Beijing spent years blocking Nvidia's best chips from reaching Chinese buyers. Now it's building a wall facing the other direction, one aimed at keeping its own AI models, training data and agent technology from leaving the country.
China's Ministry of Commerce and Ministry of Science and Technology announced on October 16 that exporters of certain artificial intelligence, biopharmaceutical and solar technologies must now register their international transfers with regulators. The two ministries are also proposing to add roughly 40 technologies across those three sectors to China's catalog of prohibited and restricted exports, with public comments open until November 15, according to a report from the Financial Times. It's the first time Beijing has moved to restrict software and trained AI systems themselves, rather than just the hardware underneath them.
For years the story of the chip war ran one direction: the US cutting China off from Nvidia's most advanced GPUs, and China scrambling to build domestic alternatives through SMIC and Huawei. This flips the frame. Washington restricts what China can buy. Beijing is now moving to restrict what China sells to everyone else, including American companies that have quietly built products on top of Chinese open-weight models. The specific target, according to the Financial Times, is Chinese open-weight AI, the models that companies anywhere can download, modify and run on their own servers without paying for API access. Alibaba's Qwen, ByteDance's Doubao, Zhipu's GLM and DeepSeek's releases have become the default cheap alternative for developers who don't want to pay OpenAI or Anthropic prices. That adoption curve has been steep. Chinese open-weight models captured 17.1% of global open-weight downloads in the year to August 2025, overtaking the United States' 15.8% share for the first time, with DeepSeek and Qwen alone accounting for 14% of that traffic.
Regulators led by the Ministry of Commerce have reportedly held closed-door meetings with Alibaba, ByteDance and Zhipu on a tiered system that would limit how much training data those companies can send overseas and whether foreign users can keep freely down the actual model weights. Access to the same models as a hosted service, through an API or a cloud subscription, would reportedly stay open. That distinction matters: Beijing isn't trying to stop the world from using Chinese AI. It wants to stop the world from owning a copy of it.
US Companies Now Run Nearly Half Their AI Traffic on Chinese Models A CNBC investigation published July 7 found Chinese-origin AI models carried at least 30% of US enterprise token volume on OpenRouter every week since February 8, peaking at 46%, up from just 4.5% in early 2025. DeepSeek and Alibaba's Qwen now lead OpenRouter's token volume outright, Moonshot AI just released its 2.8 trillion parameter Kimi K3... - US companies using Chinese AI models - why American businesses prefer cheaper AI
There's an obvious enforcement problem baked into that plan. Model weights already published are already out there, sitting on Hugging Face and in countless forks, and no regulation un-publishes a file that millions of people have already downloaded. Any restriction would bite mainly on the next generation of releases, not the ones developers are using today.
The Manus precedent is the tell #
If you want to know how serious Beijing is about this, look at what already happened to Manus. Meta announced a roughly $2 billion acquisition of the Chinese-founded AI agent startup in late December 2025. China's National Development and Reform Commission opened a probe almost immediately. In April 2026 it ordered the deal unwound, concluding it violated foreign investment and technology export rules even though Manus had already moved its headquarters and key staff to Singapore. Bloomberg reported that the NDRC's review escalated in March when authorities blocked the company's co-founders from leaving China and summoned them to Beijing for questioning. By June, Meta had cut Manus off from its internal data systems entirely and told staff to sunset the platform. That's the loophole the Financial Times says is now driving the broader push into agentic AI: a foreign buyer simply acquiring a Chinese AI company and walking out with the technology. The new catalog proposal is Beijing writing that lesson into standing policy instead of handling it case by case.
Officials have also discussed designating the leak or theft of proprietary Chinese AI technology as a national-security crime carrying harsher penalties. They've floated screening the type of foreign capital allowed to invest in domestic AI startups too. Both point the same way: tighter control over who gets a stake in, or a copy of, what China's AI companies build.
The timing isn't an accident. Taiwan has been tightening its own export rules on advanced AI chips bound for China. Micron and other chip names have taken a beating this month on fresh demand jitters. ASML is fielding its own questions about China exposure. Every piece of that fits a single, familiar story: two governments trying to control AI capability by controlling what crosses a border, just from opposite sides of the same transaction now.
Here's the part worth sitting with. For two years, the US chip embargo has been the thing forcing China to get resourceful, pushing Huawei's chip program and powering a DeepSeek-style arms race on training efficiency. Now China is applying that same instinct to its own exports. If the rule actually lands as described, the next DeepSeek-caliber model might never leave China as a downloadable file at all, only as an API call Beijing can watch and, if it wants to, shut off.
Also read: G7 Will Stage Fake AI Cyberattacks on Banks to Find the Cracks First • Accenture caps employee Claude spending at $29 a month after $1,000 bills • Taiwan Will Make It a Crime to Smuggle AI Chips Into China
China Will Let Its Top AI Firms Buy Nvidia's H200 Chips Again China plans to let Alibaba, ByteDance and DeepSeek buy Nvidia's H200 chips again, but approvals are capped below 200,000 units and reserved for training, not everyday inference work. The move, reported by The Information, follows a year of stalled U.S.-China chip diplomacy and comes as Huawei and homegrown rivals keep gaining ground inside China... - China allows Alibaba ByteDance H200 purchases - Nvidia H200 chip restrictions China eased
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.