China wants its artificial intelligence (AI) to go global. But Beijing also wants tailored control over what leaves, who acquires it, and what remains in the hands of the Chinese government. That’s the tension behind proposed revisions to technology export measures, announced on October 16. Spurring international adoption while retaining authority over strategic technology creates room for China to exert leverage over those who use or build on Chinese AI abroad.
Announcement No. 48 from the Ministry of Commerce (MOFCOM) proposed adding 23 AI technologies to its Catalogue of Technologies Prohibited or Restricted from Export. Public comments are open until November 15; pre-export registration, however, took effect immediately, even though the classifications are not yet final. The formal scope looks narrow, with MOFCOM noting most AI technology will remain freely exportable.
The registration requirement will allow MOFCOM to see transactions before they happen, while the existing licensing requirements still apply. The need for registration is triggered by a transfer of ownership abroad, a license for overseas use, or a change of owner or actual controller to a foreign entity. The requirement applies to deals already underway.
The listings reach the core technologies of qualifying advanced models, such as weights and source code, along with AI agents, AI chip design, and the interconnect and scheduling technologies for clusters of more than 10,000 accelerators. They do not cover the chips and other products made with AI technology.
The accompanying MOFCOM explanation said that offering even restricted large models through APIs, apps or web interfaces, software-as-a-service applications, or lightweight on-device deployments falls outside the scope of large-model technology export controls. Under the proposed revisions, the AI listings also exclude technology already in the public domain through open-source release or academic publication. The first open-source release of restricted technology, however, is itself an export, though one eligible for automatic licensing.
Deploying a restricted model on a customer’s servers abroad is also a technology export; it qualifies for automatic licensing when Chinese firms run it in their own overseas data centers. As under U.S. export controls, Chinese AI may go abroad as a service or an open release, but listed technology cannot be sold abroad without passing through Beijing.
China is trying to prevent another Manus controversy. Manus, an AI agent startup originally based in China that had moved its headquarters to Singapore, was acquired by the U.S. firm Meta – before the Chinese government moved to block the deal. On April 27, China’s foreign-investment security review office announced a prohibition decision requiring Meta’s acquisition of Manus to be unwound.
In that case, Beijing intervened only after the deal had closed. Under the new rules, a Manus-style sale would have to register before closing, giving regulators notice before control changes hands. The rules also signal to founders and foreign acquirers that moving headquarters abroad will not, by itself, take listed technology beyond Beijing’s reach.
Chinese authorities also reportedly barred two Manus co-founders from leaving China while the deal was under review. Separately, new State Council exit and entry rules in force since September allow exit bans on citizens whose breaches of export control or technology trade rules could endanger industrial or technological security.
Remote computing is harder to classify. When a Singapore company rents GPUs in Shanghai, the hardware stays in China, yet registration reaches any authorization of overseas use of listed technology. But MOFCOM’s carve-out for APIs and software says nothing about bare GPU rental. The rules leave this unsettled for now, and they do not rule out controls on remote access later. MOFCOM can require registration in other cases by notice.
Washington is already weighing curbs on remote access to American chips, so Beijing could apply the same logic to its own compute once it has capacity to spare.
AI Exports: What Does Beijing Gain?
If control matters this much, why let Chinese AI be exported at all?
Wide adoption of Chinese AI attracts developers, encourages complementary products, and sustains commercial relationships, above all in countries that cannot build competitive foundation models of their own. MOFCOM’s explanation explicitly supported AI capacity building in the Global South. It also promised additional facilitation for cooperation among members of the World AI Cooperation Organization (WAICO), a China-initiated body established in July by 29 founding governments. WAICO membership is not a condition of technology access, but could count in a partner’s favor when Chinese companies seek export licenses.
Automatic licensing would also cover chip designs sent abroad for fabrication and joint research using overseas resources. Because the fast track favors infrastructure Chinese firms build and operate abroad, the computing capacity and the models running on it can arrive as one package, both still owned by Chinese firms.
The narrow prohibited tier – aimed at AI-driven malicious code, models trained on China’s data, agents for defense, critical infrastructure, and government use – gives Beijing room to frame its controls as keeping AI under human control, which in turn reinforces its bid to shape AI governance through WAICO.
In one respect, China’s approach seems to resemble Pax Silica, the U.S.-led supply chain compact launched last December that now has about two dozen signatories. Both are built to reward partnership with preferential treatment.
Pax Silica, however, is weighted toward the hardware layer, with secure supply chains for minerals, semiconductors, and AI infrastructure, and co-investment among trusted partners. WAICO centers on capacity building, technology sharing, and international AI governance. Beijing now promises to ease export licensing for its members.
Indonesia, a WAICO founder still in talks over Pax Silica, said the two are “not necessarily at odds,” because WAICO focuses on AI as a tool and Pax Silica on broader digital development.
AI adoption could lead to potential leverage. Users may come to rely on subsequent model releases, training and inference know-how, integration tools, or continuing technical support. That reliance can harden into path dependence, even for certain U.S. firms that have built on Chinese open models. Authority over future controlled transfers then becomes a bargaining chip.
That leverage has limits, though. Weights already released cannot be recalled, and future open releases of restricted models would qualify for automatic licensing, which MOFCOM said would be largely filed and cleared at once. Running released weights locally, or switching to competing models, reduces reliance on Chinese providers.
The model and hardware layers can also separate. Nvidia already optimizes its platform for leading Chinese open models such as DeepSeek and Qwen, and its own filings warn that open models deployed mainly on rival platforms could reduce demand for its products. Which model a user adopts does not necessarily settle whose hardware it runs on. That could change once Chinese chips are good enough for Beijing to bind Chinese models to them by policy.
Commercial popularity does not automatically produce geopolitical control, and China’s tightened restrictions could push customers toward alternatives. Unclear rules could also lead future start-up founders to incorporate outside China from the start.
Governments and businesses that build on Chinese AI at the moment should keep the option to switch suppliers. Investors in China-origin AI firms should check for listed technology early and build registration into deal timetables. The question, now, is how far Beijing presses. If it ties its export licensing preferences more closely to WAICO endeavors, control over China’s AI stack could become a claim to lead AI governance across the Global South, possibly pushing governments in the non-Western world toward two rival stacks.
