The IAEA analogy is the tell

Hours before Donald Trump sat down with Xi Jinping in Beijing, OpenAI's VP of Global Affairs Chris Lehane went on the record with Bloomberg to propose an international body to govern frontier AI, modelled on the International Atomic Energy Agency, and, this is the part that should make you sit up, co-led by the United States and China.¹

The timing was not subtle. The framing was not subtle. "We think there's an opportunity here for the United States to lead," Lehane said, "the way the U.S. led with the IAEA."¹ That is a sentence engineered to be quoted, and it has been. I want to take it seriously, because the sentence contains, in compressed form, both the most interesting move OpenAI has made in eighteen months and the reason the proposal as drafted cannot work.

Start with what is genuinely new. Every prior multilateral AI governance attempt, Bletchley in 2023, Seoul, Paris, has either excluded China or included China as a signatory to a declaration that bound nobody to anything. The Bletchley statement was signed by 28 countries, China among them, and produced no constraint on any deployment by any party.² What Lehane proposed in Beijing is structurally different: not a Western club with the door politely ajar, but bilateral institutionalisation between the two countries that actually train frontier models at scale. That is a real departure, and it is worth saying so before pulling it apart.

It is also, I think, the only part of the proposal that survives serious examination.

The analogy does the work, and the analogy is wrong

The IAEA was established in 1957 on a specific physical premise: fissile material is scarce, expensive to produce, and trackable. You cannot enrich uranium in a garage. You cannot email a centrifuge. The safeguards regime that the IAEA enforces, inspections, material accountancy, declared facilities, works because the underlying object of concern has mass, takes industrial infrastructure to produce, and leaves isotopic signatures wherever it goes. The treaty architecture is downstream of the physics.

Model weights have none of these properties. A frontier model is a file. It can be copied an unbounded number of times at zero marginal cost. It can be transmitted across borders in seconds, hosted on infrastructure in jurisdictions that have not signed anything, fine-tuned on undeclared compute, and distilled into smaller models whose lineage is, in practice, unverifiable. Llama's weights are on Hugging Face. DeepSeek's are on Hugging Face. The horse, to borrow a phrase, has bolted; what OpenAI is proposing is a stable.

The proposal's actual enforcement hook is not weight tracking, OpenAI is not naïve enough to propose that, but compute-usage verification, linked through NIST's CAISI agreements to national safety institutes worldwide.¹ This is a more defensible mechanism, because compute, unlike weights, is still concentrated in identifiable data centres owned by a small number of hyperscalers. You can, in principle, audit GPU clusters in a way you cannot audit a 600GB file.

But "in principle" is doing heavy lifting. China's domestic chip industry, CXMT, Biren, Cambricon, exists precisely because the export-control regime since 2022 has made it strategically essential. The capability gap is real but the trajectory is not in dispute. A compute-verification regime predicated on Western hyperscalers being the only credible training venues has a shelf life measured in single-digit years, and probably fewer. Any country serious about training frontier models outside the regime will, within that window, be able to. That is the difference between fissile material and silicon: one is hard because physics; the other is hard because supply chains, and supply chains move.

Timing as strategy

I do not think OpenAI's policy team is confused about any of this. Lehane is a former Clinton White House operative and an Airbnb policy executive; he was hired into the role after the November 2023 board crisis as part of OpenAI's governance rebuild.² He knows what the IAEA does and does not do. The proposal is not a serious technical blueprint. It is a positioning document, deployed in a specific diplomatic window, designed to put OpenAI's frame on the table before any government puts its own frame down.

The goal is not to win the argument about the IAEA analogy. The goal is to be the company whose vocabulary the regulators inherit.

This is a move with a long history in regulated industries. The Basel capital accords were written, in their operational substance, by the banks they were meant to discipline. Pharma's FDA relationship has cycled through capture and counter-capture for fifty years. The pattern is consistent: when a technology is novel and the regulators have no native vocabulary, the incumbents supply the vocabulary, and the vocabulary determines what counts as compliance. Whoever defines "pre-deployment testing" defines what a competitor has to build before they can ship.

Microsoft, Google DeepMind, and xAI all signed CAISI pre-deployment testing agreements the same week as Lehane's Bloomberg interview.¹ That is not a coincidence; it is the institutional scaffolding the proposal needs in order to be plausible. The argument OpenAI is making to Washington is not "trust us"; it is "look, the structure already exists, all you have to do is internationalise it." The structure already exists because OpenAI and its peers have spent two years helping NIST build it.

Safety as market position

I want to be careful here, because the lazy version of this argument is that any company proposing safety regulation is doing so cynically, and that is not quite right. OpenAI has genuine safety researchers, has published meaningful work, and has, for all the criticism it attracts, engaged with safety institutes more seriously than its critics sometimes acknowledge. The question is not whether OpenAI cares about safety. The question is what shape of safety regime serves OpenAI's competitive position, and whether that shape is the same as the shape that minimises actual catastrophic risk.

It is not obvious that it is. OpenAI faces three competitive threats that this proposal speaks to with some precision. Anthropic has made safety its brand identity and has the strongest claim to the moral high ground in any governance conversation. Google DeepMind has the deepest pre-existing relationships with European regulators and the broadest deployment surface area. xAI now has direct political access through Musk that no other lab can match. A multilateral framework anchored to NIST standards that OpenAI has already engaged with, with a compliance baseline that reflects OpenAI's current architecture choices, is, entirely apart from anyone's intentions, a structure that raises costs for the three competitors more than it raises costs for OpenAI.

The counterpoint, which I take seriously, is that OpenAI is also the lab most associated with aggressive commercialisation, and proposing a tight governance regime carries credibility costs. If the framework is adopted and OpenAI is then seen to be the one bending it, the reputational damage is asymmetric. So the proposal is not riskless for OpenAI; it is a bet that being inside the room writing the rules outweighs the cost of being held to them. On the historical evidence from pharma and finance, that bet usually pays.

What survives

So where does this leave a reader trying to work out what to think?

The China inclusion is the part to watch. Strip away the IAEA framing, which is mostly decorative, and what remains is a serious public statement from a frontier US lab that the People's Republic of China should be inside the governance tent, not outside it. That is at odds with the deregulatory-domestically-but-hawkish-internationally posture the Trump administration has adopted, with David Sacks as AI czar and the Biden-era AI executive order rolled back.² It is at odds with eighteen months of export-control framing in which China was the adversary, full stop. Whether the proposal goes anywhere is secondary; that OpenAI was willing to say it publicly, on the day of a Trump–Xi summit, is itself a data point about where one of the most plugged-in policy shops in the industry thinks the wind is blowing.

What is on the table is not really an IAEA for AI. It is the question of whether the US and China will treat frontier AI as a domain for managed cooperation or for unmanaged competition, and OpenAI has just publicly voted for the first.

The IAEA analogy is decoration. Anyone who tells you the proposal is a serious technical blueprint has either not read it carefully or has a reason to want it to be one. The physical premises do not transfer. The enforcement mechanism, compute verification, has a shelf life bounded by China's domestic chip trajectory, which is short. Treat the analogy as marketing, not architecture.

The lobbying read is correct but incomplete. Yes, this is an incumbent shaping the regulatory terrain it wants to be regulated under. That is what incumbents do; it would be more surprising if OpenAI did not. The question worth asking is not whether OpenAI's proposal is self-interested, it is, of course it is, but whether the alternative proposals on the table are better. The Bletchley track produced declarations that bound nobody. The export-control track has accelerated Chinese domestic capability and not obviously slowed Chinese deployment. A bilateral US–China institutional frame, even one written by interested parties, is at least a different kind of failure mode to choose from.

I am not convinced the proposal as drafted will be adopted. I am fairly convinced that something with its shape will be, eventually, because nothing else credible is being proposed and the regulatory vacuum will not last. The question is which version of the proposal ends up on paper, and which company's pen writes it.

For now, the pen is in Chris Lehane's hand, in Beijing, hours before a summit. That is not nothing.

Footnotes

¶