United States — Frontier Models and GPAI
Relevant rules
Section titled “Relevant rules”| Rule | Relationship to frontier models |
|---|---|
| EO 14179 (Jan 2025) | Revokes the frontier-model reporting obligation |
| EO 14110 (2023, revoked) | Previously established the 10²⁶ FLOP reporting threshold |
| Trump AI Action Plan (Jul 2025) | A 25-page strategy + three accompanying EOs (data centres, etc.) |
| EO 14365 (Dec 2025) | Preemption of state-level frontier-AI legislation |
| OMB M-25-21/22 (Apr 2025) | Federal AI use and procurement |
| NIST AI RMF + GenAI Profile (2023 / 2024) | Voluntary frontier-model risk framework |
| California SB 53 (in force Jan 2026) | First US state frontier-AI law, 10²⁶ FLOP threshold |
| Texas TRAIGA (in force Jan 2026) | Prohibits harmful uses (not frontier-focused) |
From heavy oversight to voluntary commitments
Section titled “From heavy oversight to voluntary commitments”Biden phase (Oct 2023 to Jan 2025)
Section titled “Biden phase (Oct 2023 to Jan 2025)”EO 14110 introduced the concept of a “frontier dual-use foundation model” into federal US policy for the first time:
- Triggering threshold: ≥ 10²⁶ FLOP cumulative training compute (biological-sequence-specific: 10²³).
- Authority: information-gathering under § 705 of the Defense Production Act.
- Obligations: report to the Department of Commerce on model capabilities, red-teaming results, and safeguard measures.
- Data-centre operators: report on IaaS customers training such models.
The AI Safety Institute (AISI) was established within NIST and signed voluntary “testing and red-teaming” agreements with Anthropic, OpenAI, and others.
Trump phase (Jan 2025 onward)
Section titled “Trump phase (Jan 2025 onward)”EO 14179 revokes EO 14110; the 10²⁶ FLOP reporting obligation lost its legal basis.
Follow-on actions:
- AI Action Plan (Jul 2025): explicit “deregulate and accelerate” direction.
- AISI → CAISI (Center for AI Standards and Innovation): rebranded from “safety” to “standards”.
- Industry voluntary commitments remain (the 2023 White House Commitments did not expire with the change of administration) — but without statutory enforceability.
The actual structure of US frontier-model oversight (Apr 2026)
Section titled “The actual structure of US frontier-model oversight (Apr 2026)”Voluntary layer
Section titled “Voluntary layer”- 2023 White House Voluntary Commitments (signed by Anthropic, Google, OpenAI, Microsoft, Meta and four others).
- Frontier Model Forum (industry association: Anthropic, Google, Microsoft, OpenAI).
- NIST AI RMF + GenAI Profile as a voluntary technical baseline.
- MLCommons AI Luminate / AILuminate benchmarks.
Semi-formal layer
Section titled “Semi-formal layer”- Responsible Scaling Policies (RSPs) (Anthropic pioneered; others followed).
- Preparedness Framework (OpenAI); Frontier Safety Framework (Google DeepMind).
- Company commitments from the Seoul AI Summit (2024) and the Paris AI Summit (2025).
Residual federal layer
Section titled “Residual federal layer”- Export controls: BIS controls on AI chips / model weights under the EAR.
- National-security investment review: CFIUS on China-linked investments in model training or compute.
- Sectoral regulators (FDA, CFPB, etc.) retain their existing authority over AI, unaffected by EO 14179.
State-law layer (key change in Apr 2026)
Section titled “State-law layer (key change in Apr 2026)”Three comprehensive state AI laws take effect in 2025–2026, with SB 53 being the first US law anywhere dedicated to frontier AI:
- California SB 53 (Jan 1, 2026) — Transparency in Frontier AI Act, 10²⁶ FLOP threshold (same threshold as Biden’s EO 14110).
- Large frontier developers (revenue ≥ $500M) must publish an AI framework, issue transparency reports, and report critical safety incidents.
- Penalty: $1M / violation.
- Texas TRAIGA (Jan 1, 2026) — focused on prohibited uses; does not directly regulate frontier models.
- Colorado AI Act (Jun 30, 2026) — high-risk AI anti-discrimination.
The Dec 2025 EO 14365 threat: the Trump administration has explicitly identified state AI laws as targets, with SB 53 high on the list. Federal litigation is expected in Q1–Q2 2026. But an executive order cannot unilaterally preempt state law, and state AGs have signaled they will continue to enforce.
Structural differences vs. China and the EU
Section titled “Structural differences vs. China and the EU”- China: frontier models = large-model services, filed through CAC with TC260-003 as the yardstick — a hard gate.
- EU: dedicated GPAI chapter, 10²⁵ FLOP presumption, compliance checklist + code of conduct (GPAI CoP finalised Jul 2025).
- US: the federal layer is reduced to voluntary commitments plus deregulation; state laws introduce dedicated frontier-AI legislation for the first time (SB 53); federal-vs-state conflict is the main narrative for 2026.
Compute threshold comparison: EU 10²⁵ FLOP < US California 10²⁶ FLOP. California’s threshold is ten times stricter, which makes SB 53’s coverage narrower than the EU GPAI “systemic-risk” provision.