Anysphere, the developer of the Cursor AI coding platform, advised Home investigators it plans to discover phasing out Chinese language-developed open-weight AI fashions due to geopolitical threat, in response to a joint congressional report launched November 20, 2026.
The disclosure is an exploratory plan, not a accomplished phase-out, and the committees described no closing determination or timetable. The Home Committee on Homeland Safety and the Home Choose Committee on China printed the findings after an investigation that started in April 2026 and expanded from Airbnb and Anysphere to DoorDash on July 31, 2026.
Open-weight fashions have skilled parameters that may be downloaded, run, and modified independently of the unique developer, permitting derivatives to unfold with out a persevering with industrial relationship. The committees stated Composer 2, Cursor’s flagship proprietary mannequin, was constructed on a Moonshot AI base whose identification grew to become public solely after an impartial developer recognized it.
Cursor’s technical report identifies that base as Kimi K2.5. Composer 2 underwent continued pretraining on a code-heavy knowledge combine, adopted by large-scale reinforcement studying in life like Cursor periods utilizing the identical instruments and harness because the deployed mannequin.
Why Chinese language Fashions Grew to become Embedded
Airbnb reported that personalized variations of Alibaba’s Qwen fashions help an AI assistant dealing with roughly 80,000 visitor and host interactions every day, alongside a fine-tuned mannequin summarizing roughly 250,000 help circumstances and calls every day. Each run on U.S.-based service-provider infrastructure.
DoorDash reported 25 Chinese language mannequin households approved and deployed in a non-trivial method, outlined as incurring at the very least $5,000 in computing and serving prices. It estimated that changing one Chinese language-developed mannequin embedded in a bigger system might take as much as six months, though most different makes use of might transition nearly instantly.
The committees recognized price, customization, and near-frontier efficiency as adoption drivers, and located that america lacks a aggressive open-weight different on the worth and efficiency the market calls for.
The report compares printed API costs for a million enter tokens and a million output tokens (small items of phrases and characters used to measure mannequin utilization):
| Mannequin | Developer | Approximate Value |
|---|---|---|
| Qwen3-235B-A22B-Instruct | Alibaba | $1.15 |
| DeepSeek-R1 | DeepSeek | $2.74 |
| Kimi K2.6 | Moonshot AI | $4.95 |
| Claude Opus 4.8 | Anthropic | $30 |
| GPT-5.5 | OpenAI | $35 |
These figures don’t modify for variations in functionality, caching, batch reductions, promotions, or self-hosting prices. They illustrate the worth stress behind adoption, whereas open weights additionally let corporations fine-tune fashions on their very own knowledge and deploy them on infrastructure they management.
The safety concern is separate from internet hosting location. Operating fashions on U.S. infrastructure reduces direct knowledge publicity, however the report says that doesn’t change how the underlying mannequin was skilled or the way it might behave underneath circumstances atypical evaluations miss.
Proposed Restrictions and Alternative Dangers
The report’s 12 suggestions don’t themselves enact restrictions. They might construct on a Division of Homeland Safety order excluding DeepSeek services and products from federal civilian procurement and requiring their elimination from company programs, together with third-party merchandise constructed on DeepSeek. The chairmen advocate extending these restrictions to different Chinese language-developed fashions and their derivatives throughout federal data programs and federal contract work, with slender waivers for analysis, testing, and defensive safety. Additionally they advocate barring these fashions and derivatives from narrowly outlined high-risk critical-infrastructure capabilities, with elimination schedules tied to trusted options and momentary exceptions the place no trusted substitute exists.
To make the principles enforceable, a model-provenance customary led by the Nationwide Institute of Requirements and Know-how and a machine-readable AI invoice of supplies would doc base fashions, vital fine-tuning, and distillation sources. The requirement would cowl federal contractors and AI utilized in high-risk critical-infrastructure capabilities, and can be voluntary elsewhere. A nationwide American open-weight AI initiative would mix public computing sources, shared analysis infrastructure, voluntary licensing of American frontier-model outputs to certified open-weight builders for coaching, and federal anchor demand.
The committees warn that eradicating Chinese language fashions earlier than trusted open-weight options exist might elevate prices and focus the market amongst closed proprietary suppliers. Anysphere’s exploratory plan subsequently illustrates the bigger implementation check: restrictions want dependable lineage data to find embedded fashions and credible substitutes to keep away from buying and selling Chinese language open weights for larger prices or closed-model dependence.

