In 2013, one of the world’s most influential artificial intelligence (AI) researchers packed up his academic lab and moved to industry. It wasn’t just a career change; it was a signal, a signal that something deeper was shifting in how knowledge is created.

A decade later, that shift has become a transformation.

Artificial intelligence is no longer primarily a university-driven endeavour. It is increasingly shaped inside a handful of large, powerful firms. Debates over AI policy typically take the pace of AI innovation as given and concentrate primarily on research activity within industry. For example, André et al. (2025) and Gambacorta and Shreeti (2025) offer compelling accounts of the competitive dynamics among large technology firms in the AI sector. As AI research is shifting away from universities, the nature of innovation is changing.

For decades, universities were the beating heart of frontier research. They trained talent, produced open science, and seeded new ideas into the broader economy. But in AI, that model is breaking down. Using a new database tracking the employment histories of 42,000 AI researchers, in a recent paper we show that top talent has increasingly migrated from universities to large incumbent firms, as the salary gap between industry and academia rose more than fivefold, reaching $1.5 million in 2021 (Akcigit et al. 2026).

The allocation of AI research talent has broad implications for the pace and direction of AI innovation. Traditionally, universities have served as open knowledge platforms, generating broadly diffused ideas through open publication and the training of graduate students. Firms, facing different incentives and constraints, typically focus on proprietary innovation, limiting the flow of insights to other firms and researchers. The type of firm matters as well. Large incumbent firms, in particular, tend to focus on incremental innovation, protecting rents associated with existing technological paradigms (Akcigit and Kerr 2018, Akcigit and Goldschlag 2025). On the other hand, unlike universities, large firms can procure the massive computational resources necessary for frontier AI research.

Studying these issues requires a clear view of the allocation of AI research talent, which we provide using a new database linking published papers to administrative employer-employee data housed at the U.S. Census Bureau. The findings contribute to an emerging consensus that private AI labs are increasingly absorbing top talent (Ahmed et al. 2023, Jurowetzki et al. 2025). Unlike prior work, which either relied only on publication data or surveys, our administrative data permit more granular analysis of job transitions. Analysing the characteristics and employment dynamics of AI researchers reveals the following ten key facts:

1. By 2019, 68% of AI researchers worked in industry, up from 48% in 2001.
2. The share of AI researchers born in the US and working in industry has declined by 5.5 percentage points; this decline is almost entirely accounted for by a rise in the share of researchers from China (+3.8 percentage points) and India (+2.0 percentage points).
3.  The median age of AI researchers in industry has fallen by 2 years (from 39 to 37) but has remained flat in academia (at 42).
4. The female share of AI scientists in academia has risen by 13 percentage points (from 16% to 29%) whereas in industry it has remained relatively flat (+4 percentage points, from 19% to 23%). Academia now has greater female representation than industry.
5. In industry, the ratio of female to male average earnings has modestly widened (from 73% to 72%); in academia, it has narrowed (from 79% to 83%).
6. Since 2001, average academic real salaries declined, and AI researchers in academia were no exception.
7. Coinciding with the image recognition revolution kicked off by AlexNet (Krizhevsky et al. 2012), top 1% earnings in industry exploded from $595,000 in 2001 to $1.94 million in 2021 (measured in 2015 dollars). Top academic salaries barely budged (increasing from $301,000 to $392,000).
8. Coinciding with the publication of the transformer paper “Attention Is All You Need” (Vaswani et al. 2017), transitions from academia to industry accelerated, and a growing share of AI academics’ income began to come from secondary employment.
9. Rising job transitions from academia to industry are driven by young researchers leaving to incumbent firms (firms with greater than or equal to 1,000 employees and greater than or equal to 20 years old) and to the Professional Services and Information sectors.
10. After researchers permanently transition from academia to industry, on average their paper-writing declines (65% fewer papers per year, 30 percentage points less likely to publish a paper), patenting increases (530% more patents per year, 6 percentage points more likely to patent), and their earnings rise by 63% relative to similar job switchers within academia.

The centre of gravity has moved out of universities

Figure 1 points to a steady migration of AI researchers into industry. In 2000, fewer than half were employed there; by 2019, nearly 70% were. Research outputs move in the same direction, but not in the same way. Industry already dominated AI patenting, and its share rose further, from 86% to 95%. Its share of AI papers increased, but only from 27% to 32%. The contrast in levels is useful. Papers are generally associated with open science and broad diffusion, while patents are built to secure ownership and enable commercial exploitation (Akcigit et al. 2021). Read together, these patterns suggest that the movement of AI researchers into industry has gone hand in hand with a shift towards more commercially oriented, more easily appropriable knowledge production.

Figure 1 AI activity share

Pull factors to industry

The same period saw a sharp rise in the pay of top AI researchers, especially in industry. The left-hand panel of Figure 2 shows that compensation at the top end began to take off in the 2010s. Among the top 1 percent of industry AI earners, annual pay more than tripled between 2001 and 2021, rising from about $595,000 to more than $1.9 million in 2015 dollars. In academia, by contrast, earnings growth was much more modest: even among the top 1%, pay rose by only about 30%.

Figure 2 AI pull factors

This jump in industry pay lined up with a series of breakthroughs that changed the economics of AI. ImageNet and the ImageNet competition created a powerful new benchmark, GPUs made it possible to train much larger models, and deep learning delivered dramatic performance gains, most famously with AlexNet in 2012 (Krizhevsky et al. 2012). As data, compute, and algorithms came together, the expected returns to frontier AI research rose sharply. Firms responded by investing heavily in computing infrastructure and competing more aggressively for specialised talent.

A second shift followed after 2017. Young AI researchers became much more likely to move to large incumbent firms, while moves to smaller or newer firms stayed roughly flat (Fright-hand panel of Figure 2). That timing matched the arrival of the transformer, introduced in “Attention Is All You Need” by Alphabet researchers in 2017 (Vaswani et al. 2017). Transformer models scaled especially well with data and compute, which gave an edge to big tech firms with large proprietary datasets, expensive infrastructure, and the resources to hire top researchers. The result was not just higher pay, but a growing concentration of AI talent inside large established companies.

What happens when AI researchers leave academia?

Figure 3 shows that when AI researchers leave academia for industry, the shift is not only in who pays them, but also in the kind of research they do. We compare researchers who move from academia to industry with similar researchers who also change jobs but remain within academia. That comparison matters because it helps separate the effect of moving jobs from the effect of moving sectors.

Figure 3 Publishing and earnings after a transition from academia to industry

The results point to a clear reorientation away from open scientific work and towards proprietary innovation. In the top-left panel, researchers who move to industry become substantially less likely to publish papers and more likely to patent. By three years after the move, they are about 30 percentage points less likely to write a paper and about 6 percentage points more likely to file a patent than comparable job switchers who stay in academia. The top-right panel shows the same pattern on the intensive margin: paper production falls sharply by 65%, while patenting rises markedly 530%.

The bottom panel shows why universities have such difficulty retaining talent. Pay increases immediately after the move and stays elevated. Three years after the transition, AI researchers in industry earn approximately 63% more than comparable researchers who move within academia – a premium that is markedly larger than the roughly 23% pay increase observed for non-AI academics making similar transitions. Taken together, the figures suggest that the movement of AI talent into industry changes the private returns to research and the form that research takes, shifting activity away from publication and toward commercialisable knowledge.

Conclusion

AI policy debates often focus on compute, chips, and models. Our findings suggest that talent allocation belongs in that conversation too. AI is not just a technological revolution, it is a reorganisation of where ideas live. As talent, resources, and discovery concentrate inside a handful of firms, the balance between openness and control is shifting. The risk is not that innovation slows, but that it becomes narrower, less diffused, and more dependent on who owns the compute. The challenge ahead is clear: ensuring the AI era remains not only fast, but also open, competitive, and broadly beneficial.

Authors’ note: Any opinions and conclusions expressed herein are those of the authors and do not represent the views of the U.S. Census Bureau. The Census Bureau has ensured appropriate access and use of confidential data and has reviewed these results for disclosure avoidance protection (Project 7517031: CBDRB-FY25-CES020-005, CBDRB-FY25-CES022-004, and CBDRB-FY25CES022-005). Part of this work was completed while Nathan Goldschlag was employed by the Census Bureau.

References

Ahmed, N, M Wahed, and N C. Thompson (2023), “The Growing Influence of Industry in AI Research,” Science 379(6635): 884–886.

Akcigit, U, C A Chikis, E Dinlersoz, and N Goldschlag, “Attention (And Money) Is All You Need: Why Universities Are Struggling to Keep AI Talent,” NBER Working Paper No. 34964.

Akcigit, U and N Goldschlag (2025), “Measuring the Characteristics and Employment Dynamics of U.S. Inventors,” Journal of Economic Growth 30(2): 237-269. 

Akcigit, U, D Hanley, and N Serrano-Velarde (2021), “Back to Basics: Basic Research Spillovers, Innovation Policy, and Growth,” The Review of Economic Studies 88(1): 1-43.

Akcigit, U and W R. Kerr (2018), “Growth Through Heterogeneous Innovations,” Journal of Political Economy 126(4): 1374-1443. 

André, C, M Bétin, and P Gal (2025), “Dynamism in generative AI markets since the release of ChatGPT,” VoxEU.org, 12 July.

Bergé, L (2025), “fixest: Fast Fixed-Effects Estimations”, GitHub repository. 

Correia, S, P Guimarães, and T Zylkin (2020), “ppmlhdfe: Fast Poisson Estimation with High-Dimensional Fixed Effects,” Stata Journal 20(1): 95–115.

Gambacorta, L and V Shreeti (2025), “Big techs’ AI empire,” VoxEU.org, 16 May.

Jurowetzki, R, D S Hain, K Wirtz, and S Bianchini (2025), “Private Sector Is Hoarding AI Researchers: What Implications for Science?,” AI & Society 40(5): 4145–4152.

Krizhevsky, A, I Sutskever, and G E Hinton (2012), “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances in Neural Information Processing Systems, Vol. 25, Curran Associates.

Vaswani, A, N Shazeer, N Parmar, J Uszkoreit, L Jones, A N Gomez, L Kaiser, and I Polosukhin (2017), “Attention Is All You Need,” arXiv:1706.03762 [cs].