Everyone assumes it's an even race. It isn't. The US and China aren't spending the same amount, and they aren't paying for it the same way — and that gap decides who's exposed when the music stops.
Annual AI capital spend · 2025–26
America is making a far bigger bet
The US is pouring in roughly eight times what China is. One American private project — OpenAI's Stargate — was floated at up to $500B, larger than China's entire national AI capex for the year.
United States 2026 pace
~$700B / yr
~$3 trillion cumulative since ChatGPT · click to see how it's funded ▸
China 2025 total
~$84–98B / yr
Bank of America, 2025 · click to see how it's funded ▸
China is spending roughly one-eighth of the American pace — and getting more out of each dollar.
How the US funds its ~$700B — private capital, market-exposed
Amazon ~$220B2026 capex guidance
Microsoft ~$190B2026 capital spending
Alphabet ~$195–205B2026 guidance
Meta ~$125–145B2026 guidance
Paid for by hyperscaler cash flow (~$1.4T self-funded through 2028), record bond issuance (+974% in H1 2026) and private credit. Essentially no state financial support — and US firms draw more than ten times the private AI investment China's do.
Company earnings 2026 · Morgan Stanley · S&P Global · RAND
How China funds its ~$84–98B — state-directed, and buffered
Government — up to ~$56Bof the annual total
Internet giants — up to ~$24BAlibaba, Tencent, ByteDance, Baidu
Telecoms + special bondsthe remainder
State vehicles behind it: a ~$138B National VC Guidance Fund (AI & robotics), a ~$138B five-year Bank of China AI financing program, an ~$8.2B National AI Industry Investment Fund, and a drafted ~$295B national data-centre grid (reported, not yet enacted). Corporate pledges: Alibaba ~$53B over three years; ByteDance ~$20–25B in 2025 (plans up to ~$70B in 2026); Tencent ~$10.7B.
Why each dollar stretches further
China power~$0.08–0.14per kWh (industrial)
US power~$0.10per kWh (Virginia)
Roughly in line — China's electricity is not decisively cheaper per unit. The real edge is scale and speed of supply.
2.3× the US in total electricity generated (10,066 vs ~4,310 TWh, 2024)
+550 TWh of new demand added in 2024 alone — more than every advanced economy combined
340 GW solar + 80 GW wind added in a single year
Domestic chips (Huawei Ascend on SMIC 7nm) — forced by US export controls; cheaper, but lower-performance
Bank of America · RAND · Oxford Institute for Energy Studies · IEA · CSET, 2025–26
Model layer · usage vs revenue
Half the work, a sliver of the money
Here's the twist the raw spend hides. Chinese open-weight models now do nearly half the world's AI work — yet almost all the money still flows to the American labs, because China gives its models away for a fraction of the price. That's not weakness. It's the strategy.
Who does the worktoken usage · Aug 2026
China ~46%
US ~54%
Who gets paidmodel-layer revenue
4%
US closed models ~96%
China / open-weight US / closed
DeepSeek's API runs at roughly $0.14 per million tokens against GPT's ~$5 — on the order of ten to a hundred times cheaper.
America built a $3 trillion tower that can only be paid back by selling AI. China is deliberately driving the price of that very product toward zero.
The US bet is bigger, and it runs on private capital exposed to the market — US firms draw more than ten times the private AI investment China's do. China's is smaller and state-buffered. So the thing America most needs to earn its money back is exactly the thing its rival is giving away. Whoever "wins," the American market carries the risk.