United States Capitol, First Street Southeast, Washington, DC, USA
Cursor plans to explore dropping Chinese open-weight AI models because of “geopolitical risk”, a House investigation says. The SpaceX-owned coding platform built its flagship model, Composer 2, on Moonshot AI’s Kimi K2.5.
CNBC’s Kai Nicol-Schwarz first reported the plan from an advance copy of the report. The House Homeland Security Committee and the Select Committee on China released the joint report on Friday. It says Anysphere, Cursor’s developer, told them of its plans.
Anysphere “plans to explore phasing out PRC-developed open-weight models because of geopolitical risk,” the committees wrote.
The report describes an exploratory plan, with no decision or timetable. Cursor, SpaceX and DoorDash did not respond to CNBC’s request for comment, and Airbnb declined to comment.
Three companies under the lens
The committees opened their inquiry in April with letters to Anysphere and Airbnb, and added DoorDash in July. All three use Chinese models in different ways, the report found.
Airbnb runs customised versions of Alibaba’s Qwen on US servers. One powers an assistant handling about 80,000 guest and host conversations a day. Another summarises about 250,000 support cases and calls daily.
DoorDash reported 25 Chinese model families in real use, meaning each cost at least $5,000 to run. Swapping out one model built into a larger system could take up to six months. Most other uses could switch almost at once, it said.
For Cursor, the committees said the Chinese base became public only after an outside developer identified it. Cursor’s own technical report names Kimi K2.5 as the starting point. Cursor then trained it further on code. SpaceX completed its purchase of Cursor in August, at an implied $60bn equity value, according to CNBC.
Cheap, good and hard to trace
Cost, customisation and near-frontier performance drive adoption, the report found. The US lacks an open-weight rival at the right price. The report also says much of China’s progress rests on capabilities taken from American models. In September, OpenAI said Moonshot-linked users had tried to extract its AI reasoning. Hosting a model in the US keeps data out of China, the committees add. But it does not change how the model was trained.
The committees want to extend the federal DeepSeek ban to other Chinese developers. Their models would also be kept out of federal contract work and high-risk critical infrastructure. They also call for a model provenance standard and a national American open-weight initiative.
They warn against moving too fast. Forcing companies off Chinese models before trusted alternatives exist would raise costs, the report says. It would also hand the market to a few closed providers. It points to new open models from Reflection AI and Mistral as encouraging signs.
India’s sovereign AI lab sells DeepSeek
The same question is playing out in India. Sarvam, the startup backed by the government’s IndiaAI Mission to build a homegrown model, has begun selling access to DeepSeek and other Chinese models, Bloomberg’s Saritha Rai reported.
Its new Token Factory service hosts models from DeepSeek, Moonshot and Z.ai in India, alongside Sarvam’s own model and Google’s Gemma. It already serves more than 450bn tokens a day on DeepSeek V4.1 Flash, Sarvam said. Critics online asked why a lab built on subsidised compute was reselling Chinese models.
“We haven’t changed our motto. We continue to build AI models for India,” co-founder Pratyush Kumar told Bloomberg.
Co-founder Vivek Raghavan said the chips behind the service are “entirely funded by private money”. Keeping inference in India also keeps users’ data in the country, he added.
Sarvam has defenders. Nikhil Pahwa, MediaNama’s founder, told the site that Sarvam has to earn money in an open market. The criticism came days after Sarvam signed a deal with India’s defence ministry to build AI tools for the armed forces, MediaNama reported. The startup became India’s newest AI unicorn in June.
Get the TNW newsletter
Get the most important tech news in your inbox each week.