Cursor tells Congress it may drop Chinese open-weight models — after building its flagship coder on one
Cursor told two House committees it will explore phasing out Chinese-developed open-weight models over geopolitical risk — an awkward disclosure, given its Composer 2 coding model was built on Kimi K2.5 from Moonshot AI.
- A review, not a decision: Cursor said it will “explore phasing out” models developed in the People’s Republic of China — no commitment, no timetable, no replacement announced.
- Composer 2 runs on Kimi K2.5: Cursor built its flagship coding model directly on Moonshot AI’s weights, then added its own code-specific training on top.
- The disclosure came in a 50-page House report: a joint investigation by the Homeland Security Committee and the Select Committee on China, begun in April, that also examined Airbnb and DoorDash.
- The price gap explains the adoption: the report’s pricing data puts Alibaba’s Qwen3 at $1.15 per million tokens against $30 for Claude Opus — a 26-fold difference.
- The remedy is to build, not just ban: the report’s headline recommendation is a national American open-weight initiative, plus extending October’s DeepSeek federal ban to Chinese-developed models across federal systems and contracts.

The company behind one of the most popular AI coding tools in the world just told Congress it is considering walking away from the models its flagship product is built on. Cursor, the AI code editor owned by SpaceX, told two House committees it will explore phasing out Chinese-developed open-weight models because of geopolitical risk, according to a joint committee report released on Friday and first seen by CNBC.
The disclosure is awkward in a way that is hard to overstate. Cursor’s Composer 2 coding model was built on Kimi K2.5, a model from the Chinese lab Moonshot AI. In other words, the product at the center of the investigation is itself the example the investigators are worried about.
What Cursor told Congress#
The commitment, such as it is, is deliberately narrow. Cursor said it would explore a phaseout of models developed in the People’s Republic of China — it did not commit to one, set no timetable, and has announced no replacement base model. The distinction matters: this is a company telling lawmakers it will study an exit, not a company announcing one.
The disclosure came in a 50-page joint report from the House Committee on Homeland Security and the House Select Committee on China, the product of a bipartisan investigation that began in April into why American companies were building on Chinese foundation models. The committees, chaired by Andrew Garbarino and John Moolenaar, examined Cursor’s parent company Anysphere alongside Airbnb and DoorDash, and the report frames the issue as a supply-chain and infrastructure risk — drawing an explicit parallel with Huawei and ZTE — rather than a data-privacy complaint.
Twelve recommendations follow, and the report is blunt about the stakes. Its preface accuses Beijing of using stolen intelligence to close the AI gap, and one section warns that Chinese models are producing less secure code and autonomously building cyber exploits. The committees say the risks are ones American businesses and policymakers can no longer afford to overlook or downplay.
The awkward part: Composer 2 runs on Kimi K2.5#
Cursor’s problem is that the dependence is not incidental. Composer 2, the company’s flagship programming model, was built directly on Kimi K2.5: Moonshot’s weights went in first, and Cursor’s code-specific pretraining and reinforcement learning came after. A coding product sold to American developers as a frontier tool starts life as a Chinese lab’s model.

SpaceX completed its acquisition of Cursor in August, a deal CNBC valued at an implied $60 billion in equity. That puts a defense-adjacent owner behind a product whose core model comes from a lab in Beijing — a combination the committees clearly had in mind. The report notes that the ownership layer changes the risk calculus, even as Cursor’s own statement stops short of any promise.
None of this is secret — the Kimi K2.5 lineage is documented — but the report makes it a matter of congressional record, and it puts Cursor in the position of defending a supply chain it is now publicly reconsidering.
Why American companies picked Chinese models#
The committees do not pretend the choice was mysterious. Their investigation found three forces driving adoption: cost, the ability to customize open weights, and performance close to the frontier. The price gap is the sharpest illustration. The report’s pricing data puts Alibaba’s Qwen3 at $1.15 per million tokens against $30 for Claude Opus — a 26-fold difference that, at enterprise scale, is not a rounding error but a budget line.
Airbnb and DoorDash show what that gap buys in practice. Airbnb runs about 80,000 guest-host exchanges a day through Chinese AI, according to the report’s summary, and DoorDash told investigators it would need about six months to replace a single embedded model. Ripping a model out of a live product is measured in quarters, not days.
And the committees concede the market has a point. Their report says the United States lacks a competitive open-weight alternative at the price and performance the market demands. The dependence, in their telling, is a failure of supply as much as a failure of judgment.
What the House wants instead#
The report’s answer is to build the alternative rather than only ban the incumbent. Its headline recommendation is a national American open-weight initiative that pools public computing resources, shared infrastructure for evaluating models, and lawful access to leading American models for training — a deliberate attempt to close the price-performance gap that drove companies to Chinese weights in the first place.

On restrictions, the committees want to widen an order from October that barred DeepSeek from federal civilian agency systems, extending it to Chinese-developed models across federal information systems and work done under federal contracts. Other recommendations include model provenance standards — an AI component list disclosing which base model a product is built on — sanctions on illicit model distillers, and a mandate for the Department of Homeland Security and the cybersecurity agency CISA to move at AI speed on defense.
None of this is law yet. The report is advisory, its proposals are framed for legislative vehicles in early 2027, and it does not propose banning Chinese models from general corporate use. What it does is put a price on the status quo: keep building on Chinese weights, and expect the question to keep coming from Congress.
What to watch#
First, whether Cursor names a successor. The company has said nothing about what would replace Kimi K2.5 as Composer 2’s base, and there is no American open-weight model today that matches the price and performance the report itself says the market demands. A phaseout without a replacement is a promise without a plan.
Second, whether the report’s recommendations become law. The committees plan to advance them through legislation in early 2027, and the test will be whether the open-weight initiative gets funded — or whether only the restrictions survive.
Third, the rest of the industry. Airbnb and DoorDash were named in the same investigation, and both face the same arithmetic: cheap, capable models on one side, congressional scrutiny on the other. Cursor is the first company to tell the House it is looking for the exit. It is unlikely to be the last one asked.
The week began with American companies treating Chinese open-weight models as the cheapest path to frontier performance. It ends with Congress treating that path as a dependency to be managed — and the most popular AI coding tool in the world caught in the middle, built on the very thing it now says it may leave behind.
Sources: AI Weekly, “Cursor Tells House It May Drop Chinese Open-Weight Models” (21 Nov 2026); AI Affairs, “Cursor tells House committees it may phase out Chinese-developed AI models” (20 Nov 2026); CASRAI, “House Flags Chinese AI Models: Cursor Weighs Shift” (20 Nov 2026); Seeking Alpha via TradingView, “SpaceX’s AI-coding platform Cursor considers phasing out use of Chinese models: report” (20 Nov 2026).