The two leading US AI labs filed essentially opposite policy briefs on US-China AI governance within roughly 24 hours of each other, both timed around the Trump-Xi summit taking place in Beijing.

On May 13, OpenAI’s VP of Global Affairs Chris Lehane proposed that the United States lead the creation of a global AI governance body that includes China as a member, modeled on the International Atomic Energy Agency. The mechanism Lehane sketched runs through the Center for AI Standards and Innovation, the AI safety arm of the US Commerce Department, connecting it to the network of AI safety institutes being built in other jurisdictions around the world. OpenAI has been calling for IAEA-style AI governance since Sam Altman’s 2023 Senate testimony, but explicitly naming China as a participant, and doing so on the eve of a Beijing summit, is the newer element. It is also not obviously what the Trump White House wants to hear. Administration officials have previously signaled they would reject any framework with China as a member.

On May 14, Anthropic published a policy paper titled 2028: Two scenarios for global AI leadership, which proposes the inverse. The thesis is that 2026 represents what the paper calls “the breakaway opportunity for American AI,” and that with the right moves the United States can lock in a 12-to-24-month lead over China; without those moves, China overtakes by 2028. The three policy actions Anthropic asks for are tighter enforcement against chip smuggling, formal clarification that model distillation attacks are illegal, and a coordinated push to get trusted US AI hardware and models adopted by allied governments. Each of these is, fairly clearly, an export-control move dressed in national-interest framing.

Two Policies, Two Market Positions

What is striking about the two proposals, set side by side, is how cleanly each maps onto the company that filed it.

OpenAI has the broadest international product footprint in the AI industry. Its ChatGPT business operates across most of the developed world. Its enterprise customers include large multinationals that depend on stable cross-border data and model arrangements. An IAEA-style multilateral body is precisely the kind of governance regime in which OpenAI could expect to be a participant, perhaps the central one, and which would dampen any unilateral export-control regime that might cut its addressable market.

Anthropic, by contrast, has positioned its frontier model behind hardware and access restrictions that no other major lab has matched. Mythos Preview is available only through the restricted Project Glasswing rollout to a small number of trusted partners. The pre-market vetting regime the White House is reportedly drafting is, in significant part, a response to Anthropic’s own description of what Mythos can do. A world in which compute-export controls bite hard, in which distillation is treated as theft, and in which allied governments coordinate on hardware standards, is a world in which Anthropic’s restricted-access model is structurally advantaged.

The Trump-Xi Backdrop

Both proposals landed alongside the highest-stakes Trump-Xi meeting since the trade negotiations of the first Trump term. Each company is trying, in its own way, to influence what the United States does with its current lead in frontier AI. OpenAI is asking for that lead to be exercised through multilateral standard-setting in which China participates. Anthropic is asking for it to be exercised through technology denial in which China does not.

Neither proposal is wrong, exactly. Both can be read as good-faith reasoning about how to keep AI development safe. They are also, just as plainly, reasoning that maps onto the commercial topology of the firms doing the reasoning.

–
By the Control Plane Editorial Team