A Bill to License Frontier AI Models Above 10^26 FLOPs
Federal licensure regime for training and deploying frontier AI models above a defined compute threshold.
No person shall train or deploy an AI model exceeding 10^26 training FLOPs without a federal license issued by a new Frontier AI Authority.
Developers must notify the Authority 90 days prior to training a covered model.
$420 million for the Authority over 3 years.
Civil penalties up to $100 million per violation.
Mechanical parts, sourced & timed
Use this as your pre-round checklist. Memorize the source citation. Time yourself to the delivery target.
- Bill / Number
- S. 894 — A Bill to License Frontier AI Models Above 10^26 FLOPs
- Funding source
- $420M over 3 years for the Frontier AI Authority.
- Timeline
- Licensure regime stood up within 12 months.
- Plausible — analogous to FDA Investigational New Drug timeline.
- Enforcing agency
- Frontier AI Authority (new agency).
- New agency; constitutional under Commerce Clause + national security.
- Penalty for non-compliance
- Civil penalties up to $100M per violation.
- Source citation
- Center for AI Safety / Bengio et al. (2024), 'International Scientific Report on AI Safety' — gov.uk + ai.gov.
- Delivery time (read aloud)
- 1:15 (75s)
10^26 FLOPs is an arbitrary threshold that will cement incumbents (OpenAI, Anthropic, Google) and freeze open-source AI.
The threshold is calibrated to the Bengio et al. (2024) Bletchley report's frontier-capability frontier. Sec. 3 explicitly funds streamlined small-developer pathways. And the EU AI Act (2024) uses an even lower threshold (10^25) without measurable incumbent entrenchment after one year of implementation.