Multi-year mandatory funding for a mix of government-owned, contractor-operated, and commercial-surge compute... reversing the July 2025 White House AI Action Plan that told DoD to lean on hyperscalers.
The Department of War's FY2027 budget request asks Congress for $46 billion in multi-year mandatory funding for "Sovereign Artificial Intelligence (AI) Infrastructure" - the largest single line item the budget books carve out for defense AI compute. The figure sits inside a roughly $1.45 trillion national defense topline and a $350 billion mandatory tranche that depends on reconciliation. It is also a notably stronger ownership posture than the White House's own July 2025 AI Action Plan appeared to direct, and a public-sector counterpoint to the private AI supercluster build-out that has been reshaping the TOP500 list.
The DoW Comptroller's language is explicit about the direction of travel: a "deliberate pivot to Joint and Interagency AI investment and foundational, hybrid AI architecture with government-owned AI infrastructure." Read against the Action Plan, that language goes beyond what the executive's own infrastructure plan called for ten months earlier. Commercial cloud isn't foreclosed; the same paragraph names contractor-owned-contractor-operated capabilities and "flexible commercial surge capacity" alongside the owned infrastructure piece.
The primary-source breakdown comes from the DoW FY2027 Mandatory Funding Overview, released in April 2026. Section 7 of that document, "Next-Generation Technology and Autonomy," lays out a $102.5 billion bucket that includes the Sovereign AI line alongside $53.6 billion for "Drone Dominance," $2.4 billion for "AI-Enabled Deterrence," and $500 million for "AI Pace Setting Projects" in cyber operations, AI-enabled ISR, manufacturing, and sustainment.
The verbatim language on the $46 billion ask:
"The $46.0 billion request prioritizes enterprise-scale AI infrastructure investment to build enduring strategic advantage. The current model of fragmented, disconnected Graphics Processing Unit (GPU) clusters procured on an ad hoc basis is operationally inefficient and fiscally unsustainable. A deliberate pivot to Joint and Interagency AI investment and foundational, hybrid AI architecture with government-owned AI infrastructure eliminates redundancy, maximizes federal buying power, and delivers the integrated AI infrastructure for the Department's accelerating mission demands. Funding supports investments in foundational backbone AI infrastructure; provides strategic investment in specific contractor owned contractor operated capabilities; expands inside and outside the Continental U.S. regional compute centers to support a strategic training core and operations collocated with Combatant Commands, and builds out modular, containerized tactical compute capabilities."
Three structural commitments sit inside that paragraph. The Department wants a centrally planned procurement posture instead of program-by-program GPU buys. It will use a mix of government-owned infrastructure and "contractor owned contractor operated" arrangements - COCO, in acquisition shorthand - for regional compute centers, both inside and outside CONUS. And it wants modular, containerized compute at the tactical edge, collocated with Combatant Commands.
Three figures circulating in the public reporting all describe overlapping cuts of the same request:
The other point worth flagging early: this is reconciliation money. The $46 billion sits inside the $350 billion DoD mandatory tranche, separate from the discretionary 050 base. It requires a reconciliation bill to clear Congress. If reconciliation stalls - and reconciliation packages do stall - the Sovereign AI Infrastructure program contracts to whatever the discretionary side can absorb.
The strongest read of the FY27 ask is that DoW is going beyond what the White House's own July 2025 AI Action Plan directed, not reversing it.
Pillar II of the Action Plan, "Build American AI Infrastructure," directs DoD to:
"Prioritize DOD-led agreements with cloud service providers, operators of computing infrastructure, and other relevant private sector entities to ensure access to the advanced computing needed to train, tune, and deploy frontier AI systems in support of defense and intelligence missions."
The operative verbs are "prioritize agreements" and "ensure access" - language that points to procurement of cloud capacity rather than a prohibition on owning compute. The Plan also asks DoD, NIST, and the Center for AI Standards and Innovation (CAISI) to develop "high-security data center standards for military and intelligence usage," language about standards rather than the construction of new federal infrastructure at the scale FY27 contemplates.
The Sovereign AI Infrastructure request reads as a different posture from that emphasis. It funds government-owned racks at enterprise scale, paired with COCO regional centers and a commercial-surge layer, rather than leaning primarily on hyperscaler-provided capacity. DoW officials cite the Action Plan as policy precedent (per DefenseScoop's reporting), but the FY27 ask plainly goes further than the Plan's text on its face.
Best to call the gap tension or divergence. The Plan emphasized cloud-provider agreements; the budget emphasizes a mixed model with a heavy ownership component. Reversal would be the wrong word - the Plan never explicitly ruled out government-owned infrastructure - but the FY27 posture leans much harder on ownership than the Plan's language suggested.
Tucked into the same mandatory book is a second figure generating its own headlines: a $29.95 billion Defense Production Act Purchases mandatory appropriation, Defense-Wide, in FY27, per the FY2027 P-1 Procurement Programs exhibit. Inside Defense framed it as a "DPA surge, transforming niche."
Two clarifications matter. First, the DPA Purchases line is separate from the $46 billion Sovereign AI Infrastructure line; the two appropriations sit in different sections of the budget and should not be conflated. Second, the multiplier is roughly 30x. The FY24 DPA Title III procurement baseline was approximately $968.6 million, per the PROC_DPAP_PB_2024 budget exhibit. The FY27 mandatory $29.95 billion is a ~30-fold surge over that baseline.
The DoW description of where DPA dollars go names "kinetic weapons (including sub-tier supply chain capabilities for missiles and munitions, critical chemicals, and hypersonics), manufacturing (including castings, forgings, and other critical metal working capabilities), microelectronics (including secure advanced packaging, high-power, and radiation hardened devices), battery components." The microelectronics line is where AI-relevant industrial-base capacity sits. A separate $6.4 billion DPA Purchases slice inside the broader $48.7 billion Critical Minerals envelope covers rare earths and metallization... the materials side of the AI compute stack.
Read that against the demand-side magnitude: NVIDIA's Vera Rubin generation is reshaping 2027 data center capex for the hyperscalers themselves, and HBM allocation, not raw HBM supply, is the gating constraint on every 2026 AI infrastructure plan. The DPA surge is, in part, an attempt to insulate the federal supply chain from that pull on advanced packaging, HBM, and rad-hard variants.
While the $46 billion ask is large enough to reshape federal compute, it's modest by hyperscaler-capex standards.
Microsoft's FY2026 capital expenditure guidance, revised upward in its April 29, 2026 third-quarter release, now stands at approximately $190 billion for the fiscal year, driven in part by memory pricing. Against that figure, the entire Sovereign AI Infrastructure request is roughly three months of one hyperscaler's annual capex. Even on a generous read using the $58.5 billion total DoD AI envelope, the federal commitment is comparable to roughly four months of Microsoft's FY26 capex line alone.
That asymmetry sits inside the broader public-vs-private capex gap that has been distorting the AI capex-to-revenue picture. The federal government is now joining the buy side at notable scale, but it is not displacing the hyperscalers as the dominant marginal buyer of AI infrastructure. The reverse framing is the one worth holding alongside: in aggregate defense spend, the FY27 topline of roughly $1.45 trillion remains a multiple of any single hyperscaler's annual capex... the AI compute slice is small inside a much larger defense envelope.
Read against allied postures, the FY27 ask lands as a distinct lane rather than a copy of either the hyperscaler-rented model or the EuroHPC-style open-science model.
EuroHPC's AI Factories program, anchored by the InvestAI envelope and the Mimer services-layer model, is the civilian-adjacent peer: open-science framing on shared regional infrastructure with a services layer on top. AUKUS Pillar II covers AI, autonomy, and quantum, and enables shared U.S. compute access for the partnership. SCN has not identified a publicly disclosed defense-AI compute line item at comparable scale from an allied government, though defense-AI investment by U.S. allies is rising on multiple tracks.
The U.S. posture is now requesting government-owned AI compute alongside its hyperscaler ecosystem and alongside allied open-science compute - a third lane. The architecture in the DoW description (regional compute centers collocated with Combatant Commands, plus modular containerized tactical compute) is recognizably hub-and-spoke, but how closely it resembles a sovereign hyperscaler in practice depends on details Congress has not yet authorized.
The other thing the $46 billion creates, before any racks are racked, is regulatory scope.
The NIST 800-234 federal HPC security overlay went out as an initial public draft on May 1, 2025, with a comment period extended to August 4, 2025 and a final SP 800-234 published on May 4, 2026. It is guidance tailored to HPC environments, not a mandatory regime every DoD AI center must clear. That said, for the kind of regional compute centers the budget describes, it is the most directly relevant federal security framework on the books, and the Action Plan's call for "high-security data center standards" jointly developed by DoD, NIST, and CAISI is its regulatory complement.
How the overlay is applied to Sovereign AI Infrastructure builds matters for throughput. Whatever portion of the $46 billion lands as new federal AI compute will sit under a security framework that practitioners have been working through since 2025; the funding ask and the overlay landed in adjacent fortnights, and procurement teams will read them together.
Vendor attribution in the public FY27 budget books is thin. The $46 billion Sovereign AI Infrastructure description names no companies, only postures: "specific contractor owned contractor operated capabilities," "foundational backbone AI infrastructure," and "regional compute centers." Public FY27 documents do not map line items to vendors at the system level, and no procurement official has publicly tied the May 2026 cohort of AI partnerships to specific Sovereign AI Infrastructure execution.
What is on the record: in May 2026, the Pentagon announced AI partnerships with NVIDIA, Microsoft, and AWS, reported by Yahoo Finance under "The New AI Arsenal" headline. That announcement predates the FY27 budget submission and indicates the cohort of vendors DoD is already working with on AI compute at scale. It is not, on its own, a forecast of who wins what inside the $46 billion line. SAM.gov task orders post-enactment will be the place to look.
The existing AFRL DSRC systems on the public record - Warhawk (HPE Cray EX) and Raider (Penguin Computing TrueHPC), per the AFRL HPC Center user documentation - give a sense of who serves the classical-HPC side of the same procurement community. The Sovereign AI Infrastructure ask reads as parallel infrastructure rather than a top-up of existing High Performance Computing Modernization Program centers; DoW language about "expanding regional compute centers ... collocated with Combatant Commands" points to new builds.
The piece of FY27 framing that turns all of the above into a policy question sits at the bottom of the page.
The $46 billion is mandatory reconciliation funding. It depends on a reconciliation bill clearing both chambers. The $350 billion DoD mandatory tranche overall sits on the same legislative gate. If reconciliation stalls, the Sovereign AI Infrastructure ask reverts to whatever the discretionary 050 base can absorb, which is dramatically less. The same gate applies to the $29.95 billion DPA Purchases line and the $1 billion National Security Investment Fund seed.
Sovereign AI Infrastructure is, in other words, an industrial-policy ambition pinned to a budget instrument with a known failure mode. The Pentagon is asking Congress to fund a mixed federal AI compute posture (owned plus COCO plus surge) at enterprise scale. Congress has not yet voted on whether it wants to. If it funds the request, the U.S. enters the next decade with a distinct lane of AI compute alongside the hyperscaler and allied-science models. If it does not, the Pentagon will have spent six months articulating a posture it cannot pay for. Either outcome is now on the legislative calendar.
