Key Takeaways

  • Friedrich Hayek showed that market economies function because tacit knowledge is spread locally across millions of actors, not concentrated in a single central planner.
  • Kenneth Arrow's information paradox explains why giving away interaction exhaust destroys enterprise value: model vendors capture your tacit workflows for free.
  • A company's most defensible asset over the next five years will be its private evaluation benchmark rather than its fine-tuned models.
  • Foundation models will become interchangeable commodities that hill climb against your company's proprietary test suites.

The Economics of Tacit Knowledge

Most founders assume foundation model providers will absorb all software value. If one model gets smart enough, it does everything. Nadella rejects that premise on economic theory.

He points to Austrian economist Friedrich Hayek. An economy works because local, tacit knowledge is distributed across independent firms. If one system could hold all context, the economy would collapse into a single monopoly firm. Nadella also cites Kenneth Arrow's information paradox: you cannot know the value of information until you reveal it, but once you reveal it, you have transferred it for nothing.

“Without that, there can't be only one firm, and it can't be one intelligence,” Nadella told Harry Stebbings. The structure of business demands distributed private knowledge that external models never see.

The Value Trap of Interaction Exhaust

Most software buyers worry about leaking customer lists, source code, or confidential documents into training sets. Nadella argues that leaders miss the larger leak: day-to-day interaction exhaust.

Every time an employee clicks thumbs up, rejects a draft, or edits generated code, they provide valuable human feedback. “Now for the first time the buyer needs the protection because the buyer, even as most people think of it as my data, it's even my exhaust of using it,” Nadella noted. “Who are you creating value for? The model company, versus capturing that value for yourself.”

When your workers train external weights with their daily judgment, your company gives away its margin. The vendor learns how your best underwriters, engineers, and operators think, then sells that capability back to your competitors.

The Enterprise as an Evaluation Engine

Model architecture changes every six months. Weights get cheaper, context windows expand, and reasoning models leapfrog one another. Because models shift constantly, fine-tuning on yesterday's model is a moving target.

Nadella predicts that within five years, a company's defining intellectual property will be its private evaluation benchmark. “What is a firm? I think a firm would be an AI benchmark,” Nadella said. “Every model comes in and tries to hill climb on that eval. And that to me is how a firm works.”

In this setup, open and closed models become interchangeable compute units. You do not care whether OpenAI, Anthropic, or an open-weight model runs the task. You plug them into an internal harness that tests their outputs against your private rubric. The model that hill climbs best on your benchmark wins your workload for the week. The benchmark stays yours; the model stays a vendor commodity.

What to Do With This

Stop letting your team hit thumbs up or thumbs down inside consumer AI interfaces. Build an internal spreadsheet of your 50 most expensive edge-case decisions from last quarter, complete with the messy inputs and the exact decision your best operator made. Turn those 50 cases into an automated test suite, and run every model update against it before letting agents touch customer work.