Capabilities & Safety

FLOP Compliance Calculator

Forward FLOP calculator: facility hardware configuration to cumulative training FLOP estimate, checked against regulatory thresholds (EU AI Act 10²⁵, US EO 14110 10²⁶, China GPAI Rules 10²⁴). MFU and interconnect efficiency adjustable. Multi-jurisdiction exposure analysis based on facility location and ownership.

Could a specific AI training setup produce a model that triggers regulatory requirements? Enter a facility's hardware configuration and training duration to estimate total compute, then see which jurisdictions' thresholds would be exceeded. Checks EU AI Act, US EO 14110, and China GPAI Rules simultaneously.

⚠

Pre-training FLOP only. All current regulatory thresholds count pre-training compute. Post-training compute (RLHF, distillation, inference-time reasoning) can amplify effective capabilities beyond what pre-training FLOP alone predicts. See Ord (2025) arXiv:2503.05705; Epoch AI (2025).

2 of 2 active thresholds exceeded

EU Systemic ✗China CAC ✗

Hardware Configuration

GPU Model

GPU Count

Training Duration

days

MFU

0.40

Interconnect Eff.

0.85

Estimation Result

Daily FLOP Budget

9.52 × 10²³FLOP

Cumulative FLOP (90 days)

8.57 × 10²⁵FLOP

Estimation Bounds (cumulative)

~3.03 × 10²⁵ – 1.56 × 10²⁶FLOP

Peak Throughput32,423.9 PFLOP/s

Estimated GPU Count~32,768

Estimation Pathhardware

ASSUMPTIONS

gpu count32,768

gpu model989.5

interconnect efficiency0.85 [0.6–0.95]

mfu assumption0.4 [0.2–0.65]

Threshold Comparison (Log Scale)

8.57 × 10²⁵

EU Systemic

US EO 14110

China CAC

CA SB 1047

MIRI Mon.

MIRI Strict

10²⁰10²¹10²²10²³10²⁴10²⁵10²⁶10²⁷10²⁸

Regime	Threshold	Status	Margin	Days to Threshold
EU AI Act — Systemic Risk GPAI	1.00 × 10²⁵	IN FORCE	8.6x	~10
US Executive Order 14110 — Reporting Requirement	1.00 × 10²⁶	RESCINDED	86%	~105
China — Algorithm Registration (CAC)	1.00 × 10²⁴	IN FORCE	85.7x	~1
California SB 1047 — Safe and Secure Innovation	1.00 × 10²⁶	VETOED	86%	~105
MIRI TGT Proposal — Monitored Threshold	1.00 × 10²²	PROPOSED	8572x	<1
MIRI TGT Proposal — Strict Threshold (Prohibited)	1.00 × 10²⁴	PROPOSED	85.7x	~1

Regulatory Obligations Triggered (2)

IN FORCEEU AI Act — Systemic Risk GPAI

Regulation (EU) 2024/1689, Article 51(2) + Annex XIII · 8.6x above threshold

Model evaluation including adversarial testing (Art. 55(1)(a))
Systemic risk assessment and mitigation (Art. 55(1)(b))
Serious incident tracking and reporting to AI Office (Art. 55(1)(c))
Cybersecurity protections (Art. 55(1)(d))
Energy consumption documentation (Art. 55(1)(a), Annex XI)

Effective: 2025-08-02 · Full regulation text

IN FORCEChina — Algorithm Registration (CAC)

Interim Measures for the Management of Generative AI Services, Article 17 · 85.7x above threshold

Register algorithm with Cyberspace Administration of China (CAC)
Pre-deployment security assessment
Content labeling requirements
User data protection compliance

Effective: 2023-08-15

Methodology

Methodology v1.2.0. Estimation bounds are parameter sensitivity ranges, not confidence intervals. Dense FP16 TFLOP/s (no structured sparsity). MFU default: 0.4. Interconnect default: 0.85. Cumulative FLOP = daily rate × training days.

Methodology EU Threshold Monitor

Related analysis

Export ControlsGovernance

The regulatory regimes defining these thresholds

FLOP EngineInfrastructure

Estimate capacity to check against thresholds

Capability ScalingCapabilities

What capabilities these compute levels enable