Satya Nadella is asking the right AI question

The shift from intelligence to compounding intelligence

The most revealing sentence in Nadella’s essay may be this one:

"The real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound."

That is a subtle statement. And a profound one.

For the last two years, enterprise AI conversations have largely revolved around model capability. Which model reasons better? Which model writes better code? Which model has the largest context window? Which model tops the benchmark rankings? Those questions matter. But they implicitly assume that intelligence itself is the scarce resource. Increasingly, it isn't.

The frontier models being developed by OpenAI, Anthropic, Google, Meta, xAI, and others continue to improve at remarkable speed. Every few months, capabilities that seemed extraordinary become ordinary. The intelligence layer is becoming abundant. And when a resource becomes abundant, attention shifts to the system that organizes it.

Electricity became infrastructure. Computing became infrastructure. Networking became infrastructure. The same thing appears to be happening to intelligence. As I argued recently in "The next enterprise AI breakthrough will look obvious in retrospect," the most important question is becoming less about which model is smartest and more about how intelligence is organized, deployed, governed, measured, and continuously improved inside the enterprise. That is a fundamentally different question.

The company veteran problem

Another idea in Nadella's essay deserves attention: He argues that organizations should be able to replace a general-purpose model without losing the expertise accumulated inside their systems. His phrase is memorable: The company should retain its "company veteran" expertise.

Again, this sounds obvious... but it is surprisingly rare in today's AI architectures. Most enterprise AI initiatives still depend heavily on capabilities that live inside the model itself. Improve the model and you improve the system. Replace the model and you risk losing behavior, adaptation, and accumulated learning.

Nadella is pointing toward a different architecture: one in which the durable asset is not the model, it is the learning system surrounding the model. This is remarkably similar to what happened in previous platform transitions:

• Companies do not rebuild their ERP systems every time databases improve.
• They do not redesign their CRM strategies every time processors become faster.

The durable asset lives above the infrastructure. AI appears to be moving in the same direction: The model improves, the learning loop persists.

The return of feedback

The most striking part of Nadella's essay is that it quietly reintroduces a concept that has been strangely absent from much of the AI conversation: Feedback.

• Private evaluations
• Private reinforcement learning environments
• Improvement against business outcomes rather than benchmark scores

These ideas share a common theme: They are all mechanisms for connecting action to outcome. And that is precisely where many enterprise AI systems still struggle.

In "After the illusion: what enterprise AI must become," I argued that the industry had optimized AI to answer questions when companies actually need systems that change outcomes. The distinction sounds semantic until you realize that outputs can be generated without ever knowing whether they mattered. Outcomes cannot.

The moment a system begins measuring whether its actions moved the organization closer to its objectives, something changes: The system stops being merely generative, and it becomes adaptive. And adaptation compounds.

This is not a new idea in computer science. Systems such as DeepMind's AlphaGo and AlphaZero demonstrated years ago that feedback loops can produce extraordinary capabilities when intelligence is connected directly to objectives rather than merely to prediction. What is new is the possibility of applying similar principles to enterprises themselves.

The ecosystem question

The final section of Nadella's essay may be the most important: he argues that a world where all value accrues to a handful of foundation models is not economically or politically stable.

He's right: Every successful computing era eventually produced an ecosystem.

• The PC created software companies.
• The web created digital businesses.
• The cloud created entire industries.

The platform became valuable because value accumulated on top of it, not because all value remained trapped inside it.

This argument aligns closely with what I described in "Enterprise AI is in 1991. Where's its web?" The internet worked before the web: TCP/IP existed, email existed, FTP existed... What was missing was the layer that made those technologies consumable by ordinary organizations.

Enterprise AI today feels remarkably similar. The infrastructure is real. The capabilities are real. But the layer that allows organizations to build durable value on top of that infrastructure remains incomplete.

The companies that ultimately define the next phase of enterprise AI may not be the ones building the most powerful models: They may be the ones building the systems that allow every organization to convert intelligence into compounding institutional knowledge.

The next question

This is why I think Nadella's essay matters. Not because it provides answers, but because it asks the right question: If intelligence is becoming abundant, where does durable advantage come from?

His answer is the learning loop, and I think he's absolutely right. The next chapter of enterprise AI will not be defined by which model wins-it will be defined by which architectures allow organizations to turn human knowledge into systems that learn, improve, and compound over time.

The companies that figure that out will not simply be using AI, they will be building a new form of organizational capital. And that may turn out to be the most important asset of the AI era.