Artificial intelligence is accelerating scientific breakthroughs in fields from genomics to climate modeling, yet a growing number of researchers cannot actually use the most powerful tools. The barrier is not skepticism or lack of skill. It is a combination of scarce computing power, strict data governance rules, and a widening resource gap between universities and the private sector that threatens to concentrate AI-driven discovery in the hands of a few well-funded corporations.
The Computing Gap Between Industry and Academia
Building and running large AI models demands enormous computational resources, and most academic institutions simply do not have them. Developing large AI models requires massive datasets and massive computational power to process them, according to a study published in Nature. Many university scientists are frustrated by the limited amount of computing power available to them for AI research, as a separate Nature investigation reported, with one researcher noting that “the gap between academic and industry computing access is super important.” That frustration is backed by hard numbers: the OECD uses Top500 supercomputer rankings as a proxy for compute capacity and has warned that concentration of that capacity can imply worsening compute divides across nations and institutions alike.
The practical result is that industry now dominates frontier AI work. Nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023, according to the 2025 AI Index Report from Stanford HAI. Academia remains the top source of highly cited AI models, which means universities still produce influential ideas but increasingly lack the hardware to test them at scale. Industry is now outpacing academia in cutting-edge AI research and development, accelerating a movement of talent between sectors that further drains university labs of expertise.
Data Rules That Block Sensitive Research
Even researchers who can access computing power face a second wall: data governance restrictions that prevent them from feeding sensitive information into generative AI systems. The NIH issued guidance stating that controlled-access genomic data and derivatives cannot be shared via prompts to AI tools or with unauthorized entities. For scientists studying rare cancers, inherited diseases, or population-level genetic patterns, this effectively means the most advanced generative AI platforms are off-limits for their most valuable datasets. In most institutions, large, representative, and high-quality datasets are entirely absent or scarce, particularly for rare diseases, due to stringent data protection requirements that make it difficult to pool information across hospitals and borders.
Federal agencies face the same tension. A Government Accountability Office audit found that agencies actively adopting generative AI are simultaneously restricting its use to prevent leakage of sensitive data, with policy compliance and budget constraints standing as top obstacles. AI use cases in federal inventories rose from 571 in 2023 to 1,110 in 2024, nearly doubling demand while governance frameworks lag behind. The National Institute of Standards and Technology has tried to close that gap by defining and operationalizing generative AI risks, including privacy and security concerns, in its AI 600-1 risk management profile. But regulatory frameworks are still struggling to keep pace with the speed of AI development, creating a persistent mismatch between what researchers want to do and what institutions allow.
NAIRR: A Federal Attempt to Open the Door
The U.S. government’s flagship answer to this access problem is the National Artificial Intelligence Research Resource, a shared infrastructure pilot run by the National Science Foundation. Through a dedicated portal for NAIRR information, the program is designed to provide access to advanced computing, datasets, models, software, training, and user support specifically for researchers and educators who lack those resources. The pilot operates through a public-private partner structure, drawing contributions from technology companies alongside federal investment and aiming to lower barriers for institutions that cannot afford their own large-scale clusters.
Early results show meaningful but limited reach. The NAIRR Pilot connected more than 400 U.S. research teams, and its expansion now involves dozens of agencies and partners that are contributing compute, cloud credits, and data resources. The National Science Foundation has also highlighted the broader goal of democratizing AI research so that cutting-edge tools are not confined to a handful of companies. Still, 400 teams is a small fraction of the thousands of university labs across the country that could benefit. The program acknowledges the scale of the problem but has not yet matched it, and demand for allocations already outstrips available capacity.
Funding Channels and the Limits of Grants
For many academic groups, the only realistic path to more AI capacity is through competitive grants. Federal calls increasingly encourage proposals that integrate advanced computing and machine learning, but winning such support is far from guaranteed. Investigators are expected to navigate complex application processes, align with agency priorities, and demonstrate both technical readiness and social responsibility in their AI plans. Centralized portals such as federal grants listings and the National Science Foundation’s own research portal are intended to make opportunities more visible, yet smaller institutions often lack the administrative staff to track, interpret, and apply for these programs at scale.
Even successful grants do not always solve the underlying access problem. Awards may cover a few years of cloud credits or time on national supercomputers but leave long-term sustainability uncertain once the funding cycle ends. Some researchers rely on shared national facilities cataloged in NSF project records to piggyback on existing infrastructure, but queues can be long and usage caps strict. As AI models grow larger and more data-hungry, the gap between what a typical grant can buy and what frontier research requires is widening. Without sustained investment in shared infrastructure, one-off awards risk becoming temporary patches rather than structural solutions.
Why Access Gaps Distort Science Itself
The consequences of unequal access extend beyond individual frustration. When only well-funded labs and corporations can run frontier AI experiments, the diversity of scientific questions being asked narrows. Yale researchers have warned that AI tools can weaken the production of scientific knowledge by creating “monocultures of cognition,” where a limited set of methods, assumptions, and training backgrounds shapes what gets studied and how. Michigan State University has similarly flagged risks of potential misinformation, inaccuracies, and overreliance on automated systems in science communication, particularly when researchers cannot independently validate AI-generated analyses. If only a small group of institutions can systematically test and critique these tools, errors and biases are more likely to go unnoticed.
There are also equity implications. Under-resourced universities, including many community colleges and minority-serving institutions, are less able to offer students hands-on experience with state-of-the-art AI, even as those skills become prerequisites for scientific careers. That dynamic threatens to reinforce existing inequalities in who becomes a principal investigator, who leads major collaborations, and whose questions shape the research agenda. Efforts like NAIRR, targeted grant programs, and stricter data protection rules are all attempts to address real risks, but unless they are scaled and coordinated, they may inadvertently cement a two-tier system in which some scientists design the future of AI-driven discovery while others watch from the sidelines.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
