Three industry leaders published very different plans this month. All three want to steer the race. None of them wants to steer it the same way.
Demis Hassabis writes in A Framework for Frontier AI and the Dawning of a New Age that AGI is only a few years off. He compares it to fire or electricity, not to the internet. His worry is speed. Progress is running ahead of our understanding, and the race gives no one room to pause. His fix is a US standards body, modeled on the finance regulator FINRA and paid for mostly by industry. Labs would hand over models 30 days before release for testing on cyber, bio, and deception risks. Benchmarks would update quarterly. If things got bad, the body could coordinate a slowdown. His goal is to buy time.
Satya Nadella, in The Reverse Information Paradox, flips an old economics puzzle. The seller used to risk giving away knowledge. Now the buyer does. To make a model useful, a company must feed it its own know-how — prompts, corrections, evals. That exhaust becomes the provider's asset. Nadella wants a hard trust boundary so firms keep their own evals, memory, and tuned weights. He lists five needs: control, capability, choice, cost, and compounding. His goal is to keep value with the companies that create it.
Jensen Huang's side appears in Open Weights and American AI Leadership, an open letter NVIDIA signed alongside roughly 300 other firms. The argument is that America wins by spreading AI everywhere — factories, hospitals, farms, classrooms — not by owning one frontier model. Open weights lower cost, force competition, and stop lock-in. It also claims openness is safer than secrecy, since many eyes find flaws faster. It asks policymakers for compute access, shared datasets, no early limits on open models, and no treating distillation as theft.
All three fear a world where a few firms hold everything. All three want America setting the rules. All three insist nothing is decided yet.
Hassabis wants a gate before release. Huang wants no gate at all. Nadella skips policy and fights the battle in contracts and architecture. Distillation shows the crack — Nadella calls today's restrictions ironic, the letter defends it outright, and Hassabis would test every frontier model, open or closed.
The signature list is the story. Google, Microsoft, OpenAI, Meta, and NVIDIA all signed the open-weights letter while still selling closed models. That is hedging, not conviction. Enterprises are quietly moving toward Nadella's position because it protects their money. Hassabis's standards body has the most support in principle and the least in practice — nobody wants to be first through the gate.
My read: follow who benefits. Open weights are rising in the US because the infrastructure players want them there. AWS, Microsoft, NVIDIA, Google, and the chip suppliers all earn more when models are cheap, plentiful, and running on their hardware. Open weights deliver that. They also do something quieter. They keep labs like OpenAI and Anthropic in the application layer instead of letting them own infrastructure and applications at once. The infra companies would rather sell the road than share it with a tenant who is also building roads.
The regulator proposal reads the same way. An independent US body decides what enters the market. In practice, that is a tidy way to keep foreign models out of the domestic race while calling it safety. It may honestly be both.
The last point is the one worth holding onto. The models change fast, but so do the ideas about how to manage them. Every plan here is a snapshot of who held leverage the week it was written.
Sources