Facilities written 2026-09
Capacity
Compute capacity is estimated site by site, then read against the thresholds in statute.
Compute capacity is estimated site by site, then read against the thresholds in statute.
Estimated at sustained, whole-site operation. Article 51 draws its line at a model’s own training compute; this count measures the capacity available to run one.
Plate I Time to 10²⁵ FLOP 499 sites · none sampled out
Plate I. One point per facility, ranked left to right by the days of 24-hour sustained operation its estimated daily FLOP budget would need to accumulate 10²⁵ FLOP, with the whole-site budget allocated to a single model at the methodology's utilization factors.
The vertical axis is that duration on a log scale, bounded at one day and ten years to keep shorter durations legible; values outside those bounds sit at the edges of the plot.
The first three counts overlap: a site reaching the line in twelve days appears in the month, quarter and year counts. A site’s estimate follows the chip-count path where a chip count is disclosed and the power path otherwise.
The plot contains 499 estimable sites, and those reaching the line within a year span 29 countries. The stored catalogue snapshot holds 4,573 facilities, 511 of which carry a chip count or a power figure; the population filter then removes 11 duplicate site records and 1 chip-manufacturing record, whose power is not training capacity.
Missing sites can add capacity and estimated or planned figures can overstate what is operating, so the counts move with the recorded inputs.
Article 51 presumes high-impact capabilities for general-purpose models trained above 10²⁵ FLOP; a provider can rebut that presumption, and the Commission can designate a model on other capability criteria. Of 499 estimable sites, 178 could accumulate that total within a month under the assumptions below, the fastest in under a day.
Facility and model compute estimates read against the numerical thresholds in statute, with each instrument’s scope and status and a dated reconstruction of the EU systemic-risk register.
Per site, per regime
Choosing a regime re-derives every row against its numerical line. Calibration compares power-derived GPU counts with recorded counts; the reverse calculator solves for chips or days from the same dense FP16 specifications and shared assumptions.
Estimated daily FLOP and days to the selected line, for sites with a usable input.
EU AI Act: Systemic Risk GPAI
European Union: 10²⁵ FLOP
Regulation (EU) 2024/1689, Article 51(2) + Annex XIII · Official Journal of the European Union · Tier 1
Article 51(2): "A general-purpose AI model shall be presumed to have high impact capabilities ... when the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25." Article 52 provides for notification and a provider’s substantiated rebuttal; the Commission may also designate models on other capability or impact grounds. The July 2025 guidelines explain training-compute estimation and GPAI scope. Their indicative 10²³-FLOP criterion is not an exemption for every smaller model. EU market scope matters regardless of the provider’s location. Models placed on the market before August 2, 2025 have until August 2, 2027 to comply with the applicable GPAI obligations.
FACILITIES ≤ 30 DAYS (10²⁵ FLOP)
FACILITIES ≤ 90 DAYS
FACILITIES ≤ 1 YEAR
MODELS ABOVE EU
MODELS ABOVE 10²⁶ FLOP
The two model counts come from the 17-model frontier set behind the Frontier Models tab, and “above EU” is the part of it whose Epoch estimate clears 1025 FLOP. The Article 52(6) register lower on this page draws on a broader extract, including unreleased models and separately listed base-model rows, so it holds more rows.
499 of 499 facilities
| Facility | Country | Daily FLOP | Days to 10²⁵ FLOP | Estimation Path | Regime Exposure | Expand details |
|---|---|---|---|---|---|---|
| (†) | AE | <1 | Power: 5000 MW | No country match | ||
| US | <1 | Hardware: 700,000 GPUs (gb200) | No country match | |||
| US | <1 | Power: 2260 MW | No country match | |||
| US | <1 | Power: 2200 MW | No country match | |||
| FR | <1 | Hardware: 500,000 GPUs (gb200) | COUNTRY MATCH | |||
| US | <1 | Power: 2000 MW | No country match | |||
| IN | <1 | Hardware: 450,000 GPUs (gb200) | No country match | |||
| US | <1 | Power: 1800 MW | No country match | |||
| SA | <1 | Power: 1500 MW | No country match | |||
| US | <1 | Power: 1400 MW | No country match | |||
| US | <1 | Power: 1400 MW | No country match | |||
| US | <1 | Power: 1200 MW | No country match | |||
| US | <1 | Hardware: 300,000 GPUs (b200) | No country match | |||
| US | <1 | Power: 1000 MW | No country match | |||
| US | <1 | Power: 1000 MW | No country match | |||
| US | <1 | Hardware: 530,000 GPUs (H100 assumed) | No country match | |||
| US | <1 | Power: 750 MW | No country match | |||
| US | <1 | Hardware: 180,000 GPUs (gb200) | No country match | |||
| US | <1 | Power: 636 MW | No country match | |||
| US | <1 | Power: 631 MW | No country match | |||
| US | <1 | Hardware: 150,000 GPUs (gb200) | HQ MATCH | |||
| US | <1 | Hardware: 150,000 GPUs (gb200) | No country match | |||
| FR | <1 | Hardware: 150,000 GPUs (gb200) | COUNTRY MATCH | |||
| FR | <1 | Hardware: 150,000 GPUs (gb200) | COUNTRY MATCH | |||
| US | <1 | Power: 600 MW | No country match | |||
| US | ~1 | Power: 510 MW | No country match | |||
| SA | ~1 | Power: 500 MW | No country match | |||
| IT | ~1 | Power: 500 MW | COUNTRY MATCH | |||
| US | ~1 | Power: 500 MW | No country match | |||
| FR | ~1 | Hardware: 120,000 GPUs (gb200) | COUNTRY MATCH | |||
| US | ~1 | Hardware: 131,072 GPUs (b200) | No country match | |||
| US | ~1 | Power: 467 MW | No country match | |||
| US | ~1 | Hardware: 110,000 GPUs (gb200) | No country match | |||
| US | ~1 | Power: 450 MW | No country match | |||
| US | ~2 | Power: 369 MW | No country match | |||
| US | ~2 | Power: 340 MW | No country match | |||
| US | ~2 | Power: 339 MW | No country match | |||
| US | ~2 | Hardware: 202,224 GPUs (H100 assumed) | No country match | |||
| US | ~2 | Power: 303 MW | No country match | |||
| US | ~2 | Power: 300 MW | No country match | |||
| US | ~2 | Power: 300 MW | No country match | |||
| US | ~2 | Power: 300 MW | No country match | |||
| US | ~2 | Power: 284 MW | No country match | |||
| US | ~2 | Power: 262 MW | No country match | |||
| US | ~2 | Power: 250 MW | No country match | |||
| US | ~2 | Power: 250 MW | No country match | |||
| US | ~2 | Power: 250 MW | No country match | |||
| US | ~2 | Power: 250 MW | No country match | |||
| US | ~2 | Power: 250 MW | No country match | |||
| KR | ~2 | Hardware: 60,000 GPUs (gb200) | No country match | |||
| CN | ~2 | Power: 241.8 MW | No country match | |||
| MY | ~2 | Power: 240 MW | No country match | |||
| US | ~2 | Power: 238 MW | No country match | |||
| US | ~2 | Power: 237 MW | No country match | |||
| US | ~2 | Power: 237 MW | No country match | |||
| US | ~2 | Power: 237 MW | No country match | |||
| US | ~2 | Power: 233 MW | No country match | |||
| US | ~2 | Power: 233 MW | No country match | |||
| US | ~2 | Power: 228 MW | No country match | |||
| CN | ~3 | Power: 221 MW | No country match | |||
| US | ~3 | Power: 200 MW | No country match | |||
| US | ~3 | Power: 200 MW | No country match | |||
| US | ~3 | Power: 200 MW | No country match | |||
| US | ~3 | Power: 200 MW | No country match | |||
| US | ~3 | Hardware: 120,000 GPUs (h100_sxm5) | No country match | |||
| US | ~3 | Power: 190 MW | No country match | |||
| US | ~3 | Hardware: 45,000 GPUs (gb200) | No country match | |||
| US | ~3 | Power: 180 MW | No country match | |||
| US | ~3 | Power: 178 MW | No country match | |||
| US | ~3 | Power: 178 MW | No country match | |||
| US | ~3 | Power: 176 MW | No country match | |||
| CN | ~3 | Power: 169 MW | No country match | |||
| IN | ~3 | Power: 168 MW | No country match | |||
| US | ~3 | Hardware: 40,000 GPUs (gb200) | No country match | |||
| FR | ~3 | Hardware: 40,000 GPUs (gb200) | COUNTRY MATCH | |||
| US | ~3 | Hardware: 100,000 GPUs (h100_sxm5) | No country match | |||
| US | ~3 | Hardware: 100,000 GPUs (h100_sxm5) | No country match | |||
| US | ~3 | Hardware: 100,000 GPUs (H100 assumed) | No country match | |||
| US | ~3 | Power: 162 MW | No country match | |||
| US | ~4 | Power: 161 MW | No country match | |||
| US | ~4 | Power: 154 MW | No country match | |||
| US | ~4 | Power: 153 MW | No country match | |||
| US | ~4 | Power: 152 MW | No country match | |||
| US | ~4 | Power: 152 MW | No country match | |||
| US | ~4 | Hardware: 36,000 GPUs (gb200) | No country match | |||
| US | ~4 | Power: 144 MW | No country match | |||
| US | ~4 | Power: 141 MW | No country match | |||
| IN | ~4 | Power: 140 MW | No country match | |||
| US | ~4 | Power: 140 MW | No country match | |||
| US | ~4 | Power: 137 MW | No country match | |||
| US | ~4 | Power: 133 MW | No country match | |||
| US | ~4 | Power: 132 MW | No country match | |||
| US | ~5 | Hardware: 70,000 GPUs (h100_sxm5) | No country match | |||
| MY | ~5 | Power: 110 MW | No country match | |||
| US | ~5 | Hardware: 65,536 GPUs (h200_sxm) | No country match | |||
| US | ~5 | Hardware: 63,744 GPUs (H100 assumed) | No country match | |||
| US | ~5 | Power: 103 MW | No country match | |||
| US | ~6 | Power: 100 MW | No country match | |||
| FI | ~6 | Hardware: 60,000 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| GB | ~6 | Hardware: 23,040 GPUs (gb200) | No country match | |||
| AE | ~6 | Hardware: 55,000 GPUs (h100_sxm5) | No country match | |||
| US | ~6 | Power: 90 MW | No country match | |||
| GB | ~6 | Power: 88 MW | No country match | |||
| US | ~6 | Power: 88 MW | No country match | |||
| US | ~6 | Power: 87 MW | No country match | |||
| US | ~6 | Power: 87 MW | No country match | |||
| US | ~7 | Hardware: 20,736 GPUs (gb200) | No country match | |||
| US | ~7 | Power: 85 MW | No country match | |||
| US | ~7 | Hardware: 50,000 GPUs (h100_sxm5) | No country match | |||
| US | ~7 | Power: 82 MW | No country match | |||
| US | ~7 | Power: 77 MW | No country match | |||
| US | ~7 | Power: 77 MW | No country match | |||
| US | ~8 | Power: 75 MW | No country match | |||
| SA | ~8 | Hardware: 18,000 GPUs (gb300) | No country match | |||
| US | ~8 | Hardware: 44,544 GPUs (H100 assumed) | No country match | |||
| US | ~8 | Hardware: 42,000 GPUs (h200_sxm) | No country match | |||
| US | ~8 | Power: 68 MW | No country match | |||
| IN | ~9 | Hardware: 40,000 GPUs (H100 assumed) | No country match | |||
| US | ~9 | Power: 65 MW | No country match | |||
| MY | ~9 | Hardware: 15,428 GPUs (gb200) | No country match | |||
| US | ~9 | Power: 64 MW | No country match | |||
| IN | ~9 | Power: 60 MW | No country match | |||
| SG | ~10 | Power: 58 MW | No country match | |||
| US | ~10 | Hardware: 35,000 GPUs (h100_sxm5) | No country match | |||
| BR | ~10 | Power: 54 MW | No country match | |||
| US | ~11 | Power: 53 MW | No country match | |||
| US | ~11 | Hardware: 32,000 GPUs (h100_sxm5) | No country match | |||
| US | ~11 | Hardware: 32,000 GPUs (h100_sxm5) | No country match | |||
| IN | ~11 | Power: 50 MW | No country match | |||
| US | ~11 | Power: 50 MW | No country match | |||
| US | ~11 | Power: 50 MW | No country match | |||
| US | ~11 | Hardware: 30,000 GPUs (H100 assumed) | No country match | |||
| US | ~12 | Power: 47 MW | No country match | |||
| ~13 | Hardware: 10,496 GPUs (gb200) | HQ MATCH | ||||
| US | ~13 | Hardware: 26,000 GPUs (h100_sxm5) | No country match | |||
| US | ~13 | Hardware: 26,000 GPUs (h200_sxm) | HQ MATCH | |||
| TW | ~14 | Hardware: 10,000 GPUs (gb300) | No country match | |||
| IN | ~14 | Hardware: 25,000 GPUs (H100 assumed) | No country match | |||
| US | ~14 | Hardware: 24,576 GPUs (h100_sxm5) | No country match | |||
| US | ~14 | Hardware: 24,576 GPUs (h100_sxm5) | No country match | |||
| DE | ~14 | Hardware: 23,752 GPUs (H100 assumed) | COUNTRY MATCH | |||
| DE | ~15 | Hardware: 23,536 GPUs (h200_sxm) | COUNTRY MATCH | |||
| US | ~16 | Hardware: 22,000 GPUs (h100_sxm5) | No country match | |||
| US | ~16 | Hardware: 16,384 GPUs (mi300x) | No country match | |||
| PT | ~17 | Power: 33 MW | COUNTRY MATCH | |||
| CN | ~17 | Hardware: 20,000 GPUs (H100 assumed) | No country match | |||
| US | ~17 | Hardware: 20,000 GPUs (h100_sxm5) | No country match | |||
| US | ~17 | Hardware: 20,000 GPUs (h100_sxm5) | No country match | |||
| SA | ~17 | Hardware: 19,725 GPUs (H100 assumed)(3.8× path divergence) | No country match | |||
| US | ~18 | Power: 32 MW | No country match | |||
| US | ~18 | Hardware: 18,688 GPUs (H100 assumed) | No country match | |||
| IL | ~19 | Power: 30 MW | No country match | |||
| IN | ~19 | Power: 30 MW | No country match | |||
| CN | ~19 | Hardware: 18,000 GPUs (H100 assumed) | No country match | |||
| US | ~20 | Power: 28 MW | No country match | |||
| NO | ~21 | Hardware: 16,384 GPUs (h100_sxm5) | No country match | |||
| US | ~21 | Hardware: 16,384 GPUs (h100_sxm5) | No country match | |||
| IN | ~21 | Hardware: 16,384 GPUs (h100_sxm5) | No country match | |||
| IN | ~21 | Hardware: 16,384 GPUs (h100_sxm5) | No country match | |||
| AE | ~23 | Hardware: 15,000 GPUs (h100_sxm5) | No country match | |||
| AE | ~23 | Hardware: 15,000 GPUs (h100_sxm5) | No country match | |||
| US | ~23 | Hardware: 37,888 GPUs (mi250x) | No country match | |||
| US | ~24 | Hardware: 37,632 GPUs (mi250x) | No country match | |||
| IT | ~24 | Hardware: 5,760 GPUs (gb200) | COUNTRY MATCH | |||
| US | ~24 | Hardware: 14,400 GPUs (h100_sxm5) | No country match | |||
| CN | ~25 | Hardware: 14,000 GPUs (H100 assumed) | No country match | |||
| IT | ~25 | Hardware: 13,824 GPUs (H100 assumed) | COUNTRY MATCH | |||
| FR | ~25 | Hardware: 13,800 GPUs (H100 assumed) | COUNTRY MATCH | |||
| US | ~26 | Hardware: 10,000 GPUs (mi300x) | No country match | |||
| US | ~26 | Hardware: 10,000 GPUs (mi300x) | No country match | |||
| SA | ~27 | Hardware: 5,000 GPUs (gb200) | No country match | |||
| US | ~28 | Hardware: 12,448 GPUs (h100_sxm5) | No country match | |||
| SG | ~28 | Power: 20 MW | No country match | |||
| SG | ~28 | Power: 20 MW | No country match | |||
| FR | ~28 | Power: 20 MW | COUNTRY MATCH | |||
| US | ~28 | Power: 20 MW | No country match | |||
| FI | ~29 | Hardware: 11,912 GPUs (H100 assumed) | COUNTRY MATCH | |||
| TW | ~30 | Hardware: 4,608 GPUs (gb200) | No country match | |||
| CH | ~32 | Hardware: 10,752 GPUs (h200_sxm) | No country match | |||
| US | ~32 | Hardware: 10,752 GPUs (h100_sxm5) | No country match | |||
| US | ~33 | Hardware: 32,768 GPUs (a100_sxm) | No country match | |||
| ~34 | Hardware: 50,944 GPUs (tpu_v5e) | No country match | ||||
| US | ~34 | Hardware: 10,000 GPUs (h100_sxm5) | No country match | |||
| IN | ~34 | Hardware: 10,000 GPUs (H100 assumed) | No country match | |||
| JP | ~34 | Hardware: 10,000 GPUs (H100 assumed)(6.3× path divergence) | No country match | |||
| ~34 | Hardware: 10,000 GPUs (h100_sxm5) | No country match | ||||
| JP | ~34 | Hardware: 10,000 GPUs (h100_sxm5) | No country match | |||
| US | ~34 | Hardware: 10,000 GPUs (h100_sxm5) | No country match | |||
| US | ~38 | Hardware: 32,768 GPUs (tpu_v4) | No country match | |||
| JP | ~38 | Hardware: 4,000 GPUs (b200) | No country match | |||
| KR | ~39 | Hardware: 8,800 GPUs (h100_sxm5) | No country match | |||
| ~42 | Hardware: 8,192 GPUs (h100_sxm5) | HQ MATCH | ||||
| US | ~43 | Hardware: 8,000 GPUs (h100_sxm5) | No country match | |||
| FI | ~43 | Hardware: 8,000 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| US | ~43 | Hardware: 28,672 GPUs (tpu_v4) | No country match | |||
| US | ~44 | Hardware: 25,000 GPUs (a100_sxm) | No country match | |||
| GB | ~47 | Power: 12 MW | No country match | |||
| SE | ~52 | Hardware: 6,666 GPUs (H100 assumed) | COUNTRY MATCH | |||
| SE | ~52 | Hardware: 6,666 GPUs (H100 assumed) | COUNTRY MATCH | |||
| SE | ~52 | Hardware: 6,666 GPUs (H100 assumed) | COUNTRY MATCH | |||
| CH | ~54 | Hardware: 6,400 GPUs (h200_sxm) | No country match | |||
| ~55 | Hardware: 20,000 GPUs (a100_sxm) | No country match | ||||
| JP | ~56 | Hardware: 6,128 GPUs (h200_sxm) | No country match | |||
| CH | ~60 | Hardware: 5,704 GPUs (H100 assumed) | No country match | |||
| GB | ~63 | Hardware: 5,448 GPUs (h200_sxm) | No country match | |||
| IT | ~64 | Hardware: 13,888 GPUs (mi250x) | COUNTRY MATCH | |||
| JP | ~68 | Hardware: 2,000 GPUs (gb200) | No country match | |||
| US | ~68 | Hardware: 16,000 GPUs (a100_sxm) | No country match | |||
| US | ~69 | Hardware: 5,000 GPUs (H100 assumed)(9.3× path divergence) | No country match | |||
| SG | ~69 | Hardware: 5,000 GPUs (h100_sxm5) | No country match | |||
| IS | ~69 | Hardware: 4,992 GPUs (h200_sxm) | HQ MATCH | |||
| FI | ~75 | Hardware: 11,912 GPUs (mi250x) | COUNTRY MATCH | |||
| US | ~75 | Hardware: 4,608 GPUs (H100 assumed) | No country match | |||
| US | ~75 | Hardware: 4,608 GPUs (h100_sxm5) | No country match | |||
| ES | ~77 | Hardware: 4,480 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| ES | ~77 | Hardware: 4,480 GPUs (H100 assumed) | COUNTRY MATCH | |||
| US | ~78 | Hardware: 4,424 GPUs (h100_sxm5) | No country match | |||
| US | ~78 | Hardware: 4,400 GPUs (h100_sxm5) | No country match | |||
| IT | ~79 | Hardware: 13,824 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~79 | Hardware: 13,760 GPUs (a100_sxm) | No country match | |||
| AE | ~80 | Hardware: 4,320 GPUs (H100 assumed) | No country match | |||
| JP | ~81 | Hardware: 4,224 GPUs (H100 assumed) | No country match | |||
| FR | ~82 | Power: 6.9 MW | COUNTRY MATCH | |||
| US | ~83 | Hardware: 4,160 GPUs (H100 assumed) | No country match | |||
| US | ~83 | Hardware: 4,160 GPUs (H100 assumed) | No country match | |||
| ~83 | Hardware: 8,960 GPUs (tpu_v5p) | No country match | ||||
| US | ~84 | Hardware: 4,096 GPUs (H100 assumed) | No country match | |||
| US | ~84 | Hardware: 4,096 GPUs (H100 assumed) | No country match | |||
| FR | ~84 | Hardware: 4,096 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| FR | ~84 | Hardware: 4,096 GPUs (h200_sxm) | COUNTRY MATCH | |||
| IN | ~84 | Hardware: 4,096 GPUs (h100_sxm5) | No country match | |||
| AU | ~84 | Hardware: 4,088 GPUs (h200_sxm) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~84 | Hardware: 4,088 GPUs (h100_sxm5) | No country match | |||
| US | ~85 | Hardware: 4,032 GPUs (h200_sxm) | No country match | |||
| US | ~86 | Hardware: 4,000 GPUs (h100_sxm5) | No country match | |||
| US | ~86 | Hardware: 4,000 GPUs (H100 assumed) | No country match | |||
| US | ~86 | Hardware: 4,000 GPUs (h100_sxm5) | No country match | |||
| ~86 | Hardware: 4,000 GPUs (h100_sxm5) | No country match | ||||
| US | ~86 | Hardware: 4,000 GPUs (h200_sxm) | HQ MATCH | |||
| TH | ~86 | Hardware: 4,000 GPUs (h100_sxm5) | No country match | |||
| SE | ~86 | Hardware: 4,000 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| IN | ~86 | Hardware: 4,000 GPUs (h100_sxm5) | No country match | |||
| US | ~86 | Hardware: 4,000 GPUs (H100 assumed) | No country match | |||
| US | ~87 | Hardware: 3,000 GPUs (mi300x) | No country match | |||
| FI | ~87 | Hardware: 10,240 GPUs (mi250x) | COUNTRY MATCH | |||
| FR | ~93 | Hardware: 3,704 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| US | ~98 | Hardware: 27,648 GPUs (v100_sxm2) | No country match | |||
| IT | ~109 | Hardware: 3,170 GPUs (H100 assumed) | COUNTRY MATCH | |||
| CH | ~113 | Hardware: 3,040 GPUs (H100 assumed) | No country match | |||
| IT | ~115 | Hardware: 3,000 GPUs (H100 assumed)(3.5× path divergence) | COUNTRY MATCH | |||
| JP | ~115 | Hardware: 3,000 GPUs (h100_sxm5) | No country match | |||
| VN | ~115 | Hardware: 3,000 GPUs (h100_sxm5) | No country match | |||
| US | ~115 | Hardware: 3,000 GPUs (H100 assumed) | No country match | |||
| SA | ~122 | Hardware: 2,816 GPUs (h200_sxm) | No country match | |||
| US | ~124 | Hardware: 22,000 GPUs (v100_sxm2) | No country match | |||
| FR | ~128 | Hardware: 2,688 GPUs (h200_sxm) | COUNTRY MATCH | |||
| US | ~134 | Hardware: 2,560 GPUs (h200_sxm) | No country match | |||
| US | ~136 | Hardware: 8,000 GPUs (a100_sxm) | No country match | |||
| US | ~137 | Hardware: 2,512 GPUs (h100_sxm5) | No country match | |||
| KR | ~138 | Hardware: 2,496 GPUs (h100_sxm5) | No country match | |||
| US | ~148 | Hardware: 7,360 GPUs (a100_sxm) | No country match | |||
| FR | ~149 | Power: 3.8 MW | COUNTRY MATCH | |||
| ~151 | Hardware: 6,000 GPUs (ascend_910b) | No country match | ||||
| US | ~152 | Hardware: 7,168 GPUs (a100_sxm) | No country match | |||
| US | ~158 | Hardware: 17,280 GPUs (v100_sxm2) | No country match | |||
| JP | ~160 | Hardware: 2,156 GPUs (H100 assumed) | No country match | |||
| JP | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| US | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| US | ~168 | Hardware: 2,048 GPUs (H100 assumed) | No country match | |||
| GB | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | ||||
| ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | ||||
| NO | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| NO | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| IL | ~168 | Hardware: 2,048 GPUs (h100_sxm5) | No country match | |||
| ~168 | Hardware: 2,048 GPUs (h100_sxm5) | HQ MATCH | ||||
| FR | ~168 | Hardware: 2,048 GPUs (h200_sxm) | COUNTRY MATCH | |||
| JP | ~169 | Hardware: 2,040 GPUs (h100_sxm5) | No country match | |||
| JP | ~169 | Hardware: 2,040 GPUs (h100_sxm5) | No country match | |||
| JP | ~171 | Hardware: 2,016 GPUs (h100_sxm5) | No country match | |||
| CA | ~172 | Hardware: 2,000 GPUs (h100_sxm5) | No country match | |||
| US | ~179 | Hardware: 6,080 GPUs (a100_sxm) | No country match | |||
| FR | ~180 | Hardware: 1,908 GPUs (h200_sxm) | COUNTRY MATCH | |||
| US | ~189 | Hardware: 5,760 GPUs (a100_sxm) | No country match | |||
| US | ~201 | Hardware: 6,144 GPUs (tpu_v4) | No country match | |||
| US | ~202 | Hardware: 5,400 GPUs (a100_sxm) | No country match | |||
| JP | ~215 | Hardware: 1,600 GPUs (H100 assumed) | No country match | |||
| JP | ~224 | Hardware: 1,536 GPUs (H100 assumed) | No country match | |||
| DK | ~225 | Hardware: 1,528 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| FI | ~226 | Hardware: 1,520 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| US | ~226 | Hardware: 1,520 GPUs (H100 assumed) | No country match | |||
| RU | ~234 | Hardware: 1,472 GPUs (H100 assumed) | No country match | |||
| US | ~244 | Hardware: 4,480 GPUs (a100_sxm) | No country match | |||
| US | ~260 | Hardware: 1,000 GPUs (mi300x) | No country match | |||
| AE | ~270 | Hardware: 1,272 GPUs (h100_sxm5) | No country match | |||
| US | ~271 | Hardware: 4,032 GPUs (a100_sxm) | No country match | |||
| US | ~272 | Hardware: 10,000 GPUs (v100_sxm2) | No country match | |||
| US | ~273 | Hardware: 4,000 GPUs (a100_sxm) | No country match | |||
| ~273 | Hardware: 4,000 GPUs (a100_sxm) | No country match | ||||
| DE | ~291 | Hardware: 3,744 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~302 | Hardware: 4,096 GPUs (tpu_v4) | No country match | |||
| JP | ~307 | Hardware: 1,120 GPUs (h200_sxm) | No country match | |||
| US | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| ~336 | Hardware: 1,024 GPUs (H100 assumed) | No country match | ||||
| TH | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| US | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| US | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| US | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| IN | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| US | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| ~336 | Hardware: 1,024 GPUs (h200_sxm) | No country match | ||||
| IL | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | ||||
| TH | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| SG | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| TW | ~336 | Hardware: 1,024 GPUs (h100_sxm5) | No country match | |||
| FR | ~339 | Hardware: 1,016 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| KR | ~344 | Hardware: 1,000 GPUs (h100_sxm5) | No country match | |||
| CA | ~344 | Hardware: 1,000 GPUs (H100 assumed) | No country match | |||
| US | ~344 | Hardware: 1,000 GPUs (h100_sxm5) | No country match | |||
| KR | ~344 | Hardware: 1,000 GPUs (h100_sxm5) | No country match | |||
| US | ~344 | Hardware: 1,000 GPUs (h100_sxm5) | No country match | |||
| CA | ~344 | Hardware: 1,000 GPUs (H100 assumed) | No country match | |||
| SA | ~344 | Hardware: 7,912 GPUs (v100_sxm2) | No country match | |||
| US | ~347 | Hardware: 992 GPUs (H100 assumed) | No country match | |||
| US | ~347 | Hardware: 992 GPUs (H100 assumed) | No country match | |||
| JP | ~358 | Hardware: 960 GPUs (h100_sxm5) | No country match | |||
| IT | ~374 | Hardware: 7,280 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| US | ~379 | Hardware: 2,880 GPUs (a100_sxm) | No country match | |||
| DE | ~391 | Hardware: 880 GPUs (H100 assumed)(4.3× path divergence) | COUNTRY MATCH | |||
| DE | ~400 | Hardware: 860 GPUs (H100 assumed) | COUNTRY MATCH | |||
| CA | ~422 | Hardware: 816 GPUs (h100_sxm5) | No country match | |||
| LU | ~430 | Hardware: 800 GPUs (H100 assumed) | COUNTRY MATCH | |||
| TW | ~448 | Hardware: 768 GPUs (h100_sxm5) | No country match | |||
| US | ~463 | Hardware: 1,920 GPUs (mi250x) | No country match | |||
| US | ~487 | Hardware: 2,240 GPUs (a100_sxm) | No country match | |||
| KR | ~487 | Hardware: 2,240 GPUs (a100_sxm) | No country match | |||
| US | ~487 | Hardware: 2,240 GPUs (a100_sxm) | No country match | |||
| US | ~490 | Hardware: 5,560 GPUs (v100_sxm2) | No country match | |||
| JP | ~513 | Hardware: 5,312 GPUs (v100_sxm2) | No country match | |||
| US | ~516 | Hardware: 5,280 GPUs (v100_sxm2) | No country match | |||
| US | ~517 | Hardware: 2,112 GPUs (a100_sxm) | No country match | |||
| JP | ~533 | Hardware: 2,048 GPUs (a100_sxm) | No country match | |||
| NL | ~538 | Hardware: 640 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| NL | ~538 | Hardware: 640 GPUs (H100 assumed) | COUNTRY MATCH | |||
| BR | ~541 | Hardware: 2,016 GPUs (a100_sxm) | No country match | |||
| US | ~546 | Hardware: 2,000 GPUs (a100_sxm) | No country match | |||
| US | ~546 | Hardware: 2,000 GPUs (a100_sxm) | No country match | |||
| IN | ~565 | Power: 1 MW | No country match | |||
| US | ~591 | Hardware: 256 GPUs (b200) | No country match | |||
| CZ | ~597 | Hardware: 576 GPUs (H100 assumed) | COUNTRY MATCH | |||
| KR | ~620 | Hardware: 1,760 GPUs (a100_sxm) | No country match | |||
| US | ~623 | Hardware: 1,752 GPUs (a100_sxm) | No country match | |||
| JP | ~626 | Hardware: 4,352 GPUs (v100_sxm2) | No country match | |||
| SE | ~632 | Hardware: 544 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| FR | ~657 | Hardware: 1,352 GPUs (mi250x) | COUNTRY MATCH | |||
| US | ~672 | Hardware: 512 GPUs (H100 assumed) | No country match | |||
| US | ~672 | Hardware: 512 GPUs (H100 assumed) | No country match | |||
| US | ~683 | Hardware: 504 GPUs (h100_sxm5) | No country match | |||
| US | ~683 | Hardware: 504 GPUs (h100_sxm5) | No country match | |||
| RU | ~685 | Hardware: 1,592 GPUs (a100_sxm) | No country match | |||
| ~688 | Hardware: 500 GPUs (h100_sxm5) | No country match | ||||
| IT | ~689 | Hardware: 3,952 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| KR | ~741 | Hardware: 1,200 GPUs (mi250x) | No country match | |||
| PL | ~741 | Hardware: 464 GPUs (H100 assumed) | COUNTRY MATCH | |||
| PL | ~741 | Hardware: 464 GPUs (h200_sxm) | COUNTRY MATCH | |||
| PL | ~765 | Hardware: 450 GPUs (H100 assumed) | COUNTRY MATCH | |||
| AU | ~768 | Hardware: 448 GPUs (h100_sxm5) | No country match | |||
| FR | ~805 | Hardware: 3,384 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| SA | ~827 | Hardware: 416 GPUs (h100_sxm5) | No country match | |||
| US | ~832 | Hardware: 1,312 GPUs (a100_sxm) | No country match | |||
| US | ~832 | Hardware: 1,312 GPUs (a100_sxm) | No country match | |||
| NL | ~832 | Hardware: 1,312 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~832 | Hardware: 1,312 GPUs (a100_sxm) | No country match | |||
| US | ~859 | Hardware: 3,169 GPUs (v100_sxm2) | No country match | |||
| SE | ~860 | Hardware: 400 GPUs (H100 assumed) | COUNTRY MATCH | |||
| FR | ~875 | Hardware: 3,112 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| IT | ~908 | Hardware: 3,000 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| KR | ~974 | Hardware: 1,120 GPUs (a100_sxm) | No country match | |||
| US | ~974 | Hardware: 1,120 GPUs (a100_sxm) | No country match | |||
| PL | ~989 | Hardware: 348 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| US | ~995 | Hardware: 2,736 GPUs (v100_sxm2) | No country match | |||
| RU | ~996 | Hardware: 1,096 GPUs (a100_sxm) | No country match | |||
| RU | ~1,003 | Hardware: 1,088 GPUs (a100_sxm) | No country match | |||
| FR | ~1,010 | Hardware: 2,696 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| KR | ~1,049 | Hardware: 1,040 GPUs (a100_sxm) | No country match | |||
| KR | ~1,075 | Hardware: 320 GPUs (h100_sxm5) | No country match | |||
| PL | ~1,132 | Hardware: 304 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| AU | ~1,157 | Hardware: 768 GPUs (mi250x) | No country match | |||
| JP | ~1,176 | Hardware: 928 GPUs (a100_sxm) | No country match | |||
| US | ~1,240 | Hardware: 880 GPUs (a100_sxm) | No country match | |||
| BR | ~1,252 | Hardware: 2,176 GPUs (v100_sxm2) | No country match | |||
| US | ~1,330 | Hardware: 2,048 GPUs (v100_sxm2) | No country match | |||
| US | ~1,344 | Hardware: 256 GPUs (h100_sxm5) | No country match | |||
| ~1,344 | Hardware: 256 GPUs (h100_sxm5) | No country match | ||||
| TW | ~1,351 | Hardware: 2,016 GPUs (v100_sxm2) | No country match | |||
| LU | ~1,364 | Hardware: 800 GPUs (a100_sxm) | COUNTRY MATCH | |||
| KR | ~1,378 | Hardware: 792 GPUs (a100_sxm) | No country match | |||
| RU | ~1,378 | Hardware: 792 GPUs (a100_sxm) | No country match | |||
| IT | ~1,387 | Hardware: 248 GPUs (h100_sxm5) | COUNTRY MATCH | |||
| IS | ~1,387 | Hardware: 248 GPUs (h100_sxm5) | No country match | |||
| DE | ~1,421 | Hardware: 768 GPUs (a100_sxm) | COUNTRY MATCH | |||
| DE | ~1,421 | Hardware: 768 GPUs (a100_sxm) | COUNTRY MATCH | |||
| SI | ~1,433 | Hardware: 240 GPUs (H100 assumed) | COUNTRY MATCH | |||
| SE | ~1,451 | Hardware: 752 GPUs (a100_sxm) | COUNTRY MATCH | |||
| DE | ~1,474 | Hardware: 740 GPUs (a100_sxm) | COUNTRY MATCH | |||
| TH | ~1,550 | Hardware: 704 GPUs (a100_sxm) | No country match | |||
| BR | ~1,550 | Hardware: 704 GPUs (a100_sxm) | No country match | |||
| IN | ~1,663 | Hardware: 656 GPUs (a100_sxm) | No country match | |||
| GB | ~1,705 | Hardware: 640 GPUs (a100_sxm) | No country match | |||
| CA | ~1,716 | Hardware: 636 GPUs (a100_sxm) | No country match | |||
| DE | ~1,726 | Hardware: 632 GPUs (a100_sxm) | COUNTRY MATCH | |||
| ~1,773 | Hardware: 1,536 GPUs (v100_sxm2) | No country match | ||||
| US | ~1,783 | Hardware: 612 GPUs (a100_sxm) | No country match | |||
| US | ~1,801 | Hardware: 1,512 GPUs (v100_sxm2) | No country match | |||
| CZ | ~1,894 | Hardware: 576 GPUs (a100_sxm) | COUNTRY MATCH | |||
| AE | ~2,101 | Hardware: 1,296 GPUs (v100_sxm2) | No country match | |||
| DE | ~2,131 | Hardware: 512 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~2,131 | Hardware: 512 GPUs (a100_sxm) | No country match | |||
| US | ~2,131 | Hardware: 512 GPUs (a100_sxm) | No country match | |||
| US | ~2,179 | Hardware: 1,250 GPUs (v100_sxm2) | No country match | |||
| RU | ~2,269 | Hardware: 1,200 GPUs (v100_sxm2) | No country match | |||
| SE | ~2,273 | Hardware: 480 GPUs (a100_sxm) | COUNTRY MATCH | |||
| GB | ~2,435 | Hardware: 448 GPUs (a100_sxm) | No country match | |||
| ~2,480 | Hardware: 440 GPUs (a100_sxm) | No country match | ||||
| JP | ~2,503 | Hardware: 1,088 GPUs (v100_sxm2) | No country match | |||
| BR | ~2,503 | Hardware: 1,088 GPUs (v100_sxm2) | No country match | |||
| FR | ~2,609 | Hardware: 1,044 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| JP | ~2,659 | Hardware: 1,024 GPUs (v100_sxm2) | No country match | |||
| US | ~2,728 | Hardware: 400 GPUs (a100_sxm) | No country match | |||
| PL | ~2,841 | Hardware: 384 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~3,031 | Hardware: 360 GPUs (a100_sxm) | No country match | |||
| JP | ~3,031 | Hardware: 360 GPUs (a100_sxm) | No country match | |||
| US | ~3,039 | Hardware: 896 GPUs (v100_sxm2)(4.1× path divergence) | No country match | |||
| JP | ~3,081 | Hardware: 884 GPUs (v100_sxm2) | No country match | |||
| SG | ~3,100 | Hardware: 352 GPUs (a100_sxm) | No country match | |||
| JP | ~3,247 | Hardware: 336 GPUs (a100_sxm) | No country match | |||
| IN | ~3,247 | Hardware: 336 GPUs (a100_sxm) | No country match | |||
| US | ~3,370 | Hardware: 808 GPUs (v100_sxm2) | No country match | |||
| US | ~3,404 | Hardware: 800 GPUs (v100_sxm2) | No country match | |||
| IT | ~3,410 | Hardware: 320 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~3,410 | Hardware: 320 GPUs (a100_sxm) | No country match | |||
| GB | ~3,410 | Hardware: 320 GPUs (a100_sxm) | No country match | |||
| RU | ~3,440 | Hardware: 100 GPUs (H100 assumed) | No country match | |||
| US | ~3,453 | Hardware: 316 GPUs (a100_sxm) | No country match | |||
| JP | ~3,546 | Hardware: 768 GPUs (v100_sxm2) | No country match | |||
| CA | ~3,546 | Hardware: 768 GPUs (v100_sxm2) | No country match | |||
| BR | ~3,782 | Hardware: 720 GPUs (v100_sxm2) | No country match | |||
| NL | ~3,788 | Hardware: 288 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~3,897 | Hardware: 280 GPUs (a100_sxm) | No country match | |||
| CA | ~3,958 | Hardware: 688 GPUs (v100_sxm2) | No country match | |||
| DE | ~4,011 | Hardware: 272 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~4,133 | Hardware: 264 GPUs (a100_sxm) | No country match | |||
| AU | ~4,255 | Hardware: 640 GPUs (v100_sxm2) | No country match | |||
| US | ~4,262 | Hardware: 256 GPUs (a100_sxm) | No country match | |||
| SG | ~4,262 | Hardware: 256 GPUs (a100_sxm) | No country match | |||
| US | ~4,330 | Hardware: 252 GPUs (a100_pcie) | No country match | |||
| US | ~4,479 | Hardware: 608 GPUs (v100_sxm2) | No country match | |||
| SI | ~4,546 | Hardware: 240 GPUs (a100_pcie) | COUNTRY MATCH | |||
| US | ~4,728 | Hardware: 576 GPUs (v100_sxm2) | No country match | |||
| US | ~4,728 | Hardware: 576 GPUs (v100_sxm2) | No country match | |||
| BR | ~4,728 | Hardware: 576 GPUs (v100_sxm2) | No country match | |||
| IN | ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | |||
| US | ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | |||
| US | ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | |||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| ~5,375 | Hardware: 64 GPUs (H100 assumed) | No country match | ||||
| GB | ~5,403 | Hardware: 504 GPUs (v100_sxm2) | No country match | |||
| BR | ~5,455 | Hardware: 200 GPUs (a100_sxm) | No country match | |||
| US | ~5,455 | Hardware: 200 GPUs (a100_sxm) | No country match | |||
| US | ~5,581 | Hardware: 488 GPUs (v100_sxm2) | No country match | |||
| US | ~5,683 | Hardware: 192 GPUs (a100_sxm) | No country match | |||
| FR | ~5,683 | Hardware: 192 GPUs (a100_sxm) | COUNTRY MATCH | |||
| FR | ~5,683 | Hardware: 192 GPUs (a100_sxm) | COUNTRY MATCH | |||
| US | ~5,683 | Hardware: 192 GPUs (a100_sxm) | No country match | |||
| DE | ~5,683 | Hardware: 192 GPUs (a100_sxm) | COUNTRY MATCH | |||
| DE | ~6,079 | Hardware: 448 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| US | ~6,079 | Hardware: 448 GPUs (v100_sxm2) | No country match | |||
| JP | ~6,546 | Hardware: 416 GPUs (v100_sxm2) | No country match | |||
| RU | ~6,819 | Hardware: 160 GPUs (a100_sxm) | No country match | |||
| US | ~6,819 | Hardware: 160 GPUs (a100_sxm) | No country match | |||
| VN | ~6,819 | Hardware: 160 GPUs (a100_sxm) | No country match | |||
| BR | ~7,400 | Hardware: 368 GPUs (v100_sxm2) | No country match | |||
| KR | ~7,577 | Hardware: 144 GPUs (a100_sxm) | No country match | |||
| FI | ~8,510 | Hardware: 320 GPUs (v100_sxm2) | COUNTRY MATCH | |||
| US | ~8,510 | Hardware: 320 GPUs (v100_sxm2) | No country match | |||
| JP | ~8,510 | Hardware: 320 GPUs (v100_sxm2) | No country match | |||
| BR | ~8,729 | Hardware: 312 GPUs (v100_sxm2) | No country match | |||
| US | ~9,200 | Hardware: 296 GPUs (v100_sxm2) | No country match | |||
| US | ~10,316 | Hardware: 264 GPUs (v100_sxm2) | No country match |
Sites within a year of 10²⁵ FLOP, by country
⚠ marks a country listed in this catalogue for EU AI Act: Systemic Risk GPAI.
Estimates assume 24-hour training at dense FP16 throughput, adjusted by the methodology's MFU and interconnect factors; the power path also assumes a PUE and a GPU share of IT load. Where a record carries no mapped accelerator the estimate falls back to the H100 SXM5. Open a row for its inputs and parameter sensitivity. Methodology version 1.3.7.
(†) A marked row's daily-FLOP value is the ceiling of a published envelope; the marker carries the site's caveat, which states the operating capacity where the two differ materially.
Article 52(6)
Scrutica's search log records no Article 52(6) list in the sources checked on 21 July 2026. The table sets that dated search beside every model in the pinned Epoch extract whose estimated training compute exceeds 10²⁵ FLOP.
| Model | Provider | Training compute (Epoch est.) | Epoch publication date | Publication cohort | Code of Practice | Public notification record |
|---|---|---|---|---|---|---|
| Grok 4 | xAI | 5 × 10²⁶Speculative | 9 Jul 2025 | Published before 2 Aug 2025 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-4.5 | OpenAI | 3.8 × 10²⁶Likely | 27 Feb 2025 | Published before 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Grok 3 | xAI | 3.5 × 10²⁶Likely | 17 Feb 2025 | Published before 2 Aug 2025 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-5 | OpenAI | 6.6 × 10²⁵Speculative | 7 Aug 2025 | Published from 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026 |
| Llama 4 Behemoth (preview) | Meta AI | 5.184 × 10²⁵Likely | 5 Apr 2025 | Unreleased per Epoch‡ | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Gemini 1.0 Ultra | Google DeepMind | 5 × 10²⁵Speculative | 6 Dec 2023 | Published before 2 Aug 2025 | Signatory, full Code (listed as Google) | Not found in the enumerated sources as of 21 Jul 2026† |
| Llama Nemotron Ultra 253B | NVIDIA | 3.911 × 10²⁵Likely | 18 Mar 2025 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Composer 2.5 | Cursor | 3.87 × 10²⁵Likely | 18 May 2026 | Published from 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Llama 3.1-405B | Meta AI | 3.8 × 10²⁵Confident | 23 Jul 2024 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Claude 3.7 Sonnet | Anthropic | 3.35 × 10²⁵Likely | 24 Feb 2025 | Published before 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Grok-2 | xAI | 2.96 × 10²⁵Confident | 13 Aug 2024 | Published before 2 Aug 2025 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| Claude 3.5 Sonnet | Anthropic | 2.7 × 10²⁵Speculative | 20 Jun 2024 | Published before 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Doubao-pro | ByteDance | 2.505 × 10²⁵Likely | 28 Oct 2024 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Composer 2 | Cursor | 2.32 × 10²⁵Confident | 19 Mar 2026 | Published from 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| GPT-4 (Jun 2023) | OpenAI | 2.1 × 10²⁵Likely | 13 Jun 2023 | Published before 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-4 (Mar 2023) | OpenAI | 2.1 × 10²⁵Likely | 15 Mar 2023 | Published before 2 Aug 2025 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Nemotron-4 340B | NVIDIA | 1.8 × 10²⁵Confident | 14 Jun 2024 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Qwen3-Max | Alibaba | 1.512 × 10²⁵Speculative | 5 Sep 2025 | Published from 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Pangu Ultra | Huawei | 1.069 × 10²⁵Confident | 10 Apr 2025 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Aramco Metabrain AI | Saudi Aramco | 1.05 × 10²⁵Likely | 4 Mar 2024 | Unreleased per Epoch‡ | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Inflection-2 | Inflection AI | 1.001 × 10²⁵Confident | 22 Nov 2023 | Published before 2 Aug 2025 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
† Article 111(3) gives providers of GPAI models placed on the market before 2 August 2025 until 2 August 2027 to comply. The cohort column uses Epoch's publication date as a proxy for that placement, which the date alone leaves open, as it does each provider's own notification deadline.
‡ Article 52(1) can run before release (“...or it becomes known that it will be met”); unreleased models above the line are retained for that reason, with release status shown as Epoch records it.
Article 113 sets 2 August 2026 for the Commission's GPAI enforcement provisions.
Article 52(1) requires notification without delay and within two weeks after the Article 51(1)(a) condition is met or it becomes known that it will be met. Providers submit through EU SEND, and Article 78 protects specified confidential information, so a filing can exist without leaving any of the public traces this table reads.
Article 52(6) requires the Commission to publish and update a list of systemic-risk models, with protections for intellectual property and confidential business information. The search log below contains eleven entries checked on 21 July 2026; it is a dated record of those checks.
Rows retain the compute estimates and per-row vintages in the Epoch extract accessed on 21 July 2026. Display removes floating-point noise and caps values at four significant figures; the export preserves the raw strings. Epoch's confidence label describes the most uncertain populated field among compute, parameters and dataset size. Its documentation expresses the categories as 90% confidence intervals.
These estimates have not been reconciled model by model to the Commission's training-compute accounting guidance. The guidelines' 30% allowance governs provider reporting and leaves the Epoch estimates without a stated interval.
Article 52 permits rebuttal of the compute presumption and Commission designation on other capability criteria, so membership of the official list turns on more than an estimate above the line.
The log records the sources checked on 21 July 2026 and has not been re-run since. Open-web searches on 2026-07-21 (re-running the 2026-06-09 pass) for a Commission publication of the list, and for provider self-disclosures of Article 52(1) notifications, surfaced neither; only statutory-text mirrors and third-party reconstructions (Epoch-derived rosters) appear.
| Source | Result | Page vintage | Method |
|---|---|---|---|
| Commission AI regulatory framework policy page | No register, no link to one | Last update 7 July 2026 | direct fetch |
| Guidelines for providers of general-purpose AI models | No register; describes notification duty only | Last update 28 April 2026 | direct fetch |
| Commission FAQ on GPAI provider obligations | No register; states the 2 August 2027 rule for pre-2 August 2025 models | Last update 11 November 2025 | direct fetch |
| AI Act Service Desk, Article 52 (Commission-operated) | Quotes the 52(6) duty verbatim; links no list | No page date; footer "Official version of 13 June 2024" (statute text vintage) | direct fetch |
| Commission factpage on GPAI obligations | No register; confirms the private notification mailbox (EU-AIOFFICEGPAI-SR-PROVIDERS@ec.europa.eu) | Last update 1 August 2025 | direct fetch |
| Commission news item on the GPAI Guidelines | No register | Published 18 July 2025, updated 31 July 2025 | direct fetch |
| GPAI Code of Practice policy page | No register, no link to one | Last update 25 June 2026 | direct fetch |
| Code of Practice contents + signatory list | Signatory list present; no model register | Last update 23 April 2026 | direct fetch |
| Commission Q&A on GPAI models in the AI Act | No register; references Article 111(3) special rules | Last update 9 September 2025 | direct fetch |
| European AI Office policy page | No register; classification work described, no published list | Last update 17 July 2026 | direct fetch |
| Site-scoped searches across Commission AI Office surfaces | Zero register hits; only obligation descriptions and templates | n/a | recorded search |
Sources
Regulation (EU) 2024/1689, Articles 51, 52, 78, 111(3), 113 (tier 1, a primary source) · eur-lex.europa.eu (CELEX 32024R1689)
AI Act Service Desk, Articles 51 and 52 pages (Commission-operated; Article 52(1) and 52(6) wording verified 21 July 2026) · ai-act-service-desk.ec.europa.eu/en/ai-act/article-52
European Commission, Guidelines for providers of general-purpose AI models (EU SEND platform and notification mechanics) · digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers
European Commission, FAQ on obligations for GPAI providers (the 2 August 2027 sentence; page updated 11 November 2025) · digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers
European Commission, GPAI Code of Practice contents and signatory list (page updated 23 April 2026; checked 21 July 2026) · digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
Epoch AI, “Data on AI Models” (tier 2, a research database; CC-BY; accessed 21 July 2026; per-row vintage in the export) · epoch.ai/data/ai-models
What a FLOP count misses
In its May 2026 cyber study, AISI capped tasks at 2.5 million tokens and used a simple agent scaffold to keep comparisons stable over time; the institute says those choices suppress success rates relative to larger budgets and stronger scaffolds. Its July study varied inference budgets across several benchmarks and found newer models gaining more from the additional compute.
For a governance assessment the test conditions belong beside the score: task coverage, inference budget, available tools and scaffold. A failure under one configuration bounds that configuration alone.
Sources
AISI · “How fast is autonomous AI cyber capability advancing?” (13 May 2026): Study and methodology
AISI · “More compute, more capability: why AI agent evals need to account for test-time compute” (July 2026): Study and results
European Parliament & Council · Regulation (EU) 2024/1689, Article 51 (classification of general-purpose AI models with systemic risk): Consolidated regulation
Parameters
Every estimate above uses the hardware and power paths at these default settings. The cards also carry the cost-path parameter, which only the Estimator uses; assumptions changed there leave these defaults in place.
Additional throughput multiplier applied by the estimator alongside utilization; its default is Scrutica's choice. Published end-to-end utilization already includes communication losses, so a measured utilization rate substituted here already carries the interconnect discount.
Utilization multiplier applied to dense peak throughput. PaLM 540B reports 46.2% model FLOP utilization and 57.8% hardware FLOP utilization; the latter includes rematerialization, and both figures describe that run. The default is Scrutica's choice, and the estimator additionally applies its interconnect multiplier.
Total facility power divided by IT-equipment power. The default converts a reported facility power figure to an assumed IT load; a site's own disclosure replaces it where one exists. Use a disclosure with matching facility scope and operating period when available.
Assumed share of IT power allocated to accelerator chips. The remainder covers other IT equipment; cooling outside the IT boundary is handled through PUE. This fraction does not establish a facility’s rack composition or operating GPU count.
Assumed share of a project’s capital expenditure allocated to accelerator purchases. A project budget may cover other phases, land, buildings and infrastructure; matching the investment’s scope to the equipment being estimated is necessary before using this conversion.
The full derivation chain, with every formula, is in the methodology chapter.
The Estimator shows how one of these durations is produced: a daily FLOP budget back-solved from a disclosed chip count, a nameplate power figure or a construction budget, and how far the first two answers agree for a site disclosing both.