Satya Nadella’s LinkedIn piece presents a strategic framework arguing that enterprises must build two forms of capital simultaneously: human capital, encompassing knowledge and judgment, and token capital, the AI capability a firm owns outright. His central claim is that these reinforce each other through a feedback loop where organizational expertise compounds over time across both people and AI systems.
The framework carries practical implications for AI infrastructure ownership. Nadella warns that enterprises dependent on external model providers risk losing accumulated institutional knowledge if those providers become unavailable. He advocates for private evaluation systems, internal reinforcement learning environments, and queryable knowledge bases that remain under enterprise control, making the underlying model layer interchangeable without destroying the learning investment built on top of it.
Satya Nadella just published a piece on LinkedIn that is not a product announcement, a marketing exercise, or a keynote recap. It is a framework for how enterprises should think about the value of AI, and it lands at exactly the right moment. His argument: every company needs to build two kinds of capital simultaneously. Human capital, the knowledge, judgment, relationships, and pattern recognition of its people, and token capital, the AI capability the firm builds and owns. The critical insight is that human capital does not become less valuable as token capital grows. It becomes more valuable.
The framework, in Nadella’s own words
Nadella frames AI as the first platform shift where you can create a genuine cognitive loop between people and digital systems. Previous transitions used digital tools to enhance human work. This one creates a feedback loop where the AI absorbs organizational expertise and compounds it. His formulation: “You can offload a task, or even a job, but you can never offload your learning. The future of the firm is the ability to compound that learning across people and AI.”
The practical test he proposes is pointed: a company should be able to swap out a “generalist” model without losing the “company veteran” expertise built into its learning system. If switching from one LLM provider to another destroys your accumulated institutional knowledge, you do not own your AI capital. The model vendor does.
The building blocks he identifies are private evaluations that measure whether the model improves against outcomes that matter to the business, rather than external benchmarks. Private reinforcement learning environments that strengthen models on real traces from inside the organization. Knowledge bases that make institutional memory queryable and token usage more efficient. He calls this loop a “hill climbing machine,” and argues it compounds like no other asset. Every improved workflow generates a stronger training signal, accelerating the accumulation of tacit knowledge unique to the firm.
Why does this land differently after Fable 5?
I read Nadella’s piece the same week the Fable 5 shutdown demonstrated what happens when you do not own your AI capital. Anthropic’s most advanced model was disabled for all users worldwide by a single government letter. Every company that had built workflows around that model discovered overnight that its AI capability was rented, not owned.
Nadella’s framework is the architectural answer to that vulnerability. If your token capital lives in a learning loop you own, built on your data, evaluated on your metrics, and deployable across multiple models, then no single model shutdown can destroy it. You can switch the underlying model the way you switch cloud providers: with friction, but without catastrophic loss. That is the sovereignty test he describes, and it is exactly the test that Fable 5 proved most enterprises would fail today.
The anti-concentration argument
The part of Nadella’s piece that surprised me is how explicit he is about the political economy risk. His words: “The last thing any of us wants is a world where every company across every sector is ceding value to a few models that eat everything they see. If only a few models accrue all the values, the political economy will not tolerate it.”
He draws the globalization parallel directly: entire industrial economies were hollowed out by outsourcing, the GDP numbers looked fine on the surface, but the displacement was real, and the consequences are still being felt. He argues we should not bring that dynamic into the AI era, with a few systems capturing all the economic returns while industries see their knowledge commoditized beneath them.
This is a striking statement from the CEO of a company that sells AI models through Azure. He is arguing in public that the value of AI should flow broadly across every company and industry, rather than concentrating in a few model providers. Whether Microsoft’s actions match this rhetoric is a fair question. Still, the intellectual framework is sound and aligns with the open-source AI movement and Mistral’s European sovereignty push that I have been covering.
The connection to on-device and edge AI
Nadella’s framework has direct hardware implications. If token capital needs to be owned, not rented, then the inference infrastructure needs to be controlled by the enterprise. That points toward on-device AI, edge deployment, and private cloud rather than shared API endpoints.
The AMD EVO-X2 is the desktop version of this thesis. The Snapdragon 8 Gen 4’s on-device NPU is the mobile version. The Qualcomm AI200 data center card is the rack-scale version. And Andrew Ng’s unbiggen thesis, that smaller, better-data models running on cheaper hardware can match larger models for most enterprise use cases, is the architectural complement to Nadella’s framework: If the learning loop matters more than model size, then the hardware need not be expensive. It just needs to be yours.
Investment perspective
Nadella’s framework is the clearest articulation of what European enterprises have been circling for two years. The GDPR, the EU AI Act, and the political push for digital sovereignty are all expressions of the same concern. If your critical business capability depends on an AI infrastructure you do not control, you are exposed. Nadella’s answer, build the learning loop yourself and make the model layer swappable, is architecturally compatible with both the European regulatory framework and the commercial reality of enterprises that need AI to work regardless of geopolitical disruptions.
The irony is that Microsoft is both the enabler and the potential beneficiary of this framework. Azure’s enterprise AI platform is designed to host exactly the kinds of private evaluations, private RL, and knowledge bases that Nadella describes. Whether enterprises build their token capital on Azure, AWS, their own hardware, or some combination, the framework works. But Microsoft is clearly positioning itself as the preferred infrastructure for the compounding loop.
Nadella’s framework is not a product pitch. It is a governance model for the AI era, and it is the first one I have seen from a major CEO that takes both the technology and the political economy seriously. The thesis that human capital and token capital compound together, that the learning loop is the new IP of the firm, and that value must be distributed broadly, or the political economy will not tolerate concentration, is the most important strategic framework in enterprise AI right now. Whether Microsoft practices what Nadella preaches is a different question entirely. But the framework itself is right, and every enterprise that reads it should start building its own hill-climbing machine before someone else’s model commoditizes the knowledge underneath it.