Sunday, 22 November 2026
NEXA NEWSHD
Artificial Intelligence

SpaceX’s Cursor May Drop Chinese AI Models as Congress Urges Ban on PRC Models in Federal Systems

SpaceX’s artificial intelligence coding platform Cursor has told the United States Congress that it will explore phasing out Chinese-developed open-weight AI models from its products because of “geopolitical risk”, according to a joint investigative report released by two House committees on Friday, 20 November 2026.

The disclosure, first reported by CNBC from a copy of the report it saw before publication, places one of the most popular coding tools in the world at the centre of Washington’s growing campaign to pull Chinese AI out of the American technology stack. Cursor’s flagship coding model, Composer 2, was itself built on Kimi K2.5, a model developed by the Chinese laboratory Moonshot AI — a fact the report highlights as evidence of how deeply Chinese models have penetrated products that American developers use every day.

The report was released jointly by the House Committee on Homeland Security, chaired by Andrew R. Garbarino of New York, and the House Select Committee on China, chaired by John Moolenaar of Michigan. It follows a months-long investigation, launched in April 2026, into why American companies have adopted Chinese open-weight models, what risks that adoption creates, and whether the United States has a credible domestic alternative. Its answer, in the committees’ own words, is that the risks are ones that “American businesses, policymakers, and national security officials can no longer afford to overlook or downplay”.

Cursor has not announced a successor base model, and it has not set a timeline for any phase-out. The company, its parent Anysphere, and SpaceX — which completed its acquisition of Cursor in August — did not immediately respond to requests for comment from Reuters or CNBC, according to those outlets. That silence matters, because the report makes clear that the move under discussion is an exploration, not a completed decision. Cursor told the committees it would study how to phase the models out; it has not said it will do so, when it would do so, or what it would use instead.

What the report found

The committees’ investigation examined the provenance of Chinese open-weight models, the data and techniques used to train them, whether their capabilities were advanced through unauthorised distillation from leading American systems, and how they are being repackaged and deployed in the United States and around the world. Over seven months, the committees issued document requests, held company briefings, reviewed written responses, received classified briefings, and consulted frontier AI developers, cybersecurity executives, federal agencies, industry leaders and academic researchers.

The investigation focused on three American companies. Airbnb uses models developed by the Chinese e-commerce group Alibaba in its customer support operations. DoorDash reported that multiple families of Chinese-developed models had been authorised and deployed across its systems. And Anysphere, the company behind Cursor, built its flagship proprietary coding model on a Moonshot AI base model. The report says the committees engaged in productive discussions with all three companies about the technical, commercial and security considerations behind their choices.

The report sets out five findings. First, Chinese-developed open-weight models are already widely used across the American economy, but much of that use is hard to track. Second, cost, customisation and near-frontier performance are driving adoption, and the United States lacks a competitive open-weight alternative at the price and performance the market demands. Third, much of China’s apparent progress rests on capabilities extracted from American models, and American data-labelling firms are helping supply the data that trains Chinese models. Fourth, Beijing is using open-weight distribution, state support and low prices to expand global dependence on its models, while the terms of that dependence are beginning to change. Fifth, the risks are increasing as Chinese models become more capable, while their behaviour and security properties can differ in ways ordinary evaluations may not detect.

The scale of the dependence is striking. According to reporting on the investigation, Airbnb handles around 80,000 guest-host exchanges a day through Chinese AI. DoorDash estimated that replacing a single model would take about six months. Those figures explain why the committees describe any mandated removal as operationally disruptive rather than a simple policy switch — and why they pair their proposed restrictions with a plan to build an American alternative first.

Why Cursor matters

Cursor is not a marginal product. It is one of the most widely used AI coding assistants in the world, and its ownership now ties it directly to the centre of American power. SpaceX acquired Anysphere, the startup behind Cursor, in a $60 billion all-stock deal completed in August 2026, according to Reuters. SpaceX is developing its Grok family of AI models in collaboration with Cursor, which has been involved in their training. The company that builds the tool millions of developers rely on is now part of a defence-adjacent empire whose founder, Elon Musk, is one of the most politically connected figures in the country.

That makes the Cursor disclosure doubly significant. On one level, it is a commercial story about a product built on a Chinese foundation. On another, it is a political story about a company inside the SpaceX orbit telling Congress it is prepared to move away from Chinese technology. The report treats Cursor’s position as evidence that even companies with every commercial incentive to keep using cheap Chinese models are beginning to reckon with the geopolitical cost of doing so.

The economics, however, point the other way. Cursor’s Composer 2.5 coding model charges $0.50 per million input tokens and $2.50 per million output tokens, according to reporting on the company’s pricing. Anthropic’s Claude Opus 5.5, the kind of frontier American model Cursor would lean on if it stopped using Chinese foundations, is listed at $4 per million input tokens and $20 per million output tokens — eight times as much on output, the tokens the model actually writes. Other comparisons in the reporting are starker still: the Chinese open-weight model Qwen3 costs about $1.15 per million tokens against $30 for Claude Opus, a gap of roughly 26 times. Open weights publish a model’s parameters, so whoever hosts them captures the serving margin instead of paying a model developer a per-token fee. That is the product, not a discount.

The committees acknowledge the problem openly. Their report says the United States lacks a competitive open-weight alternative “at the price and performance the market demands”. Any phase-out that ignores that gap would raise costs for American companies and concentrate the market in the hands of a few proprietary American labs. That is why the report’s first recommendation is not a ban but a building programme.

The twelve recommendations

The chairmen put forward twelve policy recommendations. The first is to launch a national American open-weight AI initiative, pooling public computing resources, shared evaluation infrastructure and lawful access to leading American models for training. The second is to make trusted American AI the global default. The third is to establish a model provenance standard and an “AI bill of materials”, so that companies and agencies know which models are running in their systems.

The fourth recommendation is the one with the sharpest teeth: extend federal restrictions beyond DeepSeek, keep Chinese-developed AI models out of federal contract work and high-risk functions of critical infrastructure, and give agencies the visibility to enforce those restrictions. The committees want to build on an October order by the Department of Homeland Security that excluded DeepSeek from federal civilian agency systems, and to bar the use of federal cybersecurity grant funds to buy Chinese-developed models.

The remaining recommendations cover securing long-term information-sharing authorities and building a developer-level AI threat channel; evaluating Chinese models systematically and translating the results into guidance, with a single accountable federal partner for the information technology sector; preparing the Department of Homeland Security and the Cybersecurity and Infrastructure Security Agency to defend at machine speed and to secure AI agents on federal networks; making CISA the single point of submission for AI-related cyber incident reports and establishing an independent board to investigate significant incidents; holding illicit distillers accountable and closing the grey market; examining sales of American training data to Chinese AI laboratories; governing both closed and open models while avoiding overregulation; and working with allies to deepen understanding of Chinese strategy.

None of these recommendations is law. They are proposals from two committee chairmen, and the report itself notes that Congress has “too often been behind the curve on emerging technology”. But the direction of travel is unmistakable, and it is bipartisan in tone: the report frames the issue as national infrastructure, not trade policy, and draws an explicit parallel with Huawei and ZTE — the Chinese telecoms firms Washington spent years pushing out of American networks.

Distillation and the “stolen intelligence” charge

The most confrontational passage in the report concerns how Chinese laboratories are said to have built their capabilities. In the foreword, Chairmen Garbarino and Moolenaar write that federal investigations have found that major Chinese developers “used proxy networks and fraudulent accounts to harvest outputs from American frontier models and replicate their capabilities”. They add: “While America is investing billions and racing to build ‘Super Intelligence,’ Beijing is using stolen intelligence to close the gap.”

The claim is not new, but the report gives it the weight of a congressional finding. In September 2026, the National Security Agency, CISA and the Federal Bureau of Investigation released a joint advisory warning of the national security impacts of large-scale distillation efforts against American AI companies, particularly as China uses advanced AI models to increase the scale, scope and speed of its intrusions into American systems. Distillation — training a smaller model on the outputs of a larger one — is a legitimate technique in many contexts. The allegation here is that it was done at industrial scale, through deceptive means, in violation of the terms under which American models are offered.

The report also says American data-labelling firms are helping supply the data that trains Chinese models, and it recommends examining those sales. That detail matters because it shows the committees are looking at the whole supply chain, not just the models themselves: the data, the labelling, the training, the hosting and the deployment. The proposed “AI bill of materials” would make that chain visible for the first time.

There is an irony in the report’s own account of the investigation. In May, Anthropic provided committee members with one of the first live demonstrations of Mythos given to Congress, showing how advanced AI can find software vulnerabilities. The demonstration, the report says, reinforced the urgency of ensuring that federal agencies, including civilian cyber defenders, can responsibly use advanced American models to find and patch vulnerabilities. The same technology that Congress wants to keep out of Chinese hands is the technology it wants its own defenders to use.

Disclosure, not just prohibition

The Cursor report does not arrive in isolation. Reuters reported earlier in the same week that the Trump administration is considering requiring American companies to publicly disclose their use of Chinese AI models — a bid to discourage reliance on the technology without imposing new restrictions that could anger industry or Beijing. According to Reuters, the Treasury Department and the White House Office of Science and Technology Policy have discussed the plan, which would apply to companies such as Airbnb.

Disclosure and prohibition are two different tools, and the administration appears to be weighing both. A disclosure rule would name the companies using Chinese models and let market and political pressure do the work. The committees’ report goes further, urging outright exclusion from federal systems and federal contract work. The gap between the two approaches is a measure of how unsettled the policy still is — and of how much depends on whether an American open-weight alternative can actually be built.

The stakes are visible in the wider technology confrontation. The United States has tightened chip export controls, with new rules taking effect on Thursday, 20 November, that are disrupting semiconductor supply chains more broadly than expected and reaching mobile and consumer parts, according to Nikkei Asia. TSMC has paused certain shipments. A House committee has separately disclosed that nearly 900,000 advanced Nvidia and AMD chips were approved for Chinese buyers in the months to May 2026. And on Friday, 20 November, President Donald Trump told a White House press conference that he would impose a “tremendous tax” on American companies building data centres abroad. The Cursor report is one more front in the same campaign: a determination, across the executive branch and Congress, to draw a line between American and Chinese technology, and to make American companies choose a side.

Analysis: Why It Matters

The Cursor report is important not because a phase-out has happened — it has not — but because it exposes the central contradiction in Washington’s AI policy. The United States wants to keep Chinese AI out of its systems, and it wants American companies to stop using it. But the market has already made its choice, and it chose Chinese models for reasons that have nothing to do with ideology: they are cheap, they are customisable, and they are nearly as good as the best American systems. That is the committees’ own finding, and it is the most honest sentence in the report.

The price gap is the policy problem

Every policy in this space has to reckon with the arithmetic. A coding tool that runs on a Chinese base model can charge a fraction of what a tool built on a frontier American model must charge, because the Chinese weights are free to download and run. The gap is not a few per cent. On the figures reported this week, it is eightfold on output tokens for Cursor’s own product, and twenty-sixfold for the Qwen3 comparison. For a company serving millions of developers, that is the difference between a viable business and an unviable one.

This is why the report’s first recommendation matters more than its fourth. A ban on Chinese models in federal systems is enforceable by procurement rules. A ban on their use across the American economy is not, unless there is something comparably cheap to switch to. The committees propose a national American open-weight initiative — public computing power, shared evaluation infrastructure, lawful access to leading American models for training — precisely because they know that restriction without substitution will fail. The question is whether such an initiative can be built fast enough, and funded generously enough, to close a price gap that Chinese laboratories have spent years widening with state support.

There is a real risk that the policy backfires. If Chinese models are pushed out of American systems before a trusted alternative exists, costs will rise and the market will concentrate in the hands of a few proprietary American labs. That would make American AI more expensive for everyone, including the small businesses and developers who are the supposed beneficiaries of the policy. The report acknowledges this by recommending that the government “govern both closed and open models and avoid overregulation”. It is a warning against the very overreach the report’s own recommendations could produce.

A Huawei moment, but harder

The House committees draw an explicit parallel with Huawei and ZTE, and the comparison is instructive — mostly for the ways it does not hold. Huawei was a hardware problem. Its equipment sat in identifiable boxes in identifiable networks, and it could be ripped out and replaced, at great cost, by equipment from identifiable rivals. Chinese open-weight models are software. They are weights — files that can be downloaded, copied, fine-tuned and embedded inside other products. Cursor’s Composer 2 is built on Kimi K2.5, but the Kimi lineage is inside a proprietary model whose full contents are not visible to its users. Ripping that out means rebuilding the product, not swapping a box.

The enforcement problem is equally different. A model can be hosted anywhere, by anyone, and its provenance can be obscured. The report’s proposed “AI bill of materials” is an attempt to answer that: a declaration, for every system, of which models it contains and where they came from. But a bill of materials only works if companies tell the truth about it, and the history of supply-chain disclosure suggests they will not always do so. The committees know this, which is why they also call for systematic evaluation of Chinese models and a single accountable federal partner for the technology sector. Even so, the Huawei analogy may understate the difficulty by an order of magnitude.

There is also a geopolitical irony. The open-weight model ecosystem that Washington now fears was, in large part, built by the very openness the American technology industry championed. Open weights were supposed to democratise AI. They have — and the beneficiaries include Chinese laboratories that have used them, and the outputs of American frontier models, to close the gap with the leaders. The report’s charge that Beijing is “using stolen intelligence to close the gap” is, from Beijing’s perspective, simply the logic of open source working as designed. That does not make the distillation allegations untrue. It does mean the policy response has to distinguish between theft, which can be punished, and openness, which cannot be un-invented.

What to watch next

Three things will determine whether this report matters. The first is Cursor itself. The company has said it will explore a phase-out, and it has said nothing about what would replace Kimi K2.5 as the base of Composer 2, or when. If Cursor announces a successor — a Grok-based model, perhaps, given SpaceX’s collaboration with Cursor on the Grok family — that will be the moment the exploration becomes a decision. Until then, the disclosure is a signal, not a change.

The second is the Trump administration’s disclosure plan. If the Treasury Department and the White House Office of Science and Technology Policy go ahead with a rule requiring American companies to publicly disclose their use of Chinese AI models, the naming effect alone could accelerate the shift. Airbnb and DoorDash would have to say, in public, how much of their operations run on Chinese models. That is a different kind of pressure from a procurement ban, and it may prove more effective — or it may simply push the use of Chinese models further into the shadows, where it is harder to track.

The third is the national open-weight initiative. The committees propose pooling public computing resources and lawful access to leading American models for training. Whether Congress funds it, and whether the leading American labs cooperate, will decide whether there is ever a genuine alternative to the Chinese weights. Without it, the report’s restrictions will be a tax on American companies, paid in higher costs and lost competitiveness, with no compensating benefit. With it, the United States might finally have an open-weight ecosystem of its own — one that does not depend on Beijing’s generosity or its laboratories’ honesty.

The report ends with a warning that Congress risks being “light years behind” a technology advancing faster than legislation can follow. The chairmen write that they do not want to wait five years to discover that much of the world, and much of the United States, runs on AI shaped by the Chinese Communist Party, leaving policymakers with little choice but “another costly rip and replace effort”. That is a fair warning. But the report’s own evidence suggests the rip-and-replace has already begun, quietly, inside the products Americans use every day — and that the cost of reversing it will be measured not in the price of a network box, but in the price of every token a developer writes.

Sources

House Committee on Homeland Security — Chairmen Garbarino, Moolenaar Release Joint Investigative Report Urging Removal of PRC Open-Weight Models From Federal Systems, High-Risk Critical Infrastructure

Reuters — SpaceX’s Cursor weighs phasing out Chinese AI models over geopolitical risk, CNBC reports

TradingView / Seeking Alpha — SpaceX’s AI-coding platform Cursor considers phasing out use of Chinese models: report

AI Weekly — Cursor Tells House It May Drop Chinese Open-Weight Models

About the Author — Abdul Mannan

Leave a Reply

Your email address will not be published. Required fields are marked *