Part 07

The market that closed, and reopened

Industry placement for computing doctorates fell 11 points in a single cohort and has since recovered most of the way. The contraction fell hardest on the subjects closest to conventional software work, and the academic share it created has not come back down.

Doctoral cohorts 2012–2025 · measured one year out · like-for-like across cohorts

Every other page here measures a quarter of a century. This one measures the last four years, because something happened in them that the long series flattens out: the class of 2023 met a market that had closed.

51%
Into industry, 2022 cohort
40%
2023 cohort
47%
2025 cohort

Every figure below counts only positions that had begun by the midpoint of the first year out, for every cohort, so that the 2025 cohort is comparable to the rest. The chart carries the as-recorded line too, so the size of that correction is visible.

0% 12.6% 25.1% 37.7% 50.3% 62.8% 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 Industry, like-for-like Industry, as recorded Academia, like-for-like
Share of computing doctorates in an industry position one year out, by graduating cohort. The solid line is the like-for-like measure; the grey line is the same figure as recorded, which runs above it because later-starting jobs have had time to appear.

Industry placement peaked at 51% for the 2022 cohort, fell to 40% for 2023, and has recovered to 47% for 2025. The drop is the largest year-on-year move in fourteen years of cohorts.

Those 11 points split almost exactly in half. Academic placement rose 4.9 points and the share with no position recorded anywhere rose 5.0. So half of the displaced cohort took an academic position, overwhelmingly postdoctoral research rather than a faculty appointment, which is what waiting out a bad market looks like in this data. The other half stopped being visible in US career records at all, which is what leaving the country looks like.

The recovery came back out of the same place. Between 2023 and 2025 industry regained 7.1 points while the invisible share fell 7.7, almost the whole of it. Academic placement went up 0.9 and stayed up: the share entering academia has not returned to where it was before the contraction. One cohort met a closed market, and the composition of the field shifted in a way that has so far outlasted the market that caused it.

The contraction was not evenly distributed

Languages & SE -17 pp Networks -16 pp Systems & distributed -15.1 pp Security & privacy -15 pp Robotics -12.4 pp Theory -10.9 pp Scientific computing -9.8 pp Machine learning -9.7 pp HCI -9.2 pp Computer vision -9 pp NLP -8.7 pp Architecture & hardware -7.3 pp Computational biology -3.8 pp IR & data mining 1.7 pp
Change in industry placement between the 2022 cohort and the 2023–2024 cohorts, in percentage points, by dissertation subject.

The areas that lost most are the ones closest to conventional software work. Languages and software engineering, networks, and systems and distributed computing each gave up more than ten points. The areas that held are architecture and hardware, computational biology, and information retrieval and data mining. This was a software-hiring freeze rather than a computing-wide one, and a graduate’s exposure to it depended a great deal on which group their dissertation sat in.

What people were studying, meanwhile

Machine learning 13.9 pp Robotics 4.3 pp Computer vision 1.6 pp NLP 1.4 pp Scientific computing 0.7 pp HCI 0.2 pp CS education 0.2 pp Security & privacy -0.1 pp Computational biology -0.1 pp Theory -1.1 pp Graphics & visualisation -1.4 pp Databases -1.6 pp Architecture & hardware -2.6 pp Systems & distributed -3.0 pp IR & data mining -3.3 pp Languages & SE -3.4 pp Networks -5.8 pp
Change in each subject's share of doctoral dissertations, early 2010s against the early 2020s, in percentage points.

The areas students left are the same areas the market left. Theory, systems and distributed computing, languages and software engineering, and networks all shrank as a share of dissertations over the decade and lost more than ten points of industry placement in the contraction. That is not the tidy story where a shrinking field leaves better prospects for whoever stays. Both sides moved the same way.

Method separated outcomes more sharply than subject did

0% 12.7% 25.4% 38.2% 50.9% 63.6% 2021 2022 2023 2024 2025 Used a large language model Used deep learning Used neither
Industry placement one year out, by whether the dissertation used a large language model, used deep learning otherwise, or used neither. Like-for-like measure.

The ordering is the same in every year it can be measured, and it is wider than the spread across most subjects: in 2025 the gap is 15 points. Most of it sits between using deep learning and using neither, rather than between large language models and deep learning. And inside natural language processing, where these methods are now universal, the gap disappears entirely. The advantage comes from bringing the method into a field that was not already built on it.

Where the hiring went

Employer 2019–21 total 2022 2023 2024 2025
Meta Platforms, Inc. 399 140 81 175 137
Google LLC 342 89 77 86 81
Amazon.com, Inc. 178 89 53 82 77
Apple, Inc. 147 54 61 58 48
Microsoft Corp. 177 57 32 43 41
NVIDIA Corp. 67 44 26 54 44
Amazon Web Services, Inc. 79 34 27 42 37
Intel Corp. 87 40 19 16 <10
QUALCOMM, Inc. 66 33 14 16 <10
Massachusetts Institute of Technology 47 14 19 15 14
University of California, Berkeley 43 15 16 18 15
Stanford University 33 14 17 21 18

The first column is a three-year total and the rest are single years, so read down a row rather than across it.

Three things to carry away. The 2023 contraction was real and large, 11 points of industry placement in one cohort. It was not evenly spread: the software-adjacent subjects lost double digits while hardware and computational biology held. And it left a mark that the recovery has not erased, because the academic share it pushed up has stayed up.