Nadella warns enterprises they're trading business secrets for AI access

Microsoft CEO Satya Nadella has become the latest tech leader to publicly caution enterprises about the hidden costs of relying on proprietary AI models. In a recent post, Nadella articulates a concern gaining traction among venture capitalists and business leaders: that companies deploying third-party AI services sacrifice competitive advantage while enriching the model builders themselves.

The Double Payment Problem#

When organizations use AI from providers like OpenAI or Anthropic, they incur two distinct costs. The first—direct payments for token usage—is visible and expected. The second is invisible: the proprietary business knowledge they must expose to make those models genuinely useful. As Nadella frames it, "You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful."

The more sophisticated an organization's queries and feedback, the more its internal expertise seeps into these systems. Each correction the model receives becomes institutional know-how that competitors could never purchase.

The Data Extraction Loop#

These AI platforms absorb enterprise intelligence in multiple ways: through user prompts, the tools agents deploy, and crucially, the corrections people submit when models misperform. Over time, this creates a learning signal that distills an organization's competitive differentiation directly into the vendor's system. Model providers harvest this "exhaust" continuously, accumulating insights about how customers operate their businesses.

The Hypocrisy Argument#

Nadella highlights what he sees as a fundamental inconsistency in how AI labs operate. These companies legally train models on vast swaths of public internet data without restriction, claiming fair use protections. Yet they impose strict terms preventing enterprises from "distilling" their models—reverse-engineering them through careful analysis of outputs to build cheaper alternatives. Nadella argues this selective enforcement unfairly privileges model makers: "While the great innovation that comes from model providers having fair use rights to train models on public data is needed, I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation."

Microsoft's Proposed Solutions#