LLM: The U.S.–China AI Model Race Is Not Zero-Sum

Since we published our previous article comparing Chinese and U.S. foundation models, many readers have written in asking how to interpret the recent wave of Chinese open-source model developments, and whether a second “DeepSeek moment” could emerge. In response, we conducted a deep dive into this question.

The U.S. and Chinese model ecosystems are currently forming a division of labor along distinct value tiers, rather than a zero-sum substitution. U.S. models, leveraging advantages in deep reasoning, security compliance, and enterprise-grade stickiness, have secured high-ARPU knowledge work scenarios — Anthropic’s annualized revenue has reached $14 billion, OpenAI has exceeded $22 billion, growth continues to accelerate, and the vast majority of revenue derives from enterprise customers. Chinese models, meanwhile, are leveraging extreme price-performance ratios to lower the threshold for model invocation, incorporating large volumes of previously cost-prohibitive scenarios into the total addressable market and driving growth through volume and coverage. The fundamental nature of this competition is not mutual displacement, but rather a joint expansion of the global AI application market.