If your model of the AI research talent landscape assumes industry has been progressively winning the competition for PhD talent against academia, the 2024 CRA Taulbee Survey data shows the trend reversed two years ago. The share of new AI PhDs in the US and Canada going to industry fell from 77% in 2022 to 63% in 2024. The share going to academia nearly doubled in the same period to 31.6%. The share going to government rose from 0.7% in 2021 to 2% in 2024. Total new AI PhD output grew 22% from 2022 to 2024. All of the growth went to academia. The "industry brain drain from academia" narrative that dominated 2018–2022 commentary is contradicted by the most recent data.

New AI PhD employment in US/Canada by sector, 2010–2024:

The structural data is unambiguous. The absolute numbers of new AI PhDs going to academia: 76 in 2010, peaked at 154 in 2022, hit 145 in 2024. Industry numbers: 64 in 2010, peaked at 280 in 2023, dropped to 288 in 2024 (similar level). Government: small but rising from 0.7% to a 2% share. The story is not that academia is producing fewer PhDs. Academia is producing more. The story is that industry stopped scaling its PhD intake at the rate that PhD output is scaling.

New AI PhD employment share in US/Canada by sector, 2010–2024:

Two interpretations of this pattern are genuinely defensible from the same data, and the choice between them matters for talent strategy. Both are open.

Interpretation 1: industry has shifted to hiring AI talent at non-PhD levels. The argument: industry needs for AI engineering scaled enormously between 2018 and 2024, and the marginal hire moved from "PhD-trained research scientist" to "experienced software engineer with AI capability training." Industry's headcount in AI roles grew, but the composition shifted away from PhD-credentialed researchers toward applied engineers. Under this interpretation, the 63% industry share in 2024 doesn't reflect industry losing interest in AI; it reflects industry's hiring composition changing. The PhDs choosing academia are not being rejected by industry; they are choosing academia in a market where industry is hiring at non-PhD levels and academia is offering positions at the rate it always has.

Interpretation 2: academia has become structurally more attractive for AI PhDs in the past two years. The argument: academic positions in AI are now more abundant (the CRA faculty data shows the AI-related faculty count growing from 6,683 in 2024–25 to a projected 7,501 by 2026–27, a 12% growth in two years). Industry compensation for new AI PhDs may have moderated as the 2022–2023 frontier-lab hiring frenzy stabilised. Academic positions offer research independence, publication freedom, and longer-horizon work that the industrial labs may have constrained as they scaled. Under this interpretation, the trend reversal reflects an actual change in PhD preferences: academia is winning competitively, not just by default.

The data alone cannot fully adjudicate. What it can establish is that whichever interpretation is correct, the planning implication is similar: relying on "industry will hire most new AI PhDs" as a planning assumption is now reading a stale signal. The signal has been reversing for two years.

Three structural implications follow for research-track AI talent planning.

The first: academic AI research capacity is rebuilding in a way the 2018–2022 narrative did not anticipate. The CRA faculty data shows the institutional capacity to absorb the larger academic PhD intake is in place. The implication for industrial AI labs: the pipeline of PhDs available for industrial research positions has thinned at the top end while the pipeline of available applied AI engineers has thickened. Labs that depend on PhD-trained researchers for frontier work face a tighter recruiting market than they did 24 months ago.

The second: the government share, while small, has tripled. From 0.7% in 2021 to 2% in 2024 represents a meaningful absolute increase given the growing total. The US federal AI strategy includes specific PhD-level hiring intent (e.g., for the National AI Research Resource and various agency AI initiatives). The 2% share is consistent with that intent beginning to materialise. The trajectory is small but directional. Three-sector competition (industry vs academia vs government) is more visible in the 2024 data than at any prior point in the series.

The third: the talent supply variable that has not yet shown up in the data is the international PhD cohort entering and leaving the US/Canada system. The CRA Taulbee Survey covers PhDs from US and Canadian institutions, which include many international students. Whether those students stay in the region after graduation depends on visa policy and personal choice. If visa policy continues to restrict, the absolute supply of new AI PhDs from US/Canadian institutions will narrow within 24–36 months, at which point the industry-academia-government share split will play out across a smaller total.

For research talent strategy in 2026, the planning anchor needs to engage with the reversal. The "industry will absorb the marginal PhD" assumption is no longer the right default. The right default is "academia is now competing for marginal PhDs, government is entering the competition, and industry's share of the pipeline is declining." Plans built on the right default will source PhD-level research talent through more channels (including poaching from academia, which has become a more substantial source pool again).

The contested question for strategic planning teams: is the reversal a 2-year aberration that will reverse again, or a structural shift that persists through 2026–2030?

The aberration view: industrial AI labs went through a unique hiring frenzy in 2022–2023 driven by the GPT-4 / ChatGPT moment. That hiring intensity normalised in 2024 and produced the apparent share shift. Industry will resume scaling its PhD intake as it absorbs new product needs. The 2025–2026 data may show the share split returning toward 2022 levels.

The structural view: the 2018–2022 "industry wins all the PhDs" period was the unusual one, driven by a specific combination of frontier compute access, compensation scaling, and research mission alignment in the early industrial labs. As the industrial labs have matured and the compensation premium has compressed, the structural attractiveness of academia (independence, publication, breadth of work) reasserts itself. Under this view, the 2024 share split is the new equilibrium, not a temporary deviation.

The data does not yet support strong inference between these views. The reversal is only two years deep, and reversals of short duration can return. The recommendation: plan for both, with explicit signal-monitoring on the 2025–2026 data. If the structural view holds, talent strategies built for it will out-perform. If the aberration view holds, the planning cost of building for either is modest relative to the cost of mis-calibrating in the structural-view scenario.

For research-track AI talent strategy, the planning anchor needs to be flexible enough to absorb the reversal. The 2018–2022 plan that assumed industry would hire most new AI PhDs needs revision. The revised plan should treat the industry-academia-government share split as a contested and moving variable rather than as a structural certainty.

Sources

  • Primary: Stanford AI Index 2026, Chapter 7 (Education) 7.2 — hai.stanford.edu/ai-index/2026
  • PhD employment data: Computing Research Association (CRA) Taulbee Survey, 2011–2025 — US and Canadian AI PhD graduates and employment sector
  • Faculty capacity data: CRA AI-related faculty counts, 2024–2025 through projected 2026–2027
  • Policy context: US federal AI strategy PhD hiring initiatives (National AI Research Resource and agency AI hiring programmes)