Redistributed

The pipeline works like this. A student enters a public university. Taxpayers subsidize the tuition. Federal grants fund the lab. A professor — herself trained at public expense — supervises the dissertation. Five to seven years later, a newly minted AI PhD walks out the door and into a corporate lab offering four times the salary, a hundred times the compute, and a signing bonus that exceeds the annual budget of the research group that trained them.

This is not a market failure. It is the market working exactly as designed — extracting publicly funded talent into private hands. The question is whether the public gets anything back. And some institutions, as we will see, are finding ways to keep talent engaged with public-interest work even as the salary gap widens.

The Talent Pipeline

Modern AI was built in universities. Geoffrey Hinton spent decades at the University of Toronto, funded by the Canadian government, before Google recruited him in 2013. Yann LeCun built the foundations of convolutional neural networks at Bell Labs and NYU — a public university — before joining Facebook AI Research in 2013, and later leading AMI Labs under Meta. Yoshua Bengio, the exception that proves the rule, stayed at the Université de Montréal, though his students did not.

The pattern repeated at every level. The transformer architecture that powers modern AI emerged from Google Brain, but its authors were trained at Stanford, Berkeley, and the University of Toronto. The reinforcement learning breakthroughs behind AlphaGo came from DeepMind, staffed by graduates of University College London, Oxford, and Cambridge.

The AI Index Report 2025 from Stanford documents the shift in hard numbers. In 2025, industry produced the majority of notable machine learning models. A decade ago, most foundational AI research happened in universities. The talent didn't disappear — it migrated. And it took the research agenda with it.

This migration is accelerating. Between 2020 and 2025, North American AI faculty departures to industry increased significantly, according to Computing Research Association surveys. Entire subfields — large language models, multimodal systems, robotics foundation models — now have their center of gravity in corporate labs rather than universities.

The Salary Gap Is a Compute Gap

The compensation disparity is stark. A tenured AI professor at a leading public university earns $150,000 to $250,000. An AI researcher at a frontier lab earns $500,000 to $2 million or more, with equity that can multiply the total. Senior research scientists at Google DeepMind, OpenAI, and Anthropic routinely receive packages exceeding $1 million annually.

But salary is only part of the story. The real gap is compute.

A university research group applying for NSF funding might receive $300,000 over three years — enough for a small GPU cluster and a graduate student. A corporate lab can requisition thousands of H100 GPUs for a single training run. The cost of training a frontier model now exceeds $100 million. No university budget can compete.

This creates a structural impossibility. Researchers who want to work on frontier AI — the systems that define the field's direction — cannot do that work at a university. The compute doesn't exist there. As Arvind Narayanan has noted, the gap between academic and industrial compute access has become so vast that entire categories of research are now industry-exclusive (Narayanan & Kapoor, 2024).

The result is a self-reinforcing cycle. The most ambitious researchers leave for industry, where they can access the compute to do the work they care about. Their departure weakens university programs, making them less attractive to the next generation. The cycle tightens.

What Universities Lose

The consequences extend well beyond individual career decisions.

Teaching capacity. When senior AI faculty leave for industry, someone still has to teach the courses. Universities backfill with adjuncts, lecturers, or junior faculty — if they can hire at all. The students being trained in AI are increasingly taught by people who aren't doing frontier AI research. The pipeline continues to produce graduates, but the quality of training erodes.

Independent research. Corporate labs pursue research that serves commercial objectives. This is not inherently bad — some commercially motivated research produces genuine scientific advances. But it means entire categories of inquiry go unfunded. Research on AI harms, algorithmic bias, environmental costs, labor displacement — work that might embarrass the companies deploying these systems — struggles to find support.

Meredith Whittaker, now president of the Signal Foundation, documented this dynamic during her time at Google and the AI Now Institute. Researchers who studied the social costs of AI found their work unsupported or actively discouraged within corporate settings (Whittaker et al., 2018). The brain drain doesn't just move talent; it narrows the range of questions that talent is allowed to ask.

Critical perspective. Universities are supposed to be sites of independent inquiry — places where researchers can study powerful institutions without needing their permission or funding. As Kate Crawford argues in Atlas of AI, understanding AI requires examining the material conditions, labor practices, and power dynamics behind the technology (Crawford, 2021). That examination is harder to conduct from inside the companies being examined.

Public knowledge production. Academic research is published. Corporate research may or may not be. DeepMind publishes prolifically. OpenAI, despite its name, has become increasingly opaque about its methods. When research migrates to corporate labs, the public's access to knowledge about how these systems work becomes contingent on corporate communication strategies rather than scientific norms.

What Companies Gain

The private sector captures several distinct advantages from the brain drain.

Talent trained at public expense. A PhD in machine learning represents roughly $300,000 to $500,000 in public investment — tuition subsidies, research funding, faculty salaries, lab infrastructure. Companies acquire this investment by offering compensation that universities, bound by public salary structures, cannot match. The training cost is socialized; the returns are privatized.

Academic credibility. Many corporate researchers maintain university affiliations — adjunct professorships, visiting positions, advisory roles. This gives companies academic legitimacy while giving researchers a foot in both worlds. The arrangement benefits both parties but obscures the fact that the researcher's primary loyalty and the direction of their work is determined by their employer, not their university.

Research direction control. When the best researchers work in corporate labs, those labs set the research agenda for the field. What problems are worth solving, what methods are worth pursuing, what benchmarks matter — these decisions increasingly reflect commercial priorities. As Daron Acemoglu has argued, this can produce "the wrong kind of AI": systems optimized for labor replacement rather than labor augmentation, because replacement is more profitable (Acemoglu, 2021).

Recruiting pipelines. Professors who move to industry maintain relationships with their former departments. They return as guest lecturers, serve on thesis committees, and direct promising students toward their companies. The academic-industrial pipeline becomes self-perpetuating. The Frontier Lab Oligopoly benefits from a talent funnel that public institutions built and maintain.

The Global South Dimension

If the brain drain is severe within wealthy nations, it is devastating for developing ones.

Researchers trained in Nigeria, India, Brazil, and Kenya face the same salary and compute disparities — amplified by currency differences, visa incentives, and the concentration of frontier labs in a handful of US cities. A machine learning PhD from the Indian Institutes of Technology or the University of Lagos who moves to Google in Mountain View represents not just a personal career choice but a transfer of intellectual capacity from a country that invested in their education to a company that will capture the returns.

As Abeba Birhane has documented, this reproduces colonial extraction patterns in a new form: raw material (talent, data) flows from the Global South to the Global North, while finished products (models, APIs) are sold back at a premium (Birhane, 2020). The Digital Divide in AI is not only about access to technology — it is about the systematic extraction of the human capital needed to build alternatives. Initiatives like Masakhane and Lelapa AI represent deliberate efforts to build and retain AI capacity in the Global South, but they operate at a fraction of the scale of the talent outflow.

The countries that most need domestic AI capacity — to address local challenges in agriculture, healthcare, governance — are precisely the ones losing their most capable researchers to Silicon Valley. Shakir Mohamed, Png, and Isaac frame this as a question of "decolonial AI": who gets to set the research agenda, and whose problems get solved (Mohamed et al., 2020)?

What Could Change

The brain drain is not inevitable. It is produced by specific policy choices — and different choices could produce different outcomes.

Public compute infrastructure. The single most effective intervention would be giving university researchers access to compute at a scale that enables frontier work. The National AI Research Resource (NAIRR) in the United States is a step in this direction, but its current funding is modest relative to the need. France's Jean Zay supercomputer provides a model: a national compute facility dedicated to AI research, available to academic researchers on a merit basis. Public AI infrastructure would not eliminate the salary gap, but it would address the compute gap that makes academic research structurally uncompetitive.

Competitive academic compensation. Some countries are experimenting with salary top-ups for AI faculty, funded by industry partnerships or dedicated government programs. Canada's CIFAR program, which supported Hinton and Bengio during the years when neural networks were unfashionable, demonstrates that sustained public investment in researchers can keep talent in academia — and that such investment can pay extraordinary dividends, as the deep learning revolution the CIFAR program sustained has now created trillions of dollars in economic value. The UK's Turing Institute and Germany's ELLIS network are attempting similar models. The funding would need to be substantial — not matching industry salaries, but narrowing the gap enough to make the trade-off viable.

Corporate-academic partnerships with strings attached. Industry funding of university research is common, but the terms matter. Unrestricted gifts that support independent research are different from sponsored projects that direct research toward commercial applications. Policy could require that companies benefiting from publicly trained talent contribute to the institutions that trained them — a form of clawback that treats university education as a public investment deserving returns.

AI taxes dedicated to education. If AI generates extraordinary profits — and it does — a portion of those profits could be directed to the educational institutions that made them possible. This is the logic behind proposals for windfall profit taxes on AI companies, with revenue earmarked for public universities and research funding. Stock Buybacks and AI documents how current profits flow overwhelmingly to shareholders; redirecting even a fraction to the talent pipeline would transform academic AI research.

Open research mandates. Companies that hire publicly trained researchers could be required to publish the research those employees produce, or at minimum to make models and datasets available for academic use. This would not prevent brain drain, but it would ensure that the migration of talent does not also mean the enclosure of knowledge.

Comparing funding models. The NSF operates on competitive peer review with modest budgets. DARPA funds high-risk, high-reward research with program managers who bet on ideas. The EU's Horizon Europe distributes funding across member states with political considerations alongside scientific ones. Each model has different implications for talent retention. DARPA's willingness to fund ambitious research at scale has historically been most effective at keeping researchers engaged with public-interest work — because it gives them resources to do work that matters.

The Redistribution Frame

Strip away the details and the structure is clear. This is a subsidy flowing upward.

Taxpayers fund public universities. Public universities train AI researchers. AI researchers generate billions in private value. The returns flow to shareholders — not back to the public institutions that made them possible.

Mariana Mazzucato has documented the same pattern in pharmaceuticals: NIH-funded basic research produces discoveries that pharmaceutical companies monetize, with returns flowing to shareholders through pricing power and stock buybacks rather than back to the public that funded the research (Mazzucato, 2013). The AI talent pipeline follows the same logic.

The difference is that the AI version extracts not just knowledge but the knowledge-makers themselves. When a pharmaceutical company builds on NIH research, the researchers usually stay at their university. When an AI company hires the entire research group, the university loses both the knowledge and the capacity to produce more.

This is not an argument against people taking well-paying jobs. It is an argument that the public deserves a return on its investment — in the form of compute infrastructure, research funding, open knowledge, or direct financial returns. The talent pipeline was built with public money. The question is whether the public will continue to fund it without capturing any of the value it creates.

The market will not solve this on its own. Without intervention, the cycle will continue: public investment in training, private capture of talent, public loss of capacity. The countries, universities, and communities that fund the education of AI researchers will bear the costs. The companies that hire them will reap the returns.

Policy created this dynamic. Policy can change it.

Sources

Acemoglu, Daron. "Harms of AI." Cambridge Journal of Economics, 2021. Cambridge
Birhane, Abeba. "Algorithmic Colonization of Africa." SCRIPTed 17, no. 2 (2020). SCRIPTed
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
Mazzucato, Mariana. The Entrepreneurial State: Debunking Public vs. Private Sector Myths. Anthem Press, 2013. Publisher
Mohamed, Shakir, Marie-Therese Png, and William Isaac. "Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence." Philosophy & Technology 33 (2020): 659-684. arXiv
Narayanan, Arvind and Sayash Kapoor. AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference. Princeton University Press, 2024. Publisher
Stanford Institute for Human-Centered AI. "AI Index Report 2025." Stanford University, 2025. HAI
Whittaker, Meredith et al. "AI Now Report 2018." AI Now Institute, New York University, 2018. AI Now

General

The Academic Brain Drain: Who Pays to Train AI's Best Minds?