Beijing Extends Travel Curbs to Families of Top AI Staff, as Chip Tools Go Open Source

Bloomberg reported the restriction on 28 September, and Business Standard carried it the same day. Direct relatives of certain AI and chip executives now need approval from Beijing before going abroad, even for short trips. Their work is judged paramount to state security, according to people familiar with the matter cited by the outlet. It is not clear whether every family affected falls under the new rule.

China had already restricted foreign travel for top AI professionals at private firms including Alibaba and DeepSeek since earlier this year, Bloomberg News reported in May. The same report says separate rules enforcing exit bans took effect on 15 September. They cover cases from criminal investigations to possible violations of industrial and technological security law, and more names will be added over time.

The timing is not accidental. Beijing's fear of losing technology and people sharpened after Meta's $2 billion acquisition of Manus, a deal the government eventually ordered the US firm to unwind, according to the same Bloomberg account. The policy may force engineers with global ambitions to choose between staying home and leaving earlier in their careers.

Software, not just silicon

In the same week, China's AI industry moved in the opposite direction on openness. Deepseek released open-source programming tools for Huawei's Ascend chips, according to a post on its official WeChat channel, with Reuters reporting the details. The centrepiece is TileLang, a language developed by researchers at Peking University. Deepseek has used it for about a year and now calls it its main tool for work on artificial general intelligence, The New York Times reports.

The two companies also optimised a supernode, a cluster of 128 Ascend 950 chips, and Huawei "fully supported" the work, according to Deepseek.

The target is Nvidia's software moat. The Decoder notes that Nvidia's dominance rests partly on an estimated four million developers who build with CUDA. Rivals such as AMD have not crossed that ecosystem even when their hardware looked comparable on paper. Huawei's own CANN stack was the only one besides CUDA to support DeepSeek V4 on day one, according to analysis by SemiAnalysis cited in the same piece. Huawei has admitted it cannot meet demand at home and will sell fewer chips abroad. Rotating chairman Eric Xu said the company cannot accept a future that hinges on whether others are willing to sell chips to China.

SemiAnalysis also tested Jalapeño, OpenAI's inference chip, and called the CUDA moat "potentially dead" because OpenAI gets new models running on its own hardware quickly. The firm added caveats: it tested only relatively easy scenarios, and it did not run the AgentX benchmark for multistep agent tasks. In that benchmark, tested in August, Nvidia remained well ahead.

A home for models

Open weights need somewhere to live. Rest of World reports that Beijing blocked Hugging Face in 2023, opening a domestic market that Alibaba's ModelScope and OSChina's MoArk now serve. ModelScope, launched in 2022, says it hosts more than 170,000 models and 250 million users. MoArk, launched in 2023, serves some 20,000 models, according to OSChina chief executive Xu Yong.

Hugging Face still hosts more than 3 million open models, and in September Nvidia announced it was acquiring the company for $12.9 billion. In the filing for that deal, Nvidia flagged the risk that regulators could ban Chinese models from the site. Rebecca Arcesati of the Mercator Institute for China Studies told Rest of World that Chinese concerns about disrupted access are real. From Beijing's perspective, she said, it would be ideal if the whole AI stack, software tools included, were indigenised as far as possible.

Xu Yong's argument is blunter: not everyone can use a VPN all the time, so China needs a self-reliant AI ecosystem for Chinese speakers. "In the AI era, China is developing an independent ecosystem faster than in the internet era," he told Rest of World.

Export or waste

Inside China, the pressure shows up differently. Investor David Cheng spent a week in Beijing and Shanghai and wrote in Earned Intuition that the traffic was the lightest he had seen in a decade of trips. Youth unemployment stood at 17.9% as of July, he wrote, and July retail sales rose 0.6% year on year against an expected 1.5%. His through line is 内卷, involution: hundreds of labs undercutting each other on price, poaching researchers and shipping monthly, producing engineering excellence and almost no profit.

A line he heard put it better: if you don't go overseas, all that grinding was for nothing. Involution without export is wasted effort. That is the frame in which the travel curbs and the open-source releases belong together. Beijing wants the know-how to stay; the companies want the market to leave.

Not everyone reads the direction of travel the same way. The Justice Department and FBI announced court-authorised seizures of two hacking platforms, QScan and QTRouter, used by a PRC state-sponsored group called QTFY, employed by Nanjing Xinjiuwei Network Technology. Targets listed in court documents unsealed in the Southern District of California include NASA, the Federal Reserve, the Department of Energy and the US Senate. The seizure order is dated 26 August, so it is background here rather than news.

The export question also runs through hardware. Electrek reported on 30 September that Geely's 2.2MW chargers are already in service in five Chinese cities, with 15,000 sites planned by the end of next year, while EVgo's 750kW system in the US has slipped from 2026 to 2027. Two different clocks, running at two different speeds.