← Back to feed

A working paper from the National Bureau of Economic Research has put hard numbers on a shift that has been visible anecdotally for years: top AI researchers have largely left universities, and the ones who crossed over publish less, patent more, and earn dramatically more.

The study, authored by Ufuk Akcigit, Craig Chikis, Emin Dinlersoz, and Nathan Goldschlag, links academic publication records to US Census Bureau employer-employee data to track 42,000 AI researchers across 20 years. The result is the first dataset combining bibliometric and administrative microdata for this population — capturing researchers after they stop publishing, the exact moment their trajectories became most interesting.

The Compensation Split

In 2001, the top 1% of industry AI researchers earned around $595,000 per year (in 2015 dollars). By 2021, that figure had risen to $1.94 million. Top academic salaries barely moved: from $301,000 to $392,000 over the same period.

The gap between top-1% industry and top-1% academic pay widened from roughly $294,000 to $1.5 million annually — a fivefold increase. The paper notes these industry figures are likely understated because the Census data captures wage income but may not fully account for stock compensation exercised during the period.

The divergence accelerated at two clear breakpoints: the deep learning turn following AlexNet in 2012, and the transformer era after “Attention Is All You Need” in 2017.

What Happens When Researchers Cross Over

Once an AI researcher permanently moves to industry, three things happen measurably:

Output shifts from open to proprietary. Paper production declines 65% per year. Patent filings increase 530% per year. The share of AI patents held by industry rose from 86% to 95% over the study period; the share of AI papers rose only modestly, from 27% to 32%. Research that was once published — and thus available for the broader field to build on — is now filed as intellectual property.

Earnings rise 63% relative to comparable job-switchers who remain within academia.

Talent concentrates in incumbents. The researchers flowing out of universities are not going to startups. They are going to large incumbent firms — 1,000-plus employees, 20 or more years old, in Professional Services and Information. That concentration compounds existing compute-scale advantages: the organizations with the most GPUs are also capturing the most scientific talent.

The Structural Shift

The paper’s conclusion is direct: AI research in the United States has shifted from a university-centered ecosystem to one dominated by large incumbent firms. The authors identify four drivers.

First, frontier AI research is now capital-intensive in ways universities cannot match. The fixed costs of datasets, compute, and specialized hardware have moved the comparative advantage to firms with cash and hardware.

Second, when frontier researchers cluster inside firms, knowledge diffusion changes character. What used to flow as papers — readable, citable, buildable-upon — now flows as products and patents. The equilibrium may feature faster commercialization and weaker knowledge spillovers.

Third, market concentration in ideas increasingly mirrors market concentration in compute and capital. The organizations that can afford to attract and retain top researchers are the same organizations that can afford the training runs.

Fourth, demographics matter. Industry is increasingly drawing in younger and foreign-born researchers, skewing early-career and international. Academia now has higher female representation than industry, a reversal of the historical pattern, but this reflects who stayed rather than who succeeded.

The Moonlighting Signal

For researchers who remain in academia, a second trend has emerged: moonlighting. A growing share of academic AI researchers are drawing secondary income from industry employment, a pattern the paper tracks as an intermediate state before full crossover. Universities are not losing only the researchers who formally leave — they are losing significant fractions of the time and attention of those who stay.

The broader implication: the infrastructure of frontier AI research has become inseparable from industrial structure. Policy aimed at rebalancing the ecosystem — public compute provision, data access regimes, salary adjustments — faces a feedback loop where every additional year of divergence makes the rebalancing harder.