China’s Low-Cost Open-Weight AI Challenges U.S. Lead in Agentic Systems

China’s rapid advance in agentic AI stems from state-supported distillation techniques that create low-cost, open-weight models such as DeepSeek, which U.S. companies allege were trained on stolen outputs from OpenAI and Anthropic. The United States has focused on large-scale defense contracts, signing roughly $200 million deals with Anthropic, Google, OpenAI and xAI, yet disagreements over unrestricted deployment have limited adoption, leading to a blacklist of Anthropic.

Training expenses for leading U.S. models now run into hundreds of millions of dollars, while Chinese models can be trained for a few million and priced dramatically lower per token. Both sides view the competition as a battle for global ecosystem dominance, with Washington promoting the American AI Exports Program to secure hardware, software and standards abroad. Analysts note that while Chinese models match U.S. performance on many benchmarks, they lag on complex tasks such as multi-step cyber-attack simulations. The outcome will hinge on cost, openness, and trust in each nation’s AI supply chain.