Two committees of the US House of Representatives published a 50-page investigative report on Chinese open-weight models in American companies on November 20, 2026. The Committee on Homeland Security, chaired by Andrew Garbarino, and the Select Committee on China, chaired by John Moolenaar, recommend keeping AI models developed in China out of federal systems, federal contract work and high-risk functions of critical infrastructure. The case studies are Airbnb, DoorDash and Anysphere, the company behind the Cursor coding platform. The report is a political document from the Republican committee chairs, not a law. But it shows how deeply models from Alibaba, DeepSeek, Moonshot and Zhipu are already embedded in Western products, often without customers knowing.
What the report finds about Chinese open-weight models
The investigation began in April 2026. The committees requested documents, received company briefings, held classified briefings with agencies and reviewed public research. According to the Homeland Security Committee’s announcement, it produced five main findings:
- Chinese open-weight models are already widely used across the US economy, but much of that use is hard to track.
- Cost, customization and near-frontier performance drive adoption, and the US lacks a competitive open-weight alternative.
- Much of China’s apparent progress rests on capabilities extracted from American models, and US data-labeling firms help supply training data to Chinese labs.
- Beijing uses open-weight distribution, state support and low prices to build global dependence, while the terms of that dependence are starting to change.
- Risks grow as the models become more capable, while their behavior and security properties can differ in ways ordinary evaluations may not detect.
On adoption, the report cites third-party figures. On the model marketplace OpenRouter, the share of tokens from US-originating requests sent to open-weight models rose from 26 percent in September 2025 to 60 percent in August 2026, with Chinese labs accounting for most of that volume. Chinese models accounted for 41 percent of downloads on Hugging Face over the preceding year. Alibaba’s Qwen family alone has more than 151,000 derivative models on the platform.
Airbnb, DoorDash and Cursor as case studies
The report is most concrete about the three companies that handed over material. According to the full report (PDF), Airbnb runs customized versions of Alibaba’s Qwen models on US-based service provider infrastructure. They power customer support, including an AI assistant handling roughly 80,000 guest and host interactions per day and a fine-tuned model that summarizes around 250,000 support cases and calls daily.
DoorDash reported that its engineers had downloaded or evaluated models from at least six Chinese developers. 25 Chinese model families had been authorized and deployed in a non-trivial manner, defined as at least $5,000 in compute and serving costs. Replacing one model embedded in a larger system could take up to six months, while most other uses could be switched almost immediately.
The Anysphere case will interest developers most. The company built Composer 2, the flagship proprietary model in Cursor, on a Moonshot AI base model. According to the report, this only became public after an independent developer identified the underlying model. The report’s table lists Kimi K2.5 as the base. For the committees, this illustrates a broader problem: fine-tuning, continued pretraining, merging and distillation can blur a model’s origin until customers can no longer tell what a product is built on.
For context: according to the report, neither Airbnb nor DoorDash paid a Chinese developer or held a commercial agreement with one. Both run the models on US infrastructure, which substantially reduces the risk of data flowing directly to China. DoorDash also blocks connections to Chinese AI services. The committees explicitly describe their talks with the companies as “productive” and do not accuse them of wrongdoing.
Price as the main driver
The report backs up the adoption argument with published API prices. For one million input plus one million output tokens, Qwen3-235B-A22B-Instruct costs about $1.15, DeepSeek-R1 about $2.74 and Kimi K2.6 about $4.95. The same workload costs about $30 with Claude Opus 4.8 and $35 with GPT-5.5. The committees concede that these figures do not adjust for capability, caching, discounts or self-hosting. In an independent ranking of open models, the best American open-weight model placed only tenth, behind six Chinese ones.
We have covered several of these models, including Alibaba’s Qwen4-Max and Moonshot’s Kimi K3.1 Pro. The price gap has been one of their strongest selling points for months.
Distillation, military links and security questions
The tone sharpens when it comes to where the capabilities came from. US agencies assessed that six leading Chinese developers ran “industrial-scale” distillation campaigns against American frontier models, likely with Chinese government awareness. The committees rely on a joint advisory from the NSA, CISA and FBI published in September. How such attacks work and how providers defend against them is described in an OpenAI report released the same day, which we summarize in our article on adversarial distillation.
The Select Committee on China also highlights findings on military ties. Personnel associated with the 1st Bureau of the People’s Liberation Army Cyberspace Force had visited Zhipu’s Beijing headquarters on multiple occasions, most recently in August 2026. PLA procurement records also show DeepSeek, Qwen and models from Moonshot and MiniMax being built into military cyber ranges and agent systems. These claims cannot currently be verified independently; they come from the committees’ own review.
On technical risks, the report points to researchers who found that some Chinese models produced less secure code when prompts referenced politically sensitive topics or a US government user. For Zhipu’s newest open model, NIST and a US developer found that it can autonomously build cyber exploits. Its safeguards could be reduced from roughly 95 percent refusal rates to between 3 and 6 percent for an estimated $4,400. That is a general problem with open weights: whoever holds the weights can train the safety layer away.
Twelve recommendations for Congress and agencies
The report closes with twelve recommendations. The key ones:
- a national American open-weight AI initiative so companies have an affordable domestic alternative,
- a model provenance standard and an “AI bill of materials”, similar to a software bill of materials,
- extending the existing DeepSeek ban in federal civilian agencies to all China-developed models, including federal contract work and high-risk critical infrastructure functions,
- barring federal cybersecurity grant funds from being used to buy such models,
- holding illicit distillers accountable and examining sales of American training data to Chinese labs,
- making CISA the single point of submission for AI-related cyber incident reports.
Notably, the report also calls for governing both closed and open models while “avoiding overregulation”. None of the twelve recommendations calls for a general ban on Chinese models in private companies.
What it means for developers
Nothing changes immediately. These are recommendations that would only take effect through legislation or agency orders, and they do not apply to companies outside the US. The topic still matters for anyone supplying US agencies or critical infrastructure operators, and for anyone buying AI products.
The sober core of the report is transparency. Whether a model comes from China, the US or Europe matters less for privacy and security than where it runs, who controls the weights and how it was tested. Those details are often missing today. Teams that build AI features into their products should document where their models come from, inventory self-hosted weights like any other dependency and run their own security tests rather than relying on leaderboards. A mandatory model bill of materials, as the report proposes, would not be a bad idea in Europe either, regardless of which country a model comes from.



