Trump's AI Action Plan: Compute Sovereignty, Export Control Reset, and the Frontier Regulatory Vacuum

Executive Order 14179 and the Action Plan launch #

On January 23, 2025, three days into the second Trump term, the White House issued Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence. The order rescinded Biden's October 30, 2023 EO 14110, which had directed roughly 50 agency-level AI actions, mandated dual-use foundation model reporting under the Defense Production Act, and required compute and red-team disclosures. EO 14179 instructs the OSTP Director, AI and Crypto Czar David Sacks, and the National Security Advisor to deliver an AI Action Plan within 180 days, and orders agencies to revise policies inconsistent with sustaining American AI dominance.

OMB Memorandum M-25-03 (March 24, 2025) replaced Biden-era M-24-10 and M-24-18 with a single federal AI use and acquisition framework. It retains the high-impact AI category, loosens minimum risk practices, removes the chief AI officer reporting line into OMB's quarterly review, and directs agencies to prefer commercial American AI products. By June 2025, federal AI use case inventories had grown from 1,757 in 2024 to over 2,100 across 41 agencies, with the VA, HHS, and DHS holding the majority.

The AI Action Plan launched in late July 2025 alongside three executive orders covering federal LLM procurement, accelerated permitting for data center infrastructure, and an American AI export package. The 60 day RFI drew approximately 10,068 submissions, including formal filings from OpenAI, Anthropic, Google, Meta, Microsoft, Amazon, NVIDIA, and Apple. Anthropic tied federal procurement preferences to its Responsible Scaling Policy ASL tiers; OpenAI urged federal preemption of state AI laws and export carve-outs for cloud-delivered frontier models.

AI Diffusion Framework rescission and the chip export rebuild #

On January 13, 2025, outgoing Biden BIS issued the Framework for Artificial Intelligence Diffusion (90 FR 4544), an Interim Final Rule placing every country in one of three tiers. Tier 1 (18 close allies including the UK, Japan, South Korea, Germany, France, Australia, Canada, the Netherlands, Taiwan) faced no new restrictions. Tier 2 (roughly 120 countries) received per-country compute caps measured in total processing performance, stepping down through 2027. Tier 3 (China, Russia, Iran, North Korea, Belarus, Venezuela, Cuba, Syria, plus arms-embargo destinations) faced near-total presumption of denial. The rule covered NVIDIA H100, H200, and Blackwell B100, B200, GB200 accelerators, and added a Validated End User regime for hyperscaler deployments abroad.

On May 13, 2025, Commerce Secretary Howard Lutnick formally rescinded the Diffusion Framework before it took binding effect, calling it overly complex and counterproductive. BIS replaced the tiered structure with three targeted measures: guidance reaffirming that Huawei Ascend chip use anywhere risks a Foreign Direct Product Rule violation, an advisory on protecting American AI chips from diversion, and warnings against hosting Chinese training runs on US-origin GPUs. NVIDIA H20 chips, the China-compliant variant, were licensed in mid-2025, generating roughly USD 4.5 billion in revenue before a partial late-2025 revocation reimposed case-by-case review. NVIDIA recorded an inventory and purchase obligation charge of USD 4.5 billion in fiscal Q1 2026 tied to H20 license uncertainty.

The net result is a more discretionary regime that trades country-tier predictability for case-by-case Commerce, Defense, and State Department review. Gulf allies, including the UAE under the G42 Microsoft partnership and Saudi Arabia under HUMAIN, gained negotiated chip access in exchange for security commitments, ending their previous Tier 2 squeeze.

Stargate USD 500B and the CHIPS AI Centers #

On January 21, 2025, Trump joined Sam Altman (OpenAI), Masayoshi Son (SoftBank), and Larry Ellison (Oracle) at the White House to announce Stargate, a joint venture committing USD 100 billion immediately and up to USD 500 billion over four years to American AI infrastructure. SoftBank holds majority equity, OpenAI holds the operational lead, and MGX (Abu Dhabi) joins as initial equity partner alongside Microsoft, Arm, and NVIDIA as technology partners. The first campus in Abilene, Texas broke ground in early 2025 with Oracle as primary developer, targeting 1.2 GW across ten buildings. By April 2026, Stargate had announced sites in Lordstown Ohio, Milam County Texas, Doña Ana County New Mexico, and Shackelford County Texas, totaling roughly 5 GW.

The CHIPS and Science Act (August 2022) authorized USD 39 billion in semiconductor manufacturing incentives plus USD 13 billion for R&D and workforce, with USD 81 billion for the NSF over five years. Under that NSF authorization, the National AI Research Resource pilot operated through 2025 with 27 federal and private partners, and the NSF AI Research Institutes program funded 27 institutes at a cumulative commitment near USD 500 million. The Trump FY2026 budget preserved AI compute funding while cutting non-defense discretionary research, leaving CHIPS-funded AI Centers structurally intact but politically uncertain.

Strategos compute scenarios show that delivering 5 GW of new IT capacity by 2027 requires roughly USD 175 to 250 billion in capex at prevailing benchmarks of USD 35 to 50 million per megawatt for AI-grade builds. Stargate alone could absorb 15 to 20 percent of NVIDIA's projected Blackwell volume through 2027, concentrating frontier compute in a small set of operator-owned campuses.

NIST AI RMF v2 and the AI Safety Institute renaming #

NIST released AI Risk Management Framework version 2.0 on March 26, 2025, updating the original January 2023 framework. The four core functions (Govern, Map, Measure, Manage) are retained, with new content addressing generative AI deployment risks, agentic AI tool-use, supply chain provenance for training data, and organizational governance for foundation model fine-tuning. The accompanying Generative AI Profile, originally released as NIST AI 600-1 in July 2024, was integrated into the v2 main document. Federal AI Standards Coordination at NIST is led by Lina Cazaubiel, who joined as Senior Advisor for AI Standards in early 2025 and chairs the interagency working group that translates RMF guidance into procurement clauses, NDAA-mandated DoD AI controls, and bilateral standards engagement with allies through the International Network of AI Safety Institutes.

In July 2025, Commerce Secretary Lutnick renamed the US AI Safety Institute to the Center for AI Standards and Innovation (CAISI). The renaming reframed the mandate from safety evaluations and pre-deployment testing toward voluntary standards, measurement science, and competitiveness. Pre-deployment testing agreements signed in 2024 with OpenAI and Anthropic remain in force on a voluntary basis, but the Center's posture toward systemic risk evaluation, including chemical, biological, radiological, nuclear, and cybersecurity uplift testing, is being recast as a Department of Commerce competitiveness function rather than a safety regulator function. Anthropic in its RFI submission urged that pre-deployment testing remain a continuing federal capability and tied its Responsible Scaling Policy AI Safety Levels (ASL-2 through ASL-4) to suggested federal procurement preferences.

The Federal Trade Commission has continued to apply Section 5 of the FTC Act to AI cases. The 2023 Rite Aid order banned the company from using facial recognition in 200 stores for five years following discriminatory false matches. The 2024 Avant settlement penalized dark-pattern AI marketing and required USD 1.6 million in consumer redress. PrivacyCon 2025 added a dedicated AI track on synthetic media, biometric inference, and algorithmic pricing, signaling continued FTC engagement even as the Action Plan emphasizes deregulation.

State AI laws fill the federal vacuum #

With federal preemption proposed but not enacted, state legislatures became the binding floor for US AI regulation. California Senate Bill 53 (Transparency in Frontier Artificial Intelligence Act), signed by Governor Newsom on September 29, 2025 after the prior year's SB 1047 veto, requires frontier developers training models above 10^26 floating point operations to publish a Frontier AI Framework, disclose critical safety incidents to the Office of Emergency Services within 15 days, and provide whistleblower protections. Colorado Senate Bill 24-205, signed May 17, 2024 with effective date February 1, 2026, regulates high-risk AI systems in employment, lending, housing, and education through a duty of reasonable care to prevent algorithmic discrimination. Texas House Bill 149 (Texas Responsible AI Governance Act, TRAIGA), signed June 22, 2025 with effective date January 1, 2026, prohibits AI used to manipulate behavior, create child sexual abuse material, or engage in unlawful discrimination, and creates a regulatory sandbox under the Department of Information Resources.

The patchwork imposes real compliance cost. Frontier developers face California disclosure rules, Colorado pre-deployment impact assessments, Texas behavior-manipulation prohibitions, New York City Local Law 144 bias audits for hiring AI, Illinois biometric and AI hiring video laws, and a Utah AI consumer disclosure regime. Anthropic, OpenAI, Google, and Meta all maintain separate state-by-state legal review pipelines. Industry RFI submissions to the AI Action Plan asked for federal preemption of state AI laws, citing a projected USD 2 to USD 4 billion in annual compliance cost across the top ten developers if the patchwork persists through 2027.

2026 priorities and the frontier regulatory outlook #

Three vectors will define US AI governance through 2026 and into 2027. First, compute and grid. Stargate, Microsoft, Meta, Amazon, and Google together project roughly 50 to 70 GW of new US data center capacity by 2030, against a US generation interconnection queue running 2,600 GW deep with median wait times of 5 years (Lawrence Berkeley National Laboratory, 2024 queue report). The Action Plan's permitting executive order targets NEPA categorical exclusions and FERC fast-track interconnection for AI loads, but state utility commissions hold binding rate-design authority. Expect AI compute pricing to bifurcate by regional grid headroom, with ERCOT, PJM mid-Atlantic, and MISO South emerging as bottlenecks while WECC and SPP attract incremental hyperscale builds.

Second, export control posture. The rescinded Diffusion Framework leaves a more discretionary regime in which BIS, State, and Treasury negotiate bilateral compute agreements with Gulf and Indo-Pacific partners. NVIDIA H20 license oscillation in 2025 (issuance, then partial revocation) cost an estimated USD 4.5 billion in inventory writedowns and signals that case-by-case review will dominate. Frontier developers should plan for binary outcomes on individual destination countries rather than tier-based compliance regimes.

Third, frontier regulation. The Center for AI Standards and Innovation's reframed mandate, combined with the rescission of EO 14110 reporting requirements, leaves a federal void around dangerous-capability evaluation. Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework remain voluntary. California SB 53 imposes the first binding US frontier-model transparency floor. The 2026 Congressional cycle will determine whether a federal preemption bill, modeled on the bipartisan Schatz-Romney AI framework discussions, advances or whether state-led regulation hardens. Strategos clients should sequence federal procurement preparation, state compliance build-out, and export licensing playbooks rather than waiting for a unified federal standard that, on present trajectory, is unlikely to arrive before late 2027.

Sources #

• Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence
• OMB Memorandum M-25-03, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust
• America's AI Action Plan
• Framework for Artificial Intelligence Diffusion, Interim Final Rule, 90 FR 4544
• Commerce rescinds Biden AI Diffusion Rule, strengthens chip-related export controls
• OpenAI, SoftBank, Oracle announce Stargate venture committing up to USD 500B
• Artificial Intelligence Risk Management Framework, Version 2.0
• Commerce establishes Center for AI Standards and Innovation
• California SB 53 Transparency in Frontier Artificial Intelligence Act, signed by Governor Newsom
• Texas House Bill 149, Texas Responsible AI Governance Act (TRAIGA)
• Colorado SB 24-205, Consumer Protections for Artificial Intelligence
• NVIDIA Q1 FY2026 financial results, USD 4.5B H20 charge disclosure