Public Compute Commons: AI Research Infrastructure Without Hyperscaler Dependency

Compute access now shapes who gets to innovate

Frontier AI and data-intensive science are not only about algorithms. They require compute, datasets, storage, networking, technical support, and power. That means access to compute is now a participation question. A project can be scientifically strong and still be impossible without credits, allocations, secure data access, or help using specialized systems.

The private cloud market matters and should remain part of the system. But relying only on private hyperscalers creates three problems. First, public-interest research competes with commercial demand on price. Second, researchers can be locked into one provider's tools and data gravity. Third, safety, health, education, and local-government uses may lack the revenue case that commercial access rewards.

Build outward from NAIRR

NSF's National Artificial Intelligence Research Resource has moved from pilot toward a sustained national capability through a proposed Operations Center. That is exactly the right direction: coordinated national AI infrastructure that expands access beyond the largest firms and elite institutions. But NAIRR should not be treated as a single portal that solves compute scarcity by itself.

The compute commons should be federated. DOE ASCR facilities already provide access to high performance computing facilities through peer-reviewed mechanisms. National labs offer specialized hardware, scientific datasets, and domain expertise. Universities and regional consortia can support training and smaller-scale experimentation. Cloud credits can provide elastic access when public systems are not the right tool. The public role is to coordinate, audit, and fund access across those resources.

Secure enclaves matter for sensitive work

Public compute should support health, education, climate, public benefits, cybersecurity, and AI safety research. Some of that work requires sensitive data. The answer is not to keep sensitive data out of public-interest research entirely. The answer is secure enclaves: approved projects, tiered access, logging, de-identification where possible, privacy review, cybersecurity controls, and penalties for misuse.

This is where the issue connects to PRIV-01 and AI-08. A compute commons that cannot handle sensitive data will miss important public problems. A compute commons that handles sensitive data casually will lose trust and create harm. The policy has to carry both facts at once.

Allocation rules are the governance layer

Compute access is valuable. Any valuable public resource needs allocation rules. The commons should publish who can apply, what the review criteria are, how conflicts are handled, how much capacity is used, what outcomes result, and why applications are denied. It should also separate access tiers: open education and small experiments, peer-reviewed research, sensitive-data work, safety research, and dual-use projects needing additional review.

Allocation should not become a prestige subsidy. If the same institutions receive the same large allocations every cycle, the commons will fail its access mission. The system should reserve some capacity for new institutions, minority-serving institutions, community colleges, regional public universities, startups, state agencies, and public-interest nonprofits.

Energy and infrastructure costs cannot be hidden

Compute uses power, water, land, chips, and cooling. ENV-04 already says large loads should pay for grid upgrades and clean firm supply. Public compute should obey the same logic. A commons that hides data-center costs inside utility bills or local water systems would contradict the platform's own energy position.

The right approach is transparent accounting. Public compute allocations should report energy and infrastructure assumptions at the facility level where feasible. Large expansions should coordinate with grid planning and clean firm supply. The goal is not to make every researcher calculate grid impact before running a model; it is to make public infrastructure decisions honest at the program level.