The most consequential AI institution of the next decade may not be a startup, university or government laboratory. It may be the research division hidden inside a platform company whose advertising, cloud or software monopoly can absorb the bill. Google DeepMind, Anthropic, OpenAI, Microsoft Research and Meta's superintelligence organization increasingly resemble a new generation of corporate laboratories: places where scientists, engineers, proprietary data and extraordinary computing budgets are assembled under one roof because almost no one else can afford the experiment.
The historical analogy is Bell Labs, and it is tempting for a reason. AT&T's research system helped produce the transistor and information theory, along with a culture of long-horizon technical work that became shorthand for industrial science at its best. The AI companies are rebuilding part of that machine. They can move from mathematics to chips to product deployment without waiting for a grant cycle or licensing negotiation. That speed can create real public value. It can also concentrate the power to choose which questions get asked, which findings remain private and which capabilities reach the world.
The center of gravity has already moved
Figures summarized by Reuters Breakingviews show industry involvement in notable AI models rising from an historical average of about 25% to roughly 80% today. The business share of U.S. basic research, which fell to about 14% in 2004, has returned to 32%. Talent is following the capital: the share of doctorate recipients planning to enter academia has fallen from 56% in 2004 to 40% in the latest NCSES survey, with especially steep declines in mathematics and computer science.
Industry involvement in notable AI models
Why the corporate lab came back
Frontier AI is a scientific field with an industrial cost structure. Training runs need clusters, custom networking, scarce chips, large data pipelines and teams capable of operating all of it. Stanford's 2026 AI Index describes continued growth in the scale of training compute and data. Once experiments demand infrastructure measured in billions rather than laboratory benches, the organizations already earning platform-scale cash acquire a structural advantage. The lab returns because the factory and the experiment have become the same place.
The new corporate lab can fund discoveries universities cannot afford. It can also decide which discoveries the public is allowed to inspect.
The bargain is not free science
The Bell Labs story is often told as proof that concentrated companies can finance broad invention. That is true, but incomplete. Corporate research agendas are shaped by ownership, product strategy, national-security relationships and the need to defend a moat. A breakthrough can be published because it strengthens an ecosystem; another can remain internal because it threatens a product, creates liability or confers strategic advantage. The public receives extraordinary capability, but not necessarily the methods, data or freedom to reproduce it.
This matters because independent science performs jobs the frontier companies cannot credibly perform for themselves. Universities and public-interest labs evaluate claims, test harms, preserve methods, train researchers who can move between institutions and pursue questions without an immediate product path. Their role becomes more important, not less, when the most capable systems are built behind corporate access controls.
What a better settlement looks like
The answer is not to dismantle every large AI lab and hope university budgets somehow replace the compute. Nor is it to romanticize concentrated power because a few famous inventions may emerge. A workable settlement would expand public and academic compute, require stronger disclosure around evaluations and incidents, protect researcher mobility, fund independent replication and use procurement or grants to keep some foundational work in the open. The goal is to preserve the corporate lab's ability to build without allowing it to become the only institution capable of knowing.
AI may indeed revive an era of industrial invention. The question is whether society negotiates the terms while the laboratories are still being built, or discovers later that scientific progress arrived bundled with a private constitution.
- Industry now participates in ~80% of notable AI models, up from a ~25% historical average.
- PhDs heading to academia fell from 56% to 40% — talent follows the compute.
- The new Bell Labs can fund what universities can't — and keep what it wants private.
- Independent science still does the jobs labs can't: verify, replicate, test harms.
- The terms should be negotiated now, while the laboratories are still being built.