China Doesn't Need to Win the AI Race to Break It

The debate around Chinese AI has been framed as a race. Can Chinese labs catch U.S. frontier models? By most assessments, no, and the gap at the frontier may be widening.

But after a recent discussion with Goldman Sachs' Asia internet research team, I'm convinced that's the wrong question. Chinese models don't need to beat U.S. frontier models. They only need to be good enough, at a fraction of the price, to reprice the economics of AI everywhere.

The market is splitting into two tiers. Premium frontier models capture high value, complex workloads at premium pricing, and low cost inference captures everything else. And "everything else" is most of AI usage, things like inbox cleanup, summarization, drafting, and routine white collar workflows that don't necessarily need the best model in the world.

Chinese models price around the $0.20 level, and top U.S. models run $4-8. When the cost gap is 20-40x and the capability gap doesn't matter for the task, token volume migrates. Volume follows price. On some third party platforms, Chinese models already approach half of usage.

The playbook is familiar from Chinese e-commerce and EVs: subsidize adoption, undercut on price, win volume, consolidate, monetize later. In this environment, negative gross margins aren't a bug in the model so much as the strategy.

The catch is that innovation globalizes faster than profits. Many Chinese advances ship open source or open weight, including sparse attention, memory reduction, and other inference efficiency techniques. U.S. labs adopt them quickly, and they already have in their cheaper flash tier models.

The value accrues to whoever packages and distributes cheap inference, which is cloud platforms and the application layers built on top of them. So Chinese AI can compress global inference pricing without Chinese companies capturing the profit pool. The deflation gets exported even when the profits don't.

Underneath all of this is a self improvement engine. Once a model attracts real usage, it generates the data to improve its next version. Coding usage matters most, because it produces the highest quality feedback data there is.

Export controls and model access restrictions may slow specific channels, but they can't stop that flywheel if domestic Chinese adoption gets large enough. Distillation from U.S. models mattered early on, but from here, real user data matters far more.

Add it up and you get two very different scenarios for markets.

In the first, frontier leadership is everything, so premium pricing holds, U.S. hyperscaler capex earns its return, and the valuation premium on frontier labs is justified. In the second, the clearing price of inference collapses toward good enough levels, token volume migrates down tier, and the profit assumptions embedded in U.S. AI valuations, cloud capex, and semiconductor demand all need revisiting.

The frontier race decides the first world. The price of inference decides the second, and that is the one nobody is positioned for.

The positioning problem comes with a proxy problem. When investors turn bullish on China AI, they buy Alibaba or KWEB, but KWEB is internet and software exposure, not AI. The real Chinese catalysts live in domestic chips, memory, robotics, and compute infrastructure, which is exactly where constrained Nvidia access pushes incremental inference demand.

Investors have short memories, and most have already moved on from DeepSeek. So a DeepSeek 2.0 moment, likely from hardware rather than chat models, would catch the market using the wrong instruments to express the right view.

China AI isn't a single stock question or a national security footnote. It's a global macro variable. The frontier race gets the headlines, but the clearing price of inference is what reprices markets. China doesn't have to win the race to change the economics for everyone running it.