Satya Nadella at Davos: AI Diffusion Is Everything

Why Diffusion, Not Model Innovation, Determines AI's Economic Impact

Satya Nadella sits down with BlackRock CEO Larry Fink at Davos 2026 to discuss what may be the defining question of AI: how do you ensure the technology spreads fast enough to create surplus everywhere?

"If we are not talking about health outcomes, education outcomes, public sector efficiency, private sector competitiveness, we will quickly lose the social permission to use scarce energy to generate tokens." This is Nadella's clearest articulation of AI's legitimacy challenge. It's not enough to build impressive models - the benefits must diffuse broadly and quickly, or the entire enterprise loses its license to operate.

The new commodity economics: tokens per dollar per watt. Nadella argues GDP growth will be directly correlated with how efficiently countries and firms can produce and consume AI tokens. The supply side requires ubiquitous "token factories" connected to grids and networks - "just like we delivered bits, you have to deliver tokens plus bits." The demand side requires real leadership to translate tokens into outcomes.

AI is flattening information flow in organizations. The Microsoft CEO describes how preparing for Davos meetings has fundamentally changed. Instead of field teams preparing notes that trickle up through hierarchies, he asks Copilot for a 360-degree brief and immediately shares it across all functions. "It's a complete inversion of how information is flowing in the organization." Organizations structured for hierarchical information flow will need to redesign.

Firm sovereignty in AI isn't about data centers - it's about tacit knowledge. In the most provocative claim of the interview, Nadella argues the real sovereignty question nobody's discussing: "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. You're leaking enterprise value to some model company somewhere."

The barbell effect: startups and enterprises face different AI challenges. Small companies starting fresh can build knowing these tools exist. Large organizations have relationships, data, and know-how - but if they don't translate that with a new production function, they'll be stuck. "The change management challenge for large organizations is going to be bigger. The structural challenge for small organizations of how to overcome scale issues is going to be harder."

Key Takeaways

• Tokens per dollar per watt - The metric that will determine GDP growth; countries need efficient token production and consumption
• Token factories everywhere - Like electricity grids, ubiquitous AI infrastructure must be deployed globally
• Social permission required - AI must improve real outcomes (health, education, efficiency) or risk losing legitimacy
• Information flow inversion - AI flattens hierarchies; information no longer trickles up through departments
• Mindset-skillset-dataset - The formula for AI adoption: change how you think, build the skills, then engineer the context
• Enterprise sovereignty = tacit knowledge - The real sovereignty question is whether firms can embed their knowledge in models they control
• Barbell adoption - Startups adapt easily; enterprises have assets but face change management challenges
• Skilling is diffusion's limiting factor - How broadly people are skilled in AI determines how fast benefits spread

Implications for Enterprise AI Adoption

Nadella frames AI transformation as fundamentally a diffusion problem, not a technology problem. The models are advancing rapidly - the question is whether organizations, industries, and countries can absorb and translate AI capabilities into real outcomes fast enough. For enterprises, this means the competitive advantage isn't in which models you use, but in how deeply you can embed your organization's unique knowledge and workflows into AI systems. The firms that master "context engineering" - feeding AI the tacit knowledge that makes their business unique - will capture value. Those that treat AI as a generic tool will leak their competitive advantage to model providers.