"China and the US Are Running Different AI Races" is an article by Poe Zhao published in AI Frontiers on 2026-02-12. It argues that US and Chinese AI startups are optimizing for fundamentally different customers under different economic constraints, producing two distinct definitions of AI "success": the United States leading in model capability and China leading in industrial deployment.
Summary of argument
Zhao's central claim is that divergent capital environments and customer bases push US and Chinese firms toward different strategies rather than a single shared race. US firms sell capability as a product to customers who pay for it, while Chinese firms, facing scarce capital and price-sensitive customers, compete on efficiency and breadth of deployment. The article frames these as two visions of success that are measured on different metrics, so that one country can lead on capability while the other leads on economy-wide adoption.
Key claims
Capital gap
Zhao reports that US AI startups received roughly $109B in private investment in 2024, against roughly $9.3B for Chinese startups, a ratio of about 12:1. Chinese government funds only partially offset this gap: the article cites Big Fund III as having $47.5B registered, but notes a historical disbursement rate of about 50%. Chinese technology giants pursue scale through their own spending, with ByteDance at $21B and Alibaba at $53B, but Zhao argues the market is too fragmented for any single player to dominate.
Go-to-market models
US firms sell capability directly as a paid product, with examples including ChatGPT Plus at $20/mo and Copilot at $10/mo. In China, consumer AI is typically free: Baidu made Ernie 4.0 free, and ByteDance's Doubao has been free since launch, with monetization occurring instead through API access, cloud bundling, and advertising. Zhao contrasts software spending per employed person at $2,284 in the United States against $84 in China.
Constraint-driven Chinese strategy
The article identifies four strategies that Zhao attributes to operating under tighter capital and hardware constraints:
- Efficiency. DeepSeek-V3 activates only 37B of its 671B parameters per token. Zhao argues that when capital is scarce and customers are price-sensitive, efficiency expands the addressable market.
- Born-global consumer reach. MiniMax served more than 212M users across over 200 countries, with overseas markets accounting for more than 70% of revenue.
- Inference-optimized hardware. Zhao points to Biren's IPO as evidence that hardware startups can succeed by optimizing for inference rather than frontier training.
- Faster industrial deployment. The article reports that 67% of Chinese industrial firms have deployed AI in production, against 34% in the United States, roughly double.
Signals to watch
Zhao characterizes the United States as leading in frontier model capability by about a 7-month gap, while suggesting China may lead in economy-wide deployment. The article identifies inference pricing trends, revenue-mix shifts, and overseas revenue growth from Chinese consumer AI as the signals worth tracking to gauge how the two trajectories develop.
Relevance
- AI Diffusion — offers a detailed account of how AI diffuses differently under different economic constraints.
- Export Controls (AI) — Zhao frames China's constraint-driven efficiency, such as DeepSeek, as a consequence of export controls.
- Fast-Follow Problem — the cited ~7-month capability gap quantifies the current state of fast-following.
- Eight Worlds Framework — complicates a binary "racing vs. not racing" axis by describing China as racing on different metrics.
- Techno-Federalism — China's fragmented market and decentralized AI strategy bears on Wu's analysis.
- AI as Normal Technology — China's faster industrial deployment (67% vs. 34%) bears on the "slow diffusion" thesis, at least in manufacturing.