China Doesn't Need to Beat Nvidia Everywhere

The AI race is no longer only about who builds the best model. It is about who can deploy useful ones without asking permission.

The United States has been building the wrong mental model of AI power, and the error is architectural. Power, in this image, looks like a skyscraper: one extraordinary structure whose height confers dominance on whoever controls the top floors. China is denied the elevator. The strategy writes itself. It is also wrong because towers aren't how cities work. Cities derive their power from the accumulated, ordinary functioning of banks and hospitals and logistics networks and government offices, none of them extraordinary, each running well enough to do its job, together producing something no tower can match or threaten. The question for any state isn't whether it owns the summit. It's whether machine judgment has been embedded deeply enough in the tissue of everyday economic and administrative life to compound, invisibly, over years. Washington has been fighting for the penthouse but the penthouse is no longer the point.

Training is the expensive, hardware-hungry process of building a model. Think months of computation, massive clusters, and the frontier chips that export controls are designed to deny. Inference is what happens after. It's the continuous, unremarkable work of serving queries, generating outputs, running agents, powering the recommendation that surfaces when you open a shopping app in Shenzhen or the diagnostic that a state hospital runs in Chengdu. Inference is less glamorous and considerably more strategically important. A country that can train a model has a capability. A country that can run one, at scale, on infrastructure it controls, has embedded AI into its institutional nervous system. The difference is between owning a weapon and breathing.

At the frontier, American controls bite. Nvidia's best hardware remains out of reach, and the software ecosystem built around CUDA (its libraries, its toolchains and two decades of accumulated developer intuition) cannot be replicated quickly. China remains behind at the highest end, and that matters given the frontier is where efficiency gains are won before they diffuse downward. But the frontier is not the whole field and below it something different is forming.

Huawei's Ascend chips, alongside processors from Moore Threads and Cambricon, are already running Chinese inference workloads — serving existing models rather than training new ones. The task is narrower, the performance threshold lower, the failure mode more survivable. More importantly, Huawei has stopped pretending to compete with Nvidia and started building an environment in which Nvidia becomes unnecessary. CANN as the optimisation layer. MindSpore as the framework. ModelArts as the deployment platform. The claim is systemic before it is competitive. Domestic chips are already running models at Zhipu even as the hardest training workloads still depend on American hardware. Substitution has arrived first exactly where independence matters most. That sequencing is not accidental.

What produced it is worth understanding. As long as American hardware could be treated as a reliable commercial input, Chinese firms faced a straightforward calculation: use the best available tools and worry about alternatives later. The moment access became conditional (understood not as a market relationship but as a political instrument) the calculation inverted. A superior stack you cannot count on is worth less than an inferior one you can. Export controls have made Chinese AI more expensive and more difficult. They have also made the most compelling possible case for Chinese alternatives that anyone could have designed, which is less an argument against controls than a warning about their ceiling.

Recent moves to permit renewed sales of Nvidia's H200 into China make the ceiling visible. What is being managed now is not a hard cutoff but calibrated throttling which involves selective access to yesterday's hardware and a control regime that has drifted towards negotiation. A strategy calibrated mainly to the summit loses coercive force as the layers beneath it are domesticated, one workload at a time. China may remain behind at the frontier for years but if it can serve capable models at scale on hardware Washington cannot switch off then the gap at the top stops governing what happens everywhere else.

Washington is still counting floors. Denied the elevator, China builds the streets.