Unlike universities, large firms can acquire vast computational resources.
Universities are losing talented AI researchers: they are leaving for high salaries at large companies. It is not just the employer that changes – research is also shifting, reorienting towards the commercialization of innovations, while knowledge becomes locked in corporate ownership.
AI has ceased to be a field promoted by science. This field is increasingly shaped by a group of large and influential companies. In discussions about AI policy, the pace of innovation is often taken for granted, with primary attention given to research activities in the business industry. For example, a number of researchers provide compelling descriptions of the dynamics of competition among major tech companies in the AI sector. As AI research "leaves" universities, the nature of innovation is also changing.
For decades, universities have been the heart of cutting-edge research. They nurtured talent, created open science, and enriched the entire economy with new ideas. But in the field of AI, this model is breaking down. In our recent work, using a database of the career trajectories of 42,000 AI researchers in the U.S. from 2001 to 2021, we show that the best specialists are increasingly moving from universities to large companies, as the salary gap between the private sector and academia has increased fivefold.
The distribution of talented AI researchers has significant implications for the pace and direction of innovation in AI. Traditionally, universities served as platforms for open knowledge, generating widely disseminated ideas through publicly available publications and training scientists. Companies, facing various incentives and constraints, typically focus on their own innovations and restrict access to them for other firms and researchers. The type of firm also matters.
Large players tend to concentrate on incremental innovations, protecting the rents associated with existing technological paradigms. However, unlike universities, large firms can acquire vast computational resources necessary for advanced AI research.
To study these issues, a clear understanding of the distribution of talent in AI research is necessary. We provide this through a new database that links published scientific articles with administrative data on workers and employers stored at the U.S. Census Bureau. The results allow us to conclude that private AI labs are increasingly "absorbing" talented specialists. Unlike previous works that relied either solely on publication data or surveys, our administrative data allow for a more detailed analysis of transitions between workplaces. From this analysis, ten key facts emerge.
By 2019, 68% of AI researchers in the U.S. worked in business – up from 48% in 2001. The share of AI researchers born in the U.S. and working in business decreased by 5.5 percentage points; this decline is almost entirely offset by an increase in the share of researchers from China (+3.8 p.p.) and India (+2 p.p.). The average age of AI researchers in business decreased by two years – from 39 to 37 years, while it remained unchanged in academia – 42 years. The share of women among AI researchers in academia increased by 13 p.p. – from 16% to 29%, while in business it remained relatively unchanged – up 4 p.p., from 19% to 23%. The representation of women in academia is now higher than in business. In the business sector, the gap in average earnings between men and women has slightly increased (from 27% to 28%), while in science it has decreased (from 21% to 17%). Since 2001, the average real salary overall in academia has decreased, and AI researchers have not been an exception. Concurrently with the revolution in image recognition that began with the AlexNet neural network, the annual income of the top 1% of the highest-paid specialists in the industry soared more than threefold from 2001 to 2021 (in 2015 fixed prices). Meanwhile, the salaries of leading scientists during this period have changed little, and the gap with the earnings of top researchers in the corporate sector has increased from nearly twofold to fivefold, reaching $1.5 million.
With the publication of the landmark paper by researchers at Google, "Attention is All You Need," the transition from academia to business accelerated, and AI researchers began to earn more income from side jobs. The increase in transitions from academia to business is driven by young researchers moving to large, mature companies (those over 20 years old and with more than 1,000 employees), as well as to the professional services sector and the information sector. After researchers fully transition from academia to business, their average publication activity decreases: 65% fewer articles per year, and the probability of publishing a paper drops by 30 p.p. However, patent activity increases: 530% more patents per year, and the likelihood of obtaining a patent increases by 6 p.p. Earnings grow by 63% compared to similar researchers who change jobs within academia.
During the period in question, the salaries of leading AI researchers, especially in business, rose sharply. In business, the annual income of the top 1% of the highest-paid AI specialists began to rise sharply in the 2010s, and from 2001 to 2021, it increased more than threefold, from about $595,000 to $1.94 million in 2015 dollars. In academia, income growth was much more modest: even among the top 1%, salaries increased by about 30%, from $301,000 to $392,000.
This salary jump in business coincided with a series of breakthroughs that changed the economics of AI. The ImageNet project in general and the "Olympics for Neural Networks" – a competition for large-scale image recognition in ImageNet in particular – created a new powerful standard, graphics processors enabled the training of much larger models, and deep learning provided impressive performance gains, the most famous example being the AlexNet neural network of 2012. As data, computation, and algorithms converged, the expected return from advanced AI research sharply increased. Companies responded with massive investments in computational infrastructure and more aggressive competition for skilled talent.
A second shift occurred after 2017. Young AI researchers began to transition much more frequently to large, established companies, while the share of transitions to smaller or newer firms changed little. This coincided with the emergence of the transformer, introduced in the paper by researchers at Google, "Attention is All You Need" (the transformer was a new architecture for neural networks at the time, allowing the model to analyze all words in a sentence simultaneously rather than sequentially, thanks to the self-attention mechanism).
Models using transformers scaled particularly well with data and computational power, giving an advantage to large tech companies with extensive proprietary datasets, costly infrastructure, and resources to hire leading researchers. The result was not only higher salaries but also an increasing concentration of AI talent in large, mature companies.
When AI researchers transition from academia to business, it is not just their employer that changes – the very nature of their research changes.
In the journal Econs, researchers compared those who transitioned from universities to business with similar researchers who also changed jobs but remained within academia. This comparison helps to separate the effect of changing workplaces from the effect of changing sectors.
Collectively, these data suggest that the transition of AI talent from academia to business alters the private return on research and the form of research itself, shifting the focus from publications to the commercialization of knowledge.
Discussions about AI policy often focus on computational power, chips, and models. Our findings show that the distribution of talent should also be part of this discussion. AI is not just a technological revolution; it is a reorganization of where ideas "live."
As talent, resources, and discoveries concentrate in the hands of a few companies, the balance between openness and control shifts. The risk is not in slowing down innovation but in making it narrower, less widespread, and more dependent on who owns the computational resources. The task is clear: to ensure that the era of AI remains not only fast but also open, competitive, and beneficial for all.
