US Weighs New Curbs on Chinese Open-Source AI Models

The United States is reportedly planning new restrictions targeting Chinese open-source artificial intelligence models, escalating the technology competition between the world’s two largest economies. According to Semafor, US officials are considering designating Chinese models as a supply-chain risk, limiting their use by companies doing business with the US government, while also exploring tighter export controls to further constrain compute capacity in China.

The move comes at a pivotal moment. Chinese open-source AI models have surged from just 1.2% to nearly 30% of global usage in a matter of months, with the top four most-used models on the OpenRouter platform now coming from Chinese companies: DeepSeek, MiniMax, Tencent, and Xiaomi, as Fortune reported.

The Anthropic Precedent

The new restrictions would follow an unprecedented US government action in June. On June 12, 2026, the Commerce Department under Secretary Howard Lutnick issued an Is-Informed Letter (IIL) to Anthropic, ordering the company to obtain a license before any export of its most powerful models — Mythos 5 and Fable 5 — to any foreign person worldwide. According to a Mayer Brown legal analysis, this marked the first time the US government has treated an AI model itself as controlled technology under export regulations.

Anthropic responded by disabling both models for all users worldwide, stating it was not technically feasible to block access only by foreign persons on short notice. The move had immediate and far-reaching consequences.

“It is the first time that a government has ordered a model developer to restrict access to a particular model based on nationality,” Paul Triolo, a partner at DGA-Albright Stonebridge Group, told Fortune. “Companies and governments will start reconsidering how they are approaching application development based on a particular model.”

The Chinese Open-Source Surge

Rather than slowing China’s AI ambitions, the US restrictions appear to have accelerated them. Knowledge Atlas Technology (Z.ai) released its open-source GLM-5.2 model days after the Anthropic ban, with shares surging over 30% in Hong Kong trading. The company has gained more than 800% since its January debut.

“At a time when some frontier models can suddenly become unavailable, we choose to believe in a different path,” Knowledge Atlas posted on social media, according to the South China Morning Post. “Frontier intelligence should not belong to only a few people, nor be subject to withdrawal by a handful of rules at any moment.”

DeepSeek, meanwhile, closed a record funding round of approximately $7.4 billion following the Anthropic ban, as TechJournal detailed. The economics are stark: DeepSeek’s V4 Pro costs approximately $3.48 per 1 million tokens of output, while Anthropic’s restricted Fable 5 model cost $50 for the same output — a roughly 14x price difference.

Neil Shah, vice president of research at Counterpoint Research, described the dynamic bluntly: “It’s a great move for China. Obviously they’re not on the cutting edge because of the export controls, but they have their own silicon and their own software.”

Xi’s Endorsement of Open-Source AI

Chinese President Xi Jinping waded into the debate at the World AI Conference in Shanghai on July 17, endorsing open-source AI development and implicitly criticizing US efforts to protect its technological lead. As the Hindustan Times reported, Xi cast China as a champion of openness and equality in AI development, a message aimed squarely at developing nations choosing between US and Chinese technology ecosystems.

The Distillation Controversy

At the heart of the dispute lies “distillation” — a process where one AI model is trained on the outputs of another to copy its capabilities more cheaply. In April 2026, the White House Office of Science and Technology Policy accused foreign entities of conducting “industrial-scale” campaigns to distill US frontier systems. Anthropic has identified efforts by DeepSeek and two other Chinese labs to extract capabilities from its Claude platform, while OpenAI has separately warned lawmakers about similar attempts. Chinese companies have denied wrongdoing.

Sovereign AI and Market Fragmentation

The Anthropic ban has accelerated the “sovereign AI” movement, particularly across Asia. South Korea, Japan, and other nations are reconsidering their dependence on US AI models and exploring domestic alternatives. Non-Chinese rivals like Canada’s Cohere and France’s Mistral have begun marketing themselves as AI that “cannot be switched off by a single government order.”

Sung Kim, cofounder of Korean AI startup Upstage, captured the sentiment: “We need to advance our own technology as quickly as possible and become as self-reliant as we can.”

Many US startups are already voting with their wallets. As NPR reported, companies like Lindy.ai have migrated 100% of their traffic to Chinese models, citing costs that are 10x cheaper than US alternatives. Uber CEO Dara Khosrowshahi recently noted that the company “blew through” its annual AI budget in a single quarter.

Legal and Strategic Uncertainty

The unprecedented legal basis for the Anthropic IIL is already being challenged in court. An Anthropic customer that used the Fable 5 model with employees in Canada — Legion LegalTech, Corp. v. United States, No. 1:26-cv-02225 — is seeking to vacate and enjoin the directive, arguing it exceeds the government’s statutory authority.

Beijing has also weighed curbing overseas access to some of China’s top AI models, according to Reuters reporting cited by Semafor, suggesting that both sides are preparing for a more fragmented global AI landscape.

What to Watch

Several critical questions remain unanswered. Will the US formally blacklist DeepSeek or other Chinese AI companies? Can the US effectively control AI models when they can be downloaded and run locally, sidestepping any government’s restrictions? And will the sovereign AI movement permanently fragment the global market into US, Chinese, European, and other ecosystems?

As Semafor’s editorial view put it: “The hard truth this week exposed is that controlling AI models is far harder than controlling the chips they run on, and the policy tools built for one may not fit the other.”

The coming weeks will determine whether Washington can craft a strategy that achieves its security objectives without accelerating the very diversification it seeks to prevent.