China’s AI export curbs: ignore the list, watch the registration requirement
| AUTHOR | Arindam Goswami |
| DATE | November 18, 2026 |
| CATEGORIES | High Tech Geopolitics AI Export Controls Technology Denial China |
On 16th October, China published a draft which would place 40 technologies (23 of which are in the AI space) under either an export licence or prohibition and which would also require exporters to register any covered transfers. The technologies covered are things like model weights and chip designs, not AI services which are sold via apps or APIs. Using the Takshashila 2026 study of technology denial regimes (by the author and public data) this yields four findings: the AI rules do pose some risk to India, the solar rules are more impactful; the registration requirement will almost certainly outlast the list of controlled items and India’s fall in Chinese solar imports is likely to be smaller than it appears.
Published weights and access through APIs, apps and clouds are exempt. A model is covered if it meets any one of three tests: 10^26 training operations, 2 trillion parameters, or a top-ten global ranking. The compute test is weak, since it is five times the largest Chinese training run Epoch AI has estimated, ChinaTalk reports. The size and ranking tests could catch today’s leading Chinese models, and new releases may be delayed, which might affect Indian developers building on them. The compute test copies Washington’s 2023 threshold, so each side has an equal restriction to trade away, and a mutual limit on frontier-model weights around the expected 18 November Trump-Xi meeting is possible, though nothing shows it is under discussion.
These appear to be negotiation tricks and such tactics invariably lead to a settlement. Of the six escalations by Washington and Beijing since late 2024, all were either suspended or reversed within a year (median of 66 days). Domestic politics will no doubt influence some of the orders as it has done in the past, but the timing of this draft appears to be linked to the 18th November meeting rather than an election. Licensing systems have however held their ground: the US 2022 chip controls, China’s gallium licensing and the heavy rare earths licensing. Following the April 2025 rare earths licensing the magnet exports fell by about half and then recovered within three months, as seen from customs data. Registration could easily become a licensing system if the country so wished. It will remain in place until the final catalogue comes into effect so there is no real cost to Beijing in terms of delay and it does build up a record of who is exporting what. Expect a late and quite narrow text. The Takshashila study recommended a very narrow scope of controls and ChinaTalk notes that this list covers pretty much every digital chip design.
If one were to analyse the impact of controls, they have worked unevenly. China fell from 23% of Nvidia’s revenue in fiscal 2021 to 9% in fiscal 2026 (annual reports), yet Lam Research and KLA still earn about a third of revenue there, and China’s chipmaking-machinery imports nearly doubled between 2022 and 2025. Beijing now restricts purchases of H200 chips Washington approved. The study predicted adaptation and mutual denial, and a target denying itself, so a US relaxation buys Nvidia little. The lost opportunity is something to contend with, but the world has always run on states minding their own affairs. Lobbying for rules that serve all countries is sensible, but lamenting their absence doesn’t change much.
Several things in AI are worth watching. Chinese labs may rush out large open models before the final rules arrive, because anything already published is exempt. This is more likely than not. Washington’s next step is likely to be curbs on Chinese firms renting chips in data centres abroad, about half of the compute that can be restricted, the Institute for Progress estimates. India should check who its cloud customers are, as Malaysia has done, to stay a trusted host. A Chinese model with strong hacking skills is also very likely by December, going by how far Chinese models usually trail the leading ones. It would be the first test of the draft’s cyber rules, and India should test it before using it for anything critical.
Solar is where India is exposed. By March 2026, India had about 172 GW of module capacity, 27 GW of cells and roughly 2 GW of ingots and wafers. Industry estimates put the Chinese share of its equipment at 70%. The draft restricts ingot pulling, wafer slicing, polysilicon plants and advanced cell lines. India’s own mandate requires domestic cells, with net-metered and open-access projects exempt only until 31 December. China’s share of India’s cell and module imports fell from 78% in 2024 to 32% in early 2026, while Indonesia supplied $871 million in seven months against $219 million in all of 2025. Customs data show only where goods were shipped from, not who owns the plant. Chinese-owned plants have been reported in Indonesia and Laos, and registration covers changes of control, so the shift away from China may be smaller than it looks.
Going forward, there could be four pathways, weighted by judgement from the reversal patterns observed till now. In the likeliest case, the truce would get extended and the list narrowed: India should communicate its views to China’s commerce ministry before the feedback deadline, asking for automatic licences for majority Indian-owned ventures. Indian firms should order equipment now, while licences are still being granted. If Beijing uses the registration requirement quietly, through slow approvals and no announced ban, which is very likely, solar lines will most likely get held up first: India should qualify European equipment and form a buyers’ group with Indonesia and Turkey, with India the largest importers of Chinese cells. The study urged constrained countries to pool bargaining power, which India did not do in its space and nuclear cases. If the truce lapses on 10th January which doesn’t seem very likely, India should relax its cell mandate temporarily, stock spares and avoid both sides’ AI rules. If a leading Chinese lab withholds weights, which again seems very unlikely, India should build fine-tuning capacity and back open models from outside both countries.
All of this leads to a few educated guesses that countries can make: most likely is that the final catalogue will not yet be in force, since delay is free and Nvidia’s November guidance will still exclude Chinese data-centre revenue, since Beijing limits purchases. Also, India will most probably have to extend or soften the cell exemption, since domestic cells are scarce.
India can in principle object to the export controls and still be happy to accept that they can’t be changed. The AI items are the ones that require a measured response and some preparation (the cloud screening mentioned above being an example). Solar is the area in which India should be pushing. Its factories are so reliant on Chinese equipment that any delay could have an impact in the very short term. The biggest surprise will probably be that the list itself is actually of minor importance and the registration requirement is the key one. Registration is a less onerous step than a licence but it does give Beijing a record of who is exporting what to whom and allows it to intervene on a deal prior to it being completed. It also covers licences for overseas use and for changes of owner or controller. That is relevant to Chinese owned plants overseas and so is relevant to India.