Opinion

China Overtakes the US in Top AI Talent Share: 40.6% vs 34.2%

A Carnegie Endowment report shows China has edged past the US in the share of authors at top AI conferences like NeurIPS. The headline number obscures more than it reveals about where frontier research actually happens.

Font size

The Carnegie Endowment for International Peace dropped a report last week with a single striking data point: China 40.6%, US 34.2%.

It marks the first time China has surpassed the US in "share of top AI talent" by this particular metric. The story hit the front page of Hacker News, with coverage from aiweekly.co.

The number looks simple. Break it apart and it gets complicated.

What the data actually says

The Carnegie report measures something specific: the nationality and institutional affiliation of authors at top AI conferences, primarily NeurIPS.

Not GitHub stars. Not arXiv paper counts. Not H-index rankings. Carnegie chose "first authors and corresponding authors at top conferences" — a group that, in academic convention, represents "people actually doing frontier research."

40.6% vs 34.2%, a 6.4 percentage point gap on a large enough sample to be statistically meaningful. But several details tend to get glossed over:

  • The definition of "top": is it the top 20% or top 10%? Carnegie's threshold is "anyone who has published at least one paper at NeurIPS," which is not an especially high bar.
  • Nationality vs. affiliation: papers list the author's institutional affiliation, which often differs from nationality. A Chinese passport-holder doing a postdoc at MIT typically gets counted as "US-based."
  • Sample bias: NeurIPS itself skews toward the Anglophone research ecosystem. Publishing at NeurIPS from China generally requires strong English writing and often overseas experience, which already filters out a large slice of purely domestic researchers.

So the 40.6% figure roughly translates to: "share of NeurIPS-publishing AI researchers with some China connection (nationality or institution)," not "share of the world's top 100 AI researchers who are Chinese."

How fast did this overtake happen

In 2019, the ratio was roughly inverted — US 60%+, China under 30%. Exact numbers vary by source, but the order of magnitude is clear.

Five years later, the US is at 34.2% and China is at 40.6%. What happened in between?

On the US side:

  • H-1B tightening starting in 2020 under the Trump administration, with visa rejection rates climbing.
  • Chip export controls beginning late 2022, making it harder for Chinese students and researchers to access Nvidia's high-end GPUs.
  • Salary inflation at frontier labs (OpenAI, Anthropic) pulling talent out of academia — universities can't keep up.
  • A broader political climate that's turned less welcoming for Chinese researchers since 2023. Nothing official, but FBI cases and export-control enforcement speak for themselves.

On the China side:

  • Continued output from Tsinghua's Yao Class, BAAI, Fudan, and similar institutions.
  • Rising return rates of overseas PhDs — noticeably more people coming back in the past two years.
  • Domestic GPUs (Huawei Ascend, Cambricon) building workable alternatives under restricted conditions.
  • Open-source releases like DeepSeek and Qwen giving domestic researchers confidence that SOTA results are possible outside the Nvidia ecosystem.

But going from 60:30 to 34:40 in five years is not a linear story. There's at least one structural inflection point in there — possibly ChatGPT's release in late 2022 triggering a surge in Chinese government investment, or the 2023 chip ban turning "doing research in China" into something of a patriotic choice.

What the numbers leave out

Carnegie's framing carries a hidden assumption: NeurIPS authors roughly equal the world's top AI talent.

Swap in a different lens and the picture changes considerably:

Lens 1: Author nationality (Carnegie's choice)
China 40.6% > US 34.2%

Lens 2: Where people actually work (LinkedIn, company pages)
The US still leads. Core researchers at OpenAI, Anthropic, Google DeepMind, and Meta AI overwhelmingly work in the US. Many of them are Chinese by nationality, but they're doing the work inside US labs.

Lens 3: Where the breakthroughs actually happen
This one is fuzzier, but milestones like Transformer, GPT-4, Claude 3.5, and Gemini were built primarily by teams at US companies. Qwen and DeepSeek are Chinese-team achievements, but international impact is still catching up.

So 40.6% really represents the stock of China-trained AI researchers capable of publishing at NeurIPS, not "China's share of actual AI innovation." Those are very different things.

The next battlefield in the talent war

The talent competition is entering a second half, and the playbooks are diverging.

US side:

  • Loosening visa policy (STEM PhD green cards are under discussion across both Biden and Trump administrations).
  • Compute advantage — Nvidia H100/B200 supply is restricted domestically, but allies (Saudi Arabia, UAE) can still get them.
  • Startup ecosystem — compensation packages at OpenAI, Anthropic, and xAI that academia simply can't match.
  • Ironically, continued immigration tightening could accelerate the talent flow back to China.

China side:

  • Tsinghua's Yao Class continuing to expand, with selection starting at the undergraduate level.
  • Return incentives — many cities offer "Thousand Talents"-level packages.
  • Industry landing spots multiplying — large-model companies are everywhere, from ByteDance to DeepSeek to Moonshot, with compensation closing in on US top-tier offers.
  • Domestic chip ecosystems maturing — Ascend + MindSpore, Kunlunxin + PaddlePaddle. Performance gaps remain, but the stacks run.

Over the next five years, the variable that matters isn't "who trains more talent." It's who retains talent and who lets that talent produce breakthroughs. Two different problems.

Editor's take

The 40.6% vs 34.2% number sends its strongest signal not at the top, but in the middle.

The top 5% — the people actually defining AI's future — are still at US companies, and that hasn't changed much in five years. Carnegie's data pool ("anyone who has published at NeurIPS") is large enough that the number isn't being moved by the heads; it's being moved by the 40th-to-60th percentile PhD cohort.

These are researchers capable of publishing at top venues but who haven't yet landed tenure-track positions or chief scientist titles. They're the group most sensitive to policy environment — tighten the visa, bump the salaries, restrict the chips, and their flows shift.

So this number is neither proof of "China's AI rise" nor "America's AI decline." It's a cross-section of mid-tier talent migration trends.

My read:

  1. The number will keep tilting toward China, but the slope flattens — most of the people who were going to return have already returned.
  2. The real swing variable sits in 2026–2028: if US visa/immigration policy tightens further, a second wave of returns kicks in; if it loosens, this trend stalls.
  3. Carnegie's methodology has a blind spot. It undercounts the contribution of Chinese researchers working in the US to American AI, so the 34.2% US figure may be systematically understated.

One caveat: I haven't found the full Carnegie PDF — only the aiweekly.co recap and Hacker News thread. The exact methodology, especially the specific threshold for "top," needs to be checked against the original report. If Carnegie's definition differs from what I've assumed above, some of these calls need to be revised.

Source: https://aiweekly.co/alerts/carnegie-china-passes-us-as-top-ai-talent-hub-406-to-342