China's diffusion bet

The pitch on Chinese AI is usually that it is nearly as good and vastly cheaper: DeepSeek's per-token price runs 20 to 40 times below OpenAI's. JPMorgan's geopolitics unit, in an October report on the AI rivalry, buries a figure that cuts the other way. A US government evaluation last September found that comparable American systems came out roughly 35 percent cheaper end-to-end once quality — security, reliability, uptime under load — was held constant. The sticker price is low; the delivered cost is not.

That is the useful move in an otherwise familiar document. The bank's frame is that the two countries are not running the same race. The US is spending to win on frontier capability — private AI investment of $109 billion in 2024 against China's $9 billion — while China is running a state-coordinated sprint to diffuse whatever it has into every firm and agency as fast as possible, betting that adoption, not the best model, decides the outcome. The gap between the top American and Chinese models on public head-to-head rankings did narrow sharply, from roughly nine percent to under two in barely a year.

The one edge JPMorgan treats as structural is power. Chinese regional grids carry reserve margins of 80 to 100 percent — capacity built ahead of demand — against about 15 percent for US grids straining under data-center load. Models can be copied in a weekend; a decade of surplus generation cannot. If the frontier lead is real but the energy and adoption disadvantages are also real, the question that decides the outcome is not who has the best model this quarter but who can afford to run one at scale.