Washington’s chip embargo didn’t slow Chinese AI. It made Chinese AI faster.
It is true that the ninth annual World Artificial Intelligence Conference in Shanghai looked like exactly what the export-control strategy was supposed to prevent — a pageant, not a threat. Robots making cappuccinos. Furry AI pets batting eyelashes at cooing onlookers. Twenty football fields of exhibition space, much of it occupied by machines that couldn’t load a dishwasher without dropping the plate on the floor. The humanoid from Suzhou-based UniX AI, listed at $44,000, was shaky and slow, laundry sticking to its claw-like hands. This was supposed to be the showcase for a superpower’s AI ambitions, and it had the feel of a trade show running on hope and backup generators.
The trouble is that the spectacle is the wrong layer to evaluate. Underneath it, Moonshot AI’s Kimi K3 scored higher than Anthropic’s Claude Opus 4.8 on many benchmarks — the ceiling model Moonshot had access to distill. Within 48 hours, Alibaba claimed to have topped Moonshot with its own new model. These are not cappuccino robots. These are language models built on hardware that American policymakers specifically designed an export regime to prevent from being competitive.
The engineering reality, stripped of geopolitics, is simple. When you constrain hardware, you optimize software. This is the oldest lesson in constrained computing — memory was once so expensive that every byte shaped the compiler’s design choices, and the efficiency gains born of constraint outlasted the constraint itself. Chinese AI researchers, cut off from the best chips, did what engineers do under constraint: they found optimization paths that the unconstrained never bother to look for. Kimi K3’s performance against Claude Opus 4.8 is not evidence of distillation — the OpenAI executive Dean Ball acknowledged that distillation no longer explains the gap. It is evidence that a specification, even one imposed adversarially, drives innovation that abundance does not.
This is the point at which the export-control strategy reveals its foundational error. It was designed by people who think about technology the way investors think about technology — control the chokepoint, control the downstream market. The engineer reads a chokepoint differently: a constraint is a specification, and specifications drive design. Washington built a strategy on the assumption that Chinese engineers would stare at the blocked supply chain and stop. Instead, they read the blockage as a requirements document.
The political economy is not subtle. American chip companies lobbied for restrictions that protected their market position while framing the controls as national security. Nvidia’s China sales have stalled as Huawei gains ground. The controls gave Washington the feeling of having done something — which is the most dangerous product a policy apparatus can manufacture. What Cory Doctorow calls the “anti-monopoly toolkit” — competition, regulation, self-help, labor — operates in reverse here: the regulation lever, wielded by a state acting as its domestic industry’s monopolist, suppresses competition rather than enabling it. Washington’s chip restrictions were regulation in service of monopoly, not regulation against it.
What Shanghai actually demonstrated is that hardware denial has produced the outcome it was meant to prevent: a Chinese AI ecosystem that does not depend on American components, building open-source models that undercut American companies on price and availability. China’s government has decided AI is infrastructure, not a product — the AI Plus blueprint targets deployment across 90% of the economy by 2030. The open-source dimension is what should alarm American AI executives. Chinese models are gaining global traction because they are open-weight, cheaper to run, and available to anyone with a server. This is the pattern Tim Wu identifies when monopoly depends on controlling scarcity: open-weight Chinese models break exactly the scarcity Wu says sustains monopoly — the control point dissolves when the model itself is free of licensing costs. It is also the pattern that already killed American dominance in smartphones, electric vehicles, and batteries — Western companies invent the category, Chinese companies build it at cost, and the question of who leads becomes irrelevant when the product is everywhere and the margin is zero.
Moonshot had to halt new subscriptions for its premium services — computing resources strained, the chip bottleneck real. This is the one piece of evidence favoring the export-control thesis, and it cuts both ways. It proves the controls are biting. It also proves that Chinese companies are demand-constrained rather than supply-constrained, which is the position an engineer would rather be in. Shortages create workarounds. Workarounds become architecture. Architecture outlasts the policy that created it.
The export controls were enacted through executive action and agency rulemaking, not legislation. They can be reversed with a pen stroke. The open-source infrastructure now escaping the American orbit cannot.