China AI talent curbs widen race with Washington

China is tightening controls on top artificial-intelligence talent as competition with the United States broadens from chips and cloud capacity to the engineers who build the systems.

The new pressure is partly bureaucratic and partly corporate. Bloomberg reported that authorities have expanded overseas-travel approval requirements to senior AI professionals at private firms, including people connected to Alibaba and DeepSeek. Semafor described the same moment as a domestic talent war, with Beijing and Chinese tech companies trying to stop scarce expertise from leaking out of the country or across rival labs.

That makes the latest turn in the AI race less about one breakthrough model than about control over a mobile class of scientists, engineers and founders. Chips still matter. Data centres still matter. But China’s response suggests that the person who knows how to make a restricted chip useful, or how to compress a model around that constraint, is now being treated as strategic infrastructure.

People become infrastructure

Beijing’s immediate concern is not hard to see. The United States has used export controls to limit China’s access to advanced semiconductors, while American labs retain advantages in capital, cloud capacity and global recruitment. China has answered with domestic chip programmes, open-source models and industrial policy. Travel approvals add another instrument: keeping the people who can close the gap within reach of the state.

The reported rules bring private-sector AI workers closer to the treatment already associated with sensitive academic or state-linked research. According to Bloomberg’s account, some personnel now need approval from relevant authorities before foreign trips.

“they need approval from relevant authorities before embarking on overseas travel”

Source: Bloomberg News

Reuters carried the Bloomberg report as a sign that the restrictions had reached prominent private AI actors, not just formal state laboratories. That is the important boundary shift. It places commercial engineers inside a national-security perimeter, even when their day jobs are product launches, model tuning or infrastructure management.

For regulators in Beijing, the logic is defensive. If elite AI people travel freely, they attend foreign conferences, meet recruiters, advise investors and build networks that make departure easier. If their travel is slower and more visible, the state gains a measure of control over those networks. The trade-off is that AI research depends heavily on those same conferences, informal contacts and cross-border exchanges.

Companies bid up loyalty

The state is not acting alone. Chinese firms are also trying to hold their teams with money, equity and internal status. The Financial Times reported that ByteDance offered special stock tied to its AI business unit to fend off poaching. Semafor also cited an $18 million offer for a robotics chief scientist, a number that puts the contest for technical leadership closer to elite finance or sports than a normal engineering labour market.

That corporate bidding changes the read-out from a simple crackdown. Beijing can restrict some movement, but companies still have to persuade senior people that staying is worth it. Shares, bonuses and prestige are the softer side of the same retention campaign. They can work when employees believe the domestic market will keep expanding. They may look different if travel approvals start to limit collaboration or signal that exit options are closing.

This is where the policy contains its own risk. The people China most wants to retain are often the least trapped by salary alone. Top AI researchers can move between universities, start-ups, major platforms and foreign labs. Overseas-trained engineers, in particular, may read travel controls not as a patriotic appeal but as a warning about future constraints. TechCrunch argued that the borders are closing for top Chinese AI researchers, a framing that could discourage some returnees even as it keeps others in place.

Still, the short-term gains for Chinese companies are real. Retention buys continuity. It keeps model teams intact through product cycles. It limits the loss of tacit knowledge that does not appear in papers, patents or public benchmarks. That knowledge is exactly what export-control regimes struggle to measure.

The US lead narrows

The personnel fight matters because China is no longer far behind on model performance. TechCrunch, citing the Stanford AI Index, said the US-China model performance gap had fallen to 2.7 per cent from 31 per cent in 2023. That kind of narrowing makes talent mobility more politically sensitive. When a gap is large, a departing engineer may look like a loss. When the gap is small, the same departure can look like a strategic leak.

Washington has noticed the same centre of gravity. In a Politico interview, David Sacks put the scale of the talent pool bluntly.

“Something like half the world’s AI developers are Chinese.”

Source: David Sacks, Politico

That line explains why the contest cannot be reduced to Nvidia chips or export licences. Hardware bottlenecks can slow a programme. They do not erase a large engineering base. Nor do they prevent researchers from finding cheaper training methods, model distillation techniques or deployment strategies that need less advanced hardware. People are the part of the supply chain that can route around constraints.

For the United States, the policy question is the mirror image of China’s. Washington wants to deny Beijing the tools that make frontier AI easier, but it also benefits from attracting Chinese-born scientists, graduate students and founders. A broad American campaign that treats those people mainly as a security risk would hand Beijing a recruiting argument. A Chinese campaign that treats them mainly as assets to be contained could do the same for the United States.

A harder race to measure

The result is a more complicated AI race than the one usually captured in benchmark tables. Model scores can be compared. Chip shipments can be counted. Travel approvals and retention grants are harder to quantify, but they may shape who is available to solve the next bottleneck.

China’s approach also shows how closely commercial AI is now tied to state power. DeepSeek and Alibaba are not nuclear labs. ByteDance is best known globally as TikTok’s parent. Yet the talent around these firms is being discussed in the same strategic vocabulary as high-end chips and defence-adjacent science. That is a signal to investors and governments alike: the boundary between private AI competition and national technological capacity is thinning.

There is no guarantee the strategy works. Restrictions can keep people physically closer while making them professionally more restless. High compensation can retain a team while inflating costs and deepening rivalry among domestic firms. A government can slow overseas travel, but it cannot easily manufacture the open exchange that produces some of the best research.

Near term, Beijing may accept those costs. The country has watched the US use chip controls to shape the pace of Chinese AI development. Its answer is to defend the advantage it still has in people, especially engineers who know how to build under constraint. That does not settle the race. It does make clear that the next phase will be fought through passports and payrolls as much as processors.