Summary
- The Ministry of Commerce in China has initiated discussions with major AI firms like Alibaba, ByteDance, and Z.ai regarding potential restrictions on international access to the nation’s leading AI models, including unreleased versions, as reported by Reuters.
- The proposed regulation framework includes a tiered system, ranging from simple filing for basic tools to strict domestic-only access for the most sensitive advanced models.
- If China decides to limit its own open-weight AI models, this would eliminate alternative options for businesses that turned to these models after the U.S. restricted access to Anthropic and introduced barriers for GPT-5.6 in June.
The U.S. enacted its AI export restrictions in June, and it seems that China is preparing to implement similar measures as early as July.
According to Reuters, Beijing has engaged in discreet discussions over the past month with top AI companies about limiting the overseas distribution of their advanced models.
Meetings were held with Alibaba, ByteDance, and startup Z.ai to explore the possibility of restricting access to China’s most sophisticated AI systems, including those yet to be released, as reported by Reuters, which cited three anonymous sources familiar with the talks.
These discussions, convened by the Ministry of Commerce, included the consideration of imposing limits on both closed-source and open-weight models, which developers can download and modify. Officials also suggested that unauthorized disclosure or theft of proprietary AI technology might be classified as a violation of China’s national security law. Additionally, there were discussions about new regulations to control which investors could finance domestic AI startups.
The extent of any prospective restrictions remains under discussion. Two sources informed Reuters that the proposed measures might apply only to future models and not to existing ones. No specific timeline has been established, and it remains uncertain whether any regulations will be enacted.
The AI Regulatory Framework
The practical implementation of these restrictions is still unclear, but insights emerged from a summary in a Supreme People's Court journal following a May roundtable with Chinese legal experts regarding open-source AI regulation. Participants proposed a three-tier structure: basic open-source tools would require minimal government filing; more advanced technologies would undergo security reviews prior to release; and the most sensitive frontier models would either be prohibited from public release or limited to domestic use only.
This proposed framework would represent a significant shift for Chinese AI firms, whose international success has largely stemmed from their openness. For instance, Alibaba's Qwen series has gained considerable traction on Hugging Face, the largest repository for open-source AI models worldwide. ByteDance's Doubao is among the leading AI offerings in China, while Z.ai's GLM-5.2 has garnered interest from U.S. researchers for its competitive performance against top American models while offering API access at a significantly lower cost.
Should China proceed with restricting foreign access, it could increase costs for companies that have relied on Chinese models as affordable, less regulated alternatives to U.S. advanced systems. There is also apprehension among officials that Anthropic's Mythos—the cybersecurity model restricted by the Trump administration in June—could potentially be reverse-engineered and used against Chinese systems, adding urgency to the discussions.
U.S. Precedent
On June 12, Anthropic received a directive from the Commerce Department's Bureau of Industry and Security, mandating the suspension of all access to Claude Fable 5 and Mythos 5 for any foreign national, including Anthropic’s non-citizen employees. Due to the challenges in restricting access via live API endpoints by nationality, Anthropic pulled both models globally within hours, marking the first instance of the U.S. applying export controls to operational AI models rather than just the chips used for training them.
These models were reinstated on June 30, after Anthropic adjusted its safety classifiers and the Commerce Department lifted the restrictions. This pattern had already occurred four days prior with OpenAI, which released GPT-5.6 Sol, Terra, and Luna, after previewing the models with the U.S. government and initially providing access to around 20 carefully vetted partners at Washington's request.
OpenAI stated that government-gated access "shouldn't become the long-term default." An executive order from President Trump on June 2 regarding AI had already encouraged developers to voluntarily submit advanced models for federal cybersecurity reviews before public releases. A framework defining what constitutes a "covered frontier model" and the conditions for government pre-release access is expected by August 1.
Beijing's Vigilance
Chinese officials have been vigilant, particularly regarding concerns that Anthropic's Mythos—the cybersecurity model restricted in June—could be weaponized against Chinese infrastructure. There are also fears that Anthropic might be using spyware-like tactics to monitor Chinese users, which has raised further concerns in China.
‼️ BREAKING: Anthropic has embedded hidden spyware-like code in Claude Code that covertly targets Chinese users. It sends information regarding each user by injecting it into their prompt message.
Claude Code is sending information such as timezone, proxy, and potential AI Lab… pic.twitter.com/EjfwtirhES
— International Cyber Digest (@IntCyberDigest) June 30, 2026
Zhou Hongyi, the founder of Qihoo 360, explicitly called for China to develop a domestic equivalent at ISC.AI 2026 in Beijing, unveiling Tulong Feng, a homegrown AI vulnerability agent.
The proposed tiered structure from the Ministry of Commerce discussions aligns closely with how the U.S. has managed chip export controls over the past three years—a policy history that has consistently narrowed China’s technological gap rather than expanding it.
China has been tightening its regulations as well. In April, the state planning agency instructed Meta to cancel a $2 billion acquisition of AI startup Manus due to foreign investment security reviews. It also mandated that Moonshot AI and StepFun obtain government approval before accepting U.S. investment, and a broader regulatory framework released in early June increased scrutiny on cross-border transactions involving Chinese technology and data.
Closing the Escape Valve
The prevailing logic since DeepSeek R1 gained popularity in early 2025 was straightforward: U.S. restrictions on advanced AI created a natural market for Chinese open-weight models. The usage of these models on OpenRouter, a key platform for global AI distribution, surged from less than 2% in late 2024 to about 61% by mid-2026. The U.S. taking a defensive stance inadvertently granted Beijing a global distribution advantage it did not achieve through sheer technical superiority.
This market dynamic relies on the continued openness of Chinese models. If Beijing imposes restrictions on international access to its advanced systems—whether closed-source or open-weight—the escape valve would close. Z.ai’s GLM-5.2 built its appeal entirely around unrestricted access under an MIT license; limiting that distribution would undermine its value proposition.
The Lawfare analysis of June's restrictions highlighted a structural issue that applies equally to China: "national" AI restrictions do not remain confined to national borders. The U.S. order on Anthropic encompassed foreign nationals in the country, including Anthropic’s non-citizen employees. Over two-thirds of top-tier AI researchers working in the U.S. were educated abroad, with the foreign-born percentage reaching up to 70% in leading labs, according to MacroPolo data.
A nationality-based access restriction aimed at adversaries ultimately excludes the engineers necessary to resolve the vulnerabilities that prompted the restrictions in the first place. China faces a similar dilemma in reverse. ByteDance and Alibaba are already removing humanlike agent features ahead of new Chinese AI regulations set to take effect on July 15, indicating that when Beijing decides to impose restrictions, it acts without waiting for market feedback.
This situation could pose significant challenges for smaller labs and developers worldwide that rely on and contribute to open-source technologies, a field where China has thus far taken the lead.
French President Macron warned during the G7 summit that European governments would cease purchasing U.S. AI products if access could be revoked on short notice. Canadian Prime Minister Carney described the overreliance on concentrated AI resources as a “strategic mistake.” Both responses assumed that Chinese AI would serve as the unfettered alternative.