AI Professors Adapt as Closed Models and GPU Costs Shift Research to Industry
AI professors face costly GPUs, closed frontier models and US funding cuts while pursuing bias research, specialized systems and leaner model designs.
Summary
On August 10, 2026, MIT Technology Review reported on the previous week’s Schmidt Sciences AI2050 gathering in Mountain View, California, 30 miles south of San Francisco. Funded by Eric and Wendy Schmidt, AI2050 supports academics using AI and offers GPU funding; the writer disclosed receiving a Schmidt Sciences funded science communication award in 2024. Researchers said four years of large language models moved AI’s frontier from universities to companies: universities cannot afford required GPUs, Anthropic and OpenAI keep Claude and ChatGPT’s design and training details closed, and repeatedly querying OpenAI, Anthropic and Google models can be prohibitive amid reduced US federal science funding. UC Berkeley professor Nika Haghtalab compared academic AI research to biology if companies exclusively controlled CRISPR.
Johns Hopkins professor Anjalie Field avoids problems companies are likely to solve, targeting low-profit or reputationally uncomfortable questions; her recent study found language models gave less sophisticated answers to wording more commonly used by women than men. Many academics instead build specialized AI to analyze data, predict outcomes or simulate physical systems. Although Google DeepMind disbanded its Nobel Prize-winning AlphaFold protein-structure team last month, non-LLM researchers said conflating AI with energy-hungry LLMs sometimes makes climate-focused work harder to advocate for.
Several prominent academics have recently taken university leave for frontier labs, while many AI2050 fellows combine industry and academic jobs. During the past six months, OpenAI models solved real mathematics research problems, prompting experts to question humanity’s future in pure math and one fellow to fear for mathematicians’ mental health. Empirical science may resist automation because data collection is slow. Carnegie Mellon computer scientist Tim Dettmers, who makes models faster and cheaper, argues AI scientists could amplify rather than replace humans. Compute constraints are also driving smaller, more efficient models and new architectures that could produce an academic breakthrough.
Positives
- AI2050 funding helps fellows buy GPUs, a benefit several participating researchers called major.
- Anjalie Field’s study identified less sophisticated model responses to wording more commonly associated with women.
- Tim Dettmers argues AI scientists could help human researchers pursue ideas they otherwise lack time to explore.
- Compute constraints are pushing academic researchers toward smaller, cheaper models and entirely new architectures.
- Slow data collection may make empirical science harder to automate than pure mathematics.
Risks & concerns
- Anthropic and OpenAI withhold Claude and ChatGPT design and training details from outside researchers.
- Repeated queries to OpenAI, Anthropic and Google models can be prohibitively expensive as US federal science funding declines.
- Google DeepMind disbanded its Nobel Prize-winning AlphaFold team last month.
- OpenAI models solved real mathematics problems during the past six months, raising fears about human roles in pure math.
- Several prominent academics have taken university leave for frontier labs, while many AI2050 fellows hold concurrent industry positions.
- Climate AI researchers sometimes struggle for support because the public associates AI primarily with energy-hungry LLMs.
