China’s AI Export Order: What It Does and What It Means for India
| AUTHOR | Bharath Reddy |
| DATE | October 18, 2026 |
| CATEGORIES | Strategic Autonomy HTG AI |
A draft proposal by China’s ministry of commerce adds advanced artificial intelligence technologies to its export-control list. The restrictions seem to be designed more as a switch than a wall. Almost nothing changes for exporters today, but China now holds the means to monitor, slow or stop the flow of its AI technology abroad whenever it chooses.
Geopolitechs has a good translation of the MOFCOM announcement and has been used for this analysis.
The order does four things at once:
- Maps outbound technology flows. The registration requirement took effect on 16 October, before the rules themselves are final. Every technology transfer covered under its ambit must now disclose the technology, the parties, the destination, the end use and the value..
- Creates a veto while promising not to use it. Open-source releases and deployment on a company’s own overseas infrastructure get “automatic licensing”. Automatic today still means licensed: open release moves from permitted by default to permitted after registration. That route can be narrowed or withdrawn in the future.
- Cyber safety concerns are genuine but the coverage is broad. China might genuinely worry about offensive cyber capabilities but the restricted list targets a capability frontier more broadly.
- Mirrors US leverage. Washington controls chips and some controls on models. Beijing now has a matching toolkit over models, agents, and chip designs.
AI disclosure: Claude Opus 5 has been used for help with the analysis and copyediting for this piece.
What the order does
On 16 October, MOFCOM issued two things together. The first is a draft revision of the catalogue of technologies prohibited or restricted from export, adding 23 AI items within a wider list of 40 technologies (biotech and solar are also on it). The second is a temporary registration requirement, already in force, that controls the transfer of technology (weights, code, designs, know-how), but not the sale of services.
| Category | What it covers |
|---|---|
| Prohibited (no export) | Models trained on “core data”; high-performance agents for critical infrastructure, defence and core government systems; AI with offensive cyber capability (malware generation, antivirus evasion); high-accuracy detection of AI-generated content |
| Restricted (licence needed) | Foundation models: core technology (weights, checkpoints, source code, architecture, training and inference methods) of any model meeting one of three triggers: 10^26 training FLOPs, 2 trillion+ parameters, or a top-10 global ranking on a qualifying measure; also models trained on “important data”; Agents: advanced multi-step planning, tool use and multi-agent coordination; Chip design: AI chips at 14nm or below, advanced packaging, advanced memory; Clusters: 10,000+ accelerators, interconnects, cross-vendor chip scheduling; Cybersecurity: AI vulnerability discovery and attack chains; Speech and language: voiceprint, audio-fraud detection, dialect and minority-language tools |
| Not covered | Serving overseas users through APIs, apps, websites or SaaS; technology already lawfully in the public domain |
| Automatic licensing | Open-source releases of restricted models; deployment on overseas infrastructure owned by the Chinese provider |
| Registration (in force now) | All covered transfers, including licensing deals and changes in foreign ownership or control, and deals already underway but not yet completed |
What “automatic licensing” means. A licence is still legally required, but approval is meant to be quick and procedural. Licensing requirements now exist, even if the friction is near zero today.
Core and important data. Both come from China’s Data Security Law in effect since 2024, which grades data into three tiers: general, important and core.
- Important data is data whose leak or misuse could harm national security, the economy, public health or public order.
- Core data is the stricter subset tied to national security and the “lifelines” of the economy.
Why this reading
The design choices in this order make more sense as instruments of control and leverage than as purely safety measures. Here are some arguments that back up this reading.
- The criteria form a wide, discretionary net. The triggers for restriction are many and form a wide net as meeting any one is enough. Together with the top-10 ranking and “important data” criteria, they can reach almost any Chinese frontier model.
- The top-10 ranking test filters by capability, not risk. The controls track rank in global rankings, so the restrictions seem to target the Chinese frontier rather than risk. Would ranking in the top 10 on safety also trigger restriction? These details are unclear and might be addressed when the rules are framed.
- API access is unrestricted. Foreign users can use the same capabilities as a service. But this mode of access would create ongoing dependencies on the provider and not offer the same autonomy or data sovereignty as open models.
- The other controls target economically valuable know-how. Chip design at 14nm and below, advanced packaging, 10,000-accelerator clusters and cross-vendor chip scheduling carry no misuse risk. Cross-vendor scheduling is China’s workaround to US chip controls, for example mixing Nvidia and Ascend accelerators in one training run. Solar and biotech also in the same package indicate that this is a competitiveness instrument with a safety and risk framing.
- The change-of-control clause targets relocation. Registration covers deals where foreign ownership or control changes, even if no code crosses a border. This is likely aimed at Chinese AI start-ups moving their legal home abroad, as Manus did when it relocated to Singapore and was later acquired by Meta.
- Information comes before rules. Registration requirements are already in effect before the catalogue has been finalised. This allows MOFCOM to build a map of who exports what to whom.
What this means for India
For India the order impacts the terms on which Chinese open weights are available. The implications follow from that.
- Nothing changes now, but the basis for leverage exists. Developers fine-tuning Qwen or DeepSeek are unaffected: already released weights are in the public domain. Future releases might be subject to a licence Beijing can tighten.
- Diversity of jurisdictions matters as much as diversity of model vendors. Both the US and China now treat proprietary frontier weights as strategic goods and are considering options to gate open releases. A healthy open ecosystem therefore needs serious open models from outside the US and China (maybe even through a consortium of companies and countries), so that no single state holds a veto.
- Avoid lock-in to a single model or agent framework. Build model and agentic framework agnostic stacks: routing layers, standard interfaces, evaluation suites that let a base model and agentic frameworks be swapped.
- Invest in the capacity to diffuse the technology. The durable hedge is not owning a frontier model but having the talent and compute to fine-tune, distil and adapt whichever open model is best and available. This plays to India’s comparative advantage in human capital.