Microsoft and Google DeepMind agree on AI control — but not on who holds it

Two of the most powerful people in AI took to X this month to lay out competing visions for how the industry should police itself. Satya Nadella went first, on July 12, with a post called The Reverse Information Paradox. Demis Hassabis followed two days later with A Framework for Frontier AI and the Dawning of a New Age. Read together, the pair reads less like consensus and more like two companies quietly marking their turf.

Nadella's pitch is about where value gets captured. He argues enterprises effectively pay for AI twice — once in token costs, and again when their prompts, corrections, and evaluations leak proprietary know-how back into someone else's model. His answer is to own the learning loop entirely: the data, the traces, the evals, the adapted weights, the memory. Put a model-agnostic orchestration layer on top of that, he says, and the underlying model becomes a swappable commodity while the real value sits with whoever controls the orchestration. That advice happens to point enterprises straight at Azure and Foundry, where Microsoft keeps the billing, deployment, and governance no matter which model a customer picks.

Hassabis is worried about a different layer entirely — not who profits, but who signs off before a model ships. He wants a FINRA-style standards body, funded by industry but overseen by government, that tests frontier models for cyber, biological, and deception risks. Labs would submit models up to 30 days ahead of release, voluntarily at first, then as a mandatory gate for deploying in the US. Google DeepMind already runs its own internal Frontier Safety Framework, so it would likely walk into any such regime with a head start on compliance, particularly if the labs shaping the rules look a lot like itself.

Neither man is wrong, exactly. Owning your data and keeping models interchangeable is sound advice for any enterprise buyer. And an independent testing gate for frontier models is a reasonable response to real risks. But it's hard to miss that Nadella's fix reinforces the enterprise boundary Microsoft already sells into, while Hassabis's gate rewards labs with mature safety operations — which, again, describes Google DeepMind. To his credit, Hassabis's proposal isn't narrowly self-serving on paper: it would cover every frontier model regardless of origin or openness, and it calls for independent experts and open-source representatives on the oversight board, answering to the US government rather than to any single company.

What both posts really signal is that the competition has moved past benchmark scores. The fight now is over who controls the data boundary, the deployment layer, and the rules for who gets to ship a model at all. Two manifestos, two different chokepoints, and — not coincidentally — two companies staking a claim to the one they already sit closest to.