China’s Ministry of Commerce has held a series of meetings over the past month with the country’s most powerful technology companies — including Alibaba, ByteDance, and AI startup Z.ai — to discuss restricting overseas access to China’s most advanced artificial intelligence models, according to three people familiar with the discussions who were not authorized to speak publicly. The deliberations are the clearest signal yet that Beijing now treats frontier AI as a critical national asset requiring the same protective architecture it applies to semiconductors, rare earths, and aerospace technology.
What Happened
The talks, reported by Reuters and confirmed by three separate sources, involved officials from both the Ministry of Commerce — which oversees China’s export regulatory framework — and the National Development and Reform Commission, the country’s top state planning body. Their joint presence at the meetings signals that these discussions are being coordinated at a high level of the Chinese government, not handled as routine commercial regulation.
According to two of the sources, participants discussed placing limits on the most advanced AI models, covering both closed-source systems and open-weight models — those that can be downloaded, run, and customized by third parties. Officials also floated the possibility of making any leak or theft of proprietary AI technology a criminal offence under China’s national security law, one of the most serious legal instruments in the country’s statute book. A further proposal under discussion would restrict who is permitted to fund domestic AI startups, potentially closing off certain foreign investment channels.
The scope of any final rules remains undefined. Two sources said the measures may apply only to future models, not existing ones, and it is unclear when, or whether, they will formally come into force. Alibaba, ByteDance, Z.ai, the Ministry of Commerce, and the NDRC all declined to respond to requests for comment.
The Reading
Why This Is Not a Surprise
China has been tightening its grip on homegrown AI all year. In April, the NDRC ordered Meta to unwind its $2 billion acquisition of Manus, a Chinese-founded AI startup. In early June, Beijing issued sweeping new rules tightening oversight of overseas deals touching Chinese investors, technology, data, and national security concerns. Separately, authorities launched investigations into Manus and other local AI startups that had relocated abroad, probing whether they had violated existing export control laws. The Ministry of Commerce meetings are best understood as the next logical escalation in that sequence — not an isolated policy event but a consistent strategic trajectory.
Hints about the likely regulatory architecture appeared as early as May, when a roundtable of Chinese legal experts on open-source AI regulation published its conclusions in an official Supreme People’s Court journal. Participants proposed a tiered framework: basic open-source tools subject only to a simple filing requirement; more advanced technologies facing mandatory security reviews; and the most sensitive frontier models either barred from public release or restricted to domestic use only. That proposal now reads less like academic theorizing and more like a policy blueprint.
Who Says So — and What the Sources Reveal
The three sources cited by Reuters are people familiar with the discussions who declined to be identified. Their convergent accounts — on both the institutions present and the specific proposals raised — give the reporting credibility, even absent official confirmation. The silence from Alibaba, ByteDance, Z.ai, the Ministry of Commerce, and the NDRC is itself notable. In Chinese regulatory contexts, no-comment responses from major state-linked agencies often indicate that discussions are live and sensitive rather than concluded or abandoned.
Why It Matters
The commercial implications are substantial. Since DeepSeek’s R1 model demonstrated last year that Chinese AI could match or approach Western frontier performance at a fraction of the cost, Chinese models have made rapid inroads in global markets. Z.ai’s GLM-5.2, in particular, has been attracting serious attention in Silicon Valley for its combination of competitive capability and dramatically lower pricing. Alibaba’s Qwen family and ByteDance’s Doubao are already among the most widely deployed AI systems in China, with growing international footprints. Any restriction on overseas access to these models would reduce the competitive pressure they currently exert on U.S. and European incumbents — and would likely push global enterprise AI costs upward as alternatives are sought.
There is a structural irony worth naming here: the United States has spent the past two years using export controls to limit China’s access to advanced semiconductors, explicitly framing the policy as a national security imperative. Beijing is now assembling a mirror-image framework — treating its own AI software as an exportable strategic asset subject to government control rather than open-market distribution. The symmetry is not accidental. Both governments have arrived at the same conclusion: that in a world where AI capability increasingly determines economic and military outcomes, allowing unrestricted access to frontier models by adversaries is an unacceptable risk. The difference is that Washington has moved first and faster on hardware; Beijing may now be moving first on open-weight software controls, an area where Western policy remains largely undefined.
The national security dimension is not merely theoretical. Two of the Reuters sources said Chinese authorities are specifically concerned about Anthropic’s Mythos model — a cybersecurity-focused system currently restricted to “trusted” U.S. organisations after the Trump administration required that foreign nationals be blocked from accessing it — and the potential for Washington to deploy it against Chinese digital infrastructure. That concern, publicly echoed by Zhou Hongyi, the founder of cybersecurity firm 360 and a prominent voice in Chinese government-adjacent technology circles, suggests the policy impulse behind Beijing’s AI controls is at least partly defensive, not purely mercantilist. The growing consensus among regulators globally is that frontier AI models are no longer simply commercial products — they are dual-use technologies that require governance frameworks analogous to those covering weapons-grade cryptography or advanced military hardware.
The funding restriction proposal is perhaps the least discussed but potentially the most consequential element of the deliberations. Limiting who can invest in domestic AI startups would reduce the capital available to Chinese frontier labs precisely as the AI race enters its most capital-intensive phase. It would also complicate the position of international venture capital firms with existing stakes in Chinese AI companies, and could deter foreign investment in future Chinese AI ventures even where such investment might otherwise be commercially attractive.
What to Watch
The key variable is specificity. If Beijing implements a tiered framework broadly consistent with the Supreme People’s Court journal proposal — distinguishing basic open-source tools from advanced proprietary systems — it would represent a relatively surgical intervention that preserves much of the existing open-weight ecosystem while placing the most capable frontier models under state control. A broader restriction covering all advanced models, including the open-weight systems that have driven much of China’s international AI adoption, would be a more disruptive intervention with significant downstream effects on global developers, enterprises, and research institutions that have built workflows around models like Qwen.
The investigation into Manus and similar companies that moved operations overseas is also worth tracking closely. Beijing has shown a growing willingness to pursue individuals and organisations it believes have exported sensitive technology without authorisation. If those investigations result in prosecutions under the national security law framework discussed at the Ministry of Commerce meetings, it would signal that the new rules are intended to have real enforcement teeth rather than serving as soft guidance.
How China’s Proposed AI Controls Compare to U.S. and EU Approaches
| Dimension | China (Proposed) | United States (Current) | European Union (Current) |
|---|---|---|---|
| Primary instrument | Export control law + national security law | Export Administration Regulations (EAR); model-specific access restrictions | EU AI Act (risk-based classification) |
| Scope of controls | Frontier models, both closed-source and open-weight (under discussion) | Targeted model-level restrictions (e.g., Anthropic Mythos); chip export controls | High-risk AI applications; GPAI models above compute thresholds |
| Enforcement mechanism | National security law prosecutions; investment restrictions (proposed) | Nationality verification requirements; export licensing | Conformity assessments; regulatory sandboxes; fines |
| Open-weight/open-source treatment | Would restrict most advanced open-weight models (proposed tiered system) | Not yet systematically addressed | Partial exemptions for open-source; under ongoing review |
| Investment screening | Proposed restrictions on domestic AI startup funding | CFIUS reviews inbound foreign investment; OUTBOUND investment rules under development | FDI screening frameworks; no AI-specific investment rules yet |
The comparison above illustrates a significant regulatory gap: while the U.S. and EU have focused primarily on hardware controls (chips) and application-level risk classification respectively, neither has systematically addressed whether frontier open-weight AI models should be subject to export controls. China’s proposed framework — if enacted — would make Beijing the first major jurisdiction to formally restrict the overseas distribution of open-weight AI models, potentially establishing a precedent that other governments will feel pressure to follow or respond to. The AI market is already at an inflection point; a new layer of national controls on model distribution would accelerate the fragmentation of what has, until now, been a largely borderless technology ecosystem.
What This Means for the Industry
For the global AI industry, the Ministry of Commerce meetings represent a moment of genuine strategic clarification. The era in which Chinese AI companies could operate as both domestic champions and global open-source contributors — simultaneously satisfying Beijing’s ambitions for technological self-sufficiency and competing freely in international markets — may be drawing to a close. Alibaba, ByteDance, and Z.ai will face a choice between compliance with domestic controls and the international market access that has been central to their growth strategies. That is not a comfortable position, and the outcome is unlikely to be clean.
U.S. AI companies face a more complicated picture than simple competitive relief. If Chinese frontier models become harder to access globally, the immediate beneficiaries are incumbents like OpenAI, Google DeepMind, and Anthropic — but only until non-Chinese alternatives emerge to fill the gap. More importantly, the move signals that AI model distribution is becoming a geopolitical instrument in both directions, which creates pressure on Washington to develop coherent policy on open-weight model controls before Beijing sets the terms of debate.
Enterprises and research institutions that have built significant infrastructure around Chinese open-weight models face operational risk that is difficult to hedge. The uncertainty itself — the fact that restrictions may apply only to future models, or may not come into force at all — is likely to prompt procurement reviews and contingency planning among large deployers in the near term, regardless of what Beijing ultimately decides.
The deepest implication may be structural. If both the world’s two largest AI powers treat their most capable models as national assets subject to export controls, the architecture of global AI development will shift from an open, interdependent ecosystem toward something closer to the fragmented, bloc-based technology landscape that has characterized the semiconductor industry for the past decade. That would have consequences — for innovation velocity, for cost, and for the distribution of AI capability — that extend well beyond any individual company’s product roadmap. The institutions that shape those rules, not the ones that build the models, may ultimately determine who wins the AI era.
