China Just Put AI Models Themselves on an Export Control List — the Tech Cold War Moved Up the Stack
The control list just climbed from chips to models
For years, the technology contest between Washington and Beijing has been fought mostly over hardware: advanced chips, the machines that make them, and the computing power they deliver. On Friday, October 16, China proposed moving that contest one layer up the stack — to the models, agents and clusters themselves.
China’s commerce ministry, acting with the science and technology ministry, published a draft revision of its catalogue of technologies prohibited or restricted from export. The proposal adds 40 technologies across three fields: artificial intelligence, biomedicine and photovoltaic solar technology. Public comment runs until November 15. Crucially, Beijing did not wait for the consultation to finish before acting: an interim pre-export registration requirement for the covered technologies took effect the same day.
That combination — consult later, register now — is the part businesses should read twice. Even before any final rule exists, outbound transfers of the listed technologies are already visible to regulators.
What is actually on the list
The breakdown matters more than the headline number:
- 23 technologies are in AI. Reporting on the draft describes coverage of large models, intelligent agents, computing clusters, chip design and information-processing systems — in effect, the core building blocks of a modern AI industry rather than a single product category.
- 5 are in biomedicine, including areas around cell and gene therapy, small nucleic-acid drugs and pathogenic microorganisms.
- 12 are in photovoltaics, covering parts of the solar supply chain such as polysilicon, wafer and cell production, where China holds a dominant global position.
Within the 40, the draft would prohibit export of 14 technologies outright and restrict another 26, meaning those could move only under licence or approval. Examples reported from the draft include a prohibition covering AI technology for generating and mutating malicious code, and a restriction covering AI-assisted vulnerability detection — a reminder that dual-use capability, useful to defenders and attackers alike, sits at the centre of this exercise.
One carve-out deserves emphasis: technologies already in the public domain, through open-source release or published academic work, are excluded. That line will be argued over for years, because much of China’s recent AI progress — open-weight models in particular — has travelled precisely through that public channel. The draft, as described, targets proprietary technology that has not been published, not work the world can already download.
The interim registration, according to reporting on the ministry’s plan, is broad. It reaches transfers of technology ownership, overseas licensing, moving ownership to a foreign entity, and even exports that have already begun but not yet completed. In other words, deals mid-flight are not grandfathered out of sight; they are pulled into the registry.
Beijing’s own framing is deliberately calm. A commerce ministry spokesperson described the catalogue as routine trade-management work, revised repeatedly since 2001, driven by expanding uses of these technologies in sensitive fields, and not aimed at any specific country, region or entity. The stated aim is that AI technology remain safe and controllable.
You can take that framing at face value and still see the significance. Routine or not, a major AI power just asserted that a trained model, an agent system and a computing cluster are strategic assets whose departure from the country is the state’s business.
Why this is a mirror, not just a move
Anyone who has followed the last four years of US policy will recognise the shape of this. Washington built an extensive regime to keep advanced chips and chipmaking tools from flowing to China, on the argument that frontier compute is a national-security asset. Beijing objected to that logic for years — and has now adopted a version of it for its own strengths.
That mirroring is the real story. Export control is becoming the shared grammar of AI competition:
- The US model controls inputs: chips, equipment, and increasingly cloud access to compute.
- The Chinese proposal controls outputs of its own ecosystem: the models, agent technology and cluster know-how its labs and companies produce, plus the solar and biotech processes where it leads.
- Both rest on the same premise — that AI capability is too strategically important to move across borders as an ordinary commercial good.
Reporting on the draft noted that Chinese officials had already, earlier this year, discussed with domestic labs how to protect proprietary model technology, amid concern that sharing it could eventually be turned against China. Friday’s proposal turns those conversations into a formal mechanism. The concern is no longer theoretical or internal; it has a catalogue, a comment deadline and a registration desk.
The open-source question this raises — and does not settle
China’s open-weight model strategy has been one of the defining dynamics of the AI race. Releasing capable models openly won developer mindshare worldwide, pressured Western labs on price, and made Chinese AI a default ingredient in products far outside China. A control regime that excluded public-domain technology would, on its face, leave that strategy intact: publish it, and it is outside the catalogue.
But read the structure more carefully and a subtler incentive appears. If what a lab keeps private can later be restricted from export — licensed deals, overseas deployments, transfers to foreign subsidiaries — then the commercial paths for a Chinese AI company expand differently from the open path. According to the ministry’s own answer on Friday, whether international expansion is affected depends on whether it involves transferring restricted technology across the border. That is a testable, deal-by-deal question, which means it is also a legal-review question for every partnership, cloud deployment and overseas subsidiary a Chinese lab operates.
For Western companies, the mirror-image risk is dependence. Any product built on proprietary Chinese model technology delivered under licence — rather than on openly released weights — now has a visible regulatory seam in it. Procurement teams learned this lesson with chips: a component is only as reliable as the export regime standing behind it. Friday’s draft extends that lesson to models themselves.
What this means if you buy, build or invest in AI
- Map your model supply chain the way you mapped your chip supply chain. Know, for each model you depend on, whether you hold open weights you can keep running, or a licensed service whose cross-border transfer could one day need approval in Beijing, Washington, or both.
- Treat in-progress deals as exposed. The interim registration reportedly reaches transfers already under way. Any company mid-negotiation for Chinese AI technology, licensing or joint development should assume the deal may need registration now, not after November 15.
- Watch the November 15 comment deadline as a signal, not a formality. What survives consultation — especially the definitions of a covered model, agent or cluster, and how the public-domain exclusion is worded — will determine whether this is a narrow security list or a broad industrial gate.
- Expect reciprocity as the default. Each side’s controls now justify the other’s. Planning that assumes a return to free movement of frontier AI technology is planning for a world neither capital is describing.
The bottom line
Friday’s proposal is still a draft, and honest reporting requires saying so: the final catalogue may differ, licences may prove easy or hard to obtain, and the public-domain exclusion may keep the most visible channel — open models — flowing as before.
But drafts with immediate registration requirements are not trial balloons. They are how a state takes position before the argument starts. China has now said, in regulatory form, that AI models, agents and computing clusters belong on the same list as sensitive biotech processes and strategic solar technology.
The contest over AI used to be about who could get the chips. Increasingly, it is about whose permission you need to move the intelligence itself. On October 16, that permission structure got a new author — and every global AI roadmap now has to read it.