NVIDIA backs NSF’s NAIRR pilot with dedicated DGX cloud access for hundreds of research projects

NVIDIA says it has contributed to the U.S. National Science Foundation’s NAIRR pilot program by providing cloud-based compute access and hands-on onboarding for researchers running more than 700 projects nationwide. The company’s involvement is framed as a way to shorten the time from model-building to results by giving academic and research teams dedicated access to NVIDIA systems and support during their allocations.

NAIRR, the National Artificial Intelligence Research Resource, is an NSF effort designed to give researchers broader access to large-scale artificial intelligence infrastructure. NVIDIA’s blog describes its contribution as a cloud resource that provides researchers a minimum of four NVIDIA DGX nodes for at least a month, along with technical support to onboard teams and assist them across their projects.

NVIDIA ties the pilot’s output to concrete application areas. It cites work in protein prediction and infectious disease outbreak management, while also pointing to scientific simulations and materials research as beneficiaries of accelerated computing.

One example highlighted by NVIDIA is Polymathic AI, a coalition that includes the Flatiron Institute, Cambridge University, and Lawrence Berkeley National Lab. The group, according to NVIDIA, used NVIDIA GPUs and NVLink interconnect technology to scale up “fluidlike” simulations, supported by a large dataset called “Well,” which the company says will be used to train a foundation model for fluid behavior.

NVIDIA says the resulting model, named Walrus, has been made publicly available along with its data, code, and “pertained weights.” The company also describes the group’s plan to address limitations in the scale and diversity of physics-focused pretraining by exploring scaling laws intended to speed development of more capable scientific foundation models.

Energy research is another focus area in NVIDIA’s account. Researchers at the University of Michigan, led by Professor Venkat Viswanathan in aerospace engineering, are working on a model-fusion framework that combines domain-specific molecular machine learning with general-purpose large language models, with the aim of letting computational scientists explore chemical space more easily, ask chemistry questions in natural language, and identify materials for energy storage and conversion.

In that effort, NVIDIA describes a family of molecular foundation models called MIST (the Molecular Insight SMILES Transformers). It says the models were pretrained on large unlabeled molecular datasets and use a tokenizer named Smirk to better represent nuclear, electronic, geometric, isotopic, and stereochemical information. NVIDIA also says MIST models were fine-tuned on more than 400 structure-property relationships and can match or exceed state-of-the-art performance across benchmarks spanning electrochemistry, quantum chemistry, physiology, and related domains.

NVIDIA attributes the development of MIST to a 40-GPU NVIDIA DGX cluster secured through a NAIRR allocation, plus an additional 200,000 NVIDIA GPU hours on Argonne Leadership Computing Facility’s Polaris cluster. The company adds that the team used NVIDIA’s NGC PyTorch container to support reproducible GPU-accelerated development across different computing environments.

A third use case described by NVIDIA involves infectious disease monitoring. Boston University’s Hariri Institute for Computing and the Center on Emerging Infectious Diseases, in NVIDIA’s telling, is training and evaluating a large language model using NVIDIA-accelerated compute as part of an AI pipeline for BEACON (Biothreats Emergence, Analysis and Communications Network). NVIDIA says the BEACON approach is designed to extract features from online outbreak indicates and produce concise reports for downstream categorization and prioritization.

NVIDIA includes a quote from Ioannis Paschalidis, director of the Hariri Institute, who says report writing previously took several hours and that with the pipeline it takes roughly two minutes. NVIDIA also says BEACON ingests indicates from multiple sources including HealthMap, news and social media feeds, subject-matter experts, and community boards or social media, and that the output can inform clinical practice guidelines and point to gaps where more data is needed.

While NVIDIA’s description provides specific details about its compute contribution and several pilot projects, the company does not provide broader performance metrics for the overall NAIRR program, such as how much timelines or costs were reduced across all 700-plus projects. It also does not disclose the full breakdown of how many projects used NVIDIA resources, how researchers selected workloads, or how outcomes were evaluated relative to prior non-NAIRR approaches.

For what to watch next, NAIRR pilot outcomes may increasingly hinge on how quickly foundation-model and simulation pipelines move from experimentation to repeatable scientific workflows. NVIDIA’s blog indicates that more projects across universities, including Harvard, Stanford, and Colorado State University, are using NAIRR and NVIDIA infrastructure, and observers will be looking for follow-on publications, open model releases, and measured impacts on research timelines and reproducibility.