses ChatGPT, Claude, Gemini, or any other proprietary model for real work, and corrects its mistakes, refines its prompts, or lets it run agentic tasks inside your systems, you are generating exactly the kind of data Nadella is describing. Under most current terms of service, the model provider retains rights to learn from that usage. Nadella’s central complaint is that this flow of learning only runs one way: the model provider improves, and the customer has no equivalent right to study or build on what the model has learned from them in return.
As TechCrunch first reported, Nadella frames this as a matter of basic fairness. AI labs claim broad rights to train their models on publicly available data scraped from across the internet. He argues it is inconsistent for those same labs to then restrict a company’s ability to study or build on the outputs of the model that company is paying to use, a practice known as distillation. Training freely on the open internet while restricting what customers can learn from a paid model, he writes, is simply hypocritical.
Nadella Is Not the First to Raise This
Nadella’s post joins a chorus that has been building in Silicon Valley for months. Palantir chief executive Alex Karp has made a similar argument publicly, and Nadella quotes him directly: technical customers want to “own the means of production,” not hand control of their compute, models, data, and competitive edge to someone else. Venture capitalists including Jason Calacanis have voiced the same worry from the investor side, framing proprietary AI labs as potential Trojan horses that learn a customer’s business well enough to eventually compete with it.
The concern has a concrete recent example behind it. In February, Anthropic accused three Chinese AI companies, DeepSeek, Moonshot AI, and MiniMax, of running large-scale distillation campaigns against Claude, using tens of thousands of fraudulent accounts to extract more than 16 million exchanges aimed at Claude’s most valuable capabilities in reasoning, tool use, and coding. Anthropic later told US lawmakers that Alibaba’s Qwen lab ran an even larger campaign, extracting close to 29 million exchanges through roughly 25,000 fake accounts.
Those cases show distillation working in the direction AI labs object to. Nadella’s post is effectively asking why enterprises should not have an equivalent right to learn from the models they are paying for, rather than only the labs having that right over everyone else’s data. The same underlying tension, over who gets to extract value from a frontier model and on what terms, is also playing out in Washington’s export control fight with Anthropic, where the US government has already shown it is willing to treat model access itself as a lever of control.
What Nadella Wants Companies to Actually Do
Nadella’s proposed fix is less a manifesto than a checklist, and it reads like exactly what you would expect from the chief executive of a major cloud provider. He wants companies to build their own proprietary learning environments, private systems where an organisation’s prompts, corrections, and institutional context accumulate and can be used to train or fine-tune models without ever leaving the company’s control.
He also wants an orchestration layer that keeps businesses able to switch between AI models from different providers, so no single lab can lock in a customer by becoming indispensable to its workflows. And he wants companies to retain explicit ownership of their own data exhaust, meaning the prompts, feedback, and corrections that would otherwise train someone else’s model for free.
Nadella never uses the phrase “open source” in the post, but the implication runs through it. Idit Levine, chief executive of the enterprise AI infrastructure company Solo.io, told TechCrunch she is already seeing this shift with her own customers, who include T-Mobile, ADP, and SAP. After trying proprietary models, she said, businesses increasingly ask whether an open-source model run on their own premises can do most of the job at a fraction of the cost, with the added benefit that they control it outright.
Vercel, the web infrastructure company that has recently added tools for switching between AI models, told TechCrunch that open-source models accounted for 29 percent of all traffic through its AI gateway last month, though that figure comes from Vercel’s own account rather than independently published data. It is consistent, at least directionally, with the same cost-driven shift LiveAIWire has reported inside Microsoft’s own quiet move away from OpenAI and Anthropic on routine workloads.
Why Microsoft’s Own CEO Is Saying This
The most interesting thing about Nadella’s post may be who is writing it. Microsoft has invested more heavily in OpenAI than any other company on the planet, and holds a stake in Anthropic as well, a position LiveAIWire examined in detail as both labs moved toward public listings earlier this year. A warning from Nadella about proprietary AI labs quietly extracting value from their customers is, on one level, a warning about the business model of companies Microsoft has bankrolled.
It also happens to align neatly with where Microsoft makes its money. The proprietary learning environments Nadella wants companies to build need somewhere to live, and Azure, Microsoft’s cloud platform, is a natural home for exactly that infrastructure. The orchestration layer he describes, letting businesses route between different AI models rather than committing to one, is a category Microsoft is already building tools for. None of that makes the underlying argument wrong. It does mean the solution Nadella is proposing happens to route a considerable amount of enterprise AI spending through Microsoft’s own infrastructure, whichever model a company ultimately chooses to run on it.
What This Means for You
For any business currently deciding how deeply to integrate AI into its operations, Nadella’s post is a useful prompt to ask a specific question before signing a contract: what does this AI provider’s terms of service actually say about who owns the data generated by your usage, corrections, and agent activity? Most enterprise AI contracts were written before this argument became mainstream, and the answer is often that the provider retains broad rights to learn from that usage indefinitely.
The practical options Nadella is pointing toward, running open-source models on your own infrastructure, building orchestration tools that avoid dependence on a single provider, and negotiating explicit data ownership terms, are all available today, and enterprise adoption of exactly this approach is already measurably underway according to the vendors building the infrastructure for it. Whether Nadella’s intervention accelerates that shift meaningfully, or whether it is simply Microsoft positioning itself to benefit from a shift that was happening anyway, is likely to become clearer over the next year as more enterprises renegotiate their AI contracts with this argument in hand.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.