---
id: "lib-2026-03-00-why-universities-are-struggling-to-keep-ai-talent"
type: "library"
title: "Why universities are struggling to keep AI talent"
lang: "EN"
org: "NBER"
source_type: "Paper"
date: "2026-03-01"
week: "2026-W09"
url: "https://nosov.ai/library/lib-2026-03-00-why-universities-are-struggling-to-keep-ai-talent/"
original_url: "https://www.nber.org/system/files/working_papers/w34964/w34964.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "NBER working paper led by Ufuk Akcigit (University of Chicago) that links publication records to confidential US Census employer-employee data for 42,000 AI researchers; strong data on the academia-industry pay gap, but a working paper on a narrow question rather than one that shapes the wider debate."
summary_basis: "full-text"
original_minutes: "60"
tags: ["ai talent", "universities", "labor market"]
---

# Why universities are struggling to keep AI talent

Tier A (Strong evidence): NBER working paper led by Ufuk Akcigit (University of Chicago) that links publication records to confidential US Census employer-employee data for 42,000 AI researchers; strong data on the academia-industry pay gap, but a working paper on a narrow question rather than one that shapes the wider debate.

Summary based on the full text.

## In short

NBER Working Paper 34964 (March 2026) by Ufuk Akcigit and Craig Chikis (University of Chicago), Emin Dinlersoz (US Census Bureau) and Nathan Goldschlag (Economic Innovation Group). The title plays on "Attention Is All You Need". The authors link Microsoft Academic Graph publication records to Census administrative payroll data to follow 42,000 US AI researchers from 2001 to 2021. They find that AI research has moved from universities to large incumbent firms, driven by a pay gap that grew fivefold, and that researchers who move publish far less and patent far more.

## Key numbers

- $1.5 million: extra annual earnings of top 1% publishing industry AI scientists over comparable academics, a fivefold increase since 2001
- 68%: share of US AI researchers working in industry by 2019, up from 48% in 2001
- 530%: more patents per year after a permanent move from academia to industry, alongside 65% fewer papers
- 42,000: AI researchers tracked over two decades

## What it reports

- 68% of AI researchers worked in industry by 2019, up from 48% in 2001.
- Top 1% industry earnings rose from $595,000 in 2001 to $1.94 million in 2021 (2015 dollars); top academic salaries went from $301,000 to $392,000.
- Top 1% publishing industry scientists now earn $1.5 million more per year than comparable academics, five times the 2001 gap. Average real academic salaries declined since 2001.
- The share of US-born researchers in industry fell 5.5 percentage points, offset by China (+3.8 p.p.) and India (+2.0 p.p.).
- Median industry age fell from 39 to 37; academia stayed at 42.
- Female share in academia rose from 16% to 29%; in industry from 19% to 23%. The female-to-male earnings ratio narrowed in academia (79% to 83%) and widened slightly in industry (73% to 72%).
- Moves to industry accelerated after the 2017 transformer paper and are driven by young researchers going to incumbents (1,000+ employees, 20+ years old) in Professional Services and Information.
- After a permanent move: 65% fewer papers per year, 30 p.p. less likely to publish, 530% more patents per year, and earnings 63% higher than similar job switchers who stayed in academia.

## How they got it

Authors of MAG papers whose modal field is AI or adjacent were matched to Census person identifiers using an algorithm first built for patent inventors, then linked to LEHD quarterly earnings, employer age, size and sector, and demographics. The panel covers 2001-2021. Outputs are descriptive facts plus event-study comparisons of movers against similar job switchers within academia.

## Limits

- US only, and ends in 2021, before the generative AI hiring wave of 2022 onward.
- Descriptive: the paper documents the shift but does not estimate its effect on the rate of innovation.
- Researchers who never published in MAG-indexed venues are outside the sample, so applied practitioners in companies are not counted.

## Why it matters for companies

Frontier AI talent has concentrated in a small set of large, compute-rich firms that pay at a level traditional companies will not match. A manufacturing or consumer company should plan to buy AI capability through vendors and partners, and to build applied skills internally, rather than compete for research scientists. The shift from papers to patents also means less of the frontier is published openly, so companies relying on public research for know-how will see less of it. Universities under pay pressure may be more open to joint appointments and applied collaborations.

Original: https://www.nber.org/system/files/working_papers/w34964/w34964.pdf