The Lyceum: AI Intelligence Brief — May 24, 2026

The most consequential signal in the last 24 hours is silence. At T-8 days from the Section 1533 Cross-Functional Team (CFT) statutory standup deadline of June 1, 2026, no CDAO announcement, Federal Register notice, or SAM.gov vehicle confirms the team is being constituted. Assess with moderate confidence the deadline will be missed publicly; the Section 1534 sandbox task force April 1 deadline has already passed without observable standup, compounding the governance gap. The BIS replacement AI chip rule remains unpublished 377 days after rescission of the AI Diffusion Rule, leaving GCC and Tier 2 bilateral compute deals operating under catch-all enforcement rather than positive authorization. Against this static regulatory backdrop, the FY27 NDAA subcommittee markup window is open now — meaning the next round of binding AI authorities is being drafted this week, largely out of public view.

[ACQUISITION][FOUNDATION] Grok federal adoption stalled. Reuters-reviewed 2025 federal AI usage records show xAI's Grok appearing in only 3 of 400+ identified use cases, versus 234 for OpenAI-based systems, despite aggressive pricing and a previously reported $200M Pentagon contract (contested — the $200M figure traces to AI Weekly, single-source, and may reflect ceiling rather than obligated value). Decision implication: CDAO and agency acquisition leads should treat reliability and policy-domain accuracy — not unit price or brand momentum — as the binding gate in commercial foundation model selection.

[COMPUTE][DOE_NNSA] Data center siting friction materializing. Vancouver protests on May 23 against proposed AI data centers, paired with Goldman Sachs projections of U.S. data center demand doubling from 31 GW (2025) to 66 GW (2027), confirm that energy and water are now binding constraints, not externalities. Decision implication: DOE and DoD compute siting decisions for NAIRR, the Genesis Mission, and the FY27 AI Arsenal initiative must now price in community-impact and grid-interconnect risk explicitly.

[ACQUISITION][COMPUTE] FY27 "AI Arsenal" budget signal. The Pentagon's FY27 request reportedly includes ~$29.5B for an AI Arsenal initiative to centralize AI supercomputing across the joint force (contested — line-item allocation between hardware, software, and enabling infrastructure has not been published; figure traces to pre-decisional budget summaries). Decision implication: if confirmed in the President's Budget submission, this represents a strategic pivot from distributed GPU procurement toward government-owned foundational compute, reshaping vendor capture patterns.

Section 1533 of the FY2026 NDAA directs DoD to establish a CDAO-led cross-functional team to create the Department-wide framework for assessing, governing, and approving AI model development, testing, and deployment — operational by June 1, 2026 (Akin Gump analysis; CRS IF13197). No CDAO public announcement, SAM.gov notice, or congressional staff confirmation has surfaced. The Section 1534 sandbox task force — co-chaired by CDAO and DoD CIO, due April 1 — appears already missed, and it is the infrastructure layer the 1533 framework runs on. (Contested — DoD may assert existing CDAO governance guidance satisfies the statutory intent without formal CFT standup; no congressional staff have publicly accepted that reading.)

The decision consequence is concrete. The seven Pace-Setting Projects face initial demonstrations in July 2026 under CDAO (Nextgov/FCW, Feb 2026). If the CFT misses June 1 — or stands up on paper without functional leads and an assessment rubric — PSP vendors execute against no standardized evaluation criteria, and every award becomes legally contestable under the NDAA. What remains unknown: whether internal OUSD(R&E) or CAIO directives have quietly constituted the team without public notice.

HASC and SASC subcommittee marks typically circulate by late April/May (CRS IF10515). FY27 AI provisions are being drafted this week. The HASC/SASC overlap from FY26 — AI for missile defense mission planning, AI data center infrastructure on DoD installations, cybersecurity governance, and training/readiness applications (Akin Gump) — suggests FY27 will deepen rather than redirect. The watch items: whether FY27 extends the Section 1533 CFT sunset (currently 2030), whether report language penalizes the missed Section 1534 deadline, and whether the Defense Autonomous Warfare Group's $53B reconciliation authorization generates new autonomy procurement triggers. This is structural inference from the NDAA calendar, not a confirmed leak — but the 10-day window for staff-sourced reporting is where high-value confirmation appears.

When BIS rescinded the AI Diffusion Rule in May 2025, it promised to "issue a replacement rule in the future" (Gibson Dunn; Kirkland & Ellis). Twelve months later: no ANPRM, no Federal Register pre-publication, no OIRA docket entry. The administration has signaled intent to favor GCC, ASEAN, India, and Israel relative to the prior framework (Crowell & Moring), but no instrument exists. Acquisition and policy leads structuring bilateral AI compute deals with Gulf partners have only catch-all enforcement guidance to cite — no positive authorization. Compounding this, the U.S.-China AI safety dialogue confirmed May 14 by Treasury Secretary Bessent creates internal policy friction: former Biden chip-controls official Chris McGuire assesses Beijing may use the dialogue to extract concessions narrowing the U.S. lead, while other voices argue tightened controls would foreclose safety cooperation. The contradiction is unresolved and is itself shaping rule delay.

The Reuters-documented Grok adoption shortfall reframes how acquisition should evaluate "low-cost" commercial offerings. The 3-of-400 ratio against OpenAI's 234 is not a marketing problem — federal users specifically cited reliability failures on policy and regulatory queries and discomfort with marketed "unhinged" modes. General-purpose benchmarks like MMLU and HumanEval do not predict performance on the structured policy tasks that dominate USG workflows. What remains unknown: whether xAI will pursue FedRAMP High and domain-specific fine-tuning to close the gap, and whether the underlying Pentagon contract has produced any deployed capability.

[GOVERNANCE] CFT internal standup status. Whether CDAO or the CAIO has issued any non-public directive constituting the Section 1533 team. Matters because: a paper-only standup meeting the June 1 deadline while remaining operationally inert leaves PSP July demos governed by ad hoc criteria. Watch for: CDAO public release, SAM.gov support-services notice, or HASC/SASC staff inquiry to OUSD(R&E) within 30 days of the deadline.

[EXPORT] Model-weight scope of the replacement rule. Whether BIS will retain ECCN 4E091 controls on closed-weight frontier model exports (velaw.com analysis). Matters because: omission collapses the weight-export architecture and directly affects how DoD structures classified model deployments with allies. Watch for: ANPRM publication or OIRA receipt of a draft rule.

[INTEGRATION] PSP July demo evaluation standard. Whether July 2026 PSP demonstrations will be evaluated as live operational capability or proof-of-concept, and who sets the standard if the CFT is not operational. Matters because: the AI Strategy Memo's 30-day-from-public-release model parity criterion (Nextgov/FCW) has no published evaluation rubric. Watch for: DepSecDef or CDAO framing language in pre-demo communications.

[COMPUTE] FY27 AI Arsenal allocation breakdown. The $29.5B figure has no published split across hardware, software, security, and grid/siting infrastructure. Matters because: allocation determines which industrial segments capture procurement and whether energy constraints are funded as a first-order line. Watch for: FY27 budget justification books and HASC Strategic Forces/Cyber-IT subcommittee marks.

• [GOVERNANCE] If CDAO publishes no Section 1533 CFT standup announcement by June 1, 2026 → PSP vendor evaluations in July become legally contestable under the NDAA; watch for HASC/SASC staff inquiry to OUSD(R&E) or a GAO referral request within 30 days.
• [EXPORT] If BIS files a Federal Register pre-publication notice for the replacement AI chip rule before July 4, 2026 → GCC bilateral compute deal structures change immediately based on inclusion/exclusion of Tier 2 country data centers; watch the Federal Register public inspection desk daily.
• [ACQUISITION] If HASC or SASC subcommittee marks circulate in the next 10 days containing report language on the missed Section 1534 deadline → Congress has formally noticed the governance gap and FY27 enforcement teeth become probable; watch HASC markup calendar and defense trade press sourcing committee aides.
• [INTEGRATION] If DepSecDef or CDAO frames the July PSP demos as "proof-of-concept" rather than "operational" → the 30-day model parity procurement criterion is not enforceable, and CDAO authority over PSP awards is functionally weakened; watch post-demo press language.

The deadline calendar is high-confidence: Section 1533 (June 1), Section 1534 (April 1, missed), and PSP July demos are confirmed by the enacted NDAA text, CRS analyses, and the January 9 DoW AI Strategy Memo. The BIS replacement rule vacuum is verifiable by absence of any Federal Register filing. Moderate-confidence inferences include the CFT missed-standup assessment (based on absence of public signal, not confirmed internal status) and the FY27 markup-window timing (structural inference from historical NDAA cadence). The Grok $200M Pentagon contract and the $29.5B FY27 AI Arsenal figure are single-sourced and flagged contested. No item in this brief depends on Tier 3 or Tier 4 reporting.

[~1,380 words]