NSF Just Made the NAIRR Permanent — Award 2616251, $35M, and a Monthly Allocation Cycle That Is Not a Grant Competition

The National Artificial Intelligence Research Resource was built with an expiration date. When NSF launched the pilot in January 2024, it was a two-year demonstration: a stitched-together federation of federal supercomputers, industry cloud credits, datasets, and pre-trained models, held together by a small NSF team and a lot of goodwill from resource providers. The authorization ran to January 2026. After that, nothing was promised.

On September 1, 2026, NSF announced that the demonstration is over and the infrastructure is staying. The agency made a five-year, $35 million cooperative agreement — NSF award 2616251 — to the San Diego Supercomputer Center at UC San Diego to stand up and operate the NSF NAIRR Operations Center (NAIRR-OC), in partnership with the Texas Advanced Computing Center at UT Austin.

SDSC Director Frank Würthwein is the principal investigator. Co-PIs are Ashley Atkins (SDSC deputy director), Maytal Dahan (TACC's director of advanced computing interfaces), Subhashini Sivagnanam, and Mahidhar Tatineni. That personnel list is not filler — it tells you the center is being built by the people who already run national allocation systems, not by a new organization learning the job.

The number that justified the award

In roughly two and a half years, the NAIRR pilot connected about 900 research teams and educators across all 50 states, working with more than 25 private-sector organizations alongside federal and academic resource providers. NSF's own framing puts the total federated capacity in the neighborhood of 3.77 exaFLOPS.

That is a strange kind of success. Nine hundred teams is a rounding error against the number of NSF-funded investigators, and it is enormous relative to a pilot with no dedicated operations staff, no unified portal, and no support desk. The pilot proved demand existed. It did not prove the thing could scale — which is precisely the gap the operations center exists to close.

Read the award's scope of work and it is almost entirely plumbing: coordinate public and private resource providers, integrate computing with data and models and tools, operate a single national NAIRR portal, provide user support, training, and documentation, and build community feedback channels. There is workforce development in there too. There is very little new science.

That is the correct design. The binding constraint on NAIRR was never GPU count. It was that a researcher at a regional comprehensive university had no obvious front door, no one to ask when a job failed, and no confidence the resource would exist in eighteen months.

NAIRR is not a grant, and applying to it does not work like one

This is the part most researchers get wrong, and it is the single most useful thing to understand about the announcement.

You do not write a proposal, wait six months, and receive money. You request an allocation — access to compute, models, or data — through the NAIRR submission site at submit-nairr.xras.org, using a completed form and a three-page PDF. Three pages. Not fifteen.

The cycle is monthly: requests submitted by the 15th of a month are reviewed and decided by the end of the following month. There are two practical tracks — a Start-Up allocation of roughly three months, meant for benchmarking, feasibility, and getting your code running, and a full Research Project allocation of about twelve months.

Eligibility is far broader than most people assume. US-based researchers and educators at academic institutions qualify. So do nonprofits, federal agencies and FFRDCs, state, local, and tribal agencies, and startups and small businesses that hold federal grants. Graduate students can apply directly with a support letter from a faculty advisor. You must use an institutional email address, not a personal one.

The tradeoff is openness. Project results must be open and publishable, disseminated through the literature, with products made publicly available. Your project abstract, your name, and your affiliation get posted on the NAIRR website. If your work is proprietary or export-controlled, this is the wrong resource, and no amount of clever framing changes that.

Specific resources worth knowing: NVIDIA's contribution provides dedicated access to a minimum of four DGX nodes for at least a month, and a Deep Partnership track with the NSF Leadership-Class Computing Facility puts selected teams on thousands of GB200 GPUs — a scale that is otherwise unreachable without a major facility allocation.

The strategic implication for anyone writing an NSF, DOE, or NIH proposal right now: a Start-Up allocation obtained in October gives you preliminary results on real hardware by January, before the January and February deadlines. A three-page request that turns into a benchmarked pilot study is one of the highest-return uses of a graduate student's month in the entire federal system. We walk through the broader landscape of free GPU access in our guide to AI compute grants.

The operations center is one leg of a four-part build

The NAIRR-OC does not stand alone. NSF explicitly ties it to three sibling initiatives, and reading them together tells you where the agency is putting AI infrastructure money through 2027.

State and Regional AI Infrastructure Hubs (NSF 26-513) is the one with an open deadline. Roughly $100 million, about 10 awards, $4–12 million each over five years, with a hard constraint that matters enormously: one award per state or multi-state region, one proposal per institution, and any individual may serve as PI or co-PI on only one proposal. Full proposals are due November 4, 2026. Critically, NSF funds the consortium coordination, workforce development, and faculty training — the members are expected to bring the actual compute, on-premises or cloud. If your institution is not already in conversation with a state-level consortium, you are late; we covered the mechanics in our analysis of the AI Infrastructure Hubs solicitation.

Integrated Data Systems and Services (IDSS) carries roughly $83 million and builds the data backbone — the part of AI-for-science that is unglamorous and load-bearing. See our IDSS breakdown.

Unlocking Dataset Value for AI-Enabled Scientific Discovery targets the datasets themselves.

Together with the DOE's Genesis Mission, the architecture is now legible: DOE funds the mission-driven science on national lab hardware, NSF funds the shared access layer, and NAIRR-OC is the switchboard between them.

The uncomfortable context

The announcement lands inside a policy environment that is not uniformly generous to researchers. The July 2026 AI Action Plan called for exactly this kind of public-private compute investment, and OSTP's "Science: A New Golden Age" report argued for redesigning how federal research gets funded. In the same period, NSF terminated a large volume of awards, and the agency's FY2027 Graduate Research Fellowship solicitation cut expected awards to 1,500 from nearly 2,600.

Compute is being funded. People, in several programs, are not. That asymmetry is the actual story of federal AI research policy in 2026, and it has a direct strategic consequence: infrastructure access is now the cheapest thing to get and the most expensive thing to substitute for. A lab that cannot hire a third graduate student can still, for the cost of a three-page document, put the two it has on hardware that would cost six figures to rent.

The pilot's expiration date was January 2026. It is September, the resource is still running, and it now has a five-year cooperative agreement and a staffed help desk behind it. For anyone who held off applying because the whole thing looked temporary, that objection is gone — and the next monthly review deadline is the 15th.

If you are mapping which of these programs your work actually fits, and which deadlines are worth building a team around, Granted can help you sort a crowded federal AI portfolio into a short list you can act on.