Underclassic

Have a lead? Contact us!

Breaking

Trump Administration Weighs Shaming Companies That Use Chinese Open-Weight AI: What We Know

Reuters reported Wednesday, citing three people, that Treasury and the White House science office have discussed making US companies post on their websites when they use Chinese AI models. It is a discussion, not a rule. Reuters described it as a “name and shame” approach within a push against Chinese open-weight models. Our analysis, labeled as opinion below: it is government pressure on open-weight and self-hosted AI, and it steers businesses toward a few closed US labs. Jump to our take.

November 18, 2026, 12:35 p.m. ET  ·  Underclassic

Updated November 18, 2026, 1:01 p.m. ET: added background on past US moves against open-weight models and a clearly labeled opinion section. The Reuters-based reporting above is unchanged: this is a proposal under discussion, per three unnamed sources cited by Reuters. It is not a rule, and no text has been published.

The short answer

What we know

What we don't know

Who is named

Why companies use these models: price

Reuters said Chinese open-weight models are gaining ground with US companies because they are cheaper than proprietary tools from OpenAI and Anthropic and “perform well enough.” Here is what a typical job costs at list prices, using the same 50,000-tokens-in, 5,000-tokens-out task on our API pricing table (checked Nov. 18):

ModelList price per 1M tokens (in / out)50K in + 5K out
Claude Fable 5.2$8 / $40$0.60
GPT-6.2 Astra$8 / $40$0.60
Qwen4-Max (Alibaba Cloud)$1.50 / $4.50$0.098
DeepSeek V4 Pro (peak-hour rate)$1.32 / $3.96$0.086

Illustrative, not what any named company pays: list prices for standard prompts under 200K tokens, from Anthropic, OpenAI, Alibaba Cloud (Singapore) and DeepSeek. DeepSeek's off-peak rate is half. Cost = 50,000 × input price + 5,000 × output price. The models that led OpenRouter's usage ranking in October were other, cheaper Chinese models. Qwen4-Max is a closed model sold only through Alibaba's API (no weights published, per The Next Web), so it is not an open-weight model; DeepSeek V4 Pro's weights are open. Reuters' report concerns Chinese models generally, with open-weight models the stated worry.

At those prices the task costs about six to seven times less on the two Chinese models. Whether a disclosure requirement would change that math is an open question; Reuters reported no estimate of its effect. Our open-weights tokens-versus-spend report has the usage and spending data on open models more broadly.

The government's case

The government's stated worries are not frivolous. Rep. Moolenaar says Chinese models “may have embedded security risks the CCP can exploit.” In September 2025 the Commerce Department's CAISI evaluated three DeepSeek models and reported that agents built on its most secure DeepSeek model (R1-0528) were 12 times more likely than evaluated US frontier models to follow malicious instructions, that the same model responded to 94% of overtly malicious requests under a common jailbreak technique (versus 8% for US reference models), and that DeepSeek models echoed four times as many misleading CCP narratives as US ones. Those were 2025 models, not today's. Reuters also reported that OSTP Director Michael Kratsios accused China in April of stealing US labs' intellectual property on an “industrial scale,” that Treasury Secretary Scott Bessent said in July the administration could sanction Chinese AI companies found guilty of IP theft through distillation, and that China hawks fear Chinese models will make US companies dependent on Beijing. Reuters also listed American open-weight options: Meta's Muse Glimmer, Nvidia's Nemotron and Reflection AI's Beam.

Context: what Washington has already tried on open weights

Our take: this is pressure on open-weight AI

Opinion and analysis. Everything above this box is reporting. Everything below it is Underclassic's argument, not Reuters', and not established fact. Where we rely on someone else's claim, we say whose.

Our read: this is the government leaning on companies to stop running open models, with a transparency label as the lever. Reuters itself calls the plan a “name and shame” approach meant to “discourage reliance” on Chinese models, and places it inside a push to combat the spread of Chinese open-weight models. There is no law, no vote and no text. Reuters said it was not clear whether it would even go through formal rulemaking. A label never has to ban anything to work. If a company expects a public warning on its website, and maybe a letter from a House committee, for using Qwen, Kimi or DeepSeek, its lawyers can move it to a closed API long before any court tests the policy. One source told Axios that what is actually happening is “slower and more durable” than a ban, citing procurement rules, Entity List threats and public pressure campaigns aimed at US companies using Chinese models. Another source said such approaches would have “a chilling effect” on Chinese open-source tech and the US companies that rely on it. A name-and-shame label fits that pattern.

Self-hosting is in the crosshairs. Airbnb's defense is the standard one: China-origin open-source models, run only through approved US-based providers, with data kept separate. Moolenaar told Reuters that risks remain even if US companies run Chinese models on their own infrastructure. We think that sentence matters. If downloading weights and running them on your own machines, with no data leaving the building, still counts as a risk, then the objection is not about where your data goes. It is about where the model came from, and about who is allowed to run which software. Reuters did not report whether a disclosure rule would reach self-hosters. But Moolenaar's statement tells you which way the concern runs.

Why open weights matter. They are the one kind of AI you can download, run offline and keep on your own hardware, with no vendor in the loop and no prompt leaving your network. Washington has said as much. NTIA said open models cut the need to share data with third parties and spread market control beyond a few large developers. The AI Action Plan said many businesses and governments “have sensitive data that they cannot send to closed model vendors.” Open weights are also testable by outsiders, as CAISI's DeepSeek evaluation shows. That does not make them safe. You cannot read a model's behavior off its weights the way you read source code, and the CAISI results show real weaknesses in the DeepSeek models it tested. But only open weights let anyone outside the lab check.

Centralization has a price tag. On our table above, the same 50K-in, 5K-out job costs about $0.60 on Claude Fable 5.2 or GPT-6.2 Astra and about $0.086 on DeepSeek V4 Pro, whose weights are open: roughly seven times more. Qwen4-Max, at $0.098, is closed, so the cleanest open-weight comparison is DeepSeek. Cheaper closed tiers exist: OpenAI's GPT-6.2 Sol, at $2 and $10 per million tokens in and out, would run $0.15, about 1.7 times DeepSeek's cost. These are list prices for API access; self-hosting adds hardware costs of its own. See the full API pricing table. Nudging companies off these models does not send them to American open ones by default. Axios reported that US alternatives are “still limited at best” compared with cheaper, more practical Chinese models. For many businesses the practical alternative is a closed API from a handful of labs such as OpenAI, Anthropic and Google.

Who benefits. We have not seen evidence in any report that OpenAI or Anthropic lobbied for this proposal, and we saw no comment from either on it. What is on the record: one source told Axios that leading AI labs or their allies approach the administration every three to five months with ideas to ban open-source models, and Axios quoted David Sacks, an outside White House AI adviser, writing on X in July that “the leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition.” Those are claims by an anonymous source and by Sacks, not findings. TechCrunch reported that OpenAI and Anthropic have urged the administration to respond to alleged IP theft by Chinese labs, and that neither signed the open-weight letter. Companies that did sign, such as Nvidia and Microsoft, also have commercial interests, as TechCrunch noted. Whoever pushes for it, the effect would be the same: fewer practical alternatives to a few closed labs, whose prices are several times higher.

The inconsistency. The same administration's Action Plan told the government to support open models. It also said they should be “founded on American values.” Axios reports that US companies are turning to Chinese open models because they are cheaper and “just about as good.” Backing open weights in principle while shaming the companies that use the cheapest ones is hard to square.

What we would want instead. If a specific model is unsafe, publish the testing, as CAISI did, and let companies decide with the evidence in hand. Keep security requirements for federal systems, where bills like H.R. 10670 are already pending. Do not treat a download as a confession. The question we keep returning to: should running open weights on your own machine ever need a label?

The backdrop

Reuters noted the consideration comes as revelations about Nvidia CEO Jensen Huang's affair with an alleged Chinese spy have fueled renewed calls in Washington for a tougher stance on China. We covered that story in what we know about Huang's resignation. The report also came out on X fast: @MTSlive's post summarizing it had more than 300,000 views by Wednesday afternoon.

Get the daily issue

One email a day. Unsubscribe anytime.

An Underclassic breaking story, updated as reporting develops. Daily briefing: underclassic.com. Related: AI API pricing per 1M tokens; open-source AI guide; open weights: tokens vs. spend. Sources: Reuters, Nov. 18, 2026 (read via Yahoo, The Globe and Mail, Bloomberg Law); House Select Committee on China on X; House Homeland Security Committee, April 29, 2026; July 31, 2026; letters to Airbnb, Anysphere and DoorDash; Reuters, Sept. 17, 2026; Ars Technica; OpenRouter rankings; Yicai Global; The Next Web on Qwen4-Max; NTIA open-weights report, July 2024; AI Diffusion rule; AI Action Plan; NIST CAISI on DeepSeek; S. 321; H.R. 10670; Axios, July 20, 2026; TechCrunch, July 24, 2026. Reuters' own page was not readable by our tools; every Reuters claim above is attributed and was checked against the syndicated full text.