$1.5 million would be unremarkable if both researchers weren't publishing in the same journals

The gap between the highest-paid AI scientists in industry and their academic counterparts has hit $1.5 million per year. Not total compensation packages at one company. Not a one-time signing bonus. Annual W-2 earnings, verified through U.S. Census Bureau administrative records, for the top 1% of researchers who still actively publish.

The number comes from a new NBER working paper by Ufuk Akcigit, Craig Chikis, Emin Dinlersoz, and Nathan Goldschlag. The method separates it from the usual salary-survey noise: the authors matched academic publication records directly to Census Bureau employer-employee data, creating a longitudinal view of where AI scientists work, what they earn, and how the distribution has shifted over two decades.

In 2001, the same gap was roughly a fifth of what it is now.

The linkage that changes the read

Most AI salary reports rely on self-reported survey data, job posting ranges, or recruiter estimates. This paper uses administrative tax records held by the Census Bureau, linked to publication histories across AI and adjacent fields. Individual scientists are tracked across employers, sectors, and years — not snapshots, but trajectories.

The acceleration matters more than the absolute figure. The wage premium for top industry AI scientists grew roughly fivefold since 2001, with most of that concentrated after 2012 — the year AlexNet won ImageNet and deep learning shifted from a niche academic interest to an industry obsession. Before 2012, the industry premium existed but was modest enough that a tenured position with grant funding and sabbatical flexibility could plausibly compete. After 2012, the premium started compounding. By the time GPT-3 shipped in 2020, the top of the distribution had pulled away so far that the comparison stopped being about salary negotiation and started being about entirely different economic realities.

The shape of the earnings distribution is where the story gets uncomfortable for anyone trying to hire. At the 50th percentile, the industry-academia gap is real but manageable — enough to matter for a mortgage decision, not enough to reshape a career. At the 90th percentile, the gap widens into something that changes life trajectories. At the 99th percentile, the $1.5 million annual difference puts industry researchers in a compensation tier that academic institutions cannot structurally match, regardless of endowment size or grant funding.

For an ops lead at a 30-person consulting firm trying to staff an ML deployment — running fine-tuned models on Modal with Hugging Face inference endpoints — the downstream effect is concrete. A $180,000-to-$220,000 budget that felt competitive for an ML engineer in 2021 now sits below the 25th percentile for researchers with publication records. Candidates with genuine research depth — the ones who can diagnose why a fine-tune collapsed or architect a retrieval pipeline that performs under load — are increasingly priced into a bracket where only Series B+ startups and the large labs can compete.

The part nobody expected

One counterintuitive detail buried in the research: industry scientists at the top of the pay distribution are still publishing. They have not disappeared from academic conferences or stopped contributing to open research. The shift is in employer affiliation, not in research output.

This scrambles the signal most smaller firms use to evaluate ML candidates. A strong publication record used to correlate loosely with academic employment. Now it correlates just as strongly with a researcher earning seven figures at an AI lab. The resume looks the same. The salary expectation does not.

Ethan Mollick, who highlighted the paper, framed it bluntly: the researchers who remain in universities are making an increasingly expensive choice, and the ones who leave do not stop doing research. They do it under a corporate logo, often with better compute budgets and fewer grant-cycle constraints. The academic pipeline that used to produce affordable ML hires for mid-market firms now produces candidates whose reference compensation is set by OpenAI, Anthropic, and DeepMind.

The practical response for firms under 50 people is not to match Big Tech comp — that math does not close. It is to stop pretending the competition is symmetrical. Fractional ML consulting, co-development agreements with university labs where remaining researchers still need industry collaborators, or investing in upskilling existing engineers through Hugging Face courses and Modal's serverless GPU platform are all more realistic than posting a $200K ML engineer role and waiting.

The paper is available through NBER, authored by Akcigit, Chikis, Dinlersoz, and Goldschlag. The Census linkage methodology is described in detail in the working paper.