November 22, 2026CodingAgents

Cursor Built Composer 2 on Kimi. Now Congress Is Watching.

The most popular AI coding tool in America runs its house model on a Chinese base, and now it's telling Congress it might stop. A joint report from the House Homeland Security Committee and the House Select Committee on China says Anysphere, which makes Cursor, "plans to explore phasing out PRC-developed open-weight models because of geopolitical risk." Composer 2, Cursor's flagship coding model, was built on Moonshot AI's Kimi K2.5. A timing note: CNBC broke this at 14:04 UTC Friday, about eight hours before our window opened. We're running it on Saturday's TNW follow-up, which added the India angle. Report: https://chinaselectcommittee.house.gov/media/reports/joint-investigation-report-china-is-trying-to-get-us-companies-addicted-to-chinese-ai-models

Read the verb carefully. "Explore" means there's no decision and no timetable, and Cursor, SpaceX and DoorDash didn't respond to CNBC. SpaceX bought Cursor at an implied $60 billion equity value, closed in August, and Cursor has helped train Grok. So there's an obvious in-house replacement, which makes "explore" sound a lot more like "probably." The committees also say the Chinese base only became public after an outside developer spotted it. Cursor's own technical report names Kimi K2.5 as the starting point. CNBC: https://www.cnbc.com/2026/11/20/cursor-chinese-ai-models-geopolitical-risk.html

Cursor isn't alone. Airbnb runs customized Qwen models on US servers, one of them handling about 80,000 guest and host conversations a day. DoorDash reported 25 Chinese model families in real use, each costing at least $5,000 to run, and said swapping one that's built into a larger system could take up to six months. The report's diagnosis is blunt: companies adopt these models for cost, customization and near-frontier performance, and the US has no open-weight alternative at the "price and performance the market demands." Its fixes: extend October's federal DeepSeek ban to all PRC-developed models in federal systems and contracts, create a model provenance standard, and launch a "national American open-weight initiative" with public compute, shared evals and "lawful access" to US frontier models for training.

Meanwhile India is going the other way. Sarvam, the lab backed by the IndiaAI Mission to build a sovereign model, now sells DeepSeek, Moonshot and Z.ai models through its Token Factory and says it already serves more than 450 billion tokens a day on DeepSeek V4.1 Flash. TNW: https://thenextweb.com/news/cursor-chinese-ai-models-house-report-sarvam-deepseek

The real story is that the base model is now a supply-chain item, like a chip or a crypto library, and it's turning into a disclosure question. If you fine-tune your agent's model on someone else's open weights, expect a customer, a regulator or a committee to ask whose. The report also concedes the uncomfortable part: the US doesn't have a cheaper answer yet. Banning the cheap option without building one just moves the cost onto everyone building agents.
← Previous
AWS Quick Puts the Math Outside the Agent
Next β†’
China Wants to Standardize the Token, Like a Kilowatt-Hour
← Back to all articles

Comments

Loading...
>_