NSF Renewed Two AI Institutes and Let a Third Die. With No New Competition Open, Renewal Is the Only Door — and It Is Peer Reviewed.

Three of the original seven NSF AI Research Institutes reached the end of their five-year awards. Two of them are still running. One is spending down a no-cost extension and preparing to close.

That is the entire program in one sentence, and it is a very different program than the one that got announced in 2020.

In September 2026, NSF renewed the Artificial Intelligence Institute for Advances in Optimization (AI4OPT) with $20 million for a second five-year phase running through 2031, beginning October 1, 2026. In June 2026, it renewed the MIT-led Institute for Artificial Intelligence and Fundamental Interactions (IAIFI), raising annual support from $4 million to $4.98 million — roughly $24.9 million across phase two. And in 2025, it declined to renew AI2ES, the University of Oklahoma-led institute for trustworthy AI in weather, climate, and coastal oceanography, after about $20 million over five years.

Meanwhile, the door for new entrants is shut. The National AI Research Institutes program page lists no upcoming due dates, and its status reads waiting for new publication. The most recent solicitation, NSF 23-610, covered the FY2024 and FY2025 competitions and has not been reissued.

For the roughly 29 institutes in the network — spanning more than 500 collaborating institutions and over $500 million in NSF investment — renewal is not one path among several. It is the path.

Renewal is a competition, and NSF's own guidance says so

The most common misreading of an institute renewal is that it is a continuation with paperwork. It is not. NSF's published renewal advice describes a process that looks much more like a fresh competition than a no-cost extension request:

• The PI is expected to open the conversation with the cognizant program officer during the fourth funding year of the original award.
• The renewal proposal is submitted with the fourth annual project review, and no later than the end of October of the fifth funding year.
• It follows the Traditional Renewals procedures in Chapter V of the PAPPG.
• It goes to external peer review — and reviewers may be given the results of one or more of the institute's annual evaluation reports.
• Decisions normally arrive within six months of submission.
• The renewal performance period is capped at five years.

The clause that should reorganize how an institute operates is the one about annual evaluation reports. Those reports are typically treated as compliance artifacts — filed, acknowledged, forgotten. NSF is telling you they may be handed to your reviewers. That converts four years of routine reporting into the evidentiary record on which a $20 million decision gets made.

The guidance also notes the original solicitations did not promise renewal at all. NSF 20-503 and the solicitations that followed created five-year awards; the renewal opportunity was offered afterward, competitively. Nothing in the architecture guarantees a phase two.

AI2ES is the proof. Director Amy McGovern described what ends when the money does: "We had seven universities involved, and people working together that wouldn't normally work together producing amazing new results, and that's going to stop." Co-PI Philippe Tissot framed the stakes differently: "When you see the rest of the world, our competitors in other countries are not slowing down." An NSF spokesperson offered only that the agency is "deeply grateful for the groundbreaking work of all AI Institutes and remain committed to building on their success," noting that "additional award actions remain possible, subject to appropriations." The institute has one year to spend remaining funds under a no-cost extension, after which it effectively ceases to exist.

What the two survivors did differently

Read the renewal announcements side by side and a pattern emerges that has little to do with which science is more exciting.

AI4OPT did not propose a continuation. It proposed a reorganization. Phase two is structured around thrusts including AI for optimization, optimization for AI, optimization solvers, multiagent learning and optimization, and responsible AI for decision making, with a new emphasis on real-time and interactive optimization. Lead PI Pascal Van Hentenryck stays at Georgia Tech, with co-PIs Bistra Dilkina at USC and Alper Atamtürk at UC Berkeley. USC's share rose to $5.6 million — more than double its phase one funding — and its team expands to five researchers. The named application domains are energy systems, supply chains, and manufacturing: grid complexity under renewable integration, and logistics resilience.

The phase one record it brought to review was concrete and countable: more than 250 publications, 12 technology transfers, and new public datasets including Distributional MIPLIB.

IAIFI made an even sharper pivot. Phase one demonstrated that machine learning can accelerate physics discovery. Phase two inverts the arrow toward what the institute calls the "physics of AI" — using physical reasoning, physical challenges, and physical tools to understand and improve AI systems themselves. Its collaboration widened to include MIT, Harvard, Northeastern, Tufts, and Boston University.

Its evidence was overwhelmingly about people: eight postdoctoral fellows trained, three of whom secured faculty positions; an interdisciplinary PhD program that has awarded 20 doctorates since 2021; nearly 600 applications to the 2026 summer school. Managing director Marisa LaFleur made a point that reads like renewal-review advice in disguise — that "connections among the NSF AI Institutes have been as valuable as the work within them."

Both institutes also expanded education footprints rather than defending them. AI4OPT is extending its Seth Bonder computational and data science camps into the Los Angeles area alongside K-12 programming, teacher training, and faculty development.

Neither renewal argued we should keep doing what we are doing. Both argued the field moved and here is our next problem — with a countable phase one behind it and a widened institutional coalition beside it.

If you are not already inside, the AI money moved

The absence of an open AI Institutes solicitation is a real closure, not a paperwork lag. There is no pending due date to prepare against and no announced reissuance timeline. Planning a $20 million institute proposal on the assumption that NSF 23-610 returns on a predictable cycle is planning against a document that has not existed for two fiscal years.

The federal AI research dollars did not evaporate; they redistributed into instruments with different shapes and different eligibility. NSF has been building out infrastructure and access programs — the NAIRR operations center, the AI datasets program under NSF 26-512, and the TechAccess coordination hubs under NSF 26-508 — while the $1.5 billion foundational research reset reshaped how core CISE and MPS proposals get submitted at all. A team that would have written an institute preliminary proposal is better served assembling a portfolio across those instruments than waiting for a solicitation with no announced return.

For the institutes already funded, the practical instruction is narrower and more urgent. If your award began in 2021 or 2023, your renewal clock is running on a schedule NSF has published: program officer conversation in year four, proposal filed with the fourth annual review and no later than the end of October in year five, external peer review, six months to a decision.

Between now and then, the two things that most reliably distinguished a renewal from a shutdown were a countable phase one record — publications, technology transfers, datasets, degrees conferred, placements — and a phase two that named a new problem instead of extending the old one. Both survivors also broadened their institutional base rather than consolidating it.

Federal science funding rarely announces which door is closing; it just stops posting due dates on one and keeps peer-reviewing the other — and Granted is built to help research teams find the instruments that are actually open before the planning cycle assumes otherwise.