China AI export rules: DeepSeek and MiniMax clear registration

Two of China's biggest open-weight labs cleared Beijing's new AI export registration in under a week. In Seoul the same morning, the memory champion found out what the capital markets will pay for an AI-adjacent listing right now — nothing.

DeepSeek and MiniMax have completed the export registrations that Beijing's draft AI export-control rules require, and have gone ahead with open-source releases and overseas licensing, according to Chinese state media. Yuyuan Tantian, a social media account affiliated with state broadcaster CCTV, reported the registrations on Thursday evening, the South China Morning Post reported. Both companies are unnamed in any official notice; the reporting traces to a state-media account rather than a Ministry of Commerce document, and neither lab has confirmed anything on the record.

The rules behind the registration are a week old. On October 16 the Ministry of Commerce and the Ministry of Science and Technology published a draft that would impose outright export bans on 14 technologies across AI, biotechnology and solar power, plus restrictions on 26 more. The AI portion alone covers 23 areas — advanced models, AI agents and computing clusters among them. Crucially, the draft says the restrictions do not apply to technology already in the public domain, which includes open-source systems and work published in academic papers. That carve-out is why China's open-weight strategy looked untouched by the whole exercise: you cannot restrict a model by export licence if the weights are already on Hugging Face.

What the draft also did was switch on a registration requirement for covered technologies with immediate effect, even while the wider rule is still a proposal. That turns a legal exemption into an administrative step, and it is the step DeepSeek and MiniMax just took. CCTV's framing, relayed by the SCMP, is that the measures are meant to draw clearer regulatory boundaries rather than shut the door on Chinese AI exports.

Why that matters is easier to see from the outside than from Beijing. China's most consequential AI export of the past eighteen months has not been chips or data-centre contracts; it has been cheap weights. DeepSeek's V4.1-Flash-1017 landed at 77% on DeepSWE for about 20 cents a task, which we covered in detail, and MiniMax's M3.1 Flash put up 42 on the Artificial Analysis index at $0.22 a task — a price point that made US labs' mid-tier pricing look generous. Both models are downloaded and deployed outside China by developers who will never talk to either company. If registration had become a bottleneck, every one of those releases would have needed pre-clearance, and the open-weight advantage would have been rationed by permit.

Instead the two names that carry the strategy abroad are through the process within days, which is the useful signal. A registration regime that clears its most prominent applicants in a week is a list, not a gate. The state gets to know who is shipping what, and to whom, before it crosses the border; the labs get to keep shipping. That is a very Chinese settlement of the tension between wanting global adoption of your models and wanting visibility into where they go.

The counter-reading deserves to be stated plainly, because the reporting is thin. We do not know what registration obliges: whether it is per model or per release, whether it covers weights, hosted APIs, licensing deals or all three, whether there is a review period, or whether a registered company can be deregistered. It is also unclear whether this week's approvals mean Alibaba's Qwen, Zhipu and Moonshot are lining up behind them or whether DeepSeek and MiniMax are being blessed as the two the government trusts. "Clearer boundaries" is exactly what a regulator says when it has not decided how strictly to enforce them.

The likeliest outcome sits between the two: China has no interest in starving its best export story, and every interest in being able to say no. Register first, ship second, and keep the option to slow a release that goes somewhere inconvenient. For Western developers that means Chinese open weights keep arriving on schedule, with one more administrative layer behind them that nobody outside Beijing can see.

SK Hynix has scrapped the US IPO of Solidigm, its NAND subsidiary, ending a review it had run on listing the American unit to fund a new US fab. The company said the review of "potential pre-IPO and IPO options" was discontinued "in light of recent capital market conditions and share price trends," and insisted the decision "does not reflect any change in Solidigm's growth prospects or investment needs," in a disclosure on its newsroom. Bloomberg, which had reported the listing at roughly $10 billion, framed the withdrawal as a retreat under regulatory pushback.

The funding need has not gone away with the listing. Solidigm is based in the United States and needs US production capacity to serve AI-era storage demand; SK Hynix's own line was that its existing facilities cannot meet it. An IPO was the elegant version of that plan: raise outside money against a US asset, keep the parent's balance sheet out of it, and hand shareholders buybacks with the proceeds. Now SK Hynix funds the fab directly, or borrows — and the stock reversed an early 4.2% slide into a gain of more than 4% in Seoul, which the WSJ and CNBC both reported. The market's read is legible enough: no dilution, no new float competing with the parent, and no asset sold at a discount into a semiconductor correction.

The pushback mattered as much as the price. The Korea JoongAng Daily reported that the review was dropped after shareholder opposition and regulatory pressure, and The Straits Times said Seoul authorities had raised concerns about the listing. That is the part worth keeping: a US listing for a Korean company's American subsidiary is now a political question as well as a valuation one, and the answer this week was no.

Set that against yesterday's argument about memory demand. We wrote that the HBM numbers point to a 2027 glut even if Jensen Huang says otherwise — Jensen Huang says no 2029 memory glut, the HBM numbers say 2027 — and this is what that risk looks like from the financing side. The capex commitment to US capacity is being made now, at record margins, with the cheapest funding route closed. NAND and HBM are different markets with different cycles, but they are the same balance sheet, and SK Hynix just took on more of the fab bill itself.

What to watch: whether MOFCOM publishes anything formal on the AI registrations — a list, a portal, a processing time — and whether other Chinese labs say they have registered too. On the memory side, SK Hynix's next capex update, and whether the US NAND fab keeps its schedule now that the equity cheque is gone.

Should a government be able to decide which of its country's models get exported — and would you trust the answer either way? Write to us via our contact page.