Scrutica
Calculates training FLOP estimates for publicly known models against the EU AI Act's 10^25 FLOP threshold for systemic risk models, adjusting for hardware efficiency and utilization rates.
Tracks advanced AI models against the EU's systemic risk regulatory threshold (10^25 FLOPs). Models exceeding this limit face stringent reporting, safety, and cybersecurity requirements under the AI Act.
14 models exceed EU threshold (10²⁵ FLOP); 3 exceed US threshold (10²⁶ FLOP). Multi-regime analysis with facility-level FLOP estimation (hardware path and power path).The EU AI Act triggers mandatory obligations for models classified as GPAI (general-purpose AI) that are trained above 10²⁵ FLOP. Which facilities have enough compute to cross that line, and do they fall under a jurisdiction where the obligations apply?
FACILITIES ≤ 30 DAYS (10²⁵ FLOP)
~53
FACILITIES ≤ 90 DAYS
~62
FACILITIES ≤ 1 YEAR
~65
MODELS ABOVE EU
~14
MODELS ABOVE US
~3
EU AI Act — Systemic Risk GPAI
The threshold is cumulative training compute (total FLOP), not inference throughput. Article 51(2): "A general-purpose AI model shall be presumed to have high impact capabilities ... when the cumulative amount of compute used for its training measured in floating point operations is greater than 10^...
Regulation (EU) 2024/1689, Article 51(2) + Annex XIII · Official Journal of the European Union · Tier 1
Threshold-Capable Facilities by Country (within 1 year, 10²⁵ FLOP)
⚠ = facility is in a jurisdiction where EU AI Act — Systemic Risk GPAI applies
| Facility | Country | Daily FLOP | Days to 10²⁵ FLOP | Estimation Path | Regime Exposure |
|---|---|---|---|---|---|
| Stargate UAE (OpenAI/G42/Oracle) | AE | ~1.21 × 10²⁶ | <1 | Power: 5000 MW | Not subject |
| AWS Project Rainier (New Carlisle, IN) | US | ~5.33 × 10²⁵ | <1 | Power: 2200 MW | Not subject |
| Meta Richland Parish (Hyperion) | US | ~4.84 × 10²⁵ | <1 | Power: 2000 MW | Not subject |
| Stargate Abilene (OpenAI/Oracle/SoftBank) | US | ~2.91 × 10²⁵ | <1 | Power: 1200 MW | Not subject |
| Anthropic-Amazon New Carlisle | United States | ~2.65 × 10²⁵ | <1 | Power: 1092 MW | Not subject |
| Core Scientific, Inc. facility | United States | ~2.18 × 10²⁵ | <1 | Power: 900 MW | Not subject |
| QTS Richmond | United States | ~2.07 × 10²⁵ | <1 | Power: 854 MW | Not subject |
| Applied Digital facility | United States | ~1.94 × 10²⁵ | <1 | Power: 800 MW | Not subject |
| xAI Colossus 2 | United States | ~1.54 × 10²⁵ | <1 | Hardware: 530,000 GPUs (H100 assumed) | Not subject |
| Scala Data Centers S.A. facility | Brazil | ~1.45 × 10²⁵ | <1 | Power: 600 MW | Not subject |
| OpenAI Stargate Abilene | United States | ~1.43 × 10²⁵ | <1 | Power: 590 MW | Not subject |
| Microsoft Fairwater Wisconsin | United States | ~1.34 × 10²⁵ | <1 | Power: 555 MW | Not subject |
| Meta Prometheus | United States | ~1.29 × 10²⁵ | <1 | Power: 531 MW | Not subject |
| Princeton Digital Group (Singapore) Pte. Limited facility | South Korea | ~1.21 × 10²⁵ | <1 | Power: 500 MW | Not subject |
| NVIDIA facility | Saudi Arabia | ~1.21 × 10²⁵ | <1 | Power: 500 MW | Not subject |
| Microsoft Fairwater Atlanta | United States | ~1.05 × 10²⁵ | <1 | Power: 433 MW | Not subject |
| xAI Colossus 1 | United States | ~1.03 × 10²⁵ | <1 | Power: 425 MW | Not subject |
| Alphabet facility | United States | ~1.00 × 10²⁵ | <1 | Power: 413 MW | Not subject |
| Google New Albany | United States | ~9.86 × 10²⁴ | ~1 | Power: 407 MW | Not subject |
| Vantage Data Centers facility | France | ~9.69 × 10²⁴ | ~1 | Power: 400 MW | Not subject |
| EdgeConneX facility | Japan | ~8.48 × 10²⁴ | ~1 | Power: 350 MW | Not subject |
| Compass Datacenters facility | United States | ~8.48 × 10²⁴ | ~1 | Power: 350 MW | Not subject |
| Amazon Madison Mega Site | United States | ~8.26 × 10²⁴ | ~1 | Power: 341 MW | Not subject |
| CoreWeave Lancaster PA | US | ~7.27 × 10²⁴ | ~1 | Power: 300 MW | Not subject |
| CoreWeave Kenilworth NJ (NEST Campus) | US | ~6.06 × 10²⁴ | ~2 | Power: 250 MW | Not subject |
| Meta | United States | ~5.88 × 10²⁴ | ~2 | Hardware: 202,224 GPUs (H100 assumed) | Not subject |
| Alibaba Zhangbei | China | ~4.92 × 10²⁴ | ~2 | Power: 203 MW | Not subject |
| Meta Temple | United States | ~4.80 × 10²⁴ | ~2 | Power: 198 MW | Not subject |
| Meta Kuna | United States | ~4.80 × 10²⁴ | ~2 | Power: 198 MW | Not subject |
| Google Council Bluffs (East) | United States | ~4.60 × 10²⁴ | ~2 | Power: 190 MW | Not subject |
| Google Omaha | United States | ~4.58 × 10²⁴ | ~2 | Power: 189 MW | Not subject |
| Elea Digital Infraestrutura e Redes de Telecomunicações S.A. facility | Brazil | ~2.91 × 10²⁴ | ~3 | Power: 120 MW | Not subject |
| Microsoft Goodyear | United States | ~2.91 × 10²⁴ | ~3 | Hardware: 100,000 GPUs (H100 assumed) | Not subject |
| STACK Infrastructure NVA02 | United States | ~2.86 × 10²⁴ | ~3 | Power: 118 MW | Not subject |
| Google Cedar Rapids | United States | ~2.54 × 10²⁴ | ~4 | Power: 105 MW | Not subject |
| Yondr Group Holdings B.V. facility | United Kingdom | ~2.42 × 10²⁴ | ~4 | Power: 100 MW | Not subject |
| CyrusOne Allen | US | ~2.42 × 10²⁴ | ~4 | Power: 100 MW | Not subject |
| CyrusOne facility | United States | ~2.42 × 10²⁴ | ~4 | Power: 100 MW | Not subject |
| AVAIO Digital Management, LLC facility | United States | ~2.40 × 10²⁴ | ~4 | Power: 99 MW | Not subject |
| xAI Colossus Memphis | US | ~2.33 × 10²⁴ | ~4 | Hardware: 80,000 GPUs (H100 assumed) | Not subject |
| QTS Realty facility | United States | ~1.74 × 10²⁴ | ~6 | Power: 72 MW | Not subject |
| Fluidstack Lake Mariner | United States | ~1.65 × 10²⁴ | ~6 | Power: 68 MW | Not subject |
| Google Pryor (North) | United States | ~1.57 × 10²⁴ | ~6 | Power: 65 MW | Not subject |
| Data Foundry, Inc. facility | United States | ~1.45 × 10²⁴ | ~7 | Power: 60 MW | Not subject |
| ORNL Frontier (Exascale HPC) | US | ~1.10 × 10²⁴ | ~9 | Hardware: 37,888 GPUs (H100 assumed) | Not subject |
| Vantage TX1 | United States | ~1.09 × 10²⁴ | ~9 | Power: 44.8 MW | Not subject |
| Stack Infrastructure facility | United States | ~7.75 × 10²³ | ~13 | Power: 32 MW | Not subject |
| Interxion B.V. facility | Switzerland | ~5.81 × 10²³ | ~17 | Power: 24 MW | Not subject |
| Stream Data Centers, L.P. facility | United States | ~5.33 × 10²³ | ~19 | Power: 22 MW | Not subject |
| Equinix SG3 | SG | ~4.84 × 10²³ | ~21 | Power: 20 MW | Not subject |
Estimates assume continuous operation at full training capacity using dense FP16 TFLOP/s (no sparsity). Default MFU: 40%. Default interconnect efficiency: 85%. Hardware-based estimates assume H100 SXM5 where GPU model is unspecified. Power-based estimates assume PUE 1.20, 70% GPU fraction of IT load. All values are estimated. Methodology version: 1.2.0.