Academic AI researchers face funding crunch as frontier labs dominate

In Mountain View last week, MIT Technology Review's AI editor attended a convening of the Schmidt Sciences AI2050 program, an initiative funded by Eric and Wendy Schmidt that supports academics whose research involves artificial intelligence. The fellows list reads as a who's who of AI researchers, and the gathering laid bare the tensions reshaping university-based AI work.

The core problem is structural. In the past four years, AI research has reoriented around large language models, and the cutting edge has migrated from academic institutions to private companies. Universities cannot afford the GPUs required to train and operate frontier models. Even if they could, Anthropic and OpenAI keep the inner workings of Claude and ChatGPT hidden from outside researchers. As Nika Haghtalab, a computer science professor at UC Berkeley, put it during a lunch conversation, being an AI academic today is like being a biologist in a world where private companies had exclusive control over the gene-editing tool CRISPR. Experts can study how these models behave from the outside, but they cannot conduct detailed research on their design and training, nor can they influence that development themselves.

The AI2050 program does offer some relief: fellows can use funding to purchase GPUs, which several researchers identified as a major benefit of participation. But money remains a critical constraint, especially as federal scientific funding in the United States has contracted. Even researchers who don't run local models face prohibitive costs when they need to repeatedly query OpenAI's, Anthropic's, and Google's APIs in order to study them rigorously. In response, many academics are redirecting their focus. Anjalie Field, a computer science professor at Johns Hopkins, says she tries not to work on problems she thinks tech companies will solve. Companies need profit, and research questions with little commercial promise—or answers that might reflect badly on the companies—may never get industry attention. Field recently conducted a study showing that language models give less sophisticated responses to prompts phrased in ways more commonly used by women than by men. That kind of research is unlikely to emerge from Anthropic or OpenAI.

Meanwhile, a large cohort of AI academics work with specialized models rather than LLMs—tools that analyze data, make predictions, or simulate physical systems. These researchers don't necessarily compete with frontier labs, though the recent disbanding of DeepMind's AlphaFold team, which built a Nobel Prize-winning protein structure prediction model, signals instability even in high-impact specialized work. Many of these researchers raised concerns at the convening about how widespread ignorance of non-LLM AI was undermining their advocacy. Researchers building specialized AI to address climate change, for example, struggle to justify their work when the public and funders equate 'AI' with energy-intensive large language models.

The landscape is shifting as a result. Several prominent academics have taken leave from universities to join frontier labs in recent months. Many AI2050 fellows now hold industry positions alongside their academic roles. A newer threat has also emerged: OpenAI's models have solved real research problems in mathematics, prompting fears among experts about whether pure mathematicians will have a future. One fellow told the editor she was concerned about the mental health of her mathematician peers. But not all voices are pessimistic. Tim Dettmers, a computer scientist at Carnegie Mellon who works to make AI models faster and cheaper to run, argues that AI scientists will not replace humans—instead, they could make human scientists more efficient, freeing researchers to pursue wild and inspired ideas they might otherwise never pursue. Additionally, empirical science may prove harder to automate than mathematics, because data collection is intrinsically slow. And the very resource constraints that prevent academics from training frontier models may push them to innovate: discovering ways to build smaller, more efficient models, or exploring entirely new architectures. If the next major AI breakthrough comes not from a major company but from a scrappy academic lab, the outcome would be far from surprising.