{"slug": "washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are", "title": "Washington threatens to sanction Chinese AI models but American startups are already switching sides", "summary": "The Trump administration is threatening sanctions against Chinese AI labs over alleged model theft, but American startups are already switching to cheaper Chinese models. Treasury Secretary Scott Bessent said on July 21 that the administration found 'watermarks of our U.S. large language models on many of the Chinese models' and could use sanctions if theft is confirmed. Meanwhile, KuCoin data shows the share of tokens used by US companies for Chinese models rose from under 5% at the start of 2025 to 46% in April 2026, driven by pricing like DeepSeek V4 Flash at $0.14 per million tokens versus $5 for GPT-5.5.", "body_md": "*Washington is threatening Chinese AI labs over alleged model theft, but the harder problem is already in your invoices: cheaper Chinese models have become useful enough for American companies to route real work through them.*\n\nThe warning landed on July 21, and it was not vague. Treasury Secretary Scott Bessent told Fox Business that the Trump administration had found \"watermarks of our U.S. large language models on many of the Chinese models\" and said officials would examine the issue in the coming days or weeks, Bloomberg Law reported. He said Washington could use sanctions if foreign models were found to be stealing from American companies. That is the threat. The target list is messier.\n\nSome reports have framed the pressure as aimed at Chinese labs including DeepSeek, Moonshot AI and MiniMax, because those are the companies Anthropic named in February. Bessent's public comments, as reported by Bloomberg Law, Quartz, EFE and MeriTalk, did not appear to name those companies himself. That distinction matters. If you're going to punish a company for alleged theft, you need the allegation, the evidence and the responsible party lined up in public, not blurred together because the politics are convenient.\n\nAnthropic did make a specific accusation. In a February 23 post, the company said DeepSeek, Moonshot and MiniMax created about 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude, in violation of its terms of service and regional access restrictions. The company called those campaigns industrial-scale distillation attacks. Distillation itself is not exotic. A smaller model learns from the outputs of a stronger one. Used on your own models, it is routine. Used through fake accounts to copy a competitor's capabilities, it becomes a legal and commercial fight.\n\nThe White House sharpened that fight on July 22. Michael Kratsios, director of the Office of Science and Technology Policy, wrote that Moonshot AI had distilled Anthropic's Fable model for Kimi K3 and had built an internal platform to run large-scale distillation against US models while switching access methods to avoid detection, according to reports from Business Insider and Japan Times. Kratsios also alleged that Moonshot acquired Nvidia GB300-equipped servers and accessed GB300s in Thailand, hardware from Nvidia's Blackwell generation that US rules bar from sale to Chinese companies. Moonshot had not publicly answered those claims in the reports reviewed. The evidence has to travel.\n\n## Price did the work\n\nWashington's problem is that the market moved before the enforcement case did. KuCoin, citing OpenRouter data, said the share of tokens used by US companies for Chinese models rose from under 5% at the start of 2025 to 46% in April 2026, with DeepSeek at 17.6% of token traffic on the platform. That is not a press release win. It is usage.\n\nThe reason is plain. DeepSeek V4 Flash was listed at $0.14 per million tokens in the KuCoin summary, compared with $5 per million input tokens for GPT-5.5. Kimi K3's output pricing has been reported around $15 per million tokens, below Anthropic's higher-end Fable pricing in several industry comparisons. You do not need to love Beijing's industrial policy to understand why a founder with a monthly inference bill looks twice at that spread.\n\nLindy is the example to watch. The New Stack reported in June that the AI agent startup moved all of its AI agent traffic from Anthropic to DeepSeek v4 and said it was saving millions on inference. A Pondero AI brief, citing Lindy's migration writeup, put the cost reduction at roughly 90%. That is not theory. For customer support and email workflows - internal automation too - the cheaper model doesn't have to win every benchmark. It has to clear the bar your users can actually feel.\n\nDoorDash has made a similar calculation on narrower work. Fortune reported that the company was pushing lower-level tasks to Moonshot's Kimi model, with chief technology officer Andy Fang pointing to better quality at lower cost. One company can be an anecdote. Two companies plus OpenRouter token share start to look like a pattern.\n\n## Sanctions have teeth, but files move\n\nEntity List designations would still hurt. US companies would have to stop doing business with listed firms unless they received licenses, and cloud providers would face tighter compliance pressure. If your startup has quietly routed part of its product through a Chinese API because the bill is lower, you should assume procurement and legal teams will ask harder questions now - and so will security. No ban changes that.\n\nBut open-weight models are not ordinary vendors. Once weights are released, they can be downloaded, mirrored and run on infrastructure far away from the company that trained them. You can sanction Moonshot as a business. You cannot easily claw back a model file already running on a server in Singapore, Frankfurt or Virginia. That is the hard part.\n\nJensen Huang has been unusually direct on this point. Barron's reported that the Nvidia CEO supported open-weight Chinese models and joined other major technology companies in urging policymakers to avoid broad restrictions while targeting unlawful misuse. Implicator.ai reported that Huang called Chinese models including DeepSeek and Kimi K3 excellent, and argued that learning from AI is part of how the field advances. His commercial interest is obvious: more model use means more demand for Nvidia chips. His technical point still has force.\n\nThe right line is narrow, and Washington has not yet shown the public evidence needed to draw it cleanly. Anthropic's February claims are specific. Kratsios's Moonshot allegation is specific. But Bessent's sanctions threat is the one that could do real damage - and it needs fingerprints attached to it. If the US turns every cheap Chinese open-weight model into a theft case without showing those fingerprints, it gives American labs a protectionist tool and calls it national security.\n\nFrankly, that will not be enough for the startups already doing the math. If DeepSeek cuts a bill by 90%, you need more than a stern television interview to move the workload back. You need hard evidence and enforceable rules - and a domestic model market that does not punish customers for caring about cost.\n\nThe US-China AI talks expected in September may slow the first sanctions move, especially with Bessent set to lead the US side, according to reporting cited by Implicator.ai. The companies using these models should not wait for that calendar. The practical question is already on the table: which of your AI workloads can survive a forced provider change, and which ones are now tied to a model Washington may decide you cannot use?\n\n**Also read:** [CME launches cash-settled single-stock futures on Nvidia, SpaceX and 50 other US giants](https://startupfortune.com/cme-launches-cash-settled-single-stock-futures-on-nvidia-spacex-and-50-other-us-giants/) and [Alphabet's first negative free cash flow in 22 years signals the end of Big Tech's AI spending blank check](https://startupfortune.com/alphabets-first-negative-free-cash-flow-in-22-years-signals-the-end-of-big-techs-ai-spending-blank-check/) and [Andrew Ng's free graph course reignites a war over how to build AI agents](https://startupfortune.com/andrew-ngs-free-graph-course-reignites-a-war-over-how-to-build-ai-agents/)", "url": "https://wpnews.pro/news/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are", "canonical_source": "https://startupfortune.com/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are-already-switching-sides/", "published_at": "2026-07-26 20:07:32+00:00", "updated_at": "2026-07-26 20:23:27.170497+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-policy", "large-language-models", "ai-ethics"], "entities": ["Scott Bessent", "Anthropic", "DeepSeek", "Moonshot AI", "MiniMax", "Michael Kratsios", "Lindy", "DoorDash"], "alternates": {"html": "https://wpnews.pro/news/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are", "markdown": "https://wpnews.pro/news/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are.md", "text": "https://wpnews.pro/news/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are.txt", "jsonld": "https://wpnews.pro/news/washington-threatens-to-sanction-chinese-ai-models-but-american-startups-are.jsonld"}}