Satya Nadella on AI Diffusion: Tokens Are the New Commodity

The most revealing thing about this conversation isn’t any single prediction. It’s the mental model Satya Nadella uses to think about AI: tokens are a commodity, like electricity. GDP growth will correlate directly with how efficiently a country converts tokens into economic surplus. Everything else, the infrastructure investments, the skilling programs, the sovereignty debates, flows from this one premise.

The Setup

At Davos 2026, BlackRock CEO Larry Fink sits down with Nadella for a 44-minute conversation that’s less about Microsoft’s products and more about how AI restructures economies. Fink brings the investor’s lens: where’s the demand, is this a bubble, who gets left behind? Nadella brings the platform builder’s perspective: it’s about diffusion, not admiration.

The conversation has a specific texture. Fink keeps pulling toward the equity question (will AI widen the gap between rich and poor countries?), while Nadella keeps reframing it as an engineering problem with a knowable solution: make tokens cheaper, spread them everywhere, skill people up.

AI as One Continuous Arc

Nadella rejects the idea that AI is a rupture. He sees it as the latest chapter in a 70-year arc of computation: mainframes, minicomputers, client-server, web, mobile-cloud, and now AI. Each era digitizes artifacts about people, places, and things, then builds analytical and predictive power on top.

The difference this time is the abstraction level. He traces the evolution through software engineering as an example: code completions (GitHub Copilot) led to chat-based Q&A, which led to task-based agents, which led to fully autonomous coding agents that work 24/7. But even with autonomous agents, the human developer still has agency. He explicitly pushes back on the narrative that AI operates outside human control.

His analogy is striking: in the early 1980s, if you’d told people that four billion humans would wake up every morning and start typing, they’d have asked why. “We have a typist pool, that’s good enough.” But the PC created an entirely new category called knowledge work. AI, he argues, will do the same thing again, at 10x to 100x scale.

He also invokes Bill Gates’s old question: what’s the real difference between a document, a website, and an application? The answer used to be the lack of software that could transform between them. AI solves this. You write a document, say “make it a website,” and it generates the code. Say “make it an app,” and it transforms again.

The Token Economy Framework

This is the core intellectual contribution of the conversation. Nadella introduces a simple formula: tokens per dollar per watt. He claims this metric will directly correlate with GDP growth in any economy.

The logic chain:

1. Tokens are the new commodity, like electricity
2. Every economy’s job is to translate tokens into economic surplus
3. The cheaper and more abundant the tokens, the more surplus gets created
4. Therefore, the infrastructure challenge is building “token factories” everywhere, connected to the power grid and telecom networks

He notes that token pricing drops by roughly half every three months. This deflationary curve means any country or firm can plan around a commodity whose price is rapidly approaching zero. The constraint isn’t the token cost itself; it’s the surrounding infrastructure: energy, data centers, construction costs, grid modernization.

“We’ve got a new commodity, its tokens. And the job of every economy and every firm in the economy is to translate these tokens into economic growth.”

Fink validates this with a concrete data point from BlackRock: processes that used to take 12 hours to compute now take minutes. They’re managing $14 trillion across hundreds of thousands of mandates, and without AI, they simply couldn’t operate at their current scale.

The Bubble Test

When Fink raises the AI bubble question, Nadella offers a clean diagnostic:

“A telltale sign of a bubble would be if all we’re talking about are the tech firms.”

If the conversation stays on the supply side (chips, models, infrastructure spend), that’s bubble territory. The proof that it’s not a bubble is when you start hearing about an AI-accelerated drug making it through clinical trials, or a financial services firm transforming its operations. It’s not even about the “magical molecule”; it’s about AI speeding up the entire pipeline around it.

He’s more confident than cautious: AI will build on the rails of cloud and mobile, diffuse faster than previous technologies, and bend the productivity curve. But he distinguishes between two types of economic growth: capital-expenditure-driven growth (what we’re seeing now, mostly in the US) and surplus-driven growth (what needs to happen everywhere). The first is a narrow, temporary calculation. The second is the real prize.

Diffusion Is Everything

Nadella’s most pointed claim: if AI doesn’t create real economic surplus across industries and countries, “we will quickly lose even the social permission to take something like energy, which is a scarce resource, and use it to generate these tokens.”

Unlike the PC era, AI can ride the rails already laid by mobile internet. Token outputs are “pretty much available everywhere.” He cites a case from early 2023: a rural Indian farmer used a chatbot built on early GPT models to query agricultural subsidies in a local language, and even had it fill out forms. “It brought back agency to someone who perhaps didn’t have that because the technology was so much more accessible.”

On country-level adoption, he’s surprisingly optimistic. The quality of developers and startups in Jakarta, Istanbul, and Mexico City is “not that different” from Seattle or San Francisco. This is, he says, “the most uniformly spreading technology at least I’ve seen.”

But three conditions must hold for scaled adoption: a capital investment environment that attracts infrastructure spend, grid modernization (a fundamentally public-sector responsibility), and demand-side push from firms actually using the technology. The skilling dimension matters too. He draws a parallel to the PC era, when learning Excel or Word was directly tied to getting a job. AI needs the same vocational connection: pick up this skill, become a better provider of some product or service in the real economy.

The Organizational Inversion

The most practical section of the conversation. Nadella describes how AI is inverting information flow inside organizations. His Davos preparation example is revealing: preparing for 50 bilateral meetings used to involve field teams writing notes, HQ refining them, information trickling upward through departments. Nothing had changed from when he joined Microsoft in 1992 until a few years ago. Now he asks Copilot for a brief on each meeting. The AI gives him a 360-degree view: what Microsoft does as their client, what they do as Microsoft’s client, investment relationships, everything. He then shares that brief across all functions instantly.

“It’s a complete inversion of how information flows in the organization. It’s not the classic departments and trickle-up anymore.”

His formula for organizational transformation:

1. Mindset: Leaders must commit to changing the work, not just adopting the tool
2. Skill set: You can’t talk about AI in the abstract; you have to use it and learn to put guardrails around it
3. Data set (context engineering): The AI is only as good as the context you feed it. Firms’ tacit knowledge, the accumulated intelligence in departments and workflows, needs to become the AI’s context

He observes a barbell effect: small companies that start fresh adopt AI tools natively. Large organizations have advantages (relationships, data, know-how) but face massive change management challenges. Neither side can coast.

“If your rate of change doesn’t keep up with what’s possible, you’re going to get schooled by someone small being able to achieve scale because of these tools.”

On sector comparisons: the US financial sector’s adoption of AI is “night and day” faster than its earlier adoption of the cloud, partly because regulatory barriers that slowed cloud adoption don’t apply in the same way.

The European Paradox

This may be the most politically charged part of the conversation, and Nadella doesn’t hedge. His message to Europe: stop thinking defensively about sovereignty and start thinking about global competitiveness.

His argument: European firms, from German Mittelstand manufacturers to Swiss pharma and financial services, have always thrived by producing things the world needs. “When I go to a jeweler or a dentist in the United States, I’m surrounded by German Mittelstand products.” The miracle of the European economy over 200-300 years is global competitiveness, not protection.

He makes a counterintuitive point about data sovereignty: Europe should be more concerned about whether its industrial and financial companies have access to data from the US and the rest of the world, rather than obsessing over protecting European data from leaving. You’re only competitive if your output is globally competitive.

On energy: Europe’s dependence on imported power is a real constraint. The token-per-dollar-per-watt equation means expensive energy directly translates to less competitive AI. He acknowledges this without sugarcoating it: grid modernization is a public-sector responsibility that can’t be fully solved by “behind the meter” private solutions.

Corporate Sovereignty: The Overlooked Battleground

The freshest idea in the conversation. Nadella predicts that the sovereignty debate will shift from nations to firms:

“If your firm is not able to embed the tacit knowledge of the firm in a set of weights in a model that you control, by definition you have no sovereignty. That means you’re leaking enterprise value to some model company somewhere.”

The physical location of the data center is, in his words, “the least important thing.” Speed of light naturally distributes data centers globally. Encryption and key management are technically solved problems. What matters is whether you own the distilled intelligence of your own organization.

He invokes David Ricardo: comparative advantage is real, in countries and in firms. That needs to be preserved in the AI era.

His prescription: it’s a multi-model world. The winning formula is to orchestrate multiple models (closed source, open source, self-built), feed them your proprietary context, and ensure that the resulting reasoning traces and capabilities become IP you control. That’s the entire picture.

“David Ricardo was not wrong. There’s comparative advantage in countries. There is comparative advantage in firms. That needs to be preserved even in the AI era.”

One notable absence: Nadella never mentions OpenAI by name throughout the entire conversation, while repeatedly emphasizing the “multi-model world” and the importance of enterprises building their own models. This aligns with Microsoft’s recent strategy of expanding partnerships with multiple model providers.

Some Thoughts

This conversation is more useful than most Davos panels because Nadella consistently provides frameworks rather than predictions. The tokens-per-dollar-per-watt formula is simple enough to be actionable, whether you’re a country planning infrastructure or a firm planning AI adoption.

• The “commodity” framing for AI tokens is powerful because it immediately suggests the right questions: who’s the cheapest producer? Where are the bottlenecks in the supply chain? How do you ensure fair distribution?
• The bubble diagnostic is elegant: if non-tech firms aren’t the ones benefiting, it’s a bubble. If they are, it’s a platform shift.
• The corporate sovereignty argument is genuinely novel. Most sovereignty debates are about nations. Nadella’s point that firms are the more urgent unit of analysis, and that model dependence is the actual risk, reframes the entire conversation.
• The information inversion inside organizations (from hierarchical trickle-up to AI-flattened distribution) is one of those observations that sounds obvious once stated but has massive structural implications for how firms are organized.
• His refusal to frame AI as a rupture, insisting instead that it’s the latest chapter in a 70-year arc, is both historically grounded and strategically useful. It suggests the playbook from previous platform shifts still applies: diffusion determines winners, not invention.