The Lyceum: AI Intelligence Brief — May 23, 2026

Governance authority is thinning at exactly the wrong moment. DNI Tulsi Gabbard's resignation effective 30 June 2026 installs Principal Deputy Aaron Lukas as acting DNI nine days before the Section 1533 cross-functional team (CFT) statutory standup deadline of 1 June 2026 — assess with high confidence this degrades ODNI's capacity to commit to the IC AI assurance standard the AI Action Plan tasks it to co-author with CDAO and NIST. Simultaneously, GAO has concluded BIS's non-enforcement of the rescinded AI Diffusion Rule is CRA-reviewable, increasing procedural risk around a Day-376 export-control vacuum. A draft White House EO imposing a voluntary 90-day pre-release review window for frontier models was pulled hours before signing amid industry pushback (contested — Politico and Nextgov/FCW report industry lobbying as cause; no White House on-record confirmation), and DoD's FY27 request reportedly includes $29.5B for an "AI Arsenal" supercomputing initiative. Appropriations and acquisition signals are hardening while policy instruments stall. Decision-makers should expect program offices to fill the vacuum with ad hoc criteria.

[COMPUTE][ACQUISITION] DoD FY27 "AI Arsenal" — $29.5B request: DefenseScoop reports a consolidated, government-owned AI supercomputing initiative replacing fragmented GPU buys. Decision implication: shifts the JWCC/NAIRR conversation from rent-vs-buy to a centralized infrastructure program that will compete with service-level compute lines in FY27 markup.

[COMPUTE][DOE_NNSA] DOE one-year nuclear-EO review: DOE confirmed late-2025 Requests for Applications for AI data centers at Idaho National Laboratory, Oak Ridge, Paducah, and Savannah River. Decision implication: federal-site compute siting is now an active procurement lane, not signaling — NNSA/DoE planners need to clarify model-weight and telemetry security boundaries before downselects.

[INTEGRATION][ACQUISITION] AFLCMC "AI copilot" Sources Sought: Air Force market research signals an enterprise "copilots-as-a-service" construct across MAJCOMs. Decision implication: any Section 1533 framework must address persistent embedded assistants — not just discrete model approvals.

[FOUNDATION][ACQUISITION] OpenAI confidential S-1: Draft filing on 22 May positions a possible Q4 2026 listing. Decision implication: public-market disclosure will eventually expose unit economics and governance commitments that USG acquisition risk models currently infer.

[GOVERNANCE] NIST RFI on Trustworthy AI implementation: Federal Register notice moves the AI RMF from principles toward conformity assessments and sector profiles. Decision implication: responses will feed DoD/DOE ATO templates and FedRAMP baselines — primes' submissions are the leading indicator of where acquisition language is headed.

The core question for DoD AI leads: how to sequence Section 1533 CFT structuring when the IC co-anchor will be an acting principal for five to six weeks minimum. Gabbard's resignation is confirmed via Fox News exclusive and corroborated by CNBC; Lukas's acting designation is announced but unconfirmed by Senate process. Acting DNIs historically avoid signing frameworks that bind successors — assess with moderate confidence the IC AI assurance standard slips from June into fall 2026. CDAO should table any ODNI-dependent CFT deliverables before 30 June anchored in existing ODNI workstreams, or proceed unilaterally and document IC concurrence post-hoc. What remains unknown: whether the White House pairs the transition with an IC reorganization directive, and whether any draft IC AI guidance is far enough along that Lukas could issue it as continuity rather than new policy.

The question for acquisition and policy leads is whether to assume any near-term federal pre-release review regime for frontier models. Politico reports the draft EO — a voluntary 90-day review window including NSA participation — was shelved hours before signing after pushback from Zuckerberg, Musk, and other industry principals (contested — TechPolicy.Press frames this as industry capture; no White House on-record statement confirms the lobbying causation; assess with moderate confidence). This is the third governance instrument in 30 days to stall before formalization, alongside the BIS export-control vacuum and GAO's CRA finding. USG acquisition language cannot rely on a federal pre-release review baseline through at least FY26 Q4 — program offices writing AI model contracts should bake vendor-side red-teaming and evaluation commitments directly into clauses rather than wait for an executive framework. Unknown: whether a narrower EO confined to national-security agencies is being redrafted.

Congress is naming "last-mile" cloud service provider onboarding at classified levels as the binding constraint — and it is a more tractable problem than model approval. Per the HASC CITI subcommittee en bloc markup, directive language requires SECDEF/DISA to brief on rapid CSP onboarding by end of FY26 and explicitly ties the bottleneck to AI mission capabilities. Parallel report language demands enterprise-wide data ingestion tooling for LLMs across classification levels — the data architecture beneath the models. Vendors selling "AI capability" without a credible data-tagging, ingestion, and cross-domain solution story will face a steeper hill in FY27. Acquisition leads should expect appropriators to reinforce this with money or deadlines in the FY27 NDAA conference. Unknown: whether the HASC markup language survives Senate conference, and whether the $29.5B AI Arsenal request includes a dedicated onboarding/ingestion line.

Operational commanders are not aligned with the acquisition system's autonomy momentum. Admiral Frank Bradley's public "reality check" on AI in lethal decisions at SOF Week signals that operator-side and legal-review friction will persist even as compute and model dollars rise. Program managers building toward autonomous strike capabilities should plan accordingly. This is a single Tier-2 source (Orbital Today) — assess with moderate confidence on the framing, high confidence that the remarks occurred. Unknown: whether USSOCOM will codify this stance in doctrine or acquisition guidance.

[COLLECTION][GOVERNANCE] ODNI draft AI governance work-in-progress. We do not know whether a near-complete IC AI assurance standard exists that Lukas could issue as continuity. Matters because: it determines whether DoD AI leads should align to a forthcoming IC framework or proceed independently on Section 1533. Watch for: any ODNI public statement before 30 June, or AI governance references in the forthcoming 2026 Annual Threat Assessment.

[FOUNDATION] Section 1533 CFT actual standup status. Public observables (Federal Register, SAM.gov, CDAO releases) show no standup nine days from deadline; classified or quiet internal action cannot be ruled out. Matters because: program offices are currently writing ad hoc AI evaluation criteria that the CFT was supposed to standardize. Watch for: CDAO press release, SAM.gov pre-solicitation for CFT support, or HASC oversight letter.

[EXPORT] BIS response to GAO CRA finding. No public BIS posture on whether it will formalize, rescind, or re-submit the non-enforcement policy. Matters because: pending Gulf-state chip deals and licensing decisions depend on whether the current posture survives CRA review. Watch for: Federal Register ANPRM or proposed rule by 15 June 2026.

[GOVERNANCE] White House EO redraft scope. Whether a narrower national-security-only pre-release review EO is being prepared is unknown. Matters because: acquisition language for frontier models depends on whether vendors will face any federal review requirement in FY27. Watch for: OIRA EO logs or NSC readouts.

• If Lukas, as acting DNI, issues no IC AI guidance by 30 September 2026 → Section 1533 CFT calculus shifts toward DoD-unilateral framework with post-hoc IC concurrence, because acting-principal inaction will signal ODNI has ceded convening authority; watch CDAO public statements and HASC oversight correspondence.
• If BIS publishes an ANPRM or proposed rule on AI export controls addressing model weights by 15 June 2026 → export-control calculus shifts from vacuum to formal framework, because licensing predictability returns; watch Federal Register BIS docket.
• If HASC CSP-onboarding directive language survives FY27 NDAA conference with appropriations backing → AI integration calculus shifts toward onboarding-first acquisition sequencing, because the binding constraint moves from model approval to infrastructure connection; watch conference report by Q4 CY2026.
• If DOE names awardees for AI data centers at INL, Oak Ridge, Paducah, or Savannah River → federal-site compute moves from policy lane to allocable infrastructure, requiring NNSA security-boundary decisions on commercial model colocation; watch DOE award notices and NEPA filings.

Gabbard's resignation and Lukas's acting designation are high-confidence (multiple Tier-2 sources, official social media). Section 1533 statutory requirements are high-confidence via Tier-1 statutory analysis (Akin Gump, WilmerHale); the CFT non-standup is inferred from observable silence — moderate-to-high confidence. The White House EO pull is single-Tier-2-sourced (Politico) with corroboration from TechPolicy.Press and Nextgov/FCW but no on-record White House confirmation — moderate confidence on the event, low confidence on the lobbying causation. The $29.5B AI Arsenal request rests on a single DefenseScoop report and should be treated as a vendor/budget-leak signal until OSD comptroller documents surface. HASC markup language is primary-sourced and high-confidence; downstream appropriations effects are speculative.

~1,380 words