AI researchers in academia struggle to survive as the frontier moves to private labs

A journalist who arrived at the AI2050 conference hosted by Schmidt Sciences last week found hallway conversations about survival rather than flashy product launches. The event took place at a hotel in Mountain View, brought together some of the biggest names in the field, and the atmosphere reflected a sense that the game has changed. The event, funded by Eric and Wendy Schmidt, supports academic researchers working on artificial intelligence, and Full disclosure: the journalist received a 2024 science-communication award financed by Schmidt Sciences.

the frontier has moved, universities are left without GPUs

Over the past four years, research focus has shifted to large language models (LLMs), and the technological edge has migrated from universities to private companies. Universities simply cannot afford the compute power required to train and run frontier models, and even if they could, Anthropic and OpenAI do not share the internal details of Claude or ChatGPT. Nika Haghtalab, a computer-science professor at Berkeley, likened the situation to biologists operating in a world where private firms hold exclusive control of CRISPR. Researchers outside the frontier labs can study model behavior, but not the planning and training of those models.

GPU funding helps, but does not solve the problem

AI2050 offers grant funding for GPU purchases, and several researchers noted that this assistance is a significant advantage. Yet the money remains a bottleneck, especially against the backdrop of reduced federal science funding in the United States. Even scholars who do not run local models encounter high expenses when they must send repeated queries to OpenAI, Anthropic, and Google models to study them rigorously—costs that can be prohibitive for a standard academic grant.

questions companies will not ask themselves

Rather than chasing capabilities, many colleagues are turning to questions that commercial labs are unlikely to address. “I am trying not to work on problems I think will be solved by a tech company,” says Anjalie Field, a professor at Johns Hopkins. Companies need to profit, and research with low commercial upside—or that could portray companies negatively—will not attract investment. Field recently published a study showing that language models provide less sophisticated answers to phrasings that are more common among women than among men, a result hard to imagine emerging from Anthropic or OpenAI.

not just LLMs: the researchers left behind

A sizable segment of AI researchers does not work on LLMs at all. Many build specialized models for data analysis, forecasting, or simulating entire physical systems. They do not directly compete with frontier labs such as Google DeepMind’s AlphaFold team, which produced a Nobel-winning protein-structure predictor that was dismantled last month, but they face their own challenges. At the conference, participants voiced concern that the prevailing bias toward non-LLM AI harms their work. Researchers developing tools for the climate crisis, for example, find it difficult to advance when the public and media equate “AI” solely with “energy-hungry LLMs”.