NSF AI Infrastructure Hubs offer up to $12M—but not for the GPUs

The headline number is $100 million. The line that matters more is buried in the solicitation: NSF will not pay for the GPUs, cloud services, storage, networking, or other AI infrastructure.

The new NSF AI Infrastructure Hubs program is built around a trade. The U.S. National Science Foundation may provide a state or regional consortium with $4 million to $12 million over five years for coordination, research enablement, workforce development, and faculty training. The consortium must bring the money and commitments for the compute itself.

NSF expects about 10 awards and says roughly $100 million is available. Full proposals for the first cycle are due November 4, 2026, at 5 p.m. in the submitting organization’s local time. For university leaders, research-computing teams, state officials, and regional AI builders, this is not a shopping-list grant. It is a coalition deadline.

The $100 million is not ten equal checks

The official NSF solicitation anticipates 10 cooperative agreements in each award cycle. A typical proposal should request five years and a total budget between $4 million and $12 million. NSF says the award size should scale with the number of students and scientists served and with the hub’s resources.

Dividing $100 million by 10 produces a neat $10 million average, but applicants should not plan around that shortcut. The published range is wider, the total funding is subject to availability, and the program also anticipates planning grants. If a five-year hub received the low or high end and spending were spread evenly, that would be about $800,000 to $2.4 million per year. That is a simple comparison, not NSF’s disbursement schedule.

NSF will fund the people around the compute, not the compute

This is the most important distinction in the announcement. NSF says it will not fund the acquisition of computing, data, software, networking, storage, cloud services, or other AI systems and services. The state or regional consortium must show that those resources are in hand or being acquired, and that they will be suitable for the full five-year award.

NSF funding is aimed at the pieces that make infrastructure useful: consortium governance, partnerships, AI-infrastructure professionals, scientific support, workforce programs, and faculty training. Eligible staff can include systems administrators, storage and system architects, cybersecurity specialists, network and performance engineers, software engineers, trainers, and user-support experts.

That shifts the proposal from “which cluster should we buy?” to “who will use the resources, who will operate them, and what science becomes possible?” It is the same practical gap that appears in our look at a 1 MW AI compute pod: hardware density is impressive, but power, operations, access, and a real workload decide whether it becomes useful capacity.

One institution, one proposal changes the strategy

An institution may appear in no more than one proposal. If its name appears in several, NSF says only the first received will be reviewed; the rest will be returned without review. An individual also may participate in only one proposal as PI, co-PI, or senior or key personnel for each deadline.

There is another constraint: separately submitted collaborative proposals are not allowed. The partners must file one proposal through one lead organization, with subawards administered by that lead. NSF will make only one award per state or multistate region.

Those rules make coalition design an early decision, not paperwork for the final week. A community college with the right workforce program cannot casually lend its name to two competing teams. A research university cannot wait until October to discover that a state agency, philanthropic fund, or regional industry partner has already committed elsewhere.

Accredited U.S. two- and four-year institutions, including community colleges, may submit. U.S. nonprofit, nonacademic research and education organizations such as independent laboratories, observatories, museums, and professional societies also can be eligible. A region should verify its proposed lead and every participant against the full solicitation before locking the chart.

A strong hub needs a service model, not a vendor showroom

The proposal must address five elements: consortium vision and governance; compute, data, and AI infrastructure; regional partnerships; an AI-infrastructure workforce; and faculty training plus instructional materials. On-premises systems, cloud resources, or a mix are acceptable. What matters is that the resources fit the proposed research and can serve institutions of different sizes.

The NSF announcement says hubs should expand access for researchers, students, and educators while connecting training to regional job needs. They are expected to engage with the National AI Research Resource, or NAIRR, so capacity and lessons can move beyond one campus.

NVIDIA, AMD, Intel, Dell Technologies, Hangar, and the Secunda Innovation Fund are among the private and philanthropic organizations NSF says intend to support participants. In its own program announcement, NVIDIA lists training, educator support, applied learning content, technical guidance, partner platforms, and tools as possible contributions. That is useful, but the region still needs neutral governance, transparent resource allocation, and an architecture that does not become dependent on a single sales pitch.

Infrastructure questions can also collide with local constraints. Our analysis of the Texas data-center grid audit is a reminder that power requests and readiness claims need evidence, not just ambition.

A 30-day coalition test before anyone writes the proposal

A pass means the consortium can explain who receives capacity, how requests are judged, what support users get, how sensitive data is handled, and which partner pays when infrastructure costs change. A fail means the proposal is still a collection of logos around an equipment list.

There is no required letter of intent or preliminary proposal. That removes one administrative stage, but it also removes an early forcing function. Teams should create their own internal cutoff well before November 4.

My verdict: start with the missing scientific capability

The NSF AI Infrastructure Hubs program could widen access to serious AI resources beyond a few wealthy campuses. Its design is also intentionally demanding: the region must finance the infrastructure, assemble one credible coalition, and show how people will turn capacity into science, teaching, and jobs.

I would not begin with a GPU count. Begin with a researcher or student who cannot do a valuable piece of work today. Trace the compute, data, software, support, and training that would remove that barrier. Then ask whether the proposed consortium is the smallest group capable of keeping that service running for five years.

If the answer is clear, the $4 million to $12 million award can be catalytic. If the answer is vague, more partners and more hardware will only make the gap harder to see.

Go deeper

• NSF 26-513: full solicitation, eligibility, budget, and deadlines
• NSF: State and Regional AI Infrastructure Hubs announcement
• NSF: National AI Research Resource
• NVIDIA: planned support for the regional hub program

Checked August 5, 2026. Funding is subject to availability, and applicants should use the current NSF solicitation and Proposal & Award Policies & Procedures Guide when preparing a submission.