Federal AI Procurement Brief:
May–June 2026

Two months of federal AI procurement activity can be summarized in one sentence: the Defense Department stopped experimenting and started consuming, while the civilian side finally got the infrastructure it has been waiting for. In our inaugural analysis we argued that the federal AI market runs on two clocks — a defense clock measured in months and a civilian clock measured in authorization cycles. May and June 2026 provided the cleanest evidence yet for that frame. On the defense clock: CDAO quintupled Scale AI's enterprise agreement ceiling just eight months after the original award, because demand across DoD components exceeded the contract's scope. On the civilian clock: FedRAMP finalized its Consolidated Rules for 2026, formally moving FedRAMP 20x out of pilot status — the single structural change we identified as the civilian market's inflection point.

Between those two poles sits GSA's OneGov program, which quietly became the default acquisition channel for commercial AI in government. GSA officials disclosed in May that nearly 3.4 million government users across more than 120 agency organizations now have access to AI tools through OneGov agreements. That is not a procurement statistic. That is a distribution statistic — and distribution is what converts pilots into dependency, and dependency into programs of record.

On May 6, CDAO raised the ceiling of Scale AI's enterprise agreement from $100 million to $500 million — eight months after the original award. Regular readers know this publication's discipline: ceiling value is not obligation, and we do not credit vendors for paper. But a ceiling raise is a different instrument than a ceiling award. Nobody amends a contract five-fold for optics. A raise happens when task orders under the existing agreement are consuming the ceiling faster than the period of performance anticipated — which is exactly what DoD acknowledged, stating that demand across components had exceeded the original scope.

Read against our four-layer market map, this is a Layer 3 capability vendor demonstrating the program-of-record conversion motion in real time: land an enterprise agreement, let component-level demand pull consumption, then use consumption data to justify the ceiling amendment. It is the same playbook Palantir ran with Maven — at one-tenth the scale, but on the same rails. The vendors still running generic GenAI demonstrations should study the difference.

GSA's OneGov program — launched in April 2025 to negotiate government-wide discounts directly with technology OEMs — had a defining two months. The June addition of Snowflake brought discounted AI and data-cloud products to all federal agencies, with 20% off compute services and discounts scaling toward 50% as usage increases. It joined a OneGov roster that already includes the major model providers, and it arrived alongside GSA's own disclosure that 3.4 million federal users across 120+ agency organizations now reach AI tooling through these agreements.

GSA is also running the internal proof-of-concept itself: officials said in June that roughly 70% of the agency's workforce now regularly uses AI tools, which the agency equates to about 400,000 hours of automated work. The same month, GSA published its Elimination, Optimization and Automation Handbook — a step-by-step playbook for agencies to cut, streamline, and automate processes. Whatever one thinks of the methodology behind an hours-saved figure, the institutional posture is unambiguous: GSA is positioning itself as both the store and the reference customer.

The caution flag came from the procurement community itself. A widely circulated June commentary by procurement veteran Michael Garland warned of agencies "slouching toward vendor lock" — accumulating what he called emotional and cognitive incumbency in AI tools acquired at promotional OneGov prices. He is right, and the mechanism is worth naming precisely: OneGov's one-dollar-and-discount deals are customer-acquisition pricing, not sustainment pricing. The recompete on the other side of these agreements will occur after workforces have built habits, prompts, and workflows inside a specific vendor's stack. Switching costs are being installed now, at scale, for free. That is not a scandal — it is exactly what a rational vendor buys with a discount — but agencies that fail to preserve data portability and multi-vendor competency in the base period will discover in FY2028 that their leverage is gone.

In our inaugural issue we wrote that the FedRAMP 20x inflection — the moment GSA's reworked authorization process reaches broad availability — was the single most important pending structural change in the civilian AI market, and we put it at Q3–Q4 FY2026. It arrived on schedule. FedRAMP released the public preview of its Consolidated Rules for 2026 on May 4 and finalized the package by the end of June, establishing a single public rulebook, formal certification classes, and transition deadlines running through 2027. FedRAMP 20x is no longer a pilot. It is the program.

The strategic implication is unchanged from our April analysis, but the clock has now started: the window between 20x availability and market saturation is roughly 12 to 18 months. Commercial AI vendors that spent 2024–2025 locked out of the civilian market by the agency-sponsorship bottleneck can now pursue authorization on a continuous, evidence-based model built for products that ship weekly rather than annually. The vendors that prepared for this — government-grade security postures, OSCAL-ready evidence pipelines — will authorize first and set the incumbency terms for the civilian generative AI market. The ones that treated FedRAMP as a distant compliance problem will arrive at a market that has already chosen.

On June 2, the President signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." Procurement-relevant provisions: agencies are directed to strengthen federal cyber defenses on a 30-day clock, an AI cybersecurity clearinghouse is established, and the deployment of AI-enabled cybersecurity tools across federal networks is prioritized. Layered onto the administration's July 2025 AI Action Plan, the order continues the pattern that matters for this publication's readers: every major AI policy document of this administration converts, within one to two quarters, into acquisition demand — usually through existing IDIQs and enterprise agreements rather than new competed vehicles.

The 30-day cyber-defense directive is the tell. Mandates on that clock cannot be met through new procurements; they are met by exercising options and issuing task orders on paper that already exists. If you hold a cyber or AI-adjacent position on an enterprise vehicle at a CFO Act agency, the demand created by this order routes through you. If you do not, this order changed nothing about your pipeline.

In May, Brookings published its third analysis of federal AI contracting, built on Leadership Connect data. The numbers are arresting: funds obligated grew from $675 million in FY2024 to $7.2 billion in FY2026, and the potential value of awards grew from $4.6 billion to $91.8 billion — a 1,912% increase. DoD accounts for 98.9% of that potential value, holding 1,319 AI contracts against runner-up HHS's 134.

Now apply this publication's standing discipline, because the $91.8 billion figure will be quoted for the next year by people who should know better. Potential value of award is ceiling, and ceiling is not obligation. Roughly $10 billion of the DoD figure is a single instrument — Palantir's Army Enterprise Agreement — and the frontier-lab OTAs, the Scale agreement, and every IDIQ in the dataset contribute their full un-obligated ceilings to that total. The honest reading is the obligation line: $7.2 billion in FY2026, up roughly ten-fold from FY2024's $675 million. That is a real, extraordinary acceleration — it needs no ceiling inflation to be the biggest story in federal technology procurement.

Two Brookings findings deserve more attention than the headline number. First, concentration: professional/technical services (NAICS 54) and information (NAICS 51) together represent 98% of federal AI spending — the market is still buying AI as services and software, not as products. Second, breadth: only 28 of 441 federal agencies hold any AI contract at all. Defense aside, the federal AI market has not yet started. That is not a weakness in the data. That is the opportunity map.

SEWP VI goes live at scale. NASA announced roughly 2,100 awards under the sixth generation of SEWP in late June, positioning the vehicle as a one-stop acquisition channel for IT products and solutions. SEWP task orders are where a meaningful share of AI-adjacent infrastructure buying will hide from keyword-based trackers. If your market intelligence relies on FPDS keyword matches, SEWP VI just made your undercount worse.

The state-and-local flank opened. In late June, California signed a first-of-its-kind agreement making Anthropic's Claude available to every state agency and local government at steep discounts. This publication covers federal procurement, but the SLED flank matters to federal watchers for one reason: model providers are replicating the OneGov distribution playbook at the state layer, building the same habit-based incumbency one tier down. The vendors that win both tiers will be very hard to displace anywhere.

Recompete season is here. Our April thesis stands: first-generation GenAI pilots procured in FY2022–2024 under emergency task orders are expiring through FY2026–2027. The Scale ceiling raise shows what conversion looks like when it works. Watch for the pilots that go dark instead — those failures will be quieter than the wins, and they will name the next round of contestable positions.