The Strategic Compute Reserve is a proposed federal government inventory of AI compute capacity — chips, data centers, energy infrastructure, and the technical operators needed to run them — pre-positioned for national-security use. It was articulated by OpenAI in a May 27, 2026 letter to Congressional leaders and a same-day Substack post by OpenAI Global Affairs, and is framed by analogy to the Strategic Petroleum Reserve (SPR), described as federalized buffer capacity for an input the country cannot afford to be short of during a crisis.
The proposal
The OpenAI letter argues that frontier AI is now a national-security input comparable to oil in the 1970s, and that disruption — export controls in reverse, an adversary's capability surge, or a domestic capex pause — carries national-security cost the private market does not internalize. It holds that existing federal compute is insufficient, citing the Los Alamos National Laboratory Venado supercomputer, which has run OpenAI's o3 reasoning model for classified national-security research, as a pre-positioned, security-cleared compute asset operating at a scale OpenAI describes as orders of magnitude below what an AI-augmented national-security workload demands.
The proposed reserve would be stood up federally and span four layers: physical chip inventory; data-center capacity (federal or federally-leased); energy and cooling infrastructure; and cleared technical operators, the human-capital layer that OpenAI argues is binding and that the SPR analogy does not capture as cleanly. OpenAI proposes that the reserve serve specific national-security workloads — intelligence, defense, and classified research — rather than provide general commercial capacity, and does not propose it as a competing commercial cloud.
Status
As of May 28, 2026, the proposal is a policy ask rather than pending legislation. The letter was sent to House and Senate leadership, and no bill has been introduced incorporating the Strategic Compute Reserve framing. The closest existing legislative vehicle is the FY27 NDAA (FY27 National Defense Authorization Act (Chairman's Mark)), whose June 4, 2026 House Armed Services markup could absorb a Strategic Compute Reserve provision, though as of the May 26 Chairman's mark no such section is in the bill text.
The proposal sits alongside the May 22, 2026 White House approval of a $9 billion classified chip request for the CIA, NSA, and other intelligence agencies, intended to deploy frontier models on classified systems (Source: nytimes.com). OpenAI characterizes the $9 billion request as a single-agency down payment on the kind of capability the Strategic Compute Reserve would aggregate and federalize.
Debate over the SPR analogy
Several structural disanalogies have been raised between the SPR and the proposed compute reserve. The SPR stores a homogeneous input, crude oil, that does not depreciate on storage and can be drawn down rapidly; chips depreciate quickly, so a 2026 Blackwell GPU stockpiled for a 2029 crisis would be meaningfully less useful than a 2029 chip by then, implying the reserve would have to operate as a flow rather than a stock. The SPR is replenished by commercial markets, whereas a compute reserve would depend on whether Nvidia & TSMC — AI Compute Infrastructure / TSMC capacity exists to both serve the commercial market and fill the reserve, the constraint flagged by Epoch AI's May 25, 2026 compute-crunch model. And the SPR is a single federal asset, while a compute reserve would comprise at least four layers (chips, data centers, energy, operators) with different stockpiling logics, of which energy and trained operators in particular are not storable in any straightforward sense.
Positions and political alignment
OpenAI is the proposal's author and its most prominent commercial advocate.
David Sacks, the White House AI advisor, had not publicly responded as of late May 2026. Sacks has been a critic of broader federal AI regulation — he opposed the EO — Promoting Advanced AI Innovation and Security (Trump, signed June 2, 2026) in a May 21 phone call — but has supported national-security AI procurement, a posture the Strategic Compute Reserve framing fits more closely than general regulation.
John Moolenaar and the House Select Committee on China have been among the most consistent congressional voices on compute as a national-security input. The SCALE Act (SCALE Act — Rep. John Moolenaar (April 2026)) and the Strategic Compute Reserve are described as complementary instruments, one constraining adversary access and the other federalizing US-side capacity.
Scott Bessent (Treasury) has been a prominent AI-policy voice since the Mythos launch and is identified as the natural Cabinet-level home for the financing side.
Anthropic had not publicly endorsed the proposal as of 2026-05-28 but is exposed to it: Anthropic's near-final deal to make a version of Mythos available to the NSA and other intelligence agencies (per The Information, May 25) would plausibly run on the kind of capacity the reserve would aggregate.
Relationships
- proposed-by: OpenAI
- depends-on: AI Infrastructure Capex (the underlying compute supply); BIS Framework for AI Diffusion — Interim Final Rule (RESCINDED) (the export-control regime that defines what "domestic capacity" means)
- engaged-with: FY27 National Defense Authorization Act (Chairman's Mark) (the natural legislative vehicle); John Moolenaar; David Sacks; Scott Bessent; Los Alamos National Laboratory (planned)
- related: Agentic Economy; US-China AI Competition: Different Races, Different Metrics; DOD — Department of Defense (AI Deployer); Anthropic (Mythos national-security deployment); Epoch AI (compute-crunch model)
Sources
- OpenAI Global Affairs — "A Call for Strategic Compute Capacity" (Source: openaiglobalaffairs.substack.com)
- White House $9B classified chip request, NYT 2026-05-22 (Source: nytimes.com)