NSF 26-513 Opens a Second AI Hub Competition: $100M, One Award Per State, and a November 4 Deadline

Academic PIs weighing an NSF artificial intelligence play now have a hard number and a hard date: NSF 26-513, the State and Regional Artificial Intelligence Infrastructure Hubs solicitation posted to grants.gov on August 6, puts roughly $100 million behind about ten cooperative agreements, with full proposals due November 4, 2026.

The Money Pays for People, Not GPUs

The grants.gov listing for NSF 26-513 shows the opportunity posted August 6 with a November 4 close, $100,000,000 in total program funding, no cost-sharing requirement, and three assistance listings: 47.070 (Computer and Information Science and Engineering), 47.076 (STEM Education), and 47.084 (Technology, Innovation and Partnerships). NSF published the solicitation itself — "Expanding Access to Compute for Scientific Discovery" — on July 31. The title promises compute. The budget section delivers something considerably narrower, and misreading it is the fastest way to waste a fall.

NSF will fund four categories of work: consortium coordination, AI infrastructure workforce development, faculty training, and the development of instructional materials for teaching AI-enabled science. It will not fund the compute. Capital and operating costs for GPUs, storage, networking, data systems, software licensing, and cloud services are the consortium's responsibility — states, university systems, industry, philanthropy, or some blend of all four.

That distinction is the entire program. NSF 26-513 is a coordination-and-people award wrapped around hardware somebody else is buying. Individual awards run $4 million to $12 million over five years — roughly $800,000 to $2.4 million per year — issued as cooperative agreements, with about ten anticipated against $40 million to $100 million in anticipated program funding. A $12 million ceiling does not buy a frontier cluster in 2026. It buys the staff, curriculum, allocation governance, and user-support layer that makes a cluster somebody else owns actually usable by researchers who have never held an allocation on one.

For a PI at an institution without a campus GPU cluster, that is the opportunity and the trap in the same sentence. The opportunity is that NSF is explicitly trying to reach institutions outside the roughly two dozen universities that already run serious AI infrastructure — community colleges are named as eligible lead organizations, alongside four-year institutions and non-profit non-academic organizations like independent research laboratories, observatories, museums, and professional societies. The trap is that a proposal arriving without a credible, already-negotiated hardware commitment behind it is not a competitive proposal. It is a wish.

One Award Per State or Region Makes This a Consolidation Contest

The single most consequential line in the solicitation is not financial. It is geographic: only one award will be made per state or multi-state region.

That clause converts NSF 26-513 from a normal competition into something closer to a redistricting fight. If your flagship, your land-grant, your urban research university, and your community college system each submit separately, at most one of you wins and the rest have burned a submission slot. And the slots are scarce by design — an organization may participate in only one proposal per deadline, and each individual may appear as PI, co-PI, or Key Personnel on exactly one proposal per deadline. There is no hedging strategy available. You cannot appear as a co-PI on the flagship's proposal and as PI on your own.

The rational response is consolidation, and NSF clearly designed it that way. The institutions in a given state that would ordinarily compete for the same NSF dollars have a strong incentive to negotiate a single lead, a single consortium, and a subaward structure that distributes the workforce-development and faculty-training money across campuses. PIs who want a seat at that table need to be in the room during August and September, not October. By the time a lead institution's research office has drafted the project description, the Key Personnel list is closed.

The practical read for an individual investigator: decide quickly whether you are building the consortium or joining one. If you are joining, the currency is not your publication record — it is a specific, fundable role in one of the five required components, most often faculty training or curriculum development in a discipline the consortium cannot otherwise cover. Biomedical and health-science PIs are unusually well positioned here, because AI-for-science training portfolios assembled by computer science departments routinely lack anyone who can teach the domain-specific data governance, human-subjects, and reproducibility constraints that life-sciences workflows carry.

Cost Sharing Is Banned, but the Hardware Commitment Is Not Optional

Read the two rules together and the design intent becomes obvious. Voluntary committed cost sharing is prohibited — you may not offer institutional match to buy competitive advantage. But NSF does not fund the compute, which means the compute must exist or be committed by partners outside the NSF budget.

This is not a loophole; it is the mechanism. NSF is refusing to let institutions bid against each other with matching funds while simultaneously requiring that the underlying infrastructure be somebody else's line item. What reviewers will therefore weigh is not the size of a match but the credibility of partner commitments: signed letters from a state agency, a governor's technology office, a university system CIO, a utility or economic-development authority, a regional cloud or hardware vendor. National-level industry endorsements of the program — the usual roster of chip and systems vendors — are not regional commitments and should not be presented as such.

This is the same architectural logic NSF has been applying across its AI portfolio all year. When the agency spent $83 million on six Integrated Data Systems and Services awards, it was buying the connective layer rather than the compute itself — a pattern we broke down in our analysis of the IDSS awards and the national data platform strategy. NSF has decided its comparative advantage is coordination, standards, and human capacity. Hardware is increasingly somebody else's balance sheet.

Why This Is Not TechAccess, and Why Some Campuses Will Chase Both

NSF 26-513 is frequently confused with TechAccess: AI-Ready America (NSF 26-508), and the two are genuinely distinct solicitations with different money, different deadlines, and different missions.

TechAccess funds up to 56 State/Territory Coordination Hubs — one per state, the District of Columbia, and each territory — at roughly $1 million per year for three years, with a program ceiling in the $168 million to $224 million range depending on how a fourth year is exercised. Its round-one letters of intent closed June 16 with full proposals due July 16; round two carries a December 15, 2026 LOI and a January 15, 2027 full-proposal deadline. Its audience is workforce boards, small businesses, K-12 systems, and local government — AI readiness in the economy.

NSF 26-513 is about research compute. Its audience is faculty and students who need GPU hours to do science. The solicitation explicitly encourages complementary engagement between the two programs rather than treating them as alternatives, which means a state could plausibly host both a TechAccess coordination hub and an AI Infrastructure Hub, ideally with deliberate coordination between them.

For PIs, the distinction matters for a mundane reason: the one-proposal-per-person rule applies within 26-513, not across programs. Being Key Personnel on a TechAccess hub does not disqualify you here. But the reverse organizing question is real — if your institution already leads or partners in the state's TechAccess hub, that existing consortium is the obvious skeleton for the 26-513 proposal, and the relationships are already built.

The Five Required Components Are the Review Rubric

Every proposal must address five elements, and treating them as a checklist rather than a narrative is the most common way strong teams produce mediocre submissions:

1. Hub consortium stakeholders, vision, and key deliverables
2. Computing, data, and AI infrastructure description
3. Regional stakeholder partnerships
4. AI infrastructure workforce development plan
5. Faculty training and instructional materials development

Components 4 and 5 are where the NSF money actually goes, and they are the ones academic teams most often underbuild. A workforce development plan that amounts to "we will hold workshops" competes badly against one with named instructors, a defined trainee pipeline with headcounts, articulation between a community college certificate and a four-year pathway, and an evaluation design. Component 2 is where consortia lose on credibility: reviewers will read it looking for whether the described infrastructure exists, is funded, or is aspirational.

Worth remembering that this cycle is being reviewed under NSF's restructured merit review process, which shifted meaningful discretion toward program officers — the implications of which we covered in our analysis of the merit review overhaul. Early contact with the program team at AIInfrastructureHubs@nsf.gov is worth more this year than last.

Eighty Days, No Letter of Intent, No Second Chance This Cycle

There is no letter of intent and no preliminary proposal, which sounds like a relief and is actually a burden: nothing forces early alignment, and there is no low-cost signal of who else in your region is organizing. The first time you learn a peer institution filed is when the award is announced.

The November 4, 2026 deadline recurs — November 3, 2027, then the first Wednesday in November annually thereafter — so a state that misses this cycle is not locked out permanently. But with only one award available per state or region and roughly ten awards per cycle, the first-mover advantage is substantial. A region that lands an award in 2026 has effectively closed the competition for its geography for five years.

Concretely, for the next eighty days: identify whether a lead institution in your state has already begun organizing, secure a named role in one of the five components before the Key Personnel list closes, and pressure-test the hardware commitment. If nobody in your region is organizing, the fastest credible path is a conversation with your state's technology or economic-development office, because that is where the non-NSF compute dollars have to come from.

Search live NSF artificial intelligence infrastructure and compute-access solicitations on Granted to see NSF 26-513 alongside the other open AI-infrastructure opportunities your consortium could stack against it: grantedai.com/grants?q=NSF artificial intelligence infrastructure.