The Talent Ledger Flips: China Overtakes the US as the Top Destination for Elite AI Researchers

For most of the past decade, the geography of frontier AI was settled fact: the United States trained less, imported more, and employed the best. A new study from Carnegie China, published this week and first reported by The Information, upends that assumption with a single pair of numbers. Of the world’s top-tier AI researchers in 2025, 40.6% now work in China — against 34.2% in the United States. In 2022, the same methodology had the US leading 46.4% to 27.1%. Within three years, a nineteen-point American advantage has become a six-and-a-half-point deficit.

What the study actually measured

Carnegie China’s team, led by Matt Sheehan, tracked the institutional affiliations of elite AI researchers — the population that publishes at top venues such as NeurIPS, ICML, and ICLR — building on the methodology pioneered by MacroPolo’s Global AI Talent Tracker. Two findings stand out.

First, the destination flip. China is now the single most common workplace for the world’s best AI researchers, a title the US had held continuously in every prior survey wave. The drop is not an artifact of a shrinking pool: the absolute number of top researchers keeps growing. It is a shift in where they choose to work.

Second, the return-rate collapse of the “stay rate.” In the 2025 NeurIPS cohort, 69% of Chinese-origin AI researchers now work in China, up from 57% in 2022. For decades, the implicit deal of American AI was simple: the world’s brightest came to US universities, and most stayed, feeding Google, Meta, OpenAI, and a hundred startups. That pipeline is now leaking in the other direction. Carnegie also tracked 100 top Chinese-origin researchers who were at US institutions in 2019 and found that by 2025, ten had relocated to Chinese companies or universities — small in absolute terms, but each move transfers frontier know-how that took years of American investment to build.

Why researchers are going back

The report frames the shift as a “reverse brain drain” and a “strategic knowledge transfer,” and the drivers are structural rather than cyclical.

Labs that can compete. DeepSeek’s headline-grabbing efficiency results, Moonshot’s Kimi line, Alibaba’s Qwen family, and Xiaomi’s MiMo models have demonstrated that Chinese labs now work on genuinely frontier problems with frontier-scale compute. A researcher choosing between a US lab and a Chinese one in 2026 is no longer choosing between a Ferrari and a bicycle.

Money and status at home. Chinese cities and companies have spent years refining returnee packages — housing subsidies, generous equity, titles, and, critically, research budgets insulated from the funding volatility that has hammered US academic AI. Where a US professor juggles grants and GPU rationing, a returning scientist often gets a lab fully stocked on day one.

Friction in the United States. The residue of the DOJ’s China Initiative, visa uncertainty for Chinese students and scholars, and a political climate that periodically questions the loyalty of Chinese-origin scientists have raised the perceived cost of staying. None of these factors alone explains the flip; together they shift the calculus at the margin, and talent decisions are made at the margin.

A deeper domestic bench. The Economist’s March 2026 analysis of the same trend noted that by 2025, 68% of China’s top AI researchers were home-grown through their entire training path — up from 58% in 2022 — meaning China no longer needs the US pipeline to staff its labs. Some of its most celebrated systems were built by teams that never left.

Why it matters

Talent is the tightest bottleneck in AI, tighter than chips or capital. Training runs can be rented, algorithms leak across borders within weeks, but the cohort of people who can push a frontier model forward numbers in the low thousands worldwide. Every percentage point of that cohort that relocates is a direct transfer of capability that export controls cannot intercept.

The timing is uncomfortable for Washington. The entire premise of US AI strategy — technology restriction lists, chip export bans, allied coordination — assumes American labs hold a durable lead worth protecting. Carnegie’s own December 2025 analysis argued that American strength remained real but was “built on Chinese talent.” Nine months later, the foundation of that strength is measurably eroding. You cannot simultaneously restrict a country’s access to chips and expect its researchers to keep staffing your frontier labs; the study is the first clean quantitative evidence that the bill for that contradiction has come due.

There are honest caveats. Aggregate researcher counts are a lagging indicator, and employment location says nothing about where the most consequential individual breakthroughs happen. US labs retain unmatched capital markets, and the same week’s news cycle — OpenAI shipping GPT-6 Sol and Luna, Anthropic cutting Opus 5.5 prices by 40% — shows American companies still setting the commercial pace. Nor does the study capture the thousands of researchers in neither country, from London to Singapore, whose share is quietly growing.

But direction matters more than level. A metric that moved monotonically toward the US for a decade has now reversed, and the forces behind the reversal — competitive domestic labs, richer returnee packages, visa friction, and a deepening home-grown bench — are all still strengthening. The 2026 cohort data, when it arrives, will show whether 2025 was a blip or a hinge. On current evidence, it looks like a hinge.

The uncomfortable conclusion for policymakers: the AI race may not be decided by who has the most GPUs, but by who has the most people capable of using them well. On that ledger, the numbers now favor China.