When telephone giant AT&T gathered its leading scientists and engineers at 463 West Street in New York City in 1925, it couldn't have known that Bell Telephone Laboratories would ultimately be associated with 11 Nobel Prizes, the invention of the transistor in 1947 and Claude Shannon's information theory in 1948. A century later, the vast computing costs of AI are similarly pushing frontier research into the likes of Google DeepMind, Anthropic, OpenAI, Microsoft Research and Meta Superintelligence Labs. While this takes power away from universities, a new heyday for the corporate lab could come with big economic dividends.
Go back 30 years. Four AI models were published in 1996, figures by Epoch AI show. Yet only one, the AdaBoost.M2 algorithm, was developed within a corporation — funnily enough, AT&T. This was far from abnormal: between 1956 and the 2000s, only 25% of models on average were made with industry involvement. Today it's 80%.
It heralds something of a return to the 1950s, when U.S. businesses performed about 30% of all basic research, according to the National Center for Science and Engineering Statistics (NCSES). This share steadily fell as federal funding turned universities into the dominant centres of scientific inquiry. The proportion hit as low as 14% in 2004. But it's now back to 32%, driven by the growth of semiconductor, electronics and data-related industries.
There are parallels to today. The computing power, and dataset sizes, required to train frontier AI models is doubling every few months, Stanford University reports suggest, so only heavily funded AI firms can keep up. Universities still do valuable work, particularly evaluating models. But in NCSES's latest survey of doctorate recipients, only 40% said they would go into academia, compared with 56% in 2004. The largest declines were in mathematics, dropping from 73% to 39%, and computer science, from 53% to 28%.
Note also that 20th century corporate labs were borne out of an early wave of mergers that created vertically integrated monopolistic firms with huge economies of scale, kept in check by nascent antitrust law and concerns about technological obsolescence. At the turn of the century, GE accounted for 90% of lamp sales but was grappling with the expiration of Thomas Edison's original patents.
This model, of having a business monopoly funding a lab, was unpopular with both antitrust regulators. Investors, meanwhile, wanted research tied more directly to visible commercial outcomes. In the 1980s and 1990s, business concentration paused its long rise, with the top 0.1% of firms maintaining roughly 50% of sales, according to Spencer Kwon, Yueran Ma and Kaspar Zimmermann. Officials broke up AT&T, dealing a severe blow to Bell Labs, which survives inside Nokia but at a fraction of its former scale. Then the end of the Cold War weakened an ecosystem of private science backed by the Pentagon, which had seeded computer science and supported organisations such as Hughes Research Laboratories, responsible for creating the first working laser.
This contributed to a surge in patent applications, convincing officials of the model's success. Today, the main policy questions are how to increase funding to academia and promote greater university spinoffs. Yet some economists have long questioned whether innovation is as plentiful as during the postwar era, pointing to slower productivity growth.
Google's "transformer" paper in 2017 allowed OpenAI's ChatGPT to kickstart the large language model revolution five years later. It pointed to how today's hyperscalers — increasingly employing leading mathematicians alongside product engineers — could reinvent the corporate lab. But like the AT&T and GE of old, Alphabet (GOOGL.O), Microsoft (MSFT.O), Meta Platforms (META.O) and Amazon.com (AMZN.O) enjoy much deeper moats. Concentration has also increased across the U.S. economy: the top 0.1% of firms' share of revenues is now 66%, explaining a resurgence in antitrust enforcement under former President Joe Biden.
The return to the old approach, of tolerating oligopoly rents in exchange for scientific advancement, seems preferable. Particularly so as top universities shed prestige, governments back national champions and militaries are rebuilt. While recent frictions between Anthropic and Washington over autonomous weapons underscore the difficulties of public-private collaboration, they also highlight a deeper reality: AI-enabled warfare may require governments to reset the relationships with tech companies along Cold War lines, expanding beyond procurement into setting long-term research goals. Science is too important to be left solely to universities and the public sector.
