Call her Dr. Lin. She is hypothetical: no single documented 2025–2026 case matches every detail this scene needs, so rather than invent a biography and print it as reportage, this series builds her as a labeled composite from the real pattern below.1 What she carries across the border is not hypothetical at all.
This is Part 1 of Hormuz and the Weighting Game, a six-part investigation into the one-way transfer of AI advantage from the United States to China. Start with the Reader Protocol if you haven't. It names the three axes, the fourth thread, and the game theory this series argues against, not around. This part is the horizontal axis: talent. Not headcount. The direction of the marginal, elite, US-trained flow, and what a state's immigration and export-control apparatus can and cannot detain.
/ 01The Tally
Four of the best-documented departures from 2025, ranked by how much biographical detail is on the public record. None of them returned to Shenzhen. All of them are real, named, and independently corroborated.
| Researcher | US institution & tenure | Destination & role | What the new employer handed them |
|---|---|---|---|
| Wu Yonghui | Google / Google DeepMind — 17 years (2008–2025) | ByteDance Seed, VP of Foundational Research | Led Doubao 2.0, a ~1-trillion-parameter multimodal model |
| Yao Shunyu | OpenAI — ~1 year (2024–2025) | Tencent, Chief AI Scientist, age 28 | Reported ~$14M package (unconfirmed by Tencent) |
| Cao Ting | Microsoft Research Asia — multi-year, edge AI & inference systems | Tsinghua Institute for AI Industry Research | Joined an institute run by a fellow MSRA alumnus |
| Guo-Jun Qi | Microsoft / Huawei Research USA — ~10 years, IEEE Fellow | Westlake University, Hangzhou | Full-time faculty, 20-researcher lab (MAPLE) |
Notice what the last column is doing, and what it isn't. It is not a headcount. A single departure that lands a 28-year-old a Chief Scientist title at one of China's largest technology companies, or hands a 17-year Google veteran the technical leadership of a trillion-parameter model, is not equivalent to one name subtracted from a roster. The receiving institution is not settling for leftovers. It is paying up, in title, in resourcing, in reported compensation, for something more specific than a warm body: years of tacit, uncodified experience building frontier systems inside the two or three labs on earth that had already done it.
On the "cost savings" a receiving lab books when it hires a returnee rather than growing the same expertise from scratch: no single sourced figure exists for this, and this series won't invent one. The closest verified proxy is what Chinese labs are visibly willing to pay for it: Yao Shunyu's reported eight-figure package, Tencent's willingness to hand a newly created AI Infrastructure Department to a researcher one year removed from OpenAI. A lab does not pay a premium for what it could build in-house on the same timeline. The premium is the tally.
/ 02What the Crossing Actually Carries
Here is the composite, and here is exactly what it's built from. Dr. Lin spent six years at a US frontier lab. The specific number varies across the real cases behind her, from Yao Shunyu's one to Wu Yonghui's seventeen, and the Hoover Institution's study of DeepSeek's own US-experienced cohort found that length of US tenure does not predict who returns: five-year veterans go home at about the same rate as one-year visitors.3 She was not fired, not blacklisted, not the subject of an investigation. She was, in the framing several of the real departures use, drawn: Guo-Jun Qi's own words on leaving a decade in the US for Westlake were "the free-spirited atmosphere," not visa trouble; Fu Tianfan cited "China's growing investment in higher education" and proximity to family.6 Not every return in the record is push. Some are simply pull, and a series arguing a one-way transfer has to say so plainly rather than manufacture grievance where the sources report ambition.
What she carried across the border is the part no manifest lists. Not a laptop: those get imaged, or left behind, or wiped. Not a paper: those are published, searchable, already read by whoever wanted to read them. What travels is the residue of years spent debugging a training run at 3 a.m., the intuition for which hyperparameter sweep is a waste of compute and which one is the whole ballgame, the specific, unpublishable judgment that only accumulates from having been the one holding the pager when a frontier model's loss curve did something nobody in the room had a name for. AnnaLee Saxenian's research on Taiwanese and Indian engineers returning from Silicon Valley in the 1990s and 2000s gave this pattern a name before AI existed to test it: the "new argonauts" don't just bring capital home, they bring the tacit, socially embedded knowledge of how a frontier cluster actually operates (the norms, the informal networks, the pattern-matching that no patent filing or published benchmark captures), and they use it to build a second cluster that competes with, rather than merely imports from, the first.7 Shenzhen's own robotics and AI-institute buildout (more than a hundred robotics companies now clustered there, the International Digital Economy Academy staffed in part by Microsoft Research Asia alumni, Light Robotics keeping an office in the city even while headquartered in Singapore) is the receiving end of exactly that argonaut logic, whether or not any single documented case landed there on the date this scene needs.8
This is also where Keohane and Nye's distinction between sensitivity and vulnerability earns its keep, and why the Reader Protocol asked you to hold onto it. A US lab is acutely sensitive to Dr. Lin's departure: it shows up immediately, in team composition, in the Slack channel that goes quiet, in the project handed to someone two years junior. But sensitivity is not the same as vulnerability, which is the cost of the best available alternative. The US higher-education and immigration pipeline can, in principle, replace a departed researcher, at a cost and on a timeline. That is exactly why the aggregate numbers still look fine for America, and exactly why this piece has to be honest about what "one-way" actually means before it argues it.9
And there is a reason the tools built to stop this kind of transfer keep missing it. James C. Scott's diagnosis of the "high-modernist" state (legible grids, censuses, tariff schedules, export-control lists) is that it sees what can be counted and misses everything that resists counting: the informal, the tacit, the practice passed hand to hand rather than filed in a form.10 A customs officer can inspect a laptop, seize a hard drive, flag a chip shipment against an Entity List. No customs officer, no visa category, no export-control rule has ever been built to detain a trained intuition. The apparatus that spent 2025 tightening around silicon, the subject of The Stack's own teardown of the chip layer, was never going to be the apparatus that could hold onto Dr. Lin, because it was built to see hardware, and what she carries was never hardware.
The apparatus counts what it can see
Export-control lists, visa quotas, and customs declarations are built around countable, inspectable objects: a chip, a laptop, a wire transfer. Each new Entity List tightening adds precision to exactly this kind of object.
What travels is not one of those objects
What a returning researcher carries (debugging intuition, a feel for which architecture choice will scale, the tacit judgment behind a training run) was never filed anywhere an inspector could check it against a list.
The receiving cluster monetizes it in months, not years
A returnee lands inside an already-forming cluster (Shenzhen's robotics buildout, Tsinghua's AIR, Tencent's new AI Infrastructure Department) and the tacit knowledge becomes a shipped model, a funded lab, a C-suite title, faster than any visa-policy rewrite could have been debated, let alone passed.
/ 03The Honest Reckoning
Here is the steelman, stated as strongly as the sourced data allows, before this piece argues past it. MacroPolo's Global AI Talent Tracker (the field's standard-reference dataset, built from 4,622 researchers who published at NeurIPS, ICML, and ICLR) says the United States still employs 59% of the world's elite AI researchers. Of the AI researchers educated in China, 72% still work at US institutions. The tracker's own headline framing is a "5.5x" US retention edge over China.9 If the claim on the table were "the aggregate flow of AI talent runs one way, toward China," the aggregate data says the opposite, decisively, and a piece that cited only the return cases above while ignoring this would be fact-checkable-wrong on its own terms.
The bidirectional flow is real in a second sense, too, and it is happening inside the same US labs whose departures open this piece. Meta's 2025 Superintelligence Labs build-out hired Shengjia Zhao, a Chinese-origin co-creator of ChatGPT, away from OpenAI as chief scientist, along with additional researchers including Jiahui Yu and Shuchao Bi, also recruited from OpenAI.11 That is a genuine poaching war, and it captures some of the same profile of talent this piece is tracking eastward. And China's government is not a passive beneficiary quietly receiving gifts: reporting from May 2026 documents Beijing requiring government approval before top AI researchers, founders, and executives may travel abroad, advising top AI founders to avoid US travel as early as March 2025, and blocking Manus AI's co-founders from leaving the country during a review of a blocked Meta acquisition.12 A state that hoards its own talent by fiat is not the passive receiving end of a one-way gift. It is playing the same strategic game the United States is, with the same instrument: state control over the movement of trained people, just pointed inward instead of left to run on autopilot.
So the honest version of this axis is narrower than "talent flows to China." It is this: atop a much larger stock that still, decisively, favors the United States, the marginal flow among the already-elite, US-trained, frontier-lab cohort has tilted toward China since 2025 (documented in the Hoover Institution's 71% figure, in the headhunters' 30-plus relocations, in named C-suite hires a serious company does not make on a whim) at the exact moment US policy raised the price of entry (the $100,000 H-1B fee) and Chinese policy lowered it (the K-visa) within the same month.5 The more structural number sits one layer beneath the individual departures: Tsinghua's own reporting of a fall in engineering graduates applying to US PhD programs, from roughly half before the pandemic to roughly a fifth by 2026. Not stars leaving, but a pipeline that stopped refilling.4 That is the harder trend to reverse, because it never shows up as a headline departure. It shows up ten years later, as an absence no one can quite point to.
Several of the figures above (the 30-plus headhunter-assisted relocations, the Tsinghua application-rate collapse, Yao Shunyu's reported compensation) trace to secondary write-ups of a paywalled Financial Times piece this research pass could not independently retrieve, corroborated across three converging outlets but not verified against the original.4 Treat them as corroborated-but-unverified-at-source, not settled fact, until someone pulls the FT original.
One more data point complicates the push narrative further, and it belongs in the steelman rather than buried in a footnote. In 2024, Microsoft offered relocation (to the US, Australia, or Ireland) to roughly a thousand of its top China-based AI and Azure engineers. Only about a third accepted.14 That is two-thirds of a cohort declining an open door to the United States, handed to them directly by their own employer, no visa lottery or H-1B fee involved. A pure visa-friction story cannot explain that number. What it suggests instead is closer to Saxenian's argonaut logic running in reverse at the point of origin: for a meaningful share of this cohort, China is not merely tolerable, it is where they would rather build. Separately, Princeton researchers have tallied roughly fifty tenure-track scholars of Chinese descent leaving US universities for China in the first half of 2025 alone, adding to more than 850 such departures since 2011: a slower-moving, academic-track version of the same pattern, sourced with lower confidence (cited secondhand across multiple outlets, primary Princeton publication not independently located) but consistent in direction with everything else in this section.15
/ 04Who Holds the Return
Stack up this layer's ledger. The United States still holds the larger stock of the world's elite AI talent, by a wide and verified margin. That has to be said plainly, because the series that hides it isn't arguing in good faith. What the United States does not hold, on the numbers this piece could source, is the marginal direction: the specific, elite, frontier-experienced cohort that a rational lab pays a premium for, moving disproportionately toward Chinese institutions since the DeepSeek moment reset what "competitive" looks like on a fraction of the compute.3 That correction did not start with an ideology. It started with a release: a January 2025 model release that the next part of this series takes apart directly.
The energy axis this series shares with Social Physics, Article 7 explains part of why the destination matters as much as the departure: a returnee lands in an economy that can, on Cembalest's own verified numbers, absorb and retrain and redeploy at a marginal power cost the US grid is not currently positioned to match.13 But the deeper reason this axis resists the tools built to stop it is the one Scott's theory of the state predicts and Saxenian's fieldwork confirms: a government can count chips, license exports, and audit a manifest. It has no comparable instrument for a trained mind, because the state was never built to see one.
“The chip embargo had removed the hardware. But the hardware had already been mapped by a mind that no customs officer could detain.”
Dr. Lin's laptop, if she carried one across that border, got the same scrutiny every laptop gets. What she actually brought — the six years, the pager, the judgment about which loss curve meant trouble — cleared customs the moment she smiled and said she was visiting family. Part 2 follows what happens when that judgment meets an open-weight model China didn't have to steal, because the race that made it valuable started in Hangzhou, not California.