NSF's $100M AI Hubs Program Will Fund One Award Per State — Which Means the Real Competition Is Inside Your State, and It Ends November 4

There is a category of federal solicitation where the hardest work happens before anyone writes a proposal. NSF 26-513 is one of them.

The State and Regional Artificial Intelligence Infrastructure Hubs program — full title, Expanding Access to Compute for Scientific Discovery — was published July 31, 2026 and announced publicly on August 4. Approximately $100,000,000 is available. NSF anticipates 10 awards per cycle, at $4,000,000 to $12,000,000 total budget over five years each. Full proposals are due Wednesday, November 4, 2026, and on the first Wednesday in November annually thereafter.

Then there is the sentence that reorganizes everything:

Only one award per state or multi-state region.

Combined with a second limit — one proposal per organization per competition, and individuals limited to one PI or co-PI role per deadline — this is not really a national competition with ten winners. It is fifty-odd simultaneous state-level elimination tournaments, each of which has to resolve itself into a single proposal, with no formal mechanism to run that tournament and roughly six weeks left to do it.

Most institutions will not realize this until they are already drafting.

What NSF is actually buying

The most common misreading of this program is that it is a hardware purchase. It is not.

NSF's framing is that the agency funds coordination, workforce development, and faculty training — not the acquisition of the compute infrastructure itself. The theory of the program is that a state or region can assemble, from institutions that individually cannot afford frontier-scale compute, a pooled resource that collectively can. OSTP Director Michael Kratsios put the logic plainly in the announcement:

Regional partners who share in the benefits of discovery will pool their resources to unlock compute at a scale that no individual stakeholder could achieve alone.

Brian Stone, performing the duties of NSF director, framed the purpose as supporting "a national mission to advance AI for science by expanding access to critical AI resources."

Read those two statements together and the proposal you need to write becomes clearer. NSF is not evaluating whether you need compute. Everyone needs compute. NSF is evaluating whether you have assembled a governance structure capable of allocating a shared resource fairly across institutions that are not peers — a flagship R1 and a rural community college in the same consortium, competing for the same queue time.

That is an organizational design problem, and it is what the solicitation-specific review criteria actually ask about:

• Effectiveness of the governance structure for reaching diverse institutions
• Strength of workforce development plans
• Quality of regional engagement and the substance of committed partnerships
• Impact of faculty training on student access to resources
• Adequacy of the computing infrastructure's scale and design

Four of those five are about people and process. One is about the machine.

The industry partners are named, and that is unusual

NSF's announcement lists partner organizations by name: NVIDIA, AMD, Intel, Dell Technologies, Hangar, and the Secunda Innovation Fund. Having three competing silicon vendors and a major systems integrator publicly attached to a single NSF program is not typical, and it tells you something about how NSF expects the compute layer to get built.

NVIDIA published its own description of what it is contributing on August 4: training resources, educator enablement, applied learning content, technical guidance, partner platforms, and tool access. The company says it will help institutions develop degree programs, certificates and credentials in AI application areas including physical AI, automation, healthcare, energy, agriculture, manufacturing, quantum computing and cybersecurity.

Notice what is conspicuously absent from that list: hardware donation at any specified quantity, and any dollar figure at all. The published contribution is curriculum, training and enablement — real value, but not the thing that makes a hub a hub.

The strategic implication for applicants is direct. Do not build a proposal narrative that assumes a vendor will close your compute gap. The named partners give you credible workforce-development content and a plausible technical roadmap. The actual silicon, and the power and cooling to run it, still has to come from somewhere identifiable in your budget, your state's appropriation, your system's capital plan, or a partner's balance sheet — and the review criteria ask you to demonstrate that the scale and design are adequate.

Required partnerships with regional industry or non-profit entities are not optional decoration. They are a solicitation requirement, and a partner letter that promises "collaboration" without naming a resource is the kind of thing panels have learned to discount.

Who can apply, and who is being aimed at

Eligible submitters are:

• US accredited two- and four-year institutions of higher education, explicitly including community colleges
• Non-profit, non-academic organizations based in the US — the solicitation lists museums, observatories, independent research laboratories and professional societies

NSF has been explicit that the stated goal is to reach institutions that do not already have frontier compute access. That is a meaningful signal about how proposals will be read. A consortium led by the one institution in the state that already owns a large cluster, with a membership list that reads as an afterthought, is proposing to subsidize the incumbent. A consortium where the incumbent contributes capacity and the governance structure gives allocation authority to institutions that have none is proposing what the program was designed to fund.

Cost sharing: voluntary committed cost sharing is allowed but not required. In a program whose central thesis is resource pooling, "allowed but not required" should be read as an invitation. If your state legislature, system office, or industry partner has committed capital, put it in.

No letter of intent or preliminary proposal is required. That removes the usual forcing function that would have surfaced competing in-state proposals months ago. Planning grants are available under the standard PAPPG Chapter II.F.1 mechanism — worth knowing for the November 2027 cycle, less useful for this one.

How this fits the rest of the federal AI stack

NSF 26-513 does not stand alone. The announcement positions the hubs as integrating with:

• The NSF-led National AI Research Resource (NAIRR), for which NSF has now established an operations center
• The White House Genesis Mission for AI in science — which went from executive order to a $5 billion coordinating engine with 278 funded projects across 17 national labs (our breakdown)
• TechAccess: AI-Ready America, NSF 26-508, the $224 million state and territory coordination hub program covering all 56 states and territories (our analysis)

That last one matters operationally. TechAccess already funded a coordination structure in every state and territory. NSF 26-513 now funds up to ten states to attach serious compute to a coordination layer that, in many places, already exists and already has a convener.

If your state has a TechAccess coordination hub, that is almost certainly the entity that should be leading — or at minimum brokering — your 26-513 proposal. If it is not, and two of your state's universities are quietly drafting competing proposals, you have a coordination failure that the one-award-per-state rule will convert into a wasted cycle for everyone involved.

There is also a scheduling collision worth flagging: NSF 26-512, the AI Datasets program, carries the identical November 4, 2026 deadline and the same annual first-Wednesday-in-November cadence (our coverage). NSF 26-512 has no limit on proposals per PI or organization; NSF 26-513 limits individuals to one PI or co-PI role per deadline. Your senior AI faculty cannot be everywhere. Sort the role assignments now, because the constraint binds across an office's entire November portfolio, not just this one proposal.

What to do in the next six weeks

Week one: find out who else in your state is writing. This is the single highest-value action available and almost nobody does it early enough. Call the TechAccess hub. Call the state higher-education coordinating board. Call the system office. If there are two proposals forming, one of them is dead on arrival and the sooner that is known, the more usable the surviving one becomes.

Week two: decide governance before you decide architecture. The review criteria lead with governance structure effectiveness for reaching diverse institutions. Write the allocation policy first — who gets queue time, on what basis, adjudicated by whom, with what appeal — and let the technical design follow from the access model rather than the reverse.

Week three: convert partnership intentions into committed resources. "Quality of regional engagement and committed partnerships" is a scored criterion, and the operative word is committed. A letter that names a dollar figure, a hardware quantity, an FTE, or a facility beats three letters of enthusiastic support.

Weeks four through six: write the workforce case as seriously as the infrastructure case. Two of the five solicitation-specific criteria — workforce development plans and faculty training impact on student access — are about whether people can actually use the resource. This is where the named industry partners are genuinely useful, because curriculum, certificates and educator enablement are exactly what they have said they will supply.

The budget range gives you one more piece of strategic information. A request can be anywhere from $4 million to $12 million over five years, a 3x spread. NSF has not said it will award ten at the ceiling, and $100 million divided by ten awards is an average of $10 million — which means the distribution has room for a handful of large hubs and several smaller ones. If your consortium is genuinely regional and multi-state, the top of the range is defensible. If it is a single mid-size state with four institutions, a $5 million request with an airtight governance model is a better bet than a $12 million request with an aspirational one.

Ten awards. Fifty-odd states. One shot per organization, once a year, through at least November 2027. The program will be back. The question is whether your state shows up with one proposal or three.