Microsoft CEO Satya Nadella is drawing attention to a tension at the center of the AI business model: who gets to learn from whom, and who captures the value created in the process.
His argument focuses on distillation, the practice of training smaller models from the outputs of larger ones. According to the source article, providers such as OpenAI and Anthropic ban that practice in their terms of service, while also training on public data, supposedly under fair use, and learning from customer interactions.
The conflict over AI distillation
Distillation matters because it is one way smaller models can benefit from the behavior of larger systems. A smaller model can learn from outputs generated by a larger model, making the larger model a source of training signal.
The source article says providers like OpenAI and Anthropic prohibit distillation in their terms of service. It also says those restrictions have targeted mainly Chinese AI companies.
Nadella’s criticism is not simply that distillation is restricted. His point is about symmetry. He calls it "ironic" that the same providers can train on public data, supposedly under fair use, while preventing others from using their systems in a similar learning loop.
Why Nadella sees a deeper economic problem
Nadella’s broader concern is where the economic value goes. According to the source article, he argues that value can end up concentrated with infrastructure operators rather than with the companies that generate the knowledge being used.
That concern becomes sharper when AI systems are used inside companies. Every interaction can carry useful signals: corrections, ratings, usage patterns, and other feedback that show how people refine, reject, or rely on an AI answer.
Nadella describes that interaction data as "exhaust." In his view, companies are not just consuming intelligence when they use AI. They are also producing information that can improve the systems they rely on.
In consuming intelligence, you are creating intelligence.
The phrase captures the central issue. A company may pay for access to an AI tool, but the act of using the tool can also reveal internal knowledge through the feedback and behavior generated during normal work.
The reverse information paradox
Nadella calls this the "reverse information paradox." In plain terms, the customer is both buyer and contributor. The company pays money for AI access, then supplies learning signals through its use of the system.
The source article frames this as paying twice. First, companies pay with money. Then they pay again through the usage data, corrections, ratings, and other interaction traces that may help an AI provider improve its products.
That creates a strategic concern for businesses using AI systems. If a provider can learn from customer interactions, it may gain insight into company knowledge. The source article also notes that providers can potentially build competing products from what they learn.
For Nadella, the issue is not only technical. It is also about control. If the infrastructure operator captures the learning loop, the company doing the work may not capture the full value of what its own people and processes teach the AI system.
What companies are being asked to consider
The article does not describe a simple fix. It does, however, make clear what Nadella wants companies to notice: AI use is not a one-way transaction.
When a business adopts AI, it is not only receiving model outputs. It may also be generating training-relevant signals every time employees correct, rate, or otherwise interact with the system.
That makes the ownership of the learning loop a business question, not just a technical one. Companies may need to think about whether the systems they use allow them to control how their own operational knowledge is turned into future intelligence.
- Distillation lets smaller models learn from larger model outputs.
- Terms of service from providers such as OpenAI and Anthropic ban the practice, according to the source article.
- Customer interactions can create useful signals for AI providers.
- Economic value, in Nadella’s view, can concentrate with infrastructure operators instead of knowledge-generating companies.
The result is a debate about fairness, leverage, and control in the AI economy. Nadella’s argument is that the companies creating valuable knowledge should be alert to how that knowledge moves through the systems they pay to use.