China's new AI export controls deeply threaten healthcare innovation

What was published

The Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology published Announcement 48 of 2026 [1]. It proposes 40 additions to China's Catalogue of Technologies Prohibited or Restricted from Export: 23 in AI, 12 in solar manufacturing and 5 in biomedicine. Of these, 14 would be prohibited outright. The other 26 would need an export licence.

The list is published as a draft, open for comments until November 15, 2026. But one requirement did not wait. Since October 16, and until the revised catalogue takes effect, any company exporting one of the listed technologies must first register the export with MOFCOM. That covers four situations:

• assigning ownership abroad;
• licensing the technology for use abroad;
• the owner or actual controller of the technology becoming a foreign entity;
• any other case MOFCOM designates.

Exports already under way are included. The analyses published by Zhong Lun and Baker McKenzie remind clients that registering is not a licence [6][7]. The wording may still change in the final version of the rule.

The AI rules that reach life sciences

Most of the AI entries concern frontier models, chips and cybersecurity, which look far from the laboratory [4]. Looking more closely, three of them are not.

1. The exemption fits consumers, not pharma. According to MOFCOM's Q&A, using a restricted Chinese AI model remotely over the internet (through an API, an app, a website or a SaaS service) is outside the controls, which is good news [2]. So is a lightweight deployment on a local device. But deploying a covered model in your own company environment abroad is treated as an export. If the draft is adopted as it stands, that would require a licence. When the model runs on the Chinese provider's own overseas cloud, the licence could be granted automatically.

That is the reverse of how pharma, international and local alike, uses AI. A company handling patient data, clinical results or unpublished chemistry rarely sends them to an outside service. It runs the model inside its own environment. The route the rule leaves open is the one life sciences uses least. The route it controls is the one life sciences needs.

Which models are covered? A model is “caught” if it reaches any one of three thresholds: 10²⁶ operations of training compute, 2 trillion parameters, or a “globally leading” capability [3]. The last test means ranking in the global top 10 in public evaluations on any one of several dimensions. One of those dimensions is “specialised business capability”. A Chinese model that leads on biomedical tasks could therefore be caught without being a frontier model. Models already released as open weights, and anything published in academic papers, are excluded. The question is the coming generation.

2. Agents. Autonomous agents are a separate restricted category. Under the draft, one threshold is enough to trigger it [3]:

• planning that breaks an instruction into 20 subtasks;
• coordinating 60 agents at once;
• integrating 200 third-party tools.

Each threshold comes with accuracy and success rates. A Chinese provider can still sell an agent as a service. A European client that wants to host the agent's reasoning loop in its own systems would need the provider to obtain a licence. That is precisely what makes Chinese AI attractive to many buyers: low cost, open weights, and the freedom to run it in-house. For the most capable agents, the draft would put that advantage under licence. Agent platforms for research, regulatory writing or pharmacovigilance are moving towards these levels of capability.

3. The model becomes a data export. This is the entry we consider most important for the sector. Large models trained on China's “core data” would be prohibited from export. Those trained on China's “important data” would be restricted, with no size or performance threshold [3]. A model is not “caught” if important data made up less than 0.1% of its training data, under conditions such as de-identification, the data not being reproduced in the model's answers, or the data being already public.

These categories come from the Data Security Law, and the lists of what counts as “important data” are drawn up sector by sector. China has not yet published a complete list for healthcare. Population health and genomic data are among the categories most widely expected to be classified as important.

Until now, the Human Genetic Resources rules and the cross-border data rules, already strict, controlled whether the data could leave China. Under this draft, the model that learned from the data would need an export licence too. That changes the value of several kinds of project:

• a European pharma company training models in its China R&D centre on Chinese clinical or real-world data, then bringing the model to headquarters;
• a joint AI project with Chinese hospitals, whose results are meant to be used worldwide;
• a Chinese AI imaging or pathology company, trained on Chinese hospital data, selling into Europe.

One point remains unclear. The Q&A's exemption for on-device deployment (see point 1) refers to large models in general. The annex gives no such exemption for models trained on important data. How AI medical-device software with an embedded model will be treated is a question to raise in the consultation, not to assume.

In short, AI drug discovery is “caught” twice. An AI drug-discovery platform is usually built from models and agents, which fall under the AI list. It is also used to discover targets, which falls under the biotech list.