NSF Just Finished Its 2026 X-Labs Portfolio. The One-Proposal Rule Is Now a Five-Topic Chess Problem.

On September 16, 2026, the National Science Foundation announced three additional topics under the NSF X-Labs initiative, the $1.5 billion program the agency launched in May to stand up independent, milestone-driven research teams outside the standard grant architecture. NSF's own framing was that the announcement "completes the portfolio of topics for calendar year 2026."

That sentence is the news. The three new topics are Artificial Intelligence for Physical Systems — open now, with a Q&A webinar on Wednesday, October 14, 1–2 p.m. EDT — plus From Sequence to Function and Computation at the Limit of Physics, both slated for release this fall. Alongside them, the Directorate for Technology, Innovation and Partnerships issued a Request for Information on potential future X-Labs topics, with responses due October 30, 2026.

Taken together with the two May topics — Scientific Instrumentation for Sensing and Imaging, and Quantum Systems: Interconnects and Integrated Photonics — NSF now has a five-topic portfolio running against a single solicitation that stays open through May 14, 2028. And that is precisely where the strategy gets complicated, because X-Labs carries a personnel restriction that most PIs have not yet mapped across five simultaneous topics.

We covered the mechanism design of X-Labs when the first two topics dropped, in NSF X-Labs: How the $1.5B Quantum & Instrumentation Initiative Rewrites the Federal Funding Playbook. This piece is about what changes now that the portfolio is complete.

The money, restated

The award structure has not changed across topics, which is itself important — NSF is running a uniform instrument across five very different technical domains.

• Phase 0: up to $1.5 million per team, over roughly 9 to 12 months, for organizational formation — governance, milestone definition, technical roadmapping, partnership stand-up.
• Phase 1: up to $50 million per year over 24 to 36 months, available to teams that clear their Phase 0 milestones.
• Mechanism: Other Transactions, not grants and not cooperative agreements — which means the award sits outside the Federal Acquisition Regulation and outside most of 2 CFR 200.
• Written proposal: eight pages.
• Submission limits: a lead organization may submit a maximum of two written proposals per topic, and senior/key personnel may appear on only one proposal for the topic.

For calibration: a standard NSF research grant in engineering or the physical sciences runs $300,000 to $700,000 per year. An Engineering Research Center runs $5 million to $7 million per year. A $50 million annual ceiling is a category NSF has not previously offered to a team that does not exist at the time of proposal.

The per-topic personnel rule, and why five topics changes it

Read the restriction precisely: senior or key personnel may appear on only one proposal for the topic. It is scoped per topic announcement, not across the initiative.

In May, with two topics on the board — one in instrumentation, one in quantum photonics — the rule was mostly academic. The overlap between the natural applicant pools was thin. A quantum interconnects physicist was not also the obvious lead for a chemical-modality imaging platform.

With five topics, the overlaps are real, and they run right through the most competitive people in the country:

• An AI-for-robotics researcher building learned controllers for embodied systems is plausibly key personnel on both Artificial Intelligence for Physical Systems and Scientific Instrumentation for Sensing and Imaging, where AI-driven computational imaging is explicitly in scope.
• A protein design group running large structure-prediction models is squarely in From Sequence to Function — and its compute leads are equally plausible on Computation at the Limit of Physics.
• A neuromorphic or photonic computing group sits in the overlap between Computation at the Limit of Physics and Quantum Systems, since integrated photonics is load-bearing in both.
• A quantum sensing experimentalist is genuinely dual-qualified for Quantum Systems and Scientific Instrumentation.

Because the limit is per topic rather than per initiative, a single researcher can appear on one AI for Physical Systems proposal and one Sequence to Function proposal. What they cannot do is appear on two proposals within either one. So the constraint is not "pick one topic" — it is "pick one team per topic," which is a subtler and harder problem.

Three consequences follow for anyone assembling a team this fall.

First, the recruiting window is now. The named experimentalists and system architects who anchor credibility in these domains are a small population, and each of them can be spent exactly once per topic. Teams that approach a marquee PI in November, after the October 14 webinar has concentrated everyone's attention, will find that person already committed. The first-mover advantage in X-Labs is not in the proposal — it is in the roster.

Second, exclusivity has to be explicit. A conversation in which a senior collaborator says they are "happy to be involved" is worthless under this rule. Key personnel designation is binary and enforced. Teams need written confirmation of which proposal a person appears on, for which topic, before drafting the eight pages that depend on them.

Third, cross-topic positioning is legal and underused. A group that genuinely spans two topic domains can field two distinct teams — different lead organizations, different supporting casts, different technical thrusts — and appear once in each. That is not a loophole; it is how the rule is written. But it requires two credible eight-page plans, not one plan with the nouns swapped, and NSF's reviewers will notice the difference.

What "AI for Physical Systems" is actually asking for

NSF's description of the live topic: teams that "develop early-stage technologies that enable breakthroughs to accelerate entirely new forms of AI integration in physical systems, including advanced sensors, robotics, embodied systems, human-robot interfaces, and cyber-physical systems," producing "innovative platform technologies" for domains "critical to U.S. competitiveness and security" — healthcare, national defense, emergency response, advanced manufacturing, scientific discovery.

Two words in that description do heavy lifting.

"Platform." NSF is not buying an application. A proposal that promises an AI system for one surgical procedure, or one manufacturing line, or one class of drone, is misreading the ask. The unit of value is a technology layer that multiple downstream domains can build on — which is why the topic description lists five unrelated application areas rather than picking one. If your eight pages describe a single vertical, you have written a different program's proposal.

"Physical." The topic is explicitly not about frontier language models or general-purpose AI capability. It is about the gap between models that work on benchmarks and systems that work when they are wired to actuators, sensors, and human operators under real-world latency, safety, and failure conditions. Sim-to-real transfer, embodied perception, verified control, human-robot handoff, sensor fusion at the edge — that is the territory. A proposal that leads with model scale rather than physical grounding is in the wrong topic.

Set that against the two fall topics and the logic of NSF's 2026 portfolio becomes visible. None of the five topics is a basic-science gap. Protein structure prediction largely works; turning it into intentional design of predictable function is an engineering and organizational problem. The physical limits of computation are theoretically characterized; closing the gap between semiconductor practice and those limits is an engineering and organizational problem. NSF is not funding discovery here. It is funding the construction of institutions capable of executing long-horizon engineering — the thing universities are structurally bad at and startups are structurally undercapitalized for.

The October 30 RFI is a real instrument, not a formality

The RFI on future topics, with responses due October 30, 2026, is the most overlooked item in the September announcement — and the cheapest to act on.

X-Labs runs against a solicitation open through May 2028. NSF has now filled calendar year 2026. Everything after that is undetermined, and the agency is explicitly asking the community which domains to fund. A well-argued RFI response costs a few pages and no cost share, and it is the only lever most researchers will ever have on the composition of a $1.5 billion program rather than on their share of it.

What makes an RFI response useful to NSF, based on how the first five topics are framed:

• Name an organizational gap, not a research gap. Every existing topic identifies a domain where the science is roughly known and the execution vehicle does not exist. "We don't understand X yet" is the wrong pitch; "we understand X and no institution in the U.S. is structured to build it" is the right one.
• Show the platform surface. Describe what multiple downstream domains would be able to do if the technology existed. The five current topics all pass this test.
• Be specific about why a $50 million-per-year team is the right unit. If the work could be done by three coordinated $2 million grants, it is not an X-Labs topic, and saying so honestly will make the rest of your case more credible.

Groups that were not going to be competitive for the 2026 topics should treat the RFI as their actual play this fall. It is a far better use of October than a rushed eight-page proposal against an AI for Physical Systems field that will include some of the best-organized consortia in the country.

What to do before October 14

The Q&A webinar on October 14, 1–2 p.m. EDT requires registration, and it is the one place where NSF will answer mechanism questions on the record for this topic. Everything below should be settled before you sit down for it, because the webinar is more useful to people who arrive with specific problems.

Read the topic announcement itself, not the press release. The AI for Physical Systems topic announcement — posted through SAM.gov and linked from the NSF X-Labs page — carries the binding deadline, the submission mechanics, and any deviations from the May topic structure. The news release does not. Do not build a timeline from secondary coverage; NSF has not restated the proposal due date in its news announcement, and the topic announcement is the only authority on it.

Lock key personnel in writing. Per-topic exclusivity, binary enforcement, and a small pool of credible anchors. This is the gating item.

Talk to your contracting office now. An Other Transaction is not a grant. There is no familiar F&A application, no default Uniform Guidance audit posture, no standard reporting template. Terms are negotiated. Institutions whose sponsored-programs offices have never executed an OT will discover this at the worst possible moment unless someone raises it this month. The universities best positioned for X-Labs are not the ones with the strongest publication records — they are the ones with contracting offices that have done this before.

Write eight pages as a constraint, not a summary. The page limit is not a compressed version of a 25-page NSF proposal. It is a test of whether a team can state a milestone-based, multi-year engineering plan with governance and decision rights in eight pages. Teams that cannot articulate it that tightly generally cannot execute it either, and NSF designed the limit to surface exactly that.

Decide whether you are a lead or a member. Lead organizations get two proposals per topic; key personnel get one. Those are different scarcities, and a group that is the second-best lead for a topic is often the best possible partner on someone else's proposal. The team-formation game in X-Labs rewards people who figure out which of those they are early.

NSF closed the 2026 topic portfolio this week. The next lever on the program's direction closes October 30, and the next lever on your own position in it closes whenever the last uncommitted senior collaborator in your domain says yes to somebody else.