7,308 prior employers · 3,900 companies · snapshot of 21 August 2026
The frontier labs hire from 96 countries. Three in four people came from one.
We found the last job every person at OpenAI, Anthropic, Google DeepMind, xAI and Thinking Machines Lab held before they arrived, and placed that employer on the map.
The quarter that did not come through an American employer is distributed so unevenly that each lab has a recruiting signature you can read off the map, and no two of the five look alike.
People whose previous employer we could place to a country.
Arrived from outside the United States, the quarter this report is about.
Of DeepMind’s people came from a US employer, against 83% at OpenAI.
xAI’s over-representation among people arriving from Turkey.
1 · The map
Every country that feeds a frontier lab
Each person’s last completed job before joining, coloured by the country that employer calls home. Hover any country for the split across the five labs.
The United States accounts for 5,383 of the 7,308 (74%), so the shading is mostly a picture of the other 26%. No country here is the work of one employer: Britain’s 644 people come from 457 different companies, Turkey’s 51 from 47. Singapore, Hong Kong and nine other city-states and micro-territories hold 65 people between them and cannot be drawn at this resolution, so they are absent from the map and present in every figure elsewhere on this page.
2 · The finding
Each lab recruits from a different world
The bar is the share of a lab’s people who arrived from a US employer. The chips are the countries where that lab takes far more than its expected share, meaning its share of a country’s people divided by its share of all 7,308. At 1.0 times a lab takes exactly what its size predicts. At 3 times it takes three times that.
The only lab where American employers are a bare majority. 683 of its 1,403 people arrived from a non-US employer, spread across 53 other countries, and the signature is Europe.
The strongest pull in southern Europe and Latin America of any lab here. Across Turkey, Spain, Italy and Brazil together, 83 of 147 people went to xAI, which is 3.6 times its expected share.
No strong national channel anywhere. Both figures above rest on fewer than 30 people each, so neither is a pipeline so much as a coincidence that cleared the threshold.
Reaches more countries than anyone, 72 of them, and over-indexes in none. OpenAI hires the global average, at the largest scale in the set.
Seven people in the entire company arrived from a non-American employer. Fifty people across six countries cannot carry a signature and none is claimed.
Shares are of each lab’s placed previous employers: OpenAI 3,032 people, Anthropic 1,678, DeepMind 1,403, xAI 1,145, Thinking Machines 50. Multiples are shown only where a country holds at least 20 people and the lab holds at least 8 of them. Below that, one referral chain moves the multiple by a full point.
The same numbers as a table
| Lab | People | From US | US share | Countries |
|---|---|---|---|---|
| Google DeepMind | 1,403 | 720 | 51% | 54 |
| xAI | 1,145 | 743 | 65% | 61 |
| Anthropic | 1,678 | 1,365 | 81% | 51 |
| OpenAI | 3,032 | 2,512 | 83% | 72 |
| Thinking Machines | 50 | 43 | 86% | 6 |
3 · Outside the United States
Britain is the second pipeline, and it is mostly one lab’s
The eight largest non-American sources. Britain sends more people than the next four countries together, and 60% of them go to Google DeepMind.
Britain’s 644 is larger than Canada, India, France and Germany combined (502). Below Britain the gaps close fast: Turkey, Singapore and Japan sit within ten people of each other, so their order should not be read as meaningful.
4 · What follows from this
Only one of the five is not an American company with a tail
DeepMind reads like the London lab it is. At 51% American it is the only one of the five where a non-US base is large enough to change the shape of the organisation rather than decorate it. 27% of its people arrived through a British employer, and the four countries it over-indexes in after Britain are Switzerland, France, Japan and Germany. That is a European research network, and none of the other four labs has anything like it.
xAI is the only lab recruiting outside the rich-world core. DeepMind over-indexes abroad too, but every country on its list is Switzerland, France, Japan or Germany. xAI’s list runs Turkey, Spain, Italy, Brazil and India. The first four send 147 people between them and 83 of those went to a single lab holding 16% of the sample. A recruiting pattern that concentrated usually means a person rather than a policy: somebody with a network in those markets, hiring out of it. It is the finding on this page most likely to change within a year, because it rests on relationships rather than institutions.
OpenAI got the flattest map by getting the largest one. It touches 72 countries, more than any other lab, and over-indexes in none of them. Scale does that. Once a lab is hiring at the rate OpenAI is, it stops having a channel anywhere and starts having the global average everywhere, which is a different kind of advantage and a harder one to compete against.
The gap this leaves is the interesting one. Continental Europe contributes 480 people across the whole sample, which is fewer than Britain sends on its own. Any lab, sovereign programme or startup recruiting against these five is competing hardest in the handful of places they already saturate and barely at all in the rest. Nobody is contesting Turkey except xAI.
We have established the pattern, not its cause. An over-index can be a deliberate recruiting office, one well-connected employee, a language, a time zone, or an acquisition that arrived as a block. This data separates none of those, and every one of them implies something different about whether the pattern holds next year.
How we did this
- Source
- Public professional-profile employment histories, snapshot of 21 August 2026. Nobody in this sample was contacted or interviewed.
- The measure
- For every person currently listed at one of the five labs we take the most recent completed role before they joined and read the home country of that employer. That leaves 7,308 people across 96 countries, out of roughly 21,000 current employees at the five labs. The rest either list no prior employer, list one we cannot resolve to a country, or fall out at the filters below.
- Multinationals
- Somebody who joined from Google London, Microsoft Bangalore or Amazon Dublin lists an employer whose page resolves to a single American headquarters, so counting every prior employer would file all of them under the United States. We keep only employers with fewer than 5,000 people, where a company’s country is a fair proxy for where its staff sat. That removes 58% of prior roles, and it is the single largest decision on this page. Measured without it, Google DeepMind, a London lab, reads 80% American and 10% British, because Google alone supplies more prior roles than the next eleven employers combined. Measured this way it reads 51% American and 27% British.
- Students
- Universities employ people, so we keep genuine academic jobs: professors, lecturers, university staff. But profiles often list a degree or a campus society as though it were employment. We drop those by title, covering assistantships, undergraduate roles, doctoral positions, and the non-English words for intern that an English-only filter misses, such as stajyer and werkstudent. That removes 44 people, 0.6% of the sample.
- Concentration
- A national figure would be worthless if it were a single employer or one referral chain. It is not: the 7,308 people come from 3,900 distinct employers. Britain’s 644 arrive from 457 companies, Turkey’s 51 from 47, and the most concentrated country in the set still averages two people per employer.
- Known limits
- Small-employer geography tilts toward countries with the largest startup ecosystems, so the United States would lead this map on any measure and the 74% figure should be read as an upper bound rather than a point estimate. We count only people still at the five labs, so anyone who has since moved on is absent, which biases the sample toward recent arrivals. Profiles update with a lag. City-states and micro-territories holding 65 people cannot be drawn at this projection and appear in the figures but not the map.
The counts behind every chart on this page are available as a CSV, under a CC BY 4.0 licence. Reuse it freely, with a link back to this page.
How to cite this
Trayo Research. “The frontier labs hire from 96 countries. Three in four people came from one.” September 2026. https://www.trayo.ai/research/ai-talent-origins
Charts and figures may be reproduced or redrawn with attribution. If you would like a cut of the data we have not published, such as a single country broken out by employer, email [email protected] .
The next one asks why
This report shows the pattern. The follow-up tries to separate a deliberate recruiting office from one well-connected employee, which is the difference between a channel that holds and one that closes. Leave an address and we will send it when it lands. No product pitch, no weekly newsletter.