AI Agent UpdateBeijing drafts its own AI export controls — model weights are now a strategic asset
China proposes to control exports of frontier model weights, AI agents and chip design know-how — a mirror image of Washington's regime, with one elegant asymmetry: open-source models are exempt, and API access doesn't count as an export. The AI stack just got balkanized in both directions.
The News
China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) published draft Announcement No. 48 on Oct 16, proposing to add 40 technologies — 23 of them AI-related — to China's export-control catalogue. The draft's centerpiece: the weights, checkpoints, source code and architecture of frontier models, defined as those trained with at least 10^26 FLOPs of compute, holding at least 2 trillion parameters, or ranked in the global top ten.
The scope reaches well beyond model files. The draft also covers AI agents, compute clusters, and AI chip design know-how down to the ≤14nm process node, including 2.5D and 3D stacking techniques — the packaging technologies at the heart of the current HBM and accelerator race. In effect, Beijing is cataloguing the entire frontier AI stack, from silicon design to deployed agents.
The definitions carry the strategy. Hosting a restricted Chinese model on servers abroad would count as a technology export; providing API access to the same model would not. Pre-export registration requirements are already in force ahead of the rule, public comments on the draft close Nov 15, and — critically — open-sourced models are exempt from the controls entirely.
Read the structure and the intent is clear: Beijing wants to keep frontier capability — the weights themselves, which can be fine-tuned, distilled and deployed without Beijing's involvement — under state control, while letting Chinese AI services keep earning globally through APIs and letting open-source releases keep spreading Chinese technical standards through the world's developer ecosystem.
The carve-outs matter more than the controls: Beijing is taxing weight transfers, not inference — API access stays wide open, and open source stays free.
Why It Matters
This is Washington's playbook, mirrored. The US controls the export of advanced chips and the equipment to make them; China now proposes to control the export of advanced models and the know-how to design chips. The AI stack is formally balkanized in both directions — each bloc treating the other's layer of the stack as a strategic asset. The immediate commercial hit lands on sovereign-AI deals: the fast-growing business of running Chinese models on local infrastructure in third countries — Gulf states, Southeast Asia, Latin America — is precisely the weight-transfer model this draft taxes.
For US-listed names the direct revenue exposure is small — $NVDA, $AMD and $QCOM were never going to import Chinese frontier weights. The exposure is structural: every sovereign buyer now has to choose a stack whose model layer and chip layer come from the same bloc, which hardens the two ecosystems' borders and raises compliance costs for anyone selling into both. $NVDA's chart tells the week's story — down 2.4% over the past month on far bigger forces than this draft — but the draft adds a permanent, slow-moving headwind to the “one global AI market” assumption baked into long-range forecasts.
Analyst's Take
The carve-outs matter more than the controls. API access staying open means this doesn't tax inference revenue — Chinese labs keep selling intelligence by the token to the world. What it taxes is weight transfer: the sovereign deals where a government or telco takes a Chinese model in-house. That is a deliberate choice to keep the cash register open while locking the recipe in the vault — and it makes the draft far more commercially survivable for China's own champions than a blunt ban would have been.
The quiet winner is the sovereign-cloud infrastructure builder sitting outside both blocs — the firms in the Gulf, India and Southeast Asia selling neutral ground where either stack can run. And watch the open-source exemption do its work: exempting open weights isn't liberalism, it's standards strategy. Every developer building on a free Chinese model is a developer inside Beijing's technical gravity — while the frontier stays home. Comments close Nov 15; the final text's thresholds will tell us how much of this is posture and how much is law.
By The Numbers
FAQ
What does China's draft Announcement No. 48 cover?
Forty technologies proposed for China's export-control catalogue, 23 of them AI: frontier model weights, checkpoints, source code and architecture (models trained with ≥10^26 FLOPs, ≥2T parameters, or ranked in the global top ten), plus AI agents, compute clusters and AI chip design know-how at ≤14nm including 2.5D/3D stacking.
Does using a Chinese AI model abroad count as an export?
Under the draft, it depends on the form. Hosting a restricted Chinese model on servers abroad counts as a technology export. Accessing the same model through an API does not — a distinction that keeps Chinese AI services sellable globally while controlling the underlying weights.
Are open-source Chinese models affected?
No — open-sourced models are explicitly exempt. That keeps China's open-weight ecosystem (and the global developer adoption it generates) flowing freely, while closed frontier models are locked down.
When could the controls take effect?
The draft was published Oct 16 by MOFCOM and MOST, with public comments open until Nov 15. Pre-export registration requirements are already in force. A final effective date will follow the comment period and any revisions.
Source: AICoder · AI and Global Policy Project