US Chip Embargo on China: What Went Wrong

The embargo was supposed to be the chokepoint. Deny China access to advanced semiconductors, and you deny China the ability to train competitive AI models. It is, on its surface, the kind of policy logic that makes sense in a conference room: control the hardware, control the race.

Alvin Wang Graylin, who spent years operating at senior executive levels inside both the US and Chinese tech ecosystems, has a different account of what actually happened. Speaking on the Moonshots podcast with Peter Diamandis, Graylin described a double failure that should be of particular interest to anyone watching how Washington writes technology policy.

The first failure is the one that gets talked about. US export controls, by cutting Chinese labs off from the latest NVIDIA hardware, forced them to innovate under constraint. DeepSeek, Alibaba's Qwen, Kimi — these models emerged from teams that had to squeeze more performance from less compute. Graylin's contacts at Chinese GPU companies were blunt about it: "Nobody wanted to buy our stuff," they told him. "We're two or three generations behind, we were less energy efficient, but now we can't make enough because every data center in China has to buy our stuff." The embargo that was meant to freeze China's AI development instead incubated a domestic semiconductor industry that, Graylin argues, will begin exporting chips within a few years and would not have existed otherwise.

The second failure is less discussed and more structurally interesting. The embargo assumed that training and hardware are inseparable — that if you can't buy the chips, you can't run the workloads. What it missed is that training is a job, not a facility. Chinese labs without access to Blackwell-generation hardware domestically are doing their training runs in international data centers — Singapore, Europe, wherever the compute is available — and bringing the model weights back on disk. The model file moves like a liquid, as one of the podcast hosts put it. The fabs are fixed in geography; the training job is not.

Graylin put the current situation plainly: the chip controls function "more as optics than anything." That's a severe verdict, but the mechanism he describes is not in dispute — it has been widely reported that Chinese labs are routing training through third-party infrastructure to work around the restrictions. Washington's response has been to layer on more requirements: tighter know-your-customer rules for international data centers, broader restrictions, more paperwork. Graylin's read is that each additional layer increases geopolitical friction without solving the underlying technical porousness, and that escalating pressure of this kind tends to produce defensive reactions rather than strategic concessions.

The distillation frame and who pays

The other pillar of US AI export policy, alongside hardware controls, is the push to restrict model knowledge transfer — what Anthropic has termed "distillation attacks," referring to Chinese labs allegedly using API access to closed Western models to improve their own. It's a framing that has found a receptive audience in parts of Congress, and it matters enormously who gets to define the terms here, because the regulatory apparatus that follows from this frame has real compliance costs.

Graylin is skeptical that the distillation framing describes something meaningfully distinct from ordinary competitive practice. Every major lab, he argues, distills from its peers — within organizations and across them. The interesting policy question is not whether distillation happens. It clearly does, in every direction. The question is whether building a federal enforcement regime around "distillation attacks" produces any security benefit, or whether it primarily generates compliance burdens for developers and researchers who interact with multiple model providers — and creates new grounds for incumbent firms to lobby against open-weight competition.

Alibaba researchers, Graylin noted, have found evidence that US models trained on outputs that included Chinese-language queries returned responses identifying themselves as Qwen. Distillation, in other words, is not a unidirectional vulnerability. Treating it as one, and writing policy to match that framing, is precisely the kind of asymmetric regulation that benefits established players while raising barriers for everyone else. The Kimi model fracture inside the Trump administration is already showing how contested this terrain is — and how quickly companies with lobbying capacity move to shape what gets regulated and what doesn't.

What Beijing is actually doing

Graylin's most counterintuitive argument concerns the gap between Washington's assumption about Chinese AI strategy and what he actually observed in Beijing.

The assumption, baked into most US policy rationale, is that China is racing toward ASI with the same urgency — and the same timeline assumptions — driving Silicon Valley. Graylin says the behavioral evidence points elsewhere. China's Cyberspace Administration reviews every domestic model release, adding weeks or months to deployment timelines. Labs have been instructed not to purchase certain available hardware. The regulatory framework around AI is substantive: data provenance requirements, content marking rules, prohibitions on anthropomorphizing AI for children. These are not the actions of a government that believes superintelligence is eighteen months away.

"They are not behaving like they believe ASI is around the corner," Graylin said. "If they did, they would not be telling their labs don't buy the H200s the Americans are giving them."

His read of Chinese industrial strategy is that it is focused on deployment and diffusion — how to get existing AI capabilities into manufacturing, medicine, and education — rather than on a race to a capability threshold. This contrasts with the US AI Action Plan, which Graylin characterizes as supply-side: better models, better chips, American dominance. The Chinese plan, as he describes it, asks what happens after the model exists. Neither framing is complete on its own, but the gap between them shapes every policy choice downstream.

Whether one finds Graylin's characterization of Chinese intent fully persuasive or not, his core observation — that US policy is being written to counter a race condition that may not exist in the form assumed — deserves to be taken seriously. As he put it: "Having a race condition forces people to make irrational decisions."

What cooperation would actually require

September 24 has taken on additional weight. Xi Jinping is scheduled to meet with Trump at the White House that day — a meeting significant enough that Xi is skipping the UN General Assembly. Graylin is contributing to advisory work around US-China AI safety dialogues timed to the same window. He is cautiously optimistic about the possibility of a shared safety framework, citing genuine alignment between the two governments on at least one point: neither wants a superintelligence that destabilizes the other's financial system.

That's a real common interest. But shared interest is not the same as shared framework, and any policy correspondent who has watched international regulatory negotiations knows the distance between the two. Before concluding that dialogues represent anything beyond diplomatic calendar management, I'd want to see several things: an agreed definition of what "AI safety" means in this context that both sides will actually hold to, some mechanism for verifying compliance with any commitments made, and evidence that the agencies doing the negotiating have authority to bind the labs operating in their jurisdictions. Intentions at the summit level have a way of dissolving by the time they reach the rulemaking layer.

Graylin's preferred frame for the relationship — not prisoner's dilemma but stag hunt, a game theory model in which cooperation produces the best outcome for everyone but requires mutual trust to initiate — is intellectually coherent. The difficulty is that stag hunt cooperation is fragile. One party defects, the coalition collapses, and everyone retreats to individual hunting. The safety-as-luxury dynamic already visible in US labs makes it hard to sustain cooperative commitments even domestically. Extending that to a bilateral framework with a geopolitical rival requires institutional architecture that currently does not exist.

Graylin's most provocative argument is also his most structurally sound: the chip embargo created the very competitor it was meant to prevent, and doubling down on the same logic will produce more of the same result. Getting that argument in front of policymakers who are currently adding compliance layers to a regime that already routes around itself — that's the project. Whether September's dialogues are the venue for it, or theater before a harder reckoning, is the question neither Graylin nor anyone else can answer yet.

---

Samira Barnes is Buzzrag's tech policy and regulation correspondent.