The release of Moonshot AI’s Kimi K3, an open-weight model that benchmarks show performing at or above the level of Anthropic’s and OpenAI’s latest proprietary systems, has reshaped the US-China AI contest. Within days, Moonshot halted new user sign-ups, citing insufficient computing capacity, while shares of publicly traded Chinese AI labs Zhipu and MiniMax fell more than 40% and 50% respectively over the past month, as investors absorbed the economics of a business that gives away its core asset and then struggles to monetise inference.

The model’s performance has drawn a sharp response from Washington. US Treasury Secretary Scott Bessent called it “unacceptable” that Chinese open-source models may have been built on “IP theft” through distillation, a technique that extracts knowledge from a larger model. He said the administration is examining whether to sanction overseas models that steal from US companies, noting that watermarks from American large language models have been found on Chinese systems. Anthropic had previously accused Moonshot and others of accessing its Claude model at scale to train competitors.

In Beijing, the calculus is different. At the World Artificial Intelligence Conference in Shanghai, President Xi Jinping made his first appearance at the event, calling for open-source AI development and announcing a 29-country international cooperation body. Yet Chinese regulators are simultaneously consulting leading firms—ByteDance, Alibaba, Huawei, Zhipu—on tightening export controls to prevent advanced model weights, training data, and chip designs from falling into Western hands, according to the Financial Times. The proposed restrictions would block foreign acquisitions of Chinese AI startups and limit overseas semiconductor foundries from manufacturing chips based on Chinese designs.

The business pressures on open-weight labs are acute. Unlike open-source software, where support and services can generate revenue, open-weight models incur heavy per-query compute costs that typically flow to cloud providers, not the model creator. William Blair analyst Arjun Bhatia noted that inference workloads gravitate to whoever operates infrastructure most efficiently, “and this is usually not the model provider.” Zhipu lost nearly $500 million on $107 million in revenue last year; MiniMax lost $250 million on $79 million. The next milestone: US-China AI talks, led by Bessent, are expected in September before President Xi’s scheduled visit to the US on 24 September, where both sides will attempt to address the risks posed by increasingly powerful models.