MediumV Technology Geopolitics & AI28 October 2026, Wednesday
Beijing's model rule puts Chinese labs' offshore compute in doubt
Major Chinese AI labs are asking whether the draft of 16 October requires approval to train and run their best models on foreign-owned servers abroad, according to a Reuters analysis of 28 October. The comment period on the draft closes on 15 November.
The draft from the commerce and science ministries leaves overseas users free to access Chinese models through websites and the cloud. It requires approval, however, for a foreign company that takes over a model, data or know-how to run it independently. The draft sets 3 thresholds: training compute of at least 10^26 operations, at least 2 trillion parameters, or a top-10 global ranking on certain unspecified leaderboards. Law firm Han Kun asked how models that run only some of their components per task would be measured. Methods for coordinating more than 10,000 AI chips are also covered.
The question matters because offshore compute is how Chinese developers reach chips the US has banned. On 31 May the BIS confirmed that shipments of advanced chips to overseas subsidiaries of China-headquartered firms also require a licence. IDC says Nvidia's share of China's data centre chip market fell to 34% in the first half of 2026. In the worst case, labs believe they may have to maintain 2 separate versions of the same downloadable model, one for home and one for abroad.
According to Fangda, a registration requirement that took effect immediately on 16 October covers transfers placing Chinese technology under foreign control, including deals already under way. Beijing unwound Meta's acquisition of Manus, worth more than 2 billion dollars, earlier this year. The draft offers simplified approval for Chinese companies building their own facilities abroad. It does not explicitly address technology transfer to build a similar system for a foreign client. The Ministry of Commerce did not respond to Reuters' questions.
Talay assessment
Bottom line
Beijing wants to sell the service abroad while keeping the model itself at home. That makes Chinese labs' workaround of the US chip ban through offshore data centres subject to their own government's approval as well. The squeeze now comes from both sides: Washington licenses the chips, Beijing the models. The most likely path is a rule adopted at year-end that leaves thresholds vague and works through case-by-case approval.
Likely effects
- Chinese AI labsNegative1–6 months
If training on offshore servers needs approval, the last practical route to US chips becomes subject to licences from both capitals at once.
- Third-country data centresNegative1–6 months
Offshore data centres renting capacity to Chinese labs face demand uncertainty while their Chinese clients wait for model-transfer approval.
- Foreign AI investmentNegative6 months+
Because registration also covers deals under way, sales of Chinese-origin AI companies to foreign buyers risk being unwound, as with Manus.
Possibilities, ranked
- 1Vague thresholds, case-by-case approval55%
The rule is adopted at year-end, the ranking threshold stays undefined and major labs obtain approval one by one.
Watch: The final text and implementing guidance due after the 15 November comment period
- 2Offshore compute explicitly exempt25%
The final text exempts Chinese companies training their own models abroad from approval.
Watch: An exemption clause for offshore training and inference in the final text
- 3Strict scope, dual versions20%
Labs move to separate model versions for home and abroad, and offshore capacity rental slows.
Watch: DeepSeek, Moonshot or Z.AI announcing a separate international version
Probabilities are calibrated judgement based on the sources, not measurement, and are revised as new information arrives. Not investment advice.
Market reaction
Indicators affected
- Nvidia's China share▼ 34%
- Model threshold▼ 2tn parameters