How the National Artificial Intelligence Research Resource can pilot inclusive AI

Side view portrait of young black man as network engineer working with servers in data center and holding laptop
Photo credit: Shutterstock

Though the impacts of AI research and development (R&D) may be like nothing seen before, its path appears strikingly similar to the early development of previous digital technologies such as smartphones or virtual reality.

From patents to job postings, the heavy concentration of early AI activities in a few major tech hubs has reinforced a strong and geographically uneven “superstar” dynamic (see map 1). Thanks to that, many communities and demographic groups with fewer resources are being left without the talent, investments, and infrastructure needed for AI research and commercial  scale-up.

Central to this challenge is the issue of what’s called “compute”—the infrastructural elements that drive digital systems and that have been a scarce resource for many researchers. As AI models grow increasingly complex and data-intensive, the infrastructure needed to train and run these models becomes prohibitively expensive, placing a significant financial strain on academic budgets. Even top-tier research universities are struggling to keep up with the escalating demand for and costs of computational resources. The situation is even more dire at smaller and underresourced institutions, where the issue is compounded by a shortage of AI expertise.

And so federal agencies are working to create the National Artificial Intelligence Research Resource (NAIRR) aimed at democratizing AI research. The NAIRR is a proposed national initiative that aims to address many communities’ and peoples’ uneven access to AI-ready computing power, high-quality datasets, educational tools, and user support—and narrow the gaps.

To get the ball rolling, the National Science Foundation (NSF) in January launched a pilot version of the NAIRR program as a proof-of-concept for the eventual establishment of a full-scale version of the program, with contributions from 12 other agencies and 26 industry partners. As part of the launch, the architects of the pilot announced its first “allocation call:” an opportunity for researchers to request access to advanced computing resources across six different computing facilities.

All of this is an important step toward addressing the access challenge. The pilot acknowledges and begins to push back against the inherent “winner-take-most” dynamics in the competition to build local AI research hubs. However, questions abound about the NAIRR’s pilot’s capacity to meaningfully advance geographic and demographic research inclusion on AI.

Currently, few of the core facilities participating in the pilot are providing state-of-the-art graphical processing units (GPUs) that are critical for training large language models (LLMs). Out of the more than 150 proposals that the pilot received, it was only able to award 35 projects in its first round of allocations. What’s more, recent cuts to NSF’s budget raise concerns about whether the NAIRR Pilot will be funded adequately. The fear, then, is that these efforts will not be enough to ameliorate an emerging AI research map that is highly skewed in favor of geographic and demographic communities that already have access to the technology.

Two issues stand out as the NAIRR Pilot begins to sort out its AI-research inclusion agenda

All of this raises the question of what, exactly, the NAIRR Pilot should focus on as it refines the roles and activities of a longer-term, full-scale program. And here, two issues loom large.

First, the cost of access to the resources needed to carry out foundational AI research has reached new extremes. As AI models scale up to achieve more general capabilities, the demand for compute for AI research has reached a scale unavailable to many academic institutions.

The gap in access to compute resources between industry and academia is increasingly pronounced. In January 2024, Meta announced plans to acquire 350,000 of NVIDIA’s latest H100 GPUs by the end of the year. In contrast, one of the fastest supercomputers available for academic research, the Summit at Oak Ridge National Laboratory, is equipped with only about 27,000 V100 GPUs, which are already a generation older than Meta’s H100s.

The problem with this is that the high costs and scarcity of state-of-the-art GPUs keep government actors and non-elite academic institutions from querying, fine-tuning, and training LLMs to develop their own advances. (The costs strain many “elite” institutions as well.) This limitation not only slows the pace of innovation within academic circles but may also affect the diversity of ideas and applications emerging from AI research.

Yet there’s a second glaring challenge: equal access to compute resources does not always mean equal demographic participation in AI. In this vein, serious gaps remain in the availability of AI research and adoption skills at mid- and lower-tier research universities. Many of these institutions face basic shortages of tech skills, ranging from research scientists to technical support staff, a problem that can’t be solved with more state-of-the-art GPUs.

This participation concern is especially acute at Historically Black Colleges and Universities (HBCUs) and Minority Serving Institutions (MSIs), where there has been a historical underinvestment in computer science resources, program building, software development, and talent training. Researchers from these institutions, which are already underrepresented at the leading AI conferences, now risk falling behind in the widening AI divide.

Without the human capital for AI development, such as experienced faculty and research support staff, underresourced universities (and their regions) face significant barriers to entry. In light of this, the lack of training opportunities and local expertise will likely deepen geographic disparities in AI, concentrating development in already GPU-rich institutions.

What should be done and how can the NAIRR Pilot help ameliorate the situation?

Given these issues, a few initial outlines of how the NAIRR Pilot might help counter the emerging inequities of AI R&D can be envisioned.

The solution to the shortage of state-of-the-art compute resources in academia ultimately requires the acquisition or assembly of numerous large-scale GPU clusters. Cutting-edge AI research requires larger supplies of compute resources, but too few public institutions can afford the necessary infrastructure and energy. This means that the federal government needs to step up its investments in AI-suitable computational resources to close the most significant shortages of such resources and to meet related capability needs within academia. (Several bills in the Senate—such as the CREATE AI Act and the Spectrum and National Security Act—could prove helpful but appear unlikely to move this year.)

The NAIRR Pilot, meanwhile, has already demonstrated the appetite of the research community for GPUs in its first-round allocations, where most of the award recipients requested some GPU hours. However, at its current level of funding, the pilot can only address the compute issue by leveraging existing GPU resources to expand its allocations. In the short term, therefore, the pilot should facilitate wider access to the state-of-the-art computational resources available at various universities, regional facilities, and national labs. Moreover, the pilot should enable partnerships with these entities to make future opportunities such as research competitions and grants more accessible and visible to researchers.

To address the human capital challenge, meanwhile, there is no doubt that substantial investments in AI skills development are necessary. To be sure, neither the NAIRR Pilot nor any likely research resource will have the heft to close the talent inclusion gap by itself. Yet with that said, the NAIRR enterprise can play an important democratizing role by making its resources more accessible to underserved academic communities and by creating opportunities for low-resourced universities to build capacity.

At the root of the AI skills deficit resides an acute need for educational tools or resources that prepare students to engage in AI development. It’s important, therefore, that the NAIRR makes strides in bringing compute resources to the classrooms, especially those that serve underrepresented groups. The pilot is already allocating GPU resources and interactive computing notebooks for educators to teach students foundational AI concepts. The NAIRR should expand on these resources, adding AI sandboxes and teaching kits (like one developed by NVIDIA) to the mix along with other envisioned software tools and tiered technical support. Without these resources, qualified AI talent will remain in short supply.

Even outside of the classroom, the distribution of compute resources remains inequitable. A considerable proportion of the pilot’s initial allocations were granted to large research universities, which wasn’t surprising given the tight deadline for submissions. Larger, well-resourced universities that can move fast were destined to prosper under the pilot’s first-come-first-served approach. But this leaves out smaller institutions without the bandwidth to prepare and submit competitive proposals. To accommodate these institutions, the pilot should reserve resources in anticipation of their submissions and even allocate a minimum amount of compute time to all applicants if possible. This would ensure that, by the time they apply for allocations, there will still be resources available.

Additionally, the NAIRR can create pathways for smaller institutions to tap into the knowledge and resources of the broader research community by encouraging cross-institutional collaboration. In this vein, the NAIRR Pilot should prioritize proposals with collaborators from different universities in the pilot’s allocation process. To do this, the pilot could give selection preference to projects that demonstrate a meaningful and equitable partnerships across institutions. It could also issue specific calls for proposals that advance interdisciplinary research and that target areas that benefit from diverse inputs. This kind of collaboration facilitates the exchange of expertise and ideas, broadening the perspectives in and applications of AI research.

Partnerships—including one between Norfolk State University, the Yale Center for Emotional Intelligence, and the company Mainstay that developed a culturally responsive chatbot—offer an additional blueprint for connecting researchers at different universities. This way, underresourced researchers can gain valuable experience, expand their professional networks, and increase their visibility through association with established research institutions. These researchers can also offer unique viewpoints that are typically drowned out in the tech space, ultimately enriching the research process and outcomes.

While the NAIRR Pilot is just an initial step toward tackling AI access and participation concerns, it could well mark the beginning of a serious drive to address the nation’s already-stark geographic and demographic disparities in AI research. Supporting AI researchers in academia will necessarily involve more than funding research activities. Realizing the full vision of the NAIRR to this end will clearly require building and sustaining both the technical and human infrastructure that underpins these scientific endeavors. This means getting the most advanced GPU models into the hands of researchers at the bleeding edge and urgently creating capacity-building opportunities for communities that technology has historically left behind.