DARPA and NSF Just Launched AI Forge — a University-Only Forum With $750K to $3M Project Ventures, a June 22 RFI Deadline, and a Mandate to Build the National-Security AI Research Stack Frontier Labs Won't Touch

On June 1, 2026, DARPA and the National Science Foundation jointly released the AI Forge initiative through a Request for Information on SAM.gov and a fifty-page report titled Critical AI Challenges for National Security. AI Forge is a university-only research forum that will fund what the agencies are calling Project Ventures — one-year university-led research awards of roughly $750,000 to $3 million each, with multiple awards expected annually. The forum will be administered by an independent nonprofit launching in summer 2026 and will coordinate with the Center for AI Standards and Innovation (CAISI) at NIST. The RFI closes on June 22 at 5:00 PM ET. Each U.S. university — including military service academies — may submit only one authorized response.

The structural choice matters as much as the funding. DARPA and NSF are not running this through their existing grant mechanisms. They are creating a new forum, administered by a third-party nonprofit, that will function as a continuously-revisited pipeline rather than a discrete solicitation cycle. The agencies have committed to revisiting research challenges every six months. Intellectual property generated by AI Forge awards is expected to be shared among forum participants — government, industry, and academic — "preferably through open-source licensing approaches." For research universities, this is a fundamentally different contract than the patent-protection logic that has governed federal AI grants since the early Obama-era CRA programs.

DARPA program manager Matthew Marge framed the initiative as bridging "high-risk, high-reward research that requires both massive scale and deep, mission-driven work." That phrasing is a deliberate signal that AI Forge is not a competitor to OpenAI, Anthropic, or Google DeepMind. It is a complement, structured around the parts of the national-security AI research stack that those labs have no economic incentive to build.

Why a University-Only Forum, and Why Now

The three thrust areas of AI Forge — AI interpretability, AI control, and adversarial robustness — are precisely the parts of the modern AI research agenda that the leading frontier labs publish about but do not optimize for at the system level. Interpretability research at Anthropic, OpenAI, and DeepMind operates within the constraints of products that need to ship and revenue lines that need to grow. National-security AI deployment cannot operate that way. A defense system needs verifiable, bounded behavior under adversarial conditions before it goes into a forward environment. Frontier-lab interpretability work is necessary but not sufficient.

The Critical AI Challenges for National Security report makes this argument explicitly. The agencies frame interpretability as understanding "AI system behavior, decisions, and impacts beyond routine settings" — the long tail that consumer products do not need to cover. AI control is defined as "tools for verifiable evidence of reliable, bounded model behavior with human oversight." Adversarial robustness is defined as "building resilient AI that maintains integrity under deliberate attack." Each of these is a research problem with no clean commercial-product analog and a clear national-security application.

The university-only constraint follows from this analysis. Universities have the long-horizon research culture, the talent pipeline, and the publication norms that make open-licensed, foundational research feasible. Frontier labs do not. Federally Funded Research and Development Centers and national laboratories are absorbing other parts of the national-security AI portfolio under DARPA Information Innovation Office and DOE Office of Science programs. AI Forge fills a specific gap: deep, mission-aligned, open-licensed foundational research at universities, coordinated through a continuously-active forum rather than discrete grant cycles.

Reading the Award Architecture

The $750,000 to $3 million Project Ventures award size is deliberately tuned. It is large enough to fund a principal investigator, two to four postdoctoral researchers or graduate students, computing time on national-security-compatible infrastructure, and equipment for at least one full project cycle. It is not large enough to be a multi-year center grant. The one-year duration reinforces that AI Forge is built for iteration: a research team is expected to produce results inside a year, return to the forum, and either renew, redirect, or hand off the work to a transition partner.

This is closer to the DARPA Defense Sciences Office program model than to the NSF center-grant model. DARPA program managers run their portfolios on iteration cycles rather than five-year commitments. NSF has historically run on three-to-five-year center grants and individual investigator awards. The AI Forge architecture inherits the DARPA cadence and grafts NSF's university-engagement infrastructure onto it. For a research-intensive university accustomed to the NSF grant rhythm, this is a meaningful shift. Faculty preparing AI Forge proposals should plan around producing milestone results inside twelve months and structuring postdoctoral and graduate-student appointments accordingly.

The independent nonprofit administrator is the other architectural choice worth reading carefully. By moving forum administration out of DARPA and NSF and into a third-party entity, the agencies are creating a buffer that allows the forum to operate under different procurement and personnel rules than either parent agency. The nonprofit can move faster on contracting than DARPA can, and it can engage industry participants on terms that the NSF cooperative-agreement framework cannot support. The architecture is similar to how the SemiConductor Research Corporation has historically administered industry-government research consortia, and to how MITRE has administered FFRDC-style coordinating bodies. Universities should expect the nonprofit administrator's contracting terms to look more like a research consortium than a federal grant.

What a Strong RFI Response Looks Like Before June 22

The RFI is not a proposal. It is a capability submission that will populate a repository of U.S. universities interested in AI Forge participation. The agencies will use the repository to identify partners for the forum and to scope the Project Venture solicitations that follow. The RFI asks for four things: AI research expertise, computing infrastructure, national-security partnerships, and technology-transition experience.

Computing infrastructure is the easiest dimension to misjudge. AI Forge research will involve work on models and datasets that are export-controlled or classified at various levels. Universities responding to the RFI should be specific about which computing environments they can operate in — CUI-compliant Azure Government, AWS Top Secret region access through subcontract relationships, DOE Leadership Computing Facility allocations, and the like. A computer-science department running on commodity cloud infrastructure is not, by itself, a credible AI Forge partner for the national-security thrust areas.

National-security partnerships are similarly load-bearing. The RFI is signaling that the agencies want universities with active relationships across Department of War elements — Navy, Air Force, Army research labs; the Office of Naval Research; the Air Force Office of Scientific Research; combatant commands; the intelligence community's IC-CAE programs and IARPA. Universities without these existing relationships can still respond, but their response should be honest about where they are starting from and what partnerships they are building toward.

Technology-transition experience is the dimension most universities under-claim. Transition does not mean commercialization through spinouts and IP licensing — it means moving research outputs into operational use by a government customer. Universities should highlight prior DARPA TA-3 transitions, NSF I-Corps cohorts that landed at agency customers, OUSD(R&E) RIF projects, and similar pathways. This is where the open-licensing expectation has the largest downstream effect: universities accustomed to negotiating exclusive IP retention should expect AI Forge participation to require a different IP posture.

The Big Picture

AI Forge is the most significant joint DARPA-NSF research vehicle in a decade, and it lands at a moment when the federal research enterprise is under unusual structural strain. The Partnership for Public Service has reported nearly 118,000 science-related federal employee departures between September 2024 and February 2026, and project-grant obligations from science agencies fell roughly 24 percent from 2024 to 2025. Universities have responded with internal bridge programs — the Johns Hopkins Research Resilience Fund, at $60 million annually for two years, is the largest example — but bridge funding is not a substitute for a federal research vehicle that points the next decade of AI research at problems frontier labs will not solve.

The June 22 RFI deadline is binding, and the response window is narrow on purpose. Universities serious about AI Forge participation should be assembling cross-departmental responses now: computer science, electrical engineering, applied mathematics, and operations research faculty paired with research-security and contracting offices that can speak to the IP, computing-environment, and transition dimensions credibly. The forum launch in summer 2026 will set the participant roster for the first Project Venture solicitations, and missing the RFI means missing the on-ramp.