The median h-index is 55 at universities and 23 at frontier AI labs

How do universities and frontier AI labs compare on published research? We estimate a median h-index of 55 at universities and 23 at frontier labs. Career length, current publishing and job moves help explain what that difference can mean.

Key findings

• We estimate a median h-index of 55 at universities and 23 at frontier labs.
• 38 of 108 lab researchers have an h-index above the median of university peers with a similar publishing career length.
• The median citation count during 2023–2025 was 201 at universities and 305 at frontier labs.
• We count 125 moves from universities to frontier labs by 120 people.

Start with the published work

The h-index is the largest number h for which h papers each have at least h citations. The first chart compares those records across current employers.

We define frontier labs here as Google DeepMind, OpenAI, Anthropic, Meta Superintelligence Labs, xAI, Mistral AI, DeepSeek, Thinking Machines Lab and Safe Superintelligence. Google Brain counts with Google DeepMind. We assign publication records to current employers, including work from earlier jobs. People with both lab and university posts count with the labs. Meta Superintelligence Labs (5), Thinking Machines Lab (5) and Mistral AI (2) have fewer than 10 counted people. Those records remain in the frontier-lab group.

Bars for the median h-index of universities, frontier labs, other companies and of 9 organizations.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. h-index from OpenAlex; uncertain records left out.

Data table

Group or organization	Kind	Median h-index	People counted
All universities	Universities	55	1,345
All frontier labs	Frontier labs	23	108
All other companies	Other companies	36	252
All institutes and other labs	Institutes and other labs	23	120
Massachusetts Institute of Technology	Universities	65	51
Stanford University	Universities	68	47
Carnegie Mellon University	Universities	60	42
EPFL	Universities	47	31
University of California, Berkeley	Universities	64	30
University of Oxford	Universities	60	29
Google DeepMind	Frontier labs	33	67
Anthropic	Frontier labs	19	18
OpenAI	Frontier labs	14	11
Meta Superintelligence Labs	Frontier labs	fewer than 10 people	5
Thinking Machines Lab	Frontier labs	fewer than 10 people	5
Mistral AI	Frontier labs	fewer than 10 people	2

Within the lab group, Pushmeet Kohli has the highest h-index, at 100. The group median is 23. These figures use 1345 university records and 108 lab records with confident OpenAlex matches.

Bars for the median number of works a year in 2023–2025, for the same groups and organizations.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. Works per year from OpenAlex counts_by_year.

Data table

Group or organization	Kind	Median works a year, 2023–2025	People counted
All universities	Universities	11	1,345
All frontier labs	Frontier labs	3.2	108
All other companies	Other companies	3	252
All institutes and other labs	Institutes and other labs	3.7	120
Massachusetts Institute of Technology	Universities	10.7	51
Stanford University	Universities	8.3	47
Carnegie Mellon University	Universities	11.5	42
EPFL	Universities	16.7	31
University of California, Berkeley	Universities	8	30
University of Oxford	Universities	10.3	29
Google DeepMind	Frontier labs	4.3	67
Anthropic	Frontier labs	2	18
OpenAI	Frontier labs	0.7	11
Meta Superintelligence Labs	Frontier labs	fewer than 10 people	5
Thinking Machines Lab	Frontier labs	fewer than 10 people	5
Mistral AI	Frontier labs	fewer than 10 people	2

The university median is 11 works a year. The lab median is 3.2 works a year. Annual works histories are available for 1345 university people and 108 lab people. Each median needs at least 10 records. OpenAlex counts published works. Models, code and internal reports kept inside a lab remain outside these figures. Different publication incentives are a possible reason for the pattern.

How much does career length matter?

A longer career gives papers more time to collect citations. Career length is therefore a possible reason for an h-index gap. The university median is 34 years since a first paper. The lab median is 17 years. The median lab career began 17 years later. Among people with a known birth year, the university median is 1962 (624 people). The lab median is 1979 (24 people).

A scatter plot of 1451 people by years since their first paper and h-index. University people are teal and frontier-lab people are ochre.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. h-index and works per year from OpenAlex. The first paper is the first year with a work that is followed by another within 3 years.

Data table

Person at a frontier lab	Years since first paper	h-index	University people with a similar career	Their median h-index
Kyle Fish	2	3	5	2
Samuel Marks	3	2	6	3.5
Liane Lovitt	4	12	9	6
Michal Zielinski	5	2	13	7
Saffron Huang	5	11	13	7
Alejandro Escontrela	6	7	14	7
Clemens Winter	6	3	14	7
Tamay Besiroglu	6	9	14	7
Tom Henighan	6	17	14	7
Anja Hauth	7	6	18	7
Sebastian Borgeaud	7	20	18	7
Teven Le Scao	7	12	18	7
Alexander H. Liu	8	13	29	12
Clemens Meyer	8	12	29	12
Jason Wei	8	20	29	12
Jeremy Bernstein	9	6	31	14
Jiahui Yu	9	38	31	14
Paul Jäger	9	19	31	14
Thomas Hubert	9	11	31	14
Eric Wieser	10	9	41	15
Esin Durmus	10	20	41	15
Ethan Perez	10	26	41	15
Gabriel Goh	10	14	41	15
Jack Rae	10	22	41	15
Julian Schrittwieser	10	16	41	15
Long Ouyang	10	9	41	15
Mateusz Litwin	10	5	41	15
Sherry Moore	10	7	41	15
Diego de las Casas	11	13	48	20
Luca Weihs	11	19	48	20
Luke Metz	11	16	48	20
Matej Balog	11	7	48	20
Matthew Lai	11	9	48	20
Chris Olah	11	19	48	20
Aja Huang	12	7	60	21
Noam Brown	12	18	60	21
Trapit Bansal	12	14	60	21
Iason Gabriel	13	23	63	24
Jan Leike	13	21	63	24
Thang Luong	13	13	63	24
Anna Potapenko	14	14	75	27
Arvind Neelakantan	14	18	75	27
Mehdi Mirza	14	16	75	27
Michael Figurnov	14	14	75	27
Nicholas Carlini	14	48	75	27
Samuel R. Bowman	14	48	75	27
Ilge Akkaya	15	11	88	30
Andrej Karpathy	15	19	88	30
Lisa Anne Hendricks	15	33	88	30
Sander Dieleman	15	27	88	30
Lilian Weng	15	14	88	30
Josh Abramson	16	10	97	31
John Jumper	16	26	97	31
Levent Alpöge	16	6	97	31
Edward Grefenstette	17	31	103	33
Fabio Viola	17	13	103	33
Kaiming He	17	78	103	33
Yutian Chen	17	18	103	33
David Warde-Farley	18	23	109	36
Hartwig Adam	18	35	109	36
Heng-Tze Cheng	18	19	109	36
Ronen Eldan	18	23	109	36
Demis Hassabis	19	92	110	37
Jonathan T. Barron	19	68	110	37
Koray Kavukcuoglu	19	67	110	37
Sébastien Bubeck	19	42	110	37
Timothy Lillicrap	19	57	110	37
Zeb Kurth-Nelson	19	36	110	37
Allan Dafoe	20	33	114	38
Marc Deisenroth	20	41	114	38
Piotr Dollár	20	65	114	38
Shakir Mohamed	20	38	114	38
Aniruddha Kembhavi	21	37	129	41
Hado van Hasselt	21	30	129	41
Laurens van der Maaten	21	44	129	41
Sebastian Nowozin	21	44	129	41
Daan Wierstra	22	45	151	48
Dumitru Erhan	22	36	151	48
Greg S. Corrado	22	61	151	48
Meredith Ringel Morris	22	71	151	48
Noah Snavely	22	68	151	48
Pushmeet Kohli	23	100	159	50
Swarat Chaudhuri	23	33	159	50
Anton Korinek	24	42	169	54
Tara Sainath	24	65	169	54
Abbas Mehrabian	25	15	174	60
Simon Osindero	25	29	174	60
Arthur Gretton	27	54	164	61
Boaz Barak	27	49	164	61
Jason Liu	27	33	164	61
Yue Yang	27	17	164	61
Thore Graepel	29	51	178	61
Doina Precup	31	52	181	62
Shane Legg	31	26	181	62
Marcus Hutter	31	36	181	62
Olaf Ronneberger	31	51	181	62
Andrew W. Senior	32	55	191	60
Michael Isard	32	54	191	60
Trevor Blackwell	32	8	191	60
Christoph Bregler	33	34	199	62
Martin Riedmiller	33	46	199	62
Irfan Essa	36	63	218	62
Peter L Bartlett	36	69	218	62
Tino Cuéllar	36	12	218	62
Zoubin Ghahramani	37	94	213	62
Murray Shanahan	39	48	189	60
David J. Fleet	41	72	164	58
Isabelle Guyon	41	68	164	58

A missing early paper makes a publishing career appear shorter. The career can begin before or after a doctorate.

Each lab person is compared with university people whose first paper came within 2 years of theirs. There are 108 lab people with at least 5 university peers. Of these, 38 have an h-index above the peer median. 67 are below it. 3 are level with it. At the median ratio, the lab h-index is 12% below its university comparison group.

The median h-index per year of career is 1.6 at the labs. The university median is 1.6. The bars below compare the groups within career-length bands.

Pairs of bars for 4 bands of career length: 0–9 years, 6 at universities, 11 at frontier labs, 10–19 years, 29 at universities, 19 at frontier labs, 20–29 years, 54 at universities, 43 at frontier labs, 30–39 years, 61 at universities and 49.5 at frontier labs.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. h-index from OpenAlex.

Data table

Years since first paper	University people	Median h-index	Frontier-lab people	Median h-index
0–9 years	23	6	19	11
10–19 years	163	29	49	19
20–29 years	315	54	24	43
30–39 years	404	61	14	49.5

Who received citations recently?

We count citations received during 2023–2025, for work from any publication year.

Bars for the median citations received during 2023–2025, for the same groups and organizations.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. Citations received by year from OpenAlex counts_by_year.

Data table

Group or organization	People with citation histories	Median citations received, 2023–2025	People with a period works ratio	Median period citations / period works
All universities	1,345	201	1,307	6
All frontier labs	108	305	104	33
All other companies	252	112	225	9
All institutes and other labs	120	71	110	6
Massachusetts Institute of Technology	51	271	46	10
Stanford University	47	290	43	12
Carnegie Mellon University	42	170	41	5
EPFL	31	320	31	6
University of California, Berkeley	30	498	29	13
University of Oxford	29	196	29	5
Google DeepMind	67	441	65	28
Anthropic	18	290	18	24
OpenAI	11	173	9

We estimate a median of 305 citations received at frontier labs during this period. The university median is 201. Annual citation histories are available for 1345 university people and 108 lab people. Each median needs at least 10 records. The table also divides citations received by works published in the period. The ratio combines different paper sets and has limited value for measuring paper quality. EPFL and the University of California, Berkeley have a higher median citation count and more works a year than the lab group.

Where did the lab researchers come from?

To understand the lab group, we also need its earlier employers. We find 126 people with earlier university jobs among 191 current frontier-lab staff. Education is recorded for 131. Across all employers, universities had a net outflow of 670 full moves.

Of 191 current lab staff, 27 also hold a university post. Moves come from dated career records; missing positions can change the apparent balance between sectors.

Stacked columns of 125 moves from universities to frontier labs from the 2010–14 period to the 2025–26 period, split by whether the person kept the university post.

Source: People of Superintelligence career records from biographies and OpenAlex, as of 15 October 2026. Google DeepMind and Google Brain count as one lab.

Data table

Year	Person	From	To	Kept the university post
2010	Christopher Summerfield	University of Oxford	Google DeepMind	Yes
2010	Demis Hassabis	University College London	Google DeepMind	Yes
2010	Shane Legg	Università della Svizzera italiana	Google DeepMind	No
2011	Vijay Vasudevan	Carnegie Mellon University	Google DeepMind	No
2012	Aja Huang	National Taiwan Normal University	Google DeepMind	No
2012	Shakir Mohamed	University of British Columbia	Google DeepMind	No
2013	Alex Graves	University of Toronto	Google DeepMind	No
2013	Daan Wierstra	EPFL	Google DeepMind	No
2013	Koray Kavukcuoglu	Courant Institute of Mathematical Sciences	Google DeepMind	No
2013	Marc Bellemare	University of Alberta	Google DeepMind	No
2013	David Silver	University College London	Google DeepMind	No
2013	Tomas Mikolov	Brno University of Technology	Google DeepMind	No
2014	Arthur Guez	University College London	Google DeepMind	No
2014	Edward Grefenstette	University of Oxford	Google DeepMind	Yes
2014	Karen Simonyan	University of Oxford	Google DeepMind	No
2014	Nicolas Heess	University College London	Google DeepMind	No
2014	Phil Blunsom	University of Oxford	Google DeepMind	Yes
2015	Alexander Pritzel	LMU Munich	Google DeepMind	No
2015	Diederik P. Kingma	University of Amsterdam	OpenAI	Yes
2015	Andrej Karpathy	Stanford University	OpenAI	Yes
2015	Scott Reed	University of Michigan	Google DeepMind	No
2015	Timothy Lillicrap	University of Oxford	Google DeepMind	No
2015	Wojciech Zaremba	New York University	OpenAI	No
2015	Yutian Chen	University of Cambridge	Google DeepMind	No
2015	Yuval Tassa	University of Washington	Google DeepMind	No
2016	Aäron van den Oord	Ghent University	Google DeepMind	No
2016	Dario Amodei	Princeton University	OpenAI	No
2016	Bernardino Romera-Paredes	University of Oxford	Google DeepMind	No
2016	Chris J. Maddison	University of Toronto	Google DeepMind	Yes
2016	Colin Raffel	Columbia University	Google DeepMind	No
2016	Julian Schrittwieser	TU Wien	Google DeepMind	No
2016	Laurent Sifre	École Polytechnique	Google DeepMind	No
2016	Matthew Lai	Imperial College London	Google DeepMind	No
2016	Olaf Ronneberger	University of Freiburg	Google DeepMind	Yes
2016	Prafulla Dhariwal	Massachusetts Institute of Technology	OpenAI	No
2016	John Schulman	University of California, Berkeley	OpenAI	No
2016	Ashish Vaswani	University of Southern California	Google DeepMind	No
2016	Zeb Kurth-Nelson	UCL Queen Square Institute of Neurology	Google DeepMind	No
2017	Csaba Szepesvári	University of Alberta	Google DeepMind	Yes
2017	David Luan	Yale University	OpenAI	No
2017	John Jumper	University of Chicago	Google DeepMind	Yes
2017	Jan Leike	University of Oxford	Google DeepMind	Yes
2017	Murray Shanahan	Imperial College London	Google DeepMind	Yes
2017	Paul Christiano	University of California, Berkeley	OpenAI	Yes
2017	Thomas Hubert	Stanford University	Google DeepMind	No
2017	Yee Whye Teh	University of Oxford	Google DeepMind	Yes
2018	Ari Morcos	Harvard University	Google DeepMind	Yes
2018	Clemens Meyer	HEC Paris	Google DeepMind	No
2018	Fabio Viola	University of Cambridge	Google DeepMind	No
2018	Lucas Beyer	RWTH Aachen University	Google DeepMind	No
2018	Bob McGrew	Stanford University	OpenAI	No
2018	Michael Bowling	University of Alberta	Google DeepMind	Yes
2018	Nicholas Carlini	University of California, Berkeley	Google DeepMind	No
2018	Sebastian Borgeaud	University of Cambridge	Google DeepMind	No
2019	Hyeonwoo Noh	Pohang University of Science and Technology	OpenAI	Yes
2019	Hyung Won Chung	Massachusetts Institute of Technology	Google DeepMind	No
2019	Jared Kaplan	Johns Hopkins University	OpenAI	Yes
2019	Josh Abramson	University of California, Berkeley	Google DeepMind	No
2019	Manfred K. Warmuth	University of California, Santa Cruz	Google DeepMind	Yes
2019	Marcus Hutter	Australian National University	Google DeepMind	Yes
2019	Mehdi Mirza	Université de Montréal	Google DeepMind	No
2019	Michael Figurnov	National Research University Higher School of Economics	Google DeepMind	No
2019	Sebastian Ruder	Ollscoil na Gaillimhe – University of Galway	Google DeepMind	No
2019	Lilian Weng	Indiana University	OpenAI	No
2020	Anna Potapenko	National Research University Higher School of Economics	Google DeepMind	No
2020	David Warde-Farley	Université de Montréal	Google DeepMind	No
2020	Jay McClelland	Carnegie Mellon University	Google DeepMind	No
2020	Qi Liu	Hubei University of Technology	Google DeepMind	Yes
2020	Saffron Huang	Harvard University	Google DeepMind	No
2020	Richard Sutton	University of Alberta	Google DeepMind	Yes
2020	Jason Wei	Dartmouth College	Google DeepMind	Yes
2021	Aravind Srinivas	University of California, Berkeley	OpenAI	No
2021	Arthur Gretton	University College London	Google DeepMind	Yes
2021	Florian Tramèr	Stanford University	Google DeepMind	No
2021	Helen Toner	Georgetown University	OpenAI	Yes
2021	Iason Gabriel	St John's College, University of Oxford	Google DeepMind	No
2021	Ilge Akkaya	University of California, Berkeley	OpenAI	No
2021	Lisa Anne Hendricks	University of California, Berkeley	Google DeepMind	No
2021	Sam McCandlish	Stanford University	Anthropic	No
2021	Thibault “Tibo” Sottiaux	Université catholique de Louvain	Google DeepMind	No
2022	David C. Parkes	Harvard University	Google DeepMind	Yes
2022	Evan Shelhamer	University of California, Berkeley	Google DeepMind	No
2022	Dileep George	Stanford University	Google DeepMind	No
2022	Kate Larson	University of Waterloo	Google DeepMind	Yes
2022	Matej Balog	University of Cambridge	Google DeepMind	No
2022	Trapit Bansal	University of Massachusetts Amherst	OpenAI	No
2023	Ben Wang	University of Pennsylvania	OpenAI	No
2023	Danqi Chen	University of Chinese Academy of Sciences	Google DeepMind	Yes
2023	Esin Durmus	Stanford University	Anthropic	No
2023	Ethan Perez	New York University	Anthropic	No
2023	Jade Leung	University of Oxford	OpenAI	No
2023	Jimmy Ba	University of Toronto	xAI	Yes
2023	Liang Wenfeng	Technische Universität Dresden	DeepSeek	No
2023	Marc Deisenroth	University College London	Google DeepMind	Yes
2023	Arthur Mensch	École Polytechnique	Mistral AI	No
2023	Samuel R. Bowman	West Virginia University	Anthropic	Yes
2024	Boaz Barak	Harvard University	OpenAI	Yes
2024	Carroll L. Wainwright	University of California, Santa Cruz	OpenAI	No
2024	Daniel Levy	Stanford University	OpenAI	No
2024	Daniel Levy	Stanford University	Safe Superintelligence	No
2024	Eric Wieser	University of Cambridge	Google DeepMind	No
2024	Nick Ryder	University of California, Berkeley	OpenAI	No
2024	Stanislav Fort	Stanford University	Google DeepMind	No
2025	Alexander H. Liu	Massachusetts Institute of Technology	Mistral AI	No
2025	Arvind Neelakantan	University of Massachusetts Amherst	Google DeepMind	No
2025	Danqi Chen	University of Chinese Academy of Sciences	Thinking Machines Lab	No
2025	Hongyu Ren	Stanford University	OpenAI	No
2025	Hongyu Ren	Stanford University	Meta Superintelligence Labs	No
2025	Jiahui Yu	Nanjing University of Posts and Telecommunications	OpenAI	No
2025	Joe Benton	University of Oxford	Anthropic	No
2025	Ludwig Schmidt	University of Washington	Anthropic	No
2025	Nicholas Carlini	ETH Zurich	Anthropic	No
2025	Ronen Eldan	Weizmann Institute of Science	OpenAI	No
2025	Samuel Marks	Harvard University	Anthropic	No
2025	Shengjia Zhao	Stanford University	OpenAI	No
2025	Shengjia Zhao	Stanford University	Meta Superintelligence Labs	No
2025	Swarat Chaudhuri	University of Texas at Austin	Google DeepMind	No
2025	Tim Rocktäschel	University College London	Google DeepMind	No
2025	Yilun Du	Harvard University	Google DeepMind	No
2026	Abbas Mehrabian	McGill University	Google DeepMind	No
2026	Bo Li	University of Chicago	Meta Superintelligence Labs	No
2026	Jeremy Bernstein	California Institute of Technology	Thinking Machines Lab	No
2026	Kaiming He	Massachusetts Institute of Technology	Google DeepMind	No
2026	Levent Alpöge	Harvard University	Anthropic	No
2026	Sanmi Koyejo	Stanford University	Meta Superintelligence Labs	No

The records show 125 moves to frontier labs by 120 people. In 32 of these moves, the person kept the university post. The busiest period is 2015–19, with 47 moves. The list below ranks universities by recorded moves to frontier labs.

Google DeepMind received 79 of those moves (63%), the most of any lab. In 23 of its incoming moves, the person kept the university post.

1. Stanford University15Andrej Karpathy, Bob McGrew, Daniel Levy, Dileep George, Esin Durmus, Florian Tramèr, Hongyu Ren, Sam McCandlish, Sanmi Koyejo, Shengjia Zhao, Stanislav Fort and Thomas Hubert
2. University of Oxford10Bernardino Romera-Paredes, Christopher Summerfield, Edward Grefenstette, Jade Leung, Jan Leike, Joe Benton, Karen Simonyan, Phil Blunsom, Timothy Lillicrap and Yee Whye Teh
3. University of California, Berkeley9Aravind Srinivas, Evan Shelhamer, Ilge Akkaya, John Schulman, Josh Abramson, Lisa Anne Hendricks, Nicholas Carlini, Nick Ryder and Paul Christiano
4. Harvard University7Ari Morcos, Boaz Barak, David C. Parkes, Levent Alpöge, Saffron Huang, Samuel Marks and Yilun Du
5. University College London7Arthur Gretton, Arthur Guez, David Silver, Demis Hassabis, Marc Deisenroth, Nicolas Heess and Tim Rocktäschel
6. University of Cambridge5Eric Wieser, Fabio Viola, Matej Balog, Sebastian Borgeaud and Yutian Chen

The records also show 29 moves from a frontier lab to a university: Aditya Grover to Stanford University in 2016, Yutian Chen to Xi'an Jiaotong University in 2016, Chris J. Maddison to the University of Oxford in 2017, Greg S. Corrado to Stanford University in 2017, Edward Grefenstette to University College London in 2018, Ioannis Antonoglou to University College London in 2018, Jack Rae to University College London in 2018, David Silver to University College London in 2018, Timothy Lillicrap to University College London in 2018, Colin Raffel to the University of North Carolina at Chapel Hill in 2020, Thore Graepel to University College London in 2020, Jeff Clune to the University of British Columbia in 2021, Dani Yogatama to the University of Southern California in 2022, Denny Zhou to Stanford University in 2022, Florian Tramèr to ETH Zurich in 2022, Nicholas Carlini to ETH Zurich in 2022, Amanda Askell to the University of Oxford in 2023, Caglar Gulcehre to EPFL in 2023, Michal Zielinski to Polish-Japanese Academy of Information Technology in 2023, Nicolas Heess to University College London in 2023, Sebastian Ruder to the University of Washington in 2023, Yutian Chen to Carnegie Mellon University in 2023, Azalia Mirhoseini to Stanford University in 2024, Christopher Summerfield to University College London in 2024, Kristina N. Toutanova to Sofia University "St. Kliment Ohridski" in 2024, Iason Gabriel to School of Advanced Study in 2025, Iason Gabriel to the University of London in 2025, Deep Ganguli to Stanford University in 2026 and Leonidas Guibas to Department of Computer Science, Stanford University in 2026. In 18 of them the person kept the lab post.

20 people who once studied or worked at Stanford University work at a frontier lab now: Andrej Karpathy, Christoph Bregler, Daniel Levy, Dario Amodei, Douwe Kiela, Esin Durmus, Greg S. Corrado, Hongyu Ren, James Bradbury, Jared Kaplan, Jascha Sohl-Dickstein, Joanne Jang, Meredith Ringel Morris, Mike Krieger, Sam McCandlish, Samuel R. Bowman, Shengjia Zhao, Thang Luong, Thomas Hubert and Zico Kolter. The career records hold 15 moves from Stanford University straight to a frontier lab: Andrej Karpathy to OpenAI in 2015, Thomas Hubert to Google DeepMind in 2017, Bob McGrew to OpenAI in 2018, Florian Tramèr to Google DeepMind in 2021, Sam McCandlish to Anthropic in 2021, Dileep George to Google DeepMind in 2022, Esin Durmus to Anthropic in 2023, Daniel Levy to OpenAI in 2024, Daniel Levy to Safe Superintelligence in 2024, Stanislav Fort to Google DeepMind in 2024, Hongyu Ren to OpenAI in 2025, Hongyu Ren to Meta Superintelligence Labs in 2025, Shengjia Zhao to OpenAI in 2025, Shengjia Zhao to Meta Superintelligence Labs in 2025 and Sanmi Koyejo to Meta Superintelligence Labs in 2026.

14 people who once studied or worked at the Massachusetts Institute of Technology work at a frontier lab now: Adam Lerer, Alexander H. Liu, Douglas Chen, Hyung Won Chung, Irfan Essa, Matt Clifford, Nat Friedman, Paul Christiano, Prafulla Dhariwal, Sam Gross, Tamay Besiroglu, Tara Sainath, Zico Kolter and Zoubin Ghahramani. The career records hold 4 moves from the Massachusetts Institute of Technology straight to a frontier lab: Prafulla Dhariwal to OpenAI in 2016, Hyung Won Chung to Google DeepMind in 2019, Alexander H. Liu to Mistral AI in 2025 and Kaiming He to Google DeepMind in 2026.

What does it take to train a model?

We examine training resources as a possible reason to move into industry.

Two bars: 93 notable models from industry and 2 from academia in 2025.

Source: AI Index Report 2026, Chapter 1: Research and Development, Stanford HAI, 13 April 2026, from Epoch AI data.

Data table

Sector	Notable models, 2025
Industry	93
Academia	2

The Stanford AI Index reports: “In 2025, Epoch AI identified two notable AI models originating from academia, compared to 93 from industry.” (Stanford HAI). It describes academic output over the past decade: “Within academia, Tsinghua University (26), Stanford University (26), and Carnegie Mellon University (25) have been the most prolific over the past decade.” The earlier edition gives a possible reason for the sector gap: “Creating cutting-edge AI models now demands a colossal amount of data, computing power, and financial resources that are not available to academia.” (Stanford HAI)

The resources needed for a training run have been growing. Epoch AI reports: “In short, we find that the amount of compute used to train notable models has grown about 4.1x/year (90% CI: 3.7x to 4.6x) between 2010 to May 2024.” (Epoch AI) The training-cost study reports: “The analysis reveals that the amortized cost to train the most compute-intensive models has grown precipitously at a rate of 2.4x per year since 2016 (90% CI: 2.0x to 2.9x).” (arXiv 2405.21015) Epoch AI's estimates put GPT-4 training at about $79 million and Llama 3.1-405B at about $170 million (Stanford HAI).

The compute-divide study tracks academia's share of large-scale models: “In the early 2010s, around 65% of these models originated from academic labs. However, this figure fell to approximately 10% during the initial years of the 2020s.” (arXiv 2401.02452) The academic-resources survey reports: “In fact, the vast majority (85%) of respondents reported zero budget for cloud compute.” Its description makes the constraint concrete: “Our survey suggests a common range for what constitutes “academic hardware” today: 1–8 GPUs—especially RTX 3090s, A6000s, and A100s—for days (typically) or weeks (at the higher-end) at a time.” (arXiv 2410.23261)

Sources

• AI Index Report 2026, Chapter 1: Research and Development. Stanford HAI, 13 April 2026.
• AI Index Report 2025, Chapter 1: Research and Development. Stanford HAI, 7 April 2025.
• Training compute of frontier AI models grows by 4-5x per year. Epoch AI, 28 May 2024.
• The rising costs of training frontier AI models (Cottier and others). arXiv 2405.21015, 31 May 2024.
• The Compute Divide in Machine Learning: A Threat to Academic Contribution and Scrutiny? (Besiroglu and others). arXiv 2401.02452, 4 January 2024.
• $100K or 100 Days: Trade-offs when Pre-Training with Academic Resources (Khandelwal and others). arXiv 2410.23261, 30 October 2024.