Architecture should support collection of key performance indicators for evaluation of NAIRR

National AI Research Resource (NAIRR)—a shared research infrastructure that would provide AI researchers and students with significantly expanded access to computational resources, high-quality data, educational tools, and user support. This final report of the NAIRR Task Force presents a roadmap and implementation plan for a national cyberinfrastructure aimed at overcoming the access divide, reaping the benefits of greater brainpower and more diverse perspectives and experiences applied to developing the future of AI technology and its role in our society.

Such a national cyberinfrastructure also presents a unique and critical opportunity to "design in" the standards for responsible AI research practices and governance processes that uphold our priority to develop and harness these groundbreaking technologies in a manner that reinforces our Nation's democratic values and Americans' personal freedoms.

The NAIRR should be established with four measurable goals in mind, namely to (1) spur innovation, (2) increase diversity of talent, (3) improve capacity, and (4) advance trustworthy AI.

The NAIRR should meet these goals by supporting the needs of researchers and students from diverse backgrounds who are pursuing foundational, use-inspired, and translational AI research. These users should be U.S.-based or affiliated with U.S. organizations, to include academic institutions, non-profit organizations, and startups or small businesses. The NAIRR should comprise a federated set of computational, data, testbed, and software resources from a variety of providers, along with technical support and training, to meet the needs of this target user base. The specific design, implementation, and evaluation of the NAIRR should be centered around the four key goals and should support the collection of data for assessment of key indicators of system performance and success in progress toward these goals. The NAIRR administration and governance should follow a cooperative stewardship model, whereby a single Federal agency serves as the administrative home for NAIRR operations and a Steering Committee comprising principals from Federal agencies with equities in AI research drives the strategic direction of the NAIRR. A Program Management Office within the administrative home agency should provide funding and oversight for an independent Operating Entity that manages the day-to-day operations of the NAIRR. The Steering Committee, co-chaired by the National AI Initiative Office (NAIIO), would incorporate interests and perspectives from across Federal agencies in the governance of the NAIRR. These agencies should also directly support resource providers whose resources, in federation, would constitute the NAIRR. Diverse perspectives and expertise should be tapped to inform the NAIRR's operations through a User Committee, a Science Advisory Board, a Technology Advisory Board, and an Ethics Advisory Board that provide advice to the Operating Entity. The NAIRR should provide access to a federated mix of computational and data resources, testbeds, software and testing tools, and user support services via an integrated portal. Computational resources should include conventional servers, computing clusters, high-performance computing, and cloud computing, and should support access to edge computing resources and testbeds for AI R&D. Open and protected data should be made available under tiered-access protocols and co-located with computational resources. The Operating Entity should not itself operate the totality of the computer hardware that composes the NAIRR; instead, computing, along with data, testing, and training resources, should be delivered as services by partner resource providers selected through Federal agency or multi-agency funding opportunities. When fully implemented, the NAIRR should address both the capacity (ability to support a large number of users) and capability (ability to train resource-intensive AI models) needs of the AI research community. The NAIRR must be broadly accessible to a range of users and provide a platform that can be used for educational and community-building activities in order to lower the barriers to participation in the AI research ecosystem and increase the diversity of AI researchers. The NAIRR access portal and public website should provide catalogs and search and discovery tools to facilitate access to data, testbeds, and educational and training resources serving a range of experience levels. vi The NAIRR should set the standard for responsible AI research through the design and implementation of its governance processes. The NAIRR must be proactive in addressing privacy, civil rights, and civil liberties issues by integrating appropriate technical controls, policies, and governance mechanisms from its outset. The Operating Entity should work with its Ethics Advisory Board to develop criteria and mechanisms for evaluating proposed research and resources for inclusion in the NAIRR from a privacy, civil rights, and civil liberties perspective. Regular training should be required to build NAIRR users' awareness about rights, responsibilities, and best practices related to privacy, civil rights, and civil liberties in AI research, in accordance with the Blueprint for an AI Bill of Rights published by the White House Office of Science and Technology Policy in October 2022. The NAIRR should implement system safeguards in accordance with established guidelines. These guidelines include those developed by the National Institute of Standards and Technology (NIST) and the Five Safes framework: safe projects, safe people, safe settings, safe data, and safe outputs. The Operating Entity should design the NAIRR cyberinfrastructure to consist of multiple tiers, starting with two primary zones: an open science zone "NAIRR-Open" and a secure zone "NAIRR-Secure." Each zone should federate computational, network, and data resources operating in accordance with security and access-control policies that are uniform within the zone, but different between zones, reflecting the different priorities and needs of the users and resource operators. NAIRR-Open should adopt the best practices developed over two decades in the open science community; be consistent with Federal open data, open government, and research security policies; and manage access using single sign-on authentication and a resource allocation mechanism managed by the Operating Entity. NAIRR-Secure should consist of one or more secure enclaves adhering to a common set of security controls, and have the ability to support security requirements arising from legally protected data. NAIRR implementation should occur over four phases, beginning immediately after the publication of this report. In phase one, Congress should authorize and appropriate funds to establish the NAIRR. The administrative home agency and the NAIIO should coordinate the formation of the Steering Committee and stand up a Program Management Office, which will then prepare the solicitation for the Operating Entity and manage the selection process. Phased NAIRR Implementation Timeline In phase two, the Operating Entity should establish its activities and oversee creation of the NAIRR portal and user interface, building in appropriate technical and policy controls. The architecture should support collection of key performance indicators for evaluation of NAIRR progress. Resource providers should be selected via coordinated, multi-agency funding opportunities ideally released within six months of the initial Operating Entity award. In phase three, the NAIRR should achieve initial operational capability and the Operating Entity should also formalize the policies, processes, and initial technical resources to be made available to AI researchers.

Initial capabilities include (1) a portal and user support resources, (2) a mix of computational resource providers, (3) an allocation and identity system, and (4) a data publication system. In phase four, activities should transition from building out the NAIRR to establishing steady-state operations, as well as the planned evolution of NAIRR resources in response to user uptake and demand. Finally, the Task Force also presents a pilot option for the implementation of the NAIRR that would be initiated in parallel with the above phases to expedite the availability of NAIRR resources to the AI R&D community.

As envisioned, the impact of the NAIRR will be significant and far-reaching, enabling researchers to tackle problems that range from routine tasks to global challenges.

In order to achieve its vision and goals, the Task Force estimates the budget for the NAIRR as $2.6 billion over an initial six-year period. The bulk of this investment ($2.25 billion) is to fund the resources to be made accessible via the NAIRR, through appropriations to multiple Federal agencies. The Task Force estimated this budget based on recent costs of advanced computing resources as well as data, training, and software resources; estimates of usage levels to meet the current needs of the AI R&D community; and expected growth of the AI R&D community. Resource providers should be brought online every two years with a six-year lifetime, so that a new $750 million investment is made every two years to ensure that the NAIRR resources remain state-of-the-art. The Operating Entity will require between $55 million and $65 million per year to support the coordination and management of NAIRR activities. An additional $5 million per year is budgeted for external evaluation of the Operating Entity and NAIRR performance. The vision for the NAIRR laid out in this report is designed to meet the national need for increased access to the state-of-the-art resources that fuel AI innovation. The roadmap for achieving this vision builds on existing Federal investments; designs in protections for privacy, civil rights, and civil liberties; and promotes diversity and equitable access. If successful, the National AI Research Resource would transform the U.S. national AI research ecosystem and facilitate the ability to address societal-level problems by strengthening and democratizing participation in foundational, use-inspired, and translational AI R&D in the United States.