China Weighs AI Model Curbs as U.S. Firms Turn to DeepSeek

Chinese officials have met with Alibaba, ByteDance and Z.ai over the past month about possible limits on overseas access to the country's most advanced AI models, Reuters reported Tuesday, citing three people familiar with the discussions. The talks, led by China's Ministry of Commerce, covered both closed systems and more open versions, including models that have not yet been released.

The possible curbs arrive after Chinese models moved from fringe alternatives to a measurable share of U.S. AI traffic. CNBC reported that Chinese models have stayed above 30% of weekly token volume used by U.S. companies through OpenRouter since Feb. 8, compared with an 11% average over the prior 12 months. DeepSeek, Z.ai and Alibaba's Qwen are becoming cost controls for buyers at the same time Washington and Beijing are treating model access as a national security question.

For Beijing, the policy question has shifted from whether Chinese labs can compete to whether their strongest models should keep competing globally on the open terms that made them attractive.

Beijing tests a model-access line

At the meetings, officials discussed limits on the most advanced AI systems and the possibility that leaks or theft of proprietary AI technology could be treated as an offense under China's national security law. One person familiar with the talks also said officials raised possible new measures on who can fund domestic AI startups.

The scope remains unsettled. Two people said the measures may apply only to future models, and it was not immediately clear when, or whether, any rules would take effect. China's Commerce Ministry, the National Development and Reform Commission, Alibaba, ByteDance and Z.ai did not respond to requests for comment.

The companies named in the meetings sit near the center of China's model market. Alibaba's Qwen and ByteDance's Doubao are among the most widely used models in China. Z.ai, formerly Zhipu, drew U.S. developer attention in June after releasing GLM-5.2, an open-source model aimed at long-running software engineering work.

The same account put the model talks next to other recent moves by Beijing to keep homegrown AI inside the country's regulatory reach. In April, China's state planner ordered Meta to unwind its $2 billion acquisition of Chinese-founded AI startup Manus. In early June, authorities tightened oversight of overseas deals involving Chinese investors, technology, data and national security. Chinese authorities also investigated Manus and other local AI startups that had moved abroad.

The potential model-access rules would extend that logic from companies and deals to the models themselves. In a summary of a May roundtable published in an official Supreme People's Court journal, Chinese legal experts proposed a tiered system for open-source AI. Basic tools would face simple filing. More advanced technologies would face security reviews. The most sensitive frontier models could be barred from public release or restricted to domestic use.

U.S. buyers have moved first

Chinese models are gaining customers outside China on price and performance rather than on political alignment. OpenRouter's U.S. traffic data, as cited by CNBC, showed Chinese models climbing as high as 46% of weekly token use this year. The first half of 2025 had averaged 4.5%.

Kyle Chan, a fellow in the John L. Thornton China Center at Brookings, told the network that Chinese models are attractive because AI costs are rising. Companies that previously prioritized adoption over model choice are becoming more cost-conscious, he said.

Lindy, an AI startup, is one example. CEO Flo Crivello said the company moved all of its traffic from Anthropic's Claude models to DeepSeek in June and expected to save millions of dollars within months. Vercel also saw DeepSeek's share of gateway tokens climb between May and June.

Vercel saw rapid adoption of Z.ai's GLM-5.2 after its June release. Harpreet Arora, Vercel's head of agentic infrastructure, said GLM-5.2 had the fastest adoption of any model Vercel tracked in 2026, with daily token volume growing about 27 times in its first full week and customer count growing about 80 times.

OpenRouter's Justin Summerville said open-source Chinese models can be 60% to 90% cheaper than the leading Anthropic and OpenAI models. Chan said the models often operate close to the top American frontier systems while costing a fraction as much.

The possible Chinese restrictions would cut into the same release model that made those systems visible abroad. A domestic-use rule would protect capability from foreign governments and competitors while narrowing the path that carried DeepSeek, Qwen and GLM into U.S. workflows.

GLM-5.2 gives the debate a concrete object

GLM-5.2 landed within a percentage point of Anthropic's Opus 4.8 on one closely watched agentic benchmark at roughly a fifth of the cost, according to CNBC. Gabe Pereyra, co-founder of Harvey, said it was the first model where he saw open source becoming genuinely competitive with some closed-source frontier models.

Computerworld reported that Z.ai released GLM-5.2 as an MIT-licensed open-source model for long-running software engineering tasks, with a one-million-token context window and up to 131,072 output tokens. Z.ai said the model trailed Anthropic's Claude Opus 4.8 by 1% on FrontierSWE, a long-horizon coding benchmark, and edged OpenAI's GPT-5.5 by 1%.

Z.ai also said GLM-5.2 uses IndexShare, a technique that reduces per-token compute by 2.9 times at a one-million-token context length. The company said changes to the model's multi-token prediction layer increased speculative-decoding acceptance length by as much as 20%.

Pareekh Jain, CEO of Pareekh Consulting, told Computerworld that Western enterprises would still want independent benchmark validation, global customer deployments, security controls, governance and long-term support. He said the risk calculation changes if a company uses Z.ai's hosted API rather than running the open weights on its own infrastructure, because Chinese national security rules could require domestic companies to cooperate with government requests.

Washington already pulled the same lever

China is not moving in a vacuum. The Trump administration has already used export-control authority to restrict advanced U.S. models, and the Anthropic episode is now the reference case for both countries.

In mid-June, the Commerce Department ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models by any foreign national, including foreign-national Anthropic employees. Anthropic disabled access globally because it could not verify nationality in real time. The department lifted export controls on June 30, but Mythos 5 access remained limited to some U.S. organizations through Anthropic's Glasswing program while the company worked with the government on broader access.

OpenAI followed with a more controlled launch. On June 26, it announced GPT-5.6 Sol, Terra and Luna and said it was complying with a U.S. government request to limit the initial rollout to a small group of trusted partners. OpenAI said it believed in broad access and was working toward general availability in the following weeks.

The U.S. rationale is cybersecurity and foreign military use. The Reuters account framed Washington's concern around possible misuse by military intelligence users in China, Russia and other countries of concern. Anthropic's Mythos model, designed for cybersecurity professionals, became a particular concern because of its ability to find and exploit software vulnerabilities.

Chinese officials are focused on the same category of risk from the other side. Two people familiar with the discussions said authorities were deeply worried Washington might use Mythos against Chinese interests. State media and Zhou Hongyi, founder of cybersecurity firm 360, have argued that China needs its own Mythos-like capabilities.

Washington has used export-control authority to limit foreign access to the most capable American models, with China among the stated concerns. Beijing is now considering whether foreign users should reach the most capable Chinese models. Both policies can be defended as national security measures, and both make buyers think harder about models that can be switched off by a government letter.

Alibaba sits on both sides of the dispute

Alibaba is a useful case because it appears in almost every part of the story. The company was present for Beijing's discussions on possible Chinese model limits. The Associated Press reported that Rep. John Moolenaar, chair of the House Select Committee on China, asked Monumental Sports & Entertainment owner Ted Leonsis to cut business ties with Alibaba after the Pentagon designated the company a Chinese military company. Alibaba has sued to be removed from the Pentagon list.

Anthropic has accused Alibaba-affiliated operators of using Claude to train or improve Chinese models. Ars Technica reported that Anthropic told U.S. senators operators affiliated with Alibaba and Alibaba Qwen generated more than 28.8 million exchanges with Claude through almost 25,000 fraudulent accounts between April 22 and June 5. Anthropic called it the largest campaign to illicitly extract Claude capabilities that it had measured.

A separate Reuters report said Alibaba banned employees from using Claude Code at work after the tool drew scrutiny for features that could identify China-linked users. The ban came amid Anthropic's accusation that Alibaba had carried out a distillation effort. The person who spoke for that report said Alibaba employees were being told to use Qoder, Alibaba's own coding platform.

Alibaba did not respond to requests for comment in the Claude Code report or the Beijing-curbs report. The company has not publicly commented on Anthropic's accusation.

DeepSeek is building around the same constraints

SiliconANGLE, summarizing an underlying wire report, said DeepSeek has explored in-house AI accelerators for about a year. The company has contacted chip design, foundry and memory companies and is trying to hire experienced chip designers. The work remains early.

DeepSeek has relied on Nvidia and Huawei chips, with Huawei's Ascend GPUs becoming more important after U.S. restrictions on Nvidia exports to China, the publication said. The underlying report said DeepSeek does not want to become too dependent on Huawei.

DeepSeek's hardware work is separate from the Beijing meetings. It still belongs in the same commercial file: model access rules affect who can use a system, while inference chips affect how cheaply the developer can serve it.

The next fight is software distribution

The May roundtable gave the clearest public sketch of a Chinese open-source AI control system. Participants proposed basic filing for simple tools, security reviews for more advanced technologies and domestic-only restrictions for the most sensitive frontier models.

No formal rule has arrived. The scope is still being discussed, the measures may apply only to future models, and the ministries involved have not published a timetable.

Commerce Department Lifts Export Controls on Anthropic's Fable 5 and Mythos 5

Anthropic said Tuesday that the U.S. Department of Commerce has lifted the export controls it imposed on the company's Claude Fable 5 and Mythos 5 models on June 12, and that it would begin restoring

The Implicator

NSA Loses Anthropic Mythos Access After June Export-Control Order

The National Security Agency lost access to Anthropic's Mythos 5 after the Trump administration's June export-control directive forced the company to pull back its most advanced models, The New York T

The Implicator

OpenAI Restricts GPT-5.6 Release to Government-Approved Partners

OpenAI said Friday it is restricting access to GPT-5.6, its newest and most capable family of artificial-intelligence models, to a small group of partners approved by the Trump administration. The com

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai