Nadella: Companies That Only Call AI Models Are Just Working for the Model Companies

Satya Nadella recently appeared on the 20VC podcast. 20VC is a well-known venture capital podcast. In a tweet, host Harry Stebbings recalled that when he started the show at age 18 in his London bedroom, he didn't know anyone in the tech world. He had waited a long time for this chance to get the Microsoft CEO on. During Nadella's tenure at Microsoft, the company's market cap rose from $300 billion to $3.9 trillion.

In the interview, Nadella put a question that is often avoided on the table: when AI becomes a general-purpose resource, how will enterprises retain their value? He said:

“You are creating value for the model companies, not keeping that value for yourself. We need to ensure that more of the surplus value generated by using intelligence stays inside the enterprise. Every company will have a private evaluation benchmark—that is its core IP, and every model that comes in has to climb that benchmark.”

To translate: in the future, companies will not just use general-purpose models. They will use their own data, processes, and business rules to construct an evaluation benchmark that belongs only to them. A model must score high on this “exam” before it can be considered to truly understand the company. No matter how strong a general-purpose model is, it still has to sit for the test.

For readers unfamiliar with AI, an evaluation benchmark can be understood this way: AI development commonly uses public benchmark tests to measure model capabilities, such as MMLU and HumanEval. When a company organizes its internal historical data, customer cases, and business red lines into a test set, that is a private evaluation benchmark. It cannot be made public because it is core company knowledge.

This way of thinking runs counter to the mainstream narrative that “the more general AI is, the better.” Nadella's subtext is: model capabilities are becoming infrastructure that can be called upon at any time. What is truly scarce is an enterprise's ability to define “correct.” Whoever owns the private benchmark owns pricing power in the AI era.

There were other notable points in the interview. Nadella said that all AI models should be assumed to be potentially compromised, and AI systems need a “zero trust” architecture. He also questioned the reliability of chain-of-thought: a model showing its reasoning process does not mean the result is reliable. These views continue Microsoft's long-standing caution around security.

Harry Stebbings's interview itself is also a story. One commenter wrote: “You started from zero; in terms of multiples, you win.” Another joked that their company's private evaluation benchmark is “I ask whether this looks right.” Of course, that's a joke, but it points to the problem: many enterprises have not even sorted out their own data yet, and are still far from having a private evaluation benchmark.

Nadella's argument may not be entirely right. But the signal is clear enough: model companies are making models more general-purpose, while the companies that use AI best may be doing the reverse: turning their own data, rules, and evaluation standards into a private system that others cannot replicate. The future winner may not be the company with the smartest model. Companies with the most distinctive “exam” are more likely to have the edge.

Full interview video: YouTube