No matter how good the open-weight models from Chinese AI companies may be, Cursor wants to get rid of them. The SpaceX subsidiary sees geopolitical risks as a key motivation for reducing its dependence on them. This is despite the fact that its own Composer 2 would not have been possible without the Kimi K2.5 foundation model from the Chinese company Moonshot AI.
The report comes from the House Committee on Homeland Security and the House Select Committee on China and was shared with CNBC in advance. Cursor informed the committees that it wants to explore whether it can phase out the models. Airbnb and DoorDash also participated in the months-long investigation into Chinese AI models in the U.S. tech stack.
Earlier this year, it was already revealed that Cursor’s powerful coding model has Chinese origins. Composer 2 is largely based on Kimi K2.5, supplemented with reinforcement learning and coding-specific optimizations that draw on the platform’s proprietary data. According to Cursor, the model scored 61.3 on CursorBench, compared to 44.2 for the earlier Composer 1.5.
SpaceX announced the acquisition of Cursor in June, at a reported valuation of $60 billion. The deal was finalized in August. Cursor is now helping to train SpaceXAI’s Grok models.
Inexpensive and nearly at the frontier
The U.S. decision to use Chinese AI models is pragmatic. According to the report, it comes down to cost, adaptability, and performance that is close to the state of the art. The U.S. simply lacks an open-weight alternative that meets the price and performance demands of the market. The price difference is significant: In September, DeepSeek accounted for 58 percent of the tokens via Vercel’s AI Gateway, but only 6.6 percent of the spending. Airbnb specifically uses Alibaba’s Qwen for a customer service bot, while DoorDash deploys Kimi for internal tasks.
Recommendations and new rules
The committees propose a national U.S. open-weight initiative. This initiative would make public-sector computing power available, build a shared evaluation infrastructure, and consolidate access to U.S. frontier models for training. In addition, the report aims to build upon the ban on DeepSeek within federal civilian government systems announced in October. Chinese AI models would therefore no longer be permitted to run in U.S. government information systems or in the execution of federal contracts.
Washington is also moving toward digital autonomy for AI on other fronts. The Bureau of Industry and Security announced on Monday an interim final rule designed to prevent Chinese companies from gaining remote access to AI chips. This is a thorn in the side of export restrictions, as it is currently legal and easy for Chinese labs to use AI GPUs on a large scale, even though they are not actually allowed to import them into their own country.
See also: DeepSeek Super-Agent is an open-source alternative to Claude Code