Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Public-Private AI Infrastructure Must Be Able to Outlive Its Donors¶
Revised and substantially expanded July 17, 2026.
The National Artificial Intelligence Research Resource (NAIRR) pilot began in January 2024 with an unusual coalition: the National Science Foundation, other federal agencies, and dozens of private-sector and nonprofit contributors offering compute, datasets, models, platforms, training, and expertise. The coalition was correctly celebrated as evidence that the United States could mobilize resources across institutional boundaries.
Collaboration, however, is not self-governance.
When public research infrastructure depends on contributed commercial resources, the central design question is not whether industry should participate. It should. Industry operates much of the world's advanced AI infrastructure and holds expertise that a public program cannot recreate quickly. The hard question is whether the resulting system advances an independent public mission—or quietly becomes a distribution channel for the technical architectures, pricing models, and research priorities of its largest contributors.
The distinction turns on institutional design. A durable NAIRR must be able to accept private contributions without becoming dependent on any particular donor; offer researchers useful proprietary capabilities without making their work nonportable; and learn from vendors without delegating public allocation and assurance decisions to them.
In short, a national resource should be grateful for donations but architected to survive their withdrawal.
Contributions Are Not Neutral Inputs¶
The original ExecutiveGov coverage emphasized the number and breadth of participating organizations. NSF's launch announcement detailed the different forms of support: cloud and supercomputing allocations, datasets, model access, test infrastructure, privacy technologies, training, and technical assistance.
These are valuable resources. They are also strategic choices.
A cloud credit encourages researchers to learn one provider's identity, storage, orchestration, and model-service patterns. API access may permit experimentation while limiting reproducibility, inspection, or publication. A proprietary dataset can unlock research but impose restrictions on derivatives and sharing. A preconfigured environment can reduce onboarding time while hiding costs or dependencies that become visible only after the contribution ends.
None of this makes the contribution improper. It means that every resource arrives with an architecture, economic model, governance boundary, and theory of use. NAIRR should make those properties explicit so researchers and program leaders can evaluate the complete offer rather than its nominal dollar value.
Four Forms of Dependence to Avoid¶
Public-private infrastructure can become dependent in several ways.
Technical dependence¶
Research code, data pipelines, evaluation systems, and trained artifacts may rely on proprietary interfaces or services that cannot be recreated elsewhere. The immediate project succeeds, but continuation requires the same provider.
Economic dependence¶
The donated period masks the steady-state cost. A team designs around premium services and later discovers that no public funding source can sustain them. This converts a successful pilot into stranded technical debt.
Epistemic dependence¶
Providers supply not only tools but definitions of performance, risk, and best practice. If independent researchers cannot inspect or challenge those definitions, the public program may reproduce vendor claims rather than generate independent knowledge.
Agenda dependence¶
The portfolio gravitates toward questions that can be answered using contributed resources. Problems requiring different architectures, truly independent access, long-term maintenance, or scrutiny of the contributors themselves receive less support.
The answer is not to exclude proprietary resources. It is to govern each dependency consciously and maintain credible alternatives.
A Contribution Compact for NAIRR¶
Every resource provider should enter a transparent contribution compact. The exact terms will vary by resource type, but the public should be able to understand at least the following:
- Availability: amount, duration, regions, service levels, quotas, and conditions under which access may change or end;
- Eligibility and allocation: who decides which researchers receive the resource, according to what criteria, and whether the contributor may reject projects;
- Data handling: retention, logging, secondary use, training use, location, deletion, and government or researcher access to operational records;
- Intellectual property: rights in inputs, prompts, outputs, fine-tunes, code, benchmarks, and inventions;
- Publication freedom: review periods, confidentiality constraints, vulnerability-disclosure processes, and the researcher's ability to publish unfavorable results;
- Reproducibility: model and service versioning, notice of material changes, exportable logs and artifacts, and access needed to reproduce published work;
- Portability: supported data formats, interfaces, environment definitions, model export where applicable, and transition assistance;
- Security and safety: control responsibilities, incident notification, abuse handling, evaluation evidence, and red-team permissions;
- Conflicts and influence: contributor participation in governance or proposal review, recusals, and separation from allocation decisions involving competitors or the contributor's own systems;
- Exit: notice periods, data and artifact return, continuity arrangements, and treatment of ongoing projects.
Making these terms visible serves both sides. Providers can define boundaries rather than negotiate them repeatedly. Researchers can make informed design choices. Program leaders can compare the full public value and risk of contributions that otherwise look equivalent on a spreadsheet.
Portability Should Be a Research Deliverable¶
Portability is often discussed as an infrastructure feature. NAIRR should also treat it as a property of the research result.
A project does not need to run identically on every provider. Hardware, models, and managed services have legitimate differences. But the team should be able to identify which findings are general and which depend on a specific implementation. That requires versioned environments, exportable experiment metadata, documented service dependencies, and—where feasible—validation on a second substrate.
For high-value projects, NAIRR could fund a portability checkpoint near completion:
- export the relevant data and artifacts under applicable controls;
- reconstruct the workflow from its documented environment;
- replace or emulate critical proprietary dependencies;
- rerun a representative subset of experiments;
- explain performance, cost, and behavior differences;
- publish a dependency and continuation plan.
This would strengthen reproducibility and reveal hidden concentration risk. It would also produce practical evidence for public-sector buyers evaluating multi-cloud, open-model, or hybrid strategies.
Independence Requires Governance, Not Merely Diversity¶
A long list of contributors can create the appearance of independence while authority remains concentrated. NAIRR needs clear decision rights across four functions:
- Mission and portfolio: public institutions define the research objectives, access priorities, and measures of additionality.
- Resource assurance: qualified, conflict-aware reviewers assess whether contributed services are suitable for particular uses and whether their claims are adequately supported.
- Allocation: transparent processes match researchers to resources without giving donors improper control over competitors, critics, or research outcomes.
- Operations: technical teams manage identity, support, monitoring, incidents, and transitions across the federation.
Advisory participation by industry can improve each function. Final public accountability cannot be outsourced.
The CREATE AI Act introduced in the 118th Congress contemplated governance structures, an operating entity, dataset criteria, security requirements, and support for testing and evaluation. Its introduced statutory text is useful because it treats NAIRR as an institution with processes and duties rather than a voluntary directory. The reintroduced CREATE AI Act of 2025 continued that effort, including provisions for an operating entity, a free tier, security requirements, and a meaningful allocation to privacy, ethics, safety, security, risk mitigation, and trustworthiness.
Legislation will not resolve every design choice. It can establish the durable authority and accountability within which those choices are made.
Measure Concentration Before It Becomes Lock-In¶
NAIRR should publish portfolio-level measures that show whether the federation is becoming structurally dependent. Useful indicators include:
- share of compute, storage, model access, and active projects by provider;
- percentage of projects using open, exportable, or provider-specific interfaces;
- estimated continuation cost after donated access expires;
- number of projects that successfully migrate or reproduce across environments;
- provider concentration within high-priority research domains;
- frequency and duration of service changes affecting reproducibility;
- share of evaluation work performed on systems owned by the evaluator or funder;
- number of active projects affected by a contributor's exit or material term change.
These measures should not be used to impose artificial equality among providers. Some resources will be better suited or more generously supplied. The purpose is early visibility. Concentration becomes dangerous when it is discovered only after a policy change, price increase, service retirement, acquisition, or geopolitical disruption.
The Government's Unique Contribution Is Continuity¶
Private firms can often move faster than public institutions and may possess superior near-term infrastructure. The government's comparative advantage is different: it can maintain a mission across market cycles.
Research questions may take years. Cohorts of students need stable curricula. Scientific artifacts must remain accessible after a product is retired. Safety investigations may become more important precisely when commercial incentives shift. A national resource must preserve continuity of identity, metadata, provenance, allocation records, published artifacts, and institutional knowledge even as individual contributors enter and leave.
This implies that the federation's control plane should be publicly accountable and minimally dependent on a single commercial substrate. Provider-specific services can sit beneath it, but core resource discovery, project identity, lineage, policy records, portability metadata, and incident communication should remain available across transitions.
By 2026, NSF reported that NAIRR had expanded to 13 federal agencies and 28 nongovernmental partners while supporting hundreds of projects. The two-year progress update demonstrates that the federation can produce real reach. Its next proof point should be resilience: whether the infrastructure can absorb a provider's exit, preserve ongoing work, and keep the public record intact.
Collaboration Is Successful When It Produces Public Capacity¶
The deepest mistake in public-private partnership discourse is to treat collaboration itself as the outcome. Collaboration is a means. The outcome is durable public capacity: more independent knowledge, a wider and more capable research community, reusable infrastructure, and credible alternatives to concentrated private control.
That standard changes how contributions should be valued. The most important donation may not have the largest list price. It may be the one that leaves behind an open evaluation suite, a portable workflow, trained research engineers, a curated dataset with durable stewardship, or a reference architecture that other institutions can operate.
NAIRR can benefit enormously from commercial AI leadership without becoming an annex of the commercial AI market. The design principle is straightforward: accept resources in ways that increase the country's future choices.
This intersection of public mission, platform architecture, knowledge governance, and technical independence is a recurring focus of my AI and modernization work. Organizations wrestling with similar partnership and portability questions can also explore Xendev Labs or send a direct inquiry.