1. China AI export controls: what the draft proposes
China released draft rules on October 16 that would stop foreign companies from obtaining the models, data or expertise needed to operate or rebuild Chinese artificial intelligence systems on their own. Overseas customers could still use Chinese AI through websites, apps and cloud platforms. The proposal, which covers 40 technologies in total, is open for public comment until November 15, 2026, and it marks a notable turn in the global contest over who controls advanced AI.
The underlying idea is simple to state: Beijing wants the world to use its AI, but not to own it. Access can be sold abroad, while the technology itself stays at home.
Draft rules of this kind are published so that companies, researchers and other interested parties can respond before the government settles on a final version. The consultation therefore gives a window in which the wording on thresholds, definitions and exemptions could still shift, even though the interim registration requirement already applies.
2. Which technologies are on the list?
The draft is a revision to China's Catalogue of Technologies Prohibited and Restricted from Export. It was put forward by the Ministry of Commerce and the Ministry of Science and Technology. The catalogue is the official list that determines which technologies may not leave the country at all and which may leave only with a government license.
The update would add 40 technologies across three sectors. Twenty-three relate to artificial intelligence, 12 to photovoltaics, the technology behind solar panels, and five to biomedicine. The treatment differs by item: 14 of the 40 would be banned from export outright, while the remaining 26 would be allowed only with a license.
On the AI side, the draft targets what it calls essential technologies. These include the weights of large models and the know-how used to optimize their training. Model weights are the learned numerical values that give an AI system its abilities, so possessing them is essentially possessing the finished model.
The rules are aimed at systems that meet performance thresholds. These include models trained with at least 10^26 floating-point operations of computing power, and models with at least 2 trillion parameters. Both measures describe the largest, most capable systems, which means the restrictions are focused on the frontier of the field instead of on smaller, widely available models.
Export-control catalogues work as a gatekeeping tool. A technology that appears in the prohibited column cannot be sold or transferred abroad by any company, while an item in the restricted column can leave only after an exporter secures a license and shows the authorities who the recipient is and what it will do with the technology. Adding new categories to the catalogue therefore changes the legal risk for every Chinese company that deals with foreign partners.
The compute threshold deserves a short explanation. A floating-point operation is a single arithmetic calculation, and the total number performed during training is a standard way of measuring how large a training run was. A figure of 10^26 is an enormous quantity that only the most resource-intensive training runs reach. The parameter count, by contrast, describes the size of the model itself, with 2 trillion parameters placing it among the largest systems built.
3. Where does the line between access and ownership fall?
The central distinction in the draft separates two situations. In the first, a foreign customer uses a Chinese AI service. In the second, a foreign company acquires the technology that makes the service work. The first is permitted; the second is tightly controlled.
Using a service covers activities such as calling a Chinese model through an application programming interface, or subscribing to a product hosted in the cloud. The draft exempts these methods of access from export controls, so Chinese AI developers can keep selling subscriptions and API usage to customers abroad.
Acquiring the technology means obtaining items such as model weights, training data or the expertise to run a system independently. Here the draft is strict. International transfers of covered technology would require government approval, and the approval requirement extends to changes in ownership or control and to deployments in other regions.
This second category matters most for companies with global ambitions. A Chinese developer that wants to install its system in a data center overseas, or sell a stake in itself to a foreign investor, could face review that does not apply to simply selling access over the internet.
The practical result is a two-tier market. Customers abroad can keep sending requests to a Chinese model and receive answers, much as they would with any online service. What they cannot do, without approval, is take a copy of the model, study how it was trained, or reproduce it in their own facilities. The developer retains control over how the system is run, updated and priced.
This structure has an obvious appeal for the developers themselves. Keeping weights and training methods in-house protects the investment that went into building them, and selling access still produces revenue. The draft, in effect, gives that arrangement a legal basis and makes departures from it subject to government oversight.
4. What did Beijing say about its reasons?
A spokesperson for the Ministry of Commerce called the move necessary and cited national security. Controls on technology exports are commonly justified by the risk that sensitive capabilities could be misused or could strengthen a rival, and the ministry's explanation follows that pattern.
The announcement also carried an immediate practical effect. An interim registration requirement took effect on the day the draft was published. From that date until the revised catalogue takes effect, exporters must register through the commerce ministry's unified business platform before exporting any listed technology, and the requirement also reaches deals that have not yet been completed.
That interim step means companies cannot wait for the final text before changing their behavior. Anyone with pending agreements that involve covered technology already has a compliance task to complete.
Companies preparing to respond during the consultation will likely focus on definitions. Terms such as ownership, control and regional deployment can be read narrowly or broadly, and the difference determines whether an ordinary commercial arrangement, such as a joint venture or a cloud partnership that places a system on a foreign operator's servers, needs sign-off from Beijing.
5. How does this fit into the technology rivalry with the US?
Until now, most discussion of AI export controls has concerned Washington's limits on the flow of advanced chips and tools to China. The Chinese draft turns that logic around. It applies a similar approach to China's own software, data and technical expertise.
For Chinese AI developers, the draft largely formalizes a model many already follow. They sell access to their systems around the world while keeping the most valuable assets, such as weights and training methods, under their own control. Revenue from international subscriptions and API use stays available.
The picture is different for foreign companies that hoped to license Chinese models, buy into Chinese AI developers or run Chinese systems on their own hardware. These plans could face a licensing process that adds time and uncertainty, and changes in ownership or control appear to be squarely within scope.
The measure also arrives in a setting where both governments treat AI capability as a matter of national strength. Restrictions in one country tend to prompt a response from the other, and a draft like this adds a Chinese instrument to a toolbox that had been used mainly by the United States. Whether that leads to broader fragmentation of the AI market depends on how strictly the final rules are enforced and how other governments react.
6. Why do solar and biomedicine appear alongside AI?
The inclusion of photovoltaics stands out. With 12 solar-related technologies on the list, the draft extends beyond software into an industry where Chinese manufacturers already hold a strong position worldwide. Restricting the know-how behind that production would affect competitors that might want to build comparable capacity elsewhere.
Five biomedicine technologies complete the package. Together, the three sectors show that the proposal is not narrowly about AI but about a wider effort to protect fields where China regards its technical lead as strategically important.
The thresholds also shape who is affected. Because the rules reach models trained with at least 10^26 operations or with 2 trillion parameters, they fall hardest on frontier systems. Smaller models that fall below those levels would not be covered by the same AI provisions.
For solar manufacturers, the draft adds 12 technologies to the catalogue, which means transfers of that know-how abroad would fall under the same ban-or-license system that applies to the AI items. The solar entries show how the catalogue approach is being applied across industries, not only to software.
7. What happens next?
The public comment period runs until November 15, 2026, and the final text could still change. Several questions will determine how far-reaching the rules turn out to be. One is how broadly regional deployments are interpreted, since that term decides how much overseas installation requires approval. Another is whether the performance thresholds are adjusted before the catalogue is finalized.
Chinese AI companies, foreign customers and investors with exposure to Chinese systems will watch how the interim registration process works in practice. The draft leaves the access model intact, so services sold over the internet should continue as before, but the cost of owning or relocating the underlying technology is set to rise.
For investors and executives, the immediate takeaway is that the draft separates selling an AI service from transferring AI ownership, and only the latter faces new barriers. Businesses weighing contracts with Chinese AI developers should look at which side of that line their plans fall on, and whether their arrangements involve any of the 40 listed technologies before the interim registration rules apply to them.