Nadella’s AI Test: What Remains When You Swap Out Your Core Model?

What Is Token Capital?

Nadella defines two kinds of capital for every company: human capital —employees' judgment, experience, and networks; and Token capital —the AI capabilities a company builds and owns, named after the token, the smallest unit a large model processes.

The Executable Test

He asks whether a company can replace the underlying general‑purpose model without losing its proprietary expertise. If yes, the company owns its AI; if not, it is merely renting.

AI Asset Checklist

Hundreds of finely tuned prompts

A RAG knowledge base

Several fine‑tuned models

Agent orchestration logic

Internal user logs and feedback data

When swapping from GPT‑4 to Claude 4.7 and then to Qwen‑3, the impact on each asset is:

Prompt – largely wasted; prompts are model‑specific and must be rewritten.

RAG index – remains, but recall quality changes; embedding models must also be swapped.

Fine‑tune – resets to zero because it is tied to the base model.

Agent orchestration – structure can be reused, but each tool’s prompts need realignment, increasing tool‑misuse risk.

User logs and feedback – the only truly owned asset, provided it is stored in a structured way.

The first four are rented; only the last belongs genuinely to the company, yet many firms fail to capture it properly.

Learning Loop: What Nadella Really Promotes

He repeatedly mentions a "learning loop" consisting of three steps:

Convert workflows and domain knowledge into continuously trainable AI systems that evolve with each use.

Build private evaluations (private evals) tailored to real business scenarios, instead of relying on public leaderboards.

Iterate the model on internal data and interaction traces, performing proprietary reinforcement learning.

Private evals are especially scarce in China; companies often chase public benchmark scores, which do not correlate with business value.

Industry Analogy and Risk Hedge

Nadella likens the AI era to past waves of globalization that outsourced value and later caused economic backlash. He warns that if a few models monopolize knowledge, political and regulatory forces will push back.

His prescription is a risk‑hedging strategy: each company should build its own learning loop so that value remains partly within the ecosystem, avoiding a winner‑takes‑all dead‑end.

Returning to the Test

When companies honestly answer the test—"What remains after swapping all models?"—most find little left. The daily churn of new prompts, APIs, and models masks the fact that most assets are rented, not owned. Only structured user interaction data, failure retrospectives, and edge‑case handling truly belong to the company.

Try swapping every model call. What do you still have?

The answer varies, but few teams can count more than a handful of genuine assets.