The National Science Foundation's National Artificial Intelligence Research Resource pilot has backed more than 700 U.S. research projects in its first two years, running on NVIDIA DGX infrastructure that gave each grantee dedicated access to a minimum of four DGX nodes for at least a month. The program, known as NAIRR, is the federal government's most concrete attempt to give academic scientists the kind of compute that until recently sat almost entirely inside frontier AI labs. NVIDIA supplied the cloud-based hardware and onboarding support; the science came from the universities.
The breadth is the headline. NAIRR projects now span protein structure prediction, fluid-dynamics foundation models, battery chemistry, and global infectious-disease surveillance. The pilot's design — guaranteed allocations on a specific reference architecture rather than scattered credits — let labs commit to month-scale training runs that would be impractical on shared university clusters.
Polymathic AI, a coalition drawn from the Flatiron Institute, Cambridge University and Lawrence Berkeley National Lab, used NVIDIA GPUs and NVLink interconnects to build the Well, a dataset aimed at training the largest foundation model to date for fluidlike physical behavior. The resulting model, Walrus, has been released publicly along with its dataset, code and pretrained weights. The group's stated next step is to map scaling laws specifically for scientific foundation models, an area where the standard text-LLM playbook has yet to be properly translated.
Key facts
- 01The NSF's NAIRR pilot has supported over 700 research projects across the U.S. in its first two years.
- 02NVIDIA provided each researcher dedicated access to a minimum of four NVIDIA DGX nodes for at least a month.
- 03University of Michigan's MIST models were trained on a 40-GPU NVIDIA DGX cluster plus 200,000 NVIDIA GPU hours on ALCF's Polaris.
- 04Boston University's BEACON pipeline cut infectious-disease outbreak report drafting from several hours to roughly two minutes.
- 05Polymathic AI released Walrus, a publicly available foundation model for fluidlike behavior, trained on its 'Well' dataset.
At the University of Michigan, a team led by Venkat Viswanathan in the Department of Aerospace Engineering is using NAIRR compute to fuse domain-specific molecular models with general-purpose large language models. Their family of models, the Molecular Insight SMILES Transformers or MIST, was pretrained on unlabeled molecular datasets using a custom tokenizer called Smirk and then fine-tuned on more than 400 structure-property relationships across electrochemistry, quantum chemistry and physiology.
The Michigan group's compute footprint is concrete: a 40-GPU NVIDIA DGX cluster obtained through a NAIRR allocation, plus an additional 200,000 NVIDIA GPU hours on the Argonne Leadership Computing Facility's Polaris cluster. The team standardized on NVIDIA's NGC PyTorch container to keep training reproducible across the two environments. The practical aim is to let computational chemists query chemical space in natural language and surface candidate materials for next-generation batteries and aviation-scale energy storage.
Boston University's Hariri Institute for Computing and Center on Emerging Infectious Diseases is training an LLM on NVIDIA accelerated compute for an outbreak-monitoring system called BEACON, the Biothreats Emergence, Analysis and Communications Network. The model ingests signals from HealthMap, news feeds, social media and field communications, then drafts structured outbreak reports for clinicians and public-health analysts. Internationally deployed doctors, government organizations and academic researchers are already using BEACON in the field.
The time savings are the most legible benefit of the NAIRR pilot to date. Other universities active on NAIRR include Harvard, Stanford and Colorado State, each running smaller projects under the same DGX-allocation model. The pilot is still formally a pilot — the question Congress and NSF face now is whether to turn it into a standing program, and at what scale.
The skeptical read is that 700 projects across two years is a modest output relative to the cost of building a permanent national AI research infrastructure, and that the most ambitious science still happens inside well-funded private labs with budgets an order of magnitude larger than anything an NSF allocation provides. Critics inside the academic-computing community have also flagged that month-long allocations are too short for the largest pretraining runs, forcing groups like Michigan's to stitch together NAIRR time with separate DOE supercomputer hours.
For NVIDIA, the NAIRR pilot is a quiet but strategically useful program: it standardizes American academic AI research on the DGX reference architecture and the NGC software stack, the same combination Jensen Huang has been positioning as the default substrate for what he calls agentic scientific AI. If NAIRR graduates from pilot to permanent program — and the early project list gives Congress a credible case to do so — the downstream effect is a generation of U.S.-trained researchers whose first serious model was built on NVIDIA hardware, with all the lock-in that implies.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
