Hook
Last month, Satya Nadella stood on stage and told the world that enterprises are giving away their most valuable asset—not cash, not tokens, but the very knowledge their employees generate while using AI. "Companies are paying for the API calls," he said, "but they're also handing over the proprietary thinking that makes their business unique." This was not a quiet engineering memo. It was a strategic bomb thrown into the heart of the model supplier economy. And as a protocol PM who has watched centralized platforms extract value from user data for years, I saw something deeper: the exact same pattern that DeFi was built to break.
Context
The current AI API model is a perfect reproduction of the Web2 surveillance economy. You pay per token—that's the visible cost. But the invisible cost is far larger: every prompt you type, every correction your engineers make, every evaluation you run becomes a training signal for the model supplier. OpenAI, Anthropic, and others have built their business on what I call the "reverse information paradox." They claim the right to train on public internet content, yet they restrict customers from using model outputs to train their own systems. Meanwhile, they quietly ingest your enterprise's internal reasoning patterns to improve their next release. This is not an accident of engineering; it is a deliberate business architecture designed to centralize value at the model layer. Sound familiar? It should. It's the same logic that made Facebook and Google rich off user-generated content.
Core
Let me deconstruct this through a blockchain lens. In DeFi, we solved the "user-as-product" problem by separating the protocol from the application. Uniswap V4's hooks, for example, allow developers to build custom logic on top of the core exchange without sacrificing sovereignty. The model supplier's architecture is the opposite: they want to own both the base model and the feedback loop. But Nadella is right to sound the alarm, and ironically, his prescription aligns perfectly with the decentralization thesis.
Based on my experience auditing over 40 token models during the ICO era, I can tell you that the fundamental issue here is incentive misalignment. The model supplier's revenue (API tokens) grows when enterprises use their API more. But their model quality also grows when enterprises provide high-quality feedback (evaluations, corrections, preferences). So they are effectively double-dipping: charging you for access while using your expertise to improve their product, which they then sell to your competitors. In blockchain terms, this is a protocol that extracts MEV from its own liquidity providers.
What makes this particularly insidious is the asymmetry of learning. When your team spends 1000 hours crafting prompts and building evaluation datasets, that represents a massive capital investment—what Nadella calls "human capital." But under the current model, you cannot take that capital with you. If you switch from GPT-4 to Llama 4 tomorrow, you lose all the accumulated improvement. Your prompts, your fine-tuning weights, your agent trajectories—they are locked inside the model supplier's infrastructure. True ownership begins where the server ends. And right now, your AI learning assets live on their server, not yours.
Now, let's talk about the technical solution Nadella proposed: "Separate the orchestration layer from the model. Own your evaluation, your memory, your fine-tuning weights." This is essentially a call for modular, composable AI architecture. But here's the critical point he conveniently omits: Microsoft wants you to own those layers on Azure, which is still a centralized platform. He's offering you a slightly better cage, not an open field.
As a protocol PM who has watched cross-chain bridges bleed $2.5 billion to hacks, I know that code is law, but incentives are the judge. The real solution is not to move your AI assets from OpenAI's server to Microsoft's server. It is to put them on a decentralized, permissionless infrastructure where you control access, provenance, and value sharing. Imagine an on-chain registry where your enterprise's evaluation dataset is tokenized as an NFT you own, and any model that wants to use it must pay you royalties. Imagine a DAO where enterprises collectively govern the terms under which their feedback can train a public base model. That is the vision that Nadella's warning should inspire, not just another Azure upsell.
Contrarian
Before we get too excited, let me apply the pragmatism test. The most vocal advocates for "owning your AI learning" are the platform vendors themselves—Microsoft, Google, Amazon. They want you to buy their orchestration tools, their vector databases, their monitoring dashboards. The model suppliers (OpenAI, Anthropic) will fight back by offering "data-private" tiers where they promise not to train on your inputs, but those tiers will cost significantly more. The net result could be a widening gap between enterprises that can afford to build their own AI stack and those that cannot. Surprisingly, the most equitable outcome might actually come from open-source models like Llama or Mistral, combined with decentralized storage and compute. But even that requires technical sophistication most small businesses lack.
This brings us to the second blind spot: the assumption that enterprises can effectively "own" their learning assets without new forms of governance. If every company builds its own evaluation dataset and fine-tuned model, we lose the network effects of shared improvement. The open-source community thrives because of shared benchmarks and collaborative refinement. A hyper-fragmented landscape of proprietary knowledge islands could slow overall AI progress. Debate is the compiler for better consensus. We need protocols that enable selective sharing and collective benefit, not just siloed ownership.
Takeaway
Nadella's warning is a gift to the decentralization movement, even if he didn't intend it that way. He has publicly validated what blockchain advocates have been saying for years: centralized platforms extract value from user participation under the guise of convenience. The question now is not whether enterprises should own their AI learning assets—that battle is already won in principle. The question is whether they will build those assets on open, programmable infrastructures that give them true sovereignty, or on the next generation of proprietary cages. The market will decide, but I know where I'm placing my bet. After all, disruption is the baseline, not the goal. The goal is liberation.