The Latest in U.S.-China AI Competition

Council on Foreign Relations  |  Michael Froman, Chris McGuire

Chinese artificial intelligence firm Moonshot AI released its Kimi K3 open-weight model on July 16, 2026, directly challenging American technological dominance in frontier software development. A joint U.S.-UK government assessment indicates the model performs comparably to Anthropic's Claude Fable but remains approximately six months behind in critical cyber capabilities.

To achieve this rapid advancement, the Beijing-based developer utilized banned U.S. hardware in Thailand and executed large-scale distillation attacks to extract proprietary training data from leading American laboratories. Although the model is free to download, its massive size requires expensive cloud-hosted U.S. chips to run, highlighting China's persistent reliance on Western computing infrastructure. In response, Washington is considering Information and Communications Technology and Services regulations to block domestic businesses and cloud providers from hosting or transacting with Chinese AI entities. Ultimately, restricting access to American hardware remains the primary mechanism to widen the technological gap and secure sufficient time to establish robust safety standards.

Comment

The deployment of distillation attacks to train Moonshot AI's Kimi K3 model exposes a structural vulnerability in the API-based delivery of frontier systems like Anthropic's Claude Fable. By querying these public interfaces, Chinese developers can bypass the capital-intensive pre-training phase. This subsidises domestic development using American capital and reduces the efficacy of physical export controls on Nvidia graphics processing units. Consequently, the technological frontier between Washington and Beijing becomes highly porous, transforming proprietary software advantages into temporary leads.

To read in details click here