Academic AI research is being reshaped by a simple problem: the most powerful systems are now expensive to build and difficult for outsiders to inspect. At a convening of the Schmidt Sciences AI2050 program in Mountain View, California, researchers described a field where universities still matter deeply, but no longer control many of the tools setting the agenda.
The result is not a retreat from AI research. It is a negotiation over what academics can study, what they can build, and where their work can still change the direction of the technology.
The frontier has moved away from campus
Over the past four years, AI research has reorganized itself around large language models. That shift has changed the balance of power between universities and private companies. The cutting edge of AI is now largely inside organizations with the money, computing infrastructure, and closed systems needed to train and operate frontier models.
Universities face two linked barriers. First, they generally cannot afford the GPUs required to train and run models at the frontier. Second, even when academics can study commercial systems from the outside, companies such as Anthropic and OpenAI do not provide access to the inner details of Claude or ChatGPT.
That matters because outside researchers can observe how a system behaves, but they cannot fully examine how it was designed or trained. They also cannot directly steer those design and training choices. Nika Haghtalab, a computer science professor at UC Berkeley, compared the situation to biology in a world where private companies had exclusive control over CRISPR.
The comparison captures the tension facing AI academics. A major scientific tool exists, it is transforming the field, and many of the people best positioned to scrutinize it are kept at a distance from its core mechanisms.
Funding helps, but access remains uneven
The Schmidt Sciences AI2050 program supports academics whose work involves AI, and some fellows can use funding from the program to buy GPUs. Researchers at the convening described that support as a major benefit. But funding remains a central pressure point, especially in the United States, where federal scientific funding has been reduced.
The cost problem is not limited to researchers who train or run local models. Even scholars who study commercial systems through repeated queries can face prohibitive expenses when they need to test OpenAI’s, Anthropic’s, and Google’s models rigorously.
This creates a difficult research environment. Careful evaluation often requires repeated experiments, comparisons, and enough scale to make findings meaningful. If each interaction with a model costs money, the price of independent scrutiny can rise quickly.
For university labs, these constraints shape research choices. Some questions become harder to pursue because they require too much compute. Others become dependent on access to systems controlled by companies that may not share enough information for deeper scientific work.
Academics are choosing questions companies may not ask
Some researchers are responding by focusing on questions that frontier labs are unlikely to prioritize. Anjalie Field, a computer science professor at Johns Hopkins, said, “I try not to work on problems that I think are gonna be solved by a tech company.”
That strategy reflects a practical divide. Companies need to make money, and some research questions may have little obvious profit potential. Others could produce results that make companies look bad, which gives industry less incentive to pursue them.
Field’s own work shows why independent research matters. She conducted a study finding that language models give less sophisticated responses to prompts phrased in ways more commonly used by women than by men. The source article notes that it is difficult to imagine that kind of research coming out of Anthropic or OpenAI.
This is one of the clearest roles for academic AI research in the current era. Universities can investigate harms, gaps, and social consequences that may sit outside a company’s business priorities. They can also ask whether systems work equally well for different people, use cases, and forms of language.
AI research is bigger than LLMs
The public conversation often treats AI and large language models as if they are the same thing. Many academics at the AI2050 convening do not work with LLMs at all. Some build specialized AI models that analyze data, make predictions, or simulate entire physical systems.
These researchers are not necessarily competing with frontier labs in the same way as language-model researchers. Their work can be tied to scientific domains and practical problems rather than general-purpose chat systems. But they face a different challenge: explaining what they do in a climate where many people assume AI means energy-intensive LLMs.
That misunderstanding can make advocacy harder. Researchers building specialized AI tools to help address climate change, for example, may have to first correct assumptions about what kind of AI they are using and why it matters.
The source article also points to uncertainty inside industry research itself. Google DeepMind’s AlphaFold team, which built a Nobel Prize–winning model that predicts the structures of proteins, was disbanded last month. That example underscores how quickly priorities can shift even around high-profile scientific AI work.
The academic role is changing, not disappearing
The pressure on universities is already changing careers. Several prominent academics have recently taken leave from their universities to join frontier labs, and many AI2050 fellows hold industry positions alongside academic jobs.
Another concern has emerged in the past six months: OpenAI’s models have solved a number of real research problems in mathematics. Some experts now worry about whether humans have a future in pure math, and one fellow said she was concerned about the mental health of mathematician peers.
Still, the outlook is not only negative. Empirical science may be much harder to automate than mathematics because collecting data is intrinsically slow. Tim Dettmers, a computer scientist at Carnegie Mellon who works to make AI models faster and cheaper to run, sees AI mathematicians and scientists as potentially useful rather than simply threatening.
In that view, AI scientists would not replace humans. They could make human researchers more efficient, giving them more room to pursue ambitious ideas they might otherwise never have had time to explore.
Resource limits may also push academics toward important technical advances. If university researchers cannot train frontier models, they have strong incentives to make models smaller, cheaper, and more efficient, or to explore new architectures entirely. The next major AI breakthrough could still come from an academic lab working under constraints rather than from a major company with the largest systems.