Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
NAIRR Should Be a Public Option for AI Research¶
Revised and substantially expanded July 17, 2026.
When the National Science Foundation launched the National Artificial Intelligence Research Resource (NAIRR) pilot in January 2024, the most visible contribution was compute. That was understandable. Frontier AI had made accelerators, cloud capacity, and model access strategic resources, and academic researchers often could not obtain them at the scale available inside the largest technology companies.
But describing NAIRR primarily as a way to distribute GPU time understates both the market failure and the institutional opportunity.
The scarce resource in advanced AI research is not one thing. It is a capability stack: compute, high-quality data, models, secure environments, engineering support, evaluation infrastructure, legal and governance expertise, reproducible workflows, and communities able to learn from one another. A researcher with cloud credits but no data rights, platform engineering, safety-evaluation support, or path to sustain a working prototype has received purchasing power—not research capability.
NAIRR should therefore be understood as a public option for AI research infrastructure: a national mechanism through which questions with high scientific or public value can be investigated even when they do not align with the incentives, risk tolerance, or product road maps of a few dominant firms.
That is a more ambitious mission than access. It is also the mission most likely to produce lasting public value.
The Market Failure Is About Research Direction¶
The original NSF launch announcement described a partnership among federal agencies and nongovernmental contributors providing compute, datasets, models, software, training, and user support. The breadth matters because concentration in AI affects not only who can run an experiment, but which experiments are considered practical.
When the infrastructure is controlled by organizations that monetize models, cloud consumption, advertising, enterprise software, or proprietary data, their incentives naturally shape the ecosystem. This does not require misconduct. Product road maps favor work that can be commercialized. Internal researchers enjoy access to systems and telemetry unavailable to outsiders. External programs tend to offer the resources a provider already knows how to deliver. Questions that challenge a platform's business model, require long time horizons, serve small populations, or demand unusually strong transparency may struggle to attract support.
The result is a form of agenda concentration. Researchers theoretically remain free to ask any question, but the cost and availability of infrastructure make some questions far easier to pursue than others.
A public option changes the feasible set. It can support:
- independent testing of widely deployed models;
- research on low-resource languages and underserved communities;
- noncommercial scientific and public-interest foundation models;
- secure work with health, infrastructure, environmental, or government data;
- reproducibility studies that do not produce a new product;
- interpretability, robustness, privacy, and safety research whose value accrues broadly;
- alternatives to architectures and assumptions favored by incumbent providers.
The strategic benefit is epistemic diversity. A country that depends on a small number of firms not only for AI products but for the evidence used to evaluate those products has created a fragile knowledge system.
Access Must Be Measured as a Completed Research Journey¶
Infrastructure programs often measure inputs because inputs are legible: accelerator hours awarded, accounts created, datasets listed, or institutions represented. Those measures are useful, but they can conceal unequal conversion into outcomes.
A well-resourced laboratory can turn a compute allocation into a result because it already has machine-learning engineers, research software practices, data pipelines, security staff, and grant administrators. A teaching-focused university, small nonprofit, or interdisciplinary public-interest team may spend much of the award period trying to configure the environment, move data, understand costs, or recruit the necessary expertise.
Equal allocations can therefore reproduce unequal capability.
NAIRR should evaluate the whole research journey:
- Discovery. Can an eligible researcher understand what resources exist and whether they fit the question?
- Application. Can a credible interdisciplinary team apply without possessing the grant-writing machinery of a major research institution?
- Onboarding. Can the team reach a working environment quickly, with identities, quotas, data, models, and support configured?
- Execution. Can it monitor spend, reproduce runs, manage artifacts, obtain technical help, and adjust as results emerge?
- Evaluation. Does it have access to representative benchmarks, red-team methods, domain experts, and secure testing environments?
- Dissemination. Can it publish results, code, data, or model artifacts consistent with law, safety, licensing, and research-security obligations?
- Continuation. Can a valuable project move to another funding source or infrastructure provider without being stranded?
The unit of access is not an account. It is a team's ability to complete that journey and produce credible, reusable knowledge.
Four Layers of a Real National Research Resource¶
The NAIRR capability stack can be organized into four layers.
1. Heterogeneous infrastructure¶
Researchers need more than a single commercial cloud pattern. Useful infrastructure includes national-laboratory and university supercomputers, public and private clouds, testbeds, edge systems, secure enclaves, storage, high-speed networking, and emerging accelerators. Diversity permits comparative research and reduces the risk that the national resource simply trains users into one vendor's architecture.
The abstraction layer should make common tasks portable without pretending that all hardware and services are identical. Reproducible environment definitions, standard job interfaces, open telemetry, artifact registries, and exportable data formats are more valuable than a lowest-common-denominator portal.
2. Governed data and models¶
Data access is often a harder constraint than compute. NAIRR needs curated open resources, controlled-access datasets, clear rights and restrictions, lineage records, versioning, and incident procedures. Sensitive research should be enabled through governed environments rather than prohibited by default or made artificially “open.”
The same applies to models. Researchers should know the license, provenance, intended uses, evaluation evidence, interface, version, data-handling behavior, and exit options associated with each service or artifact. Model access that cannot support reproducibility or independent evaluation is a demonstration resource, not necessarily scientific infrastructure.
3. Research operations and assurance¶
Many teams need help converting a research design into a reliable computational workflow. Shared services should cover experiment tracking, cost controls, secure software supply chains, privacy-enhancing technologies, model and dataset documentation, evaluation harnesses, red-team support, and reproducibility review.
This layer is especially important for trustworthy AI. Safety research cannot be relegated to a topical funding category while the rest of the platform lacks lineage, access control, monitoring, and incident response. Trustworthiness must be an operating property of the resource.
4. Human capability and community¶
Training and user support are not auxiliary benefits. They determine whether access broadens participation. Research facilitators, domain specialists, data stewards, security engineers, and responsible-AI practitioners can help teams choose resources, design feasible studies, and interpret results.
The strongest community model would make successful teams contributors: reusable environments, evaluation suites, datasets, curricula, and lessons learned should flow back into the shared resource where appropriate. NAIRR then becomes a compounding knowledge system rather than a sequence of one-time allocations.
What Two Years of Operation Suggest¶
By March 2026, NSF reported that NAIRR had supported more than 600 research and education projects and 6,000 students across every state, the District of Columbia, and Puerto Rico. The program had also grown to include 13 federal partners and 28 nongovernmental contributors. Those figures indicate meaningful reach and demand.
They should be treated as the beginning of evaluation, not the conclusion.
The next questions are about additionality and conversion:
- Which projects would not have happened without NAIRR?
- Which institutions and investigators became first-time participants in advanced AI research?
- How long did teams wait between award and productive use?
- Which resources were oversubscribed, underused, or abandoned during onboarding?
- What share of projects produced reproducible artifacts, publications, trained people, follow-on funding, or operational public value?
- Did teams remain portable across providers, or become dependent on donated proprietary services?
- Which safety, privacy, and security failures occurred, and how did the infrastructure learn from them?
- Did the portfolio expand the range of research questions, or mainly subsidize work already favored by the market?
NSF's FY 2026 Annual Evaluation Plan appropriately recognized the need to evaluate lessons from the pilot. The most valuable evaluation will combine portfolio-level measures with detailed studies of the research journey, including teams that did not complete it.
A Public Option Needs an Explicit Theory of Additionality¶
NAIRR should not compete with commercial services merely by giving away the same services for free. Its legitimacy rests on additionality: producing capabilities and knowledge that the private market would underprovide.
That suggests a practical allocation hierarchy. Highest priority should go to work that has strong scientific or public value and one or more of the following characteristics:
- lacks an obvious commercial sponsor;
- requires independent access to evaluate powerful systems;
- serves communities or domains poorly represented in existing resources;
- needs controlled data or specialized infrastructure unavailable through ordinary accounts;
- creates open tools, benchmarks, standards, or datasets that improve the broader ecosystem;
- develops researchers and institutions that have been structurally excluded from advanced AI work;
- tests high-risk or high-consequence systems under appropriate governance.
This does not exclude commercially relevant research. It distinguishes a national research resource from a general-purpose subsidy.
The Novel Opportunity: NAIRR as a Learning System¶
The strongest version of NAIRR is not a static catalog. It is a learning system that observes where research repeatedly fails and turns those failures into shared capability.
If many teams struggle to evaluate retrieval-augmented systems, NAIRR can develop a reusable evaluation harness. If privacy review stalls health projects, it can create reference architectures and agreements for controlled analysis. If smaller institutions cannot operate distributed training, it can fund research engineering as a shared service. If contributed datasets lack provenance, it can establish machine-readable lineage requirements. If proprietary model interfaces frustrate reproducibility, it can favor resources with version stability and exportable evidence.
This is how infrastructure compounds. Each project should leave the next project with a better paved road—without constraining legitimate methodological diversity.
The January 2024 launch was therefore important for a reason larger than its impressive coalition. It created a place where the United States can decide that research capacity, independent evaluation, and scientific pluralism are public goods. Whether NAIRR fulfills that promise will depend less on the number of resources in its catalog than on whether a wider range of people can turn important questions into durable knowledge.
I work on this same boundary between infrastructure, knowledge systems, responsible AI, and mission outcomes. More of that work is available through my portfolio and research writing. If you are designing an AI platform or public-sector capability that must broaden participation without sacrificing rigor, you can connect with me on LinkedIn.