The Federal AI Build-Out Has Two Clocks, and Buyers Keep Reading the Wrong One

The policy scaffolding arrived fast. On July 23, 2025 the White House released "Winning the Race: America's AI Action Plan" alongside three executive orders, organized around three pillars: accelerating AI innovation, building American AI infrastructure, and leading in international AI diplomacy and security. The plan enumerates more than 90 federal policy actions. Four months later, on November 24, 2025, the administration issued an executive order launching the DOE-led 'Genesis Mission,' mobilizing all 17 National Laboratories to build an integrated AI, supercomputing, and quantum discovery platform, with milestones to identify federal computing resources by February 22, 2026 and demonstrate initial operating capability by August 21, 2026.

For most buyers, the operative document is neither of those. It is OMB Memorandum M-25-21, issued April 3, 2025, which required CFO Act agencies to designate Chief AI Officers within 60 days and convene Agency AI Governance Boards within 90 days. The harder deadline is now behind many teams: agencies had until April 15, 2026 to bring every 'high-impact' AI system into compliance or shut it down, and must begin publicly reporting annual AI use-case inventories in 2026.

The through-line is that the mandates set direction and deadlines, but they do not resolve the physical and authorization work underneath. A governance board can be stood up in ninety days. A classified enclave cannot.

The fast track is real but narrow. GAO's own audit, GAO-26-107859, published April 13, 2026, reviewed 13 AI acquisitions across 44 contracts at DOD, DHS, GSA, and VA and found that none of the four agencies' policies require systematically collecting and sharing lessons learned through the OMB-directed repository. GAO issued four recommendations; all four agencies concurred. The same report notes federal agencies more than doubled AI use from 2023 to 2024, and that Congress appropriated roughly $1.7 billion for AI efforts government-wide.

Where classification enters, the ground shifts under the vendor list itself. As of May 2026, DoD is expanding classified and frontier-model work to eight companies for IL6 and IL7 environments, explicitly excluding Anthropic after DoD designated the company a supply-chain risk amid a dispute over Claude's use in military operations. Anthropic has sued the Pentagon. The original 'frontier AI' contracts awarded in summer 2025, to OpenAI, Anthropic, Google, and xAI, were each worth up to $200 million. On the unclassified side, DoD's GenAI.mil platform launched December 2025 at Impact Level 5 for CUI and had roughly 1.3 million users as of early May 2026.

The lesson for buyers is that a model on today's approved list is not necessarily on tomorrow's once the environment classification rises. The software-layer speed does not transfer.

For agencies, the longer pole is no longer GPU allocation. It is electricity. DOE's push to open federal land for gigawatt-scale data centers made this explicit: after an April 2025 RFI identified 16 candidate sites, DOE selected four for developer partnerships, Idaho National Laboratory, Oak Ridge Reservation, Paducah Gaseous Diffusion Plant, and Savannah River Site, with Argonne cited as able to host a future 1,000 MW 'data park' targeting operations by 2028. DOE wants construction starting by the end of 2026 and operations by the end of 2027.

The grid itself is now the object of federal intervention. On June 18, 2026 FERC issued 'show cause' orders under Federal Power Act Section 206 to all six regional grid operators, directing them to justify or reform interconnection rules for 'large loads,' single-site customers with peak load of 50 MW or more, a definition that squarely covers AI data centers. Operators have 60 days to justify current tariffs or file reforms. A DOE directive from October 2025 separately proposed asserting federal jurisdiction over new large electric loads above 20 megawatts, in exchange for operational flexibility such as occasional curtailment during grid stress.

Hardware still matters, but it has slipped down the critical path. Blackwell-generation GPU lead times ran roughly three to seven months in early-to-mid 2026, with the binding constraint shifting from chip fabrication (TSMC CoWoS packaging allocated through mid-2027) to High Bandwidth Memory supply. For most agencies, substation capacity and the power-usage effectiveness of 1990s-era buildings will govern the schedule long before a memory order does.