Capacity Estimator — Hardware, Power, and Cost-Based Compute Capacity Methodology

Gringras, David

Scrutica

Facilities written 2026-09

Capacity

Compute capacity is estimated site by site, then read against the thresholds in statute.

Estimate daily compute

Compare recorded sites

Use a disclosed chip count, power capacity, or construction budget.

Default-parameters derivation (Hardware path)

10,000 H100 SXM5 GPUs at the documented defaults (interconnect efficiency 0.85, MFU 0.4). This derivation is fixed; the calculator below, once its scripts load, exposes every parameter as an adjustable slider.

Hardware-path FLOP derivation steps at default parameters (10,000 H100 SXM5)
Step	Formula	Result
Peak FP16	gpu_count × per_gpu_fp16_tflops	9.89 × 10¹⁸ FLOP/s
Effective FP16	peak × interconnect_efficiency (0.85)	8.41 × 10¹⁸ FLOP/s
Training Throughput	effective × mfu_assumption (0.4)	3.36 × 10¹⁸ FLOP/s
Daily FLOP Budget	training × 86,400	2.91 × 10²³ FLOP/day

At this daily budget, the EU AI Act 10²⁵ FLOP threshold is reached in approximately 34.4 days. Adjust GPU count, model, interconnect efficiency, and MFU in the interactive calculator below.

Chip-count estimates of daily compute exceed power-based estimates at 279 of 314 sites

Plate I Two estimates of one site’s daily FLOP budget 314 sites · none sampled out

estimate from the chip count estimate from the power figure the disagreement between the two

279within 1.5×

151.5–2×

142–3×

6beyond 3×

Spacing between neighbouring sites carries no information; only their order does.

Every facility disclosing both a chip count and a power figure is drawn: 314 of the 511 that disclose either, across 32 countries. The wider figure counts chip-manufacturing sites too, because a fab’s two disclosures check against each other like any other site’s; Thresholds drops them and states a smaller population, because a fab’s power draw is production load and the power path would read it as training compute.

The larger estimate is 1.15× the smaller at the median and 9.3× in the most divergent pair. For NVIDIA SATURN V Volta, the recorded 264 chips and 0.2 MW produce estimates that differ by under 0.05 per cent using the V100 SXM2 specification.

Facility records as of 2026-09-06.

The two paths agree within a factor of two at 294 of 314 facilities. Each pair uses the same accelerator model: the one identified in the site’s record, or H100 SXM5 when no model is recognised. Differences can reflect the assumed power overhead and GPU share of IT power, or disclosures about different stages of construction (an announced fleet alongside first-phase power capacity, for example). Agreement depends on these assumptions; disagreement alone cannot identify an incorrect disclosure.

Configure an estimate

How this is derived

Check

Published hardware and compute for frontier training runs

These 10 frontier-class training runs (≥ 10²⁵ FLOP) provide reference figures for interpreting an estimate. The table pairs available hardware disclosures with Epoch’s compute estimates, or a lab’s compute figure where no Epoch record is available.

Model	Reported hardware	Training compute (FLOP)	Epoch confidence	Source provenance
Grok 4 xAI · Jul 2025	200,000 GPUs (class undisclosed) via Epoch derivation	5.0e+26 FLOP T2	Speculative	Epoch (T2)
Grok 3 xAI · Feb 2025	NVIDIA H100 SXM5 80GB × 80,000 via Epoch derivation	3.5e+26 FLOP T2	Likely	Epoch (T2)
Llama 4 Behemoth (preview) Meta Platforms, Inc. · Apr 2025	32,000 GPUs (class undisclosed) Lab paper / announcement	5.2e+25 FLOP T2	Likely	Lab disclosure (T4) Epoch (T2)
Gemini 1.0 Ultra Google DeepMind · Dec 2023	Google TPUv4 Lab paper / announcement	5.0e+25 FLOP T2	Speculative	Lab disclosure (T4) Epoch (T2)
Llama 3.1-405B Meta Platforms, Inc. · Jul 2024	NVIDIA H100 80GB × 16,384 Lab paper / announcement	3.8e+25 FLOP T2	Confident	Lab disclosure (T1) Epoch (T2)
GPT-4 (Jun 2023) OpenAI · Jun 2023	NVIDIA A100 SXM4 40 GB × 25,000 via Epoch derivation	2.1e+25 FLOP T2	Likely	Epoch (T2)
GPT-4 (Mar 2023) OpenAI · Mar 2023	NVIDIA A100 SXM4 40 GB × 25,000 via Epoch derivation	2.1e+25 FLOP T2	Likely	Lab disclosure (T4) Epoch (T2)
Nemotron-4 340B Nvidia · Jun 2024	NVIDIA H100 80GB SXM5 × 6,144 Lab paper / announcement	1.8e+25 FLOP T2	Confident	Lab disclosure (T4) Epoch (T2)
Pangu Ultra Huawei Technologies · Apr 2025	Huawei Ascend NPUs × 8,192 Lab paper / announcement	1.1e+25 FLOP T2	Confident	Lab disclosure (T4) Epoch (T2)
Inflection-2 Inflection AI · Nov 2023	NVIDIA H100 × 5,000 Lab paper / announcement	1.0e+25 FLOP T2	Confident	Lab disclosure (T1) Source unavailable when checked 2026-09-02 (HTTP 404). Flagged citation Epoch (T2)

Source: Epoch AI, CC BY 4.0 · lab-primary rows from each lab’s canonical paper / system card / announcement (URLs above). Table: frontier_model_runs.

Parameters

Shared assumptions

The calculator and Thresholds view use these five default parameters. Adjust them with the calculator’s sliders. Methodology version 1.3.7, dated 2026-09-05.

Shared assumptions5 parameters · methodology v1.3.7

Interconnect Efficiency0.85

range 0.6–0.95Hardware / Power / Cost paths

Scrutica modeling assumption

Model FLOP Utilization (MFU)0.4

range 0.2–0.5Hardware / Power / Cost paths

Scrutica modeling assumption; benchmark context: PaLM, Table 3

Power Usage Effectiveness (PUE)1.15

range 1.05–1.45Power path

Scrutica modeling assumption

GPU Fraction of IT Load0.49

range 0.35–0.65Power path

Scrutica modeling assumption

GPU Fraction of Capex0.45

range 0.3–0.6Cost path

Scrutica modeling assumption

The full derivation chain, with every formula, is in the methodology chapter.

The register behind Plate I

Compare estimates by facility

For each site, compare both estimates with the chip count and power capacity used to calculate them. Filter by country or degree of disagreement; the URL preserves your filters and sort order.

Divergence

Country

Sort

314 facilities

Facility	Country	Divergence	Inputs
NVIDIA SATURN V Volta	US	1.00× 0% normalized	264 GPUs 0.2 MW
Meta 2017 V100 Cluster	US	1.00× 0% normalized	22,000 GPUs 15 MW
Sakura's B200s Phase 2	JP	1.00× 0% normalized	10,000 GPUs 16 MW
OpenAI/Microsoft Atlanta	US	1.01× 1% normalized	300,000 GPUs 700 MW
EuroHPC Karolina	CZ	1.01× 1% normalized	576 GPUs 0.5 MW
University of Edinburgh DiRAC Tursa	GB	1.02× 2% normalized	448 GPUs 0.4 MW
xAI Fulton Georgia	US	1.02× 2% normalized	12,448 GPUs 20 MW
NVIDIA Circe	US	1.02× 2% normalized	576 GPUs 0.4 MW
University of Florida HiPerGator 3.0 Superpod	US	1.02× 2% normalized	1,120 GPUs 1 MW
Lawrence Berkeley NL NERSC Perlmutter	US	1.03× 3% normalized	7,168 GPUs 7 MW
Simon Fraser University/Compute Canada Cedar	CA	1.03× 3% normalized	768 GPUs 0.6 MW
EuroHPC Leonardo	IT	1.03× 3% normalized	13,824 GPUs 13 MW
University of Tokyo Wisteria/BDEC-01 (Aquarius)	JP	1.04× 4% normalized	360 GPUs 0.3 MW
TotalEnergies Pangea III	FR	1.04× 4% normalized	3,384 GPUs 2 MW
Petrobras Pegasus (Pégaso)	BR	1.05× 5% normalized	2,016 GPUs 2 MW
Atos Spartan2	FR	1.05× 5% normalized	192 GPUs 0.2 MW
Cineca Marconi-100	IT	1.05× 5% normalized	3,952 GPUs 3 MW
Taiwania 2	TW	1.06× 6% normalized	2,016 GPUs 1 MW
NVIDIA SATURN V Phase 3	US	1.06× 6% normalized	5,560 GPUs 4 MW
IBM DYEUS	US	1.06× 6% normalized	296 GPUs 0.2 MW
Laboratório Nacional de Computação Científica Santos Dumont	BR	1.06× 6% normalized	368 GPUs 0.2 MW
Microsoft Azure ND v2 Largest Stated	US	1.06× 6% normalized	800 GPUs 0.5 MW
SberCloud Christofari	RU	1.06× 6% normalized	1,200 GPUs 0.8 MW
SENAI CIMATEC Ogbon Cimatec/Petrobras	BR	1.06× 6% normalized	312 GPUs 0.2 MW
AWS Fast BERT Training	US	1.06× 6% normalized	2,048 GPUs 1 MW
Preferred Networks MN-2	JP	1.06× 6% normalized	1,024 GPUs 0.7 MW
SAKURA Internet 2019 V100 Supercomputer	JP	1.06× 6% normalized	768 GPUs 0.5 MW
TACC Longhorn	US	1.06× 6% normalized	448 GPUs 0.3 MW
CSC Puhti	FI	1.06× 6% normalized	320 GPUs 0.2 MW
Microsoft Research Hypercluster	US	1.06× 6% normalized	576 GPUs 0.4 MW
University of Tsukuba Cygnus	JP	1.06× 6% normalized	320 GPUs 0.2 MW
Lawrence Livermore NL Sierra	US	1.06× 6% normalized	17,280 GPUs 12 MW
Ubilink.AI Supercomputer	TW	1.06× 6% normalized	1,024 GPUs 2 MW
Lawrence Livermore NL RZAdams	US	1.06× 6% normalized	512 GPUs 0.8 MW
Eni HPC4 Phase 2	IT	1.06× 6% normalized	3,000 GPUs 2 MW
FZJ JURECA	DE	1.07× 6% normalized	768 GPUs 0.7 MW
Nebius Kansas City Phase 2	US	1.07× 7% normalized	26,000 GPUs 40 MW
AWS EC2 P4d	US	1.07× 7% normalized	4,000 GPUs 3 MW
Continental DGX Supercomputer	DE	1.07× 7% normalized	448 GPUs 0.3 MW
Immunity Bio Cluster	US	1.07× 7% normalized	320 GPUs 0.2 MW
Nagoya University Flow Type II (Furo)	JP	1.07× 7% normalized	884 GPUs 0.6 MW
NCI Australia Gadi	AU	1.07× 7% normalized	640 GPUs 0.4 MW
Argonne NL Sophia (Formerly Theta)	US	1.07× 7% normalized	192 GPUs 0.2 MW
Azure OpenAI GPT-3 Cluster	US	1.07× 7% normalized	10,000 GPUs 7 MW
Japan Atomic Energy Agency and Quantum and Radiological Science and Technology HPE SGI8600	JP	1.07× 7% normalized	1,088 GPUs 0.7 MW
Leonardo SpA davinci-1	IT	1.07× 7% normalized	320 GPUs 0.3 MW
Microsoft Azure Immunity Bio	US	1.07× 7% normalized	1,250 GPUs 0.8 MW
MIT Supercloud	US	1.07× 7% normalized	488 GPUs 0.3 MW
Naver DGX Superpod	KR	1.07× 7% normalized	1,120 GPUs 1 MW
Oracle 2020 A100 Cluster	US	1.07× 7% normalized	512 GPUs 0.4 MW
NVIDIA In-house DGX A100 Cluster	US	1.07× 7% normalized	400 GPUs 0.3 MW
RPI Supercomputer 2	US	1.07× 7% normalized	608 GPUs 0.4 MW
Lawrence Livermore NL Lassen Phase 2	US	1.07× 7% normalized	3,169 GPUs 2 MW
Petrobras Dragão	BR	1.08× 8% normalized	2,176 GPUs 2 MW
Max-Planck-Gesellschaft Raven	DE	1.09× 8% normalized	768 GPUs 0.7 MW
Colossus 2	US	1.09× 8% normalized	530,000 GPUs 946 MW
Calcul Québec Narval	CA	1.09× 9% normalized	636 GPUs 0.5 MW
EuroHPC MeluXina	LU	1.09× 9% normalized	800 GPUs 0.7 MW
EuroHPC Vega	SI	1.09× 9% normalized	240 GPUs 0.2 MW
ND A100 v4	US	1.09× 9% normalized	2,000 GPUs 2 MW
Argonne NL Polaris	US	1.09× 9% normalized	2,240 GPUs 2 MW
Commissariat a l'Energie Atomique Topaze	FR	1.09× 9% normalized	192 GPUs 0.2 MW
Dell Rattler	US	1.09× 9% normalized	192 GPUs 0.2 MW
High-Performance Computing Center Stuttgart Hawk	DE	1.09× 9% normalized	192 GPUs 0.2 MW
Indiana University Big Red 200	US	1.09× 9% normalized	256 GPUs 0.2 MW
Los Alamos NL Chicoma	US	1.09× 9% normalized	512 GPUs 0.4 MW
Osaka University SQUID	JP	1.09× 9% normalized	336 GPUs 0.3 MW
Samsung SSC-21	KR	1.09× 9% normalized	1,760 GPUs 2 MW
U.S. Army Jean	US	1.09× 9% normalized	280 GPUs 0.2 MW
Yandex Lyapunov	RU	1.09× 9% normalized	1,096 GPUs 0.9 MW
MTS Grom	RU	1.09× 9% normalized	160 GPUs 0.1 MW
NVIDIA Cambridge-1	GB	1.09× 9% normalized	640 GPUs 0.5 MW
Recursion BioHive-1	US	1.09× 9% normalized	320 GPUs 0.3 MW
Saudi Aramco Dammam-7	SA	1.09× 9% normalized	7,912 GPUs 5 MW
SberCloud Christofari Neo	RU	1.09× 9% normalized	792 GPUs 0.7 MW
SiDi IARA	BR	1.09× 9% normalized	200 GPUs 0.2 MW
Tesla Auto-Labeling Cluster	US	1.09× 9% normalized	1,752 GPUs 2 MW
Tesla Training Cluster	US	1.09× 9% normalized	4,032 GPUs 3 MW
Texas A&M Grace	US	1.09× 9% normalized	200 GPUs 0.2 MW
Vingroup VinAI Research Superpod	VN	1.09× 9% normalized	160 GPUs 0.1 MW
Yandex Chervonenkis	RU	1.09× 9% normalized	1,592 GPUs 1 MW
Yandex Galushkin	RU	1.09× 9% normalized	1,088 GPUs 0.9 MW
Microsoft Azure Pioneer-EUS	US	1.09× 9% normalized	1,312 GPUs 1 MW
Microsoft Azure Pioneer-SCUS	US	1.09× 9% normalized	1,312 GPUs 1 MW
Microsoft Azure Pioneer-WEU	NL	1.09× 9% normalized	1,312 GPUs 1 MW
Microsoft Azure Pioneer-WUS2	US	1.09× 9% normalized	1,312 GPUs 1 MW
Microsoft Azure Voyager-EUS2	US	1.09× 9% normalized	2,112 GPUs 2 MW
University of Minnesota Agate	US	1.09× 9% normalized	264 GPUs 0.2 MW
Karlsruher Institut für Technologie HoreKa	DE	1.10× 9% normalized	740 GPUs 0.6 MW
NSCC ASPIRE 2A Phase 2	SG	1.10× 10% normalized	352 GPUs 0.3 MW
AGH Cyfronet Athena	PL	1.10× 10% normalized	384 GPUs 0.3 MW
Ezra-1 Stability AI AWS Cluster	—	1.10× 10% normalized	4,000 GPUs 3 MW
IBM Vela	US	1.10× 10% normalized	2,880 GPUs 2 MW
Indiana University Bloomington Jetstream2	US	1.10× 10% normalized	360 GPUs 0.3 MW
Microsoft GPT-4 cluster	US	1.10× 10% normalized	25,000 GPUs 21 MW
TACC Lonestar6	US	1.10× 10% normalized	252 GPUs 0.2 MW
Aleph Alpha alpha ONE	DE	1.10× 10% normalized	512 GPUs 0.4 MW
Tesla A100 Cluster Phase 2	US	1.10× 10% normalized	7,360 GPUs 6 MW
Microsoft Azure Meta AI Rental	US	1.10× 10% normalized	5,400 GPUs 5 MW
GENCI Adastra	FR	1.10× 10% normalized	1,352 GPUs 2 MW
G42 Microsoft 100 MW UAE Cluster	AE	1.11× 10% normalized	55,000 GPUs 100 MW
Calcul Québec Béluga	CA	1.11× 10% normalized	688 GPUs 0.4 MW
Nebius ISEG	FI	1.11× 11% normalized	1,520 GPUs 2 MW
JUWELS-Booster	DE	1.11× 11% normalized	3,744 GPUs 3 MW
ExxonMobil Discovery 5	US	1.12× 11% normalized	2,000 GPUs 2 MW
CoreWeave Muskogee	US	1.13× 12% normalized	40,000 GPUs 100 MW
NVIDIA Selene Phase 2	US	1.13× 12% normalized	4,480 GPUs 5 MW
TU Dresden AlphaCentauri	DE	1.13× 12% normalized	272 GPUs 0.3 MW
DeepL Mercury	SE	1.13× 12% normalized	544 GPUs 0.8 MW
Microsoft Azure Eagle	US	1.13× 12% normalized	14,400 GPUs 21 MW
AWS EC2 P5 UltraClusters	US	1.13× 12% normalized	20,000 GPUs 29 MW
Google TPUv5e	—	1.13× 12% normalized	50,944 GPUs 24 MW
Imbue 10k Cluster	US	1.13× 12% normalized	10,000 GPUs 15 MW
Lambda Labs H100/H200	US	1.13× 12% normalized	32,000 GPUs 46 MW
Microsoft Azure ND H100 v5 VM	US	1.13× 12% normalized	4,000 GPUs 6 MW
Mistral Large Model Implied Cluster	—	1.13× 12% normalized	4,000 GPUs 6 MW
NHN Cloud's National AI Data Center	KR	1.13× 12% normalized	1,000 GPUs 1 MW
NVIDIA Eos Phase 2	US	1.13× 12% normalized	4,608 GPUs 7 MW
Oracle OCI Supercluster A100s	US	1.13× 12% normalized	32,768 GPUs 27 MW
Scaleway Nabuchodonosor	FR	1.13× 12% normalized	1,016 GPUs 1 MW
Softbank SuperPOD	JP	1.13× 12% normalized	2,048 GPUs 2 MW
Tesla 10k H100 Cluster	US	1.13× 12% normalized	10,000 GPUs 15 MW
Together AI H100 Cluster	US	1.13× 12% normalized	4,424 GPUs 6 MW
Amazon Titan training cluster	US	1.13× 12% normalized	13,760 GPUs 11 MW
Google Hypercomputer TPU v5p pod	—	1.13× 12% normalized	8,960 GPUs 10 MW
MosaicML MPT training cluster	—	1.13× 12% normalized	440 GPUs 0.4 MW
MPT-30B training cluster	—	1.13× 12% normalized	256 GPUs 0.4 MW
NAVER Corp Sejong	KR	1.13× 12% normalized	2,240 GPUs 2 MW
NEC Corp Japan Supercomputer	JP	1.13× 12% normalized	928 GPUs 0.8 MW
NVIDIA Helios	—	1.13× 12% normalized	1,024 GPUs 1 MW
SK Telecom Titan Phase 2	KR	1.13× 12% normalized	1,040 GPUs 0.9 MW
AIRAWAT-PSAI Phase 2	IN	1.13× 12% normalized	656 GPUs 0.5 MW
hessian.AI fortytwo	DE	1.13× 12% normalized	632 GPUs 0.5 MW
KT SuperPOD	KR	1.13× 12% normalized	792 GPUs 0.7 MW
Meta Research SuperCluster (RSC-1) Phase 2	US	1.13× 12% normalized	16,000 GPUs 13 MW
Meta Research SuperCluster 2 (RSC-2)	US	1.13× 12% normalized	8,000 GPUs 7 MW
NSC Berzelius Phase 2	SE	1.13× 12% normalized	752 GPUs 0.6 MW
KAUST Shaheen-III	SA	1.15× 14% normalized	2,816 GPUs 5 MW
RPI AiMOS	US	1.15× 14% normalized	1,512 GPUs 0.9 MW
Meta GenAI 2024a	US	1.15× 14% normalized	24,576 GPUs 35 MW
Meta GenAI 2024b	US	1.15× 14% normalized	24,576 GPUs 35 MW
Saudi Aramco Tuwaiq-1	SA	1.15× 14% normalized	416 GPUs 0.6 MW
Wroclaw Centre for Networking and Supercomputing Lem	PL	1.15× 14% normalized	304 GPUs 0.4 MW
Core42 SuperPOD	AE	1.15× 14% normalized	1,272 GPUs 2 MW
CSIRO Virga	AU	1.15× 14% normalized	448 GPUs 0.6 MW
Denvr Dataworks H100	US	1.15× 14% normalized	1,024 GPUs 1 MW
GreenNode Bangkok Cluster	TH	1.15× 14% normalized	1,024 GPUs 1 MW
IBM Blue Vela	US	1.15× 14% normalized	1,024 GPUs 1 MW
LeptonAI H100 Cluster	US	1.15× 14% normalized	1,024 GPUs 1 MW
Meta 100k	US	1.15× 14% normalized	100,000 GPUs 143 MW
Neevcloud cluster 1	IN	1.15× 14% normalized	1,024 GPUs 1 MW
NexGen Cloud Hyperstack AQ Compute Supercomputer	NO	1.15× 14% normalized	16,384 GPUs 23 MW
Northern Data Group Njored Taiga Cloud Island 5	GB	1.15× 14% normalized	2,048 GPUs 3 MW
Northern Data Group Taiga Cloud Island 3	—	1.15× 14% normalized	2,048 GPUs 3 MW
Northern Data Group Taiga Cloud Island 4	—	1.15× 14% normalized	2,048 GPUs 3 MW
Northern Data Group Taiga Cloud NO1 Island 1	NO	1.15× 14% normalized	2,048 GPUs 3 MW
Northern Data Group Taiga Cloud NO1 Island 2	NO	1.15× 14% normalized	2,048 GPUs 3 MW
NVIDIA Israel-1 Phase 2	IL	1.15× 14% normalized	2,048 GPUs 3 MW
OpenAI/Microsoft Goodyear Arizona	US	1.15× 14% normalized	100,000 GPUs 143 MW
Oracle OCI MI300x	US	1.15× 14% normalized	16,384 GPUs 25 MW
Oracle OCI Supercluster H100s	US	1.15× 14% normalized	16,384 GPUs 23 MW
Oracle OCI Supercluster H200s	US	1.15× 14% normalized	65,536 GPUs 94 MW
Ori Global Cloud H100 Cluster	—	1.15× 14% normalized	1,024 GPUs 1 MW
PCSS Poznan Proxima	PL	1.15× 14% normalized	348 GPUs 0.5 MW
Sesterce H100s Phase 2	FR	1.15× 14% normalized	4,096 GPUs 6 MW
Sesterce Nordics	—	1.15× 14% normalized	8,192 GPUs 12 MW
SIAM AI HGX	TH	1.15× 14% normalized	1,024 GPUs 1 MW
SoftBank CHIE-2	JP	1.15× 14% normalized	2,040 GPUs 3 MW
SoftBank CHIE-3	JP	1.15× 14% normalized	2,040 GPUs 3 MW
Tesla Earnings Call Claim	US	1.15× 14% normalized	35,000 GPUs 50 MW
Ahrefs Yep1	US	1.15× 14% normalized	504 GPUs 0.7 MW
Andreessen Horowitz Oxygen	US	1.15× 14% normalized	20,000 GPUs 29 MW
BNY Mellon Supercomputer	—	1.15× 14% normalized	500 GPUs 0.7 MW
CoreWeave H200s	US	1.15× 14% normalized	42,000 GPUs 60 MW
Fastweb NeXXt AI Factory	IT	1.15× 14% normalized	248 GPUs 0.4 MW
Hut 8 H100 Cluster	US	1.15× 14% normalized	1,000 GPUs 1 MW
Iris Energy Prince George cluster	CA	1.15× 14% normalized	816 GPUs 1 MW
Opera Iceland KEF-1 SuperPOD	IS	1.15× 14% normalized	248 GPUs 0.4 MW
Recursion BioHive-2	US	1.15× 14% normalized	504 GPUs 0.7 MW
Vultr Chicago Cluster	US	1.15× 14% normalized	3,000 GPUs 5 MW
TSUBAME4.0	JP	1.16× 15% normalized	960 GPUs 1 MW
Gcore data center Phase 2	KR	1.16× 15% normalized	1,000 GPUs 2 MW
University of Cambridge Wilkes-3	GB	1.16× 15% normalized	320 GPUs 0.3 MW
Inflection AI Cluster	US	1.17× 15% normalized	22,000 GPUs 31 MW
Novo Nordisk Gefion	DK	1.17× 15% normalized	1,528 GPUs 3 MW
Project Ceiba Phase 2	US	1.17× 16% normalized	20,736 GPUs 50 MW
ExxonMobil Discovery 6	US	1.17× 16% normalized	4,032 GPUs 6 MW
Google A3 VMs	US	1.17× 16% normalized	26,000 GPUs 36 MW
Nebius Finland Phase 2	FI	1.17× 16% normalized	60,000 GPUs 84 MW
Saudi Data & AI Authority Sovereign AI factory	SA	1.17× 16% normalized	5,000 GPUs 12 MW
Sesterce Valence	FR	1.17× 16% normalized	40,000 GPUs 96 MW
together.ai 36k GB200s	US	1.17× 16% normalized	36,000 GPUs 86 MW
AIST ABCI 3.0	JP	1.17× 16% normalized	6,128 GPUs 9 MW
AIST ABCI-Q	JP	1.17× 16% normalized	2,048 GPUs 3 MW
CEA EXA1-HE Phase 3	FR	1.17× 16% normalized	2,688 GPUs 4 MW
Foxconn Hon Hai Kaohsiung Supercomputer	TW	1.17× 16% normalized	4,608 GPUs 11 MW
Horizon Compute Baobab Phase 2	US	1.17× 16% normalized	2,048 GPUs 3 MW
iGenius Colosseum	IT	1.17× 16% normalized	5,760 GPUs 14 MW
JCAHPC Miyabi	JP	1.17× 16% normalized	1,120 GPUs 2 MW
KDDI Sharp Sakai	JP	1.17× 16% normalized	2,000 GPUs 5 MW
MITRE Federal AI Sandbox	US	1.17× 16% normalized	256 GPUs 0.4 MW
PanaAI AUS AISF	AU	1.17× 16% normalized	4,088 GPUs 6 MW
S. Korea 6th national supercomputer	KR	1.17× 16% normalized	8,800 GPUs 12 MW
Sesterce Pegasus	—	1.17× 16% normalized	10,496 GPUs 25 MW
Sesterce Synapse Phase 2	FR	1.17× 16% normalized	4,096 GPUs 6 MW
TensorWave MI300X Cluster 1 Phase 2	US	1.17× 16% normalized	10,000 GPUs 15 MW
TensorWave MI300X Cluster 2	US	1.17× 16% normalized	10,000 GPUs 15 MW
Voltage Park Location 5	US	1.17× 16% normalized	4,088 GPUs 6 MW
Voltage Park Location 6	US	1.17× 16% normalized	4,088 GPUs 6 MW
Voltage Park Texas Phase 2	US	1.17× 16% normalized	4,088 GPUs 6 MW
Voltage Park Utah	US	1.17× 16% normalized	4,088 GPUs 6 MW
Voltage Park Virginia	US	1.17× 16% normalized	4,088 GPUs 6 MW
Voltage Park Washington	US	1.17× 16% normalized	4,088 GPUs 6 MW
Yotta Shakti Cloud D1	IN	1.17× 16% normalized	16,384 GPUs 23 MW
Yotta Shakti Cloud NM1 Phase 2	IN	1.17× 16% normalized	16,384 GPUs 23 MW
YTL AI Johor	MY	1.17× 16% normalized	15,428 GPUs 37 MW
CoreWeave LiquidLab	US	1.17× 16% normalized	4,000 GPUs 6 MW
Corvex B200s	US	1.17× 16% normalized	256 GPUs 0.5 MW
FPT AI Factory Japan	JP	1.17× 16% normalized	3,000 GPUs 4 MW
FPT AI Factory Vietnam	VN	1.17× 16% normalized	3,000 GPUs 4 MW
Huawei Pangu Ultra MoE 910Bs	—	1.17× 16% normalized	6,000 GPUs 5 MW
Magic G4 Google Cloud Rental	US	1.17× 16% normalized	8,000 GPUs 11 MW
Nebius ISEG2	IS	1.17× 16% normalized	4,992 GPUs 7 MW
NVIDIA CoreWeave Eos-DFW Rumored Phase 2	US	1.17× 16% normalized	32,000 GPUs 45 MW
OneAsia OBON Clusters	TH	1.17× 16% normalized	4,000 GPUs 6 MW
Oracle OCI Supercluster B200s	US	1.17× 16% normalized	131,072 GPUs 262 MW
Poolside 10k Cluster	—	1.17× 16% normalized	10,000 GPUs 14 MW
Quebec 2k H100 Cluster	CA	1.17× 16% normalized	2,000 GPUs 3 MW
Samsung SSC4	KR	1.17× 16% normalized	2,496 GPUs 3 MW
SoftBank Planned B200 Superpod	JP	1.17× 16% normalized	4,000 GPUs 8 MW
Sustainable Metal Cloud Singapore Phase 2	SG	1.17× 16% normalized	5,000 GPUs 7 MW
Sweden 4k H100 Cluster	SE	1.17× 16% normalized	4,000 GPUs 6 MW
University of Bristol Isambard-AI	GB	1.17× 16% normalized	5,448 GPUs 8 MW
TACC Frontera	US	1.18× 16% normalized	808 GPUs 0.5 MW
Princeton Della Phase 2	US	1.18× 17% normalized	612 GPUs 0.7 MW
NVIDIA Tethys	US	1.19× 18% normalized	160 GPUs 0.1 MW
Pawsey Supercomputing Centre Setonix	AU	1.21× 19% normalized	768 GPUs 0.8 MW
G42 Microsoft 30 MW UAE Cluster A	AE	1.22× 20% normalized	15,000 GPUs 30 MW
G42 Microsoft 30 MW UAE Cluster B	AE	1.22× 20% normalized	15,000 GPUs 30 MW
Telangana Yotta Hyderbad AI City Cluster Phase 2	IN	1.22× 20% normalized	25,000 GPUs 50 MW
NSTDA Supercomputer Center (ThaiSC) LANTA	TH	1.22× 20% normalized	704 GPUs 0.5 MW
Mare Nostrum 5	ES	1.23× 20% normalized	4,480 GPUs 6 MW
Microsoft Goodyear	US	1.23× 21% normalized	100,000 GPUs 202 MW
KT Internal MI250 Cluster	KR	1.23× 21% normalized	1,200 GPUs 1 MW
Oak Ridge NL Frontier	US	1.24× 21% normalized	37,632 GPUs 40 MW
University of Illinois NCSA Delta	US	1.24× 22% normalized	880 GPUs 0.7 MW
NFDG Andromeda Phase 2	US	1.24× 22% normalized	4,400 GPUs 6 MW
Foxconn Big Innovation Cloud AI factory	TW	1.26× 23% normalized	10,000 GPUs 26 MW
Microsoft Explorer-WUS3	US	1.27× 24% normalized	1,920 GPUs 2 MW
Reliance Industries Supercomputer	IN	1.27× 24% normalized	450,000 GPUs 1,000 MW
Applied Digital Ellendale Possible Phase 3	US	1.27× 24% normalized	180,000 GPUs 400 MW
NVIDIA DGX SuperPOD 2019	—	1.27× 24% normalized	1,536 GPUs 0.9 MW
Eni HPC5	IT	1.27× 24% normalized	7,280 GPUs 4 MW
Petrobras Atlas	BR	1.28× 24% normalized	1,088 GPUs 1 MW
Eni HPC6	IT	1.29× 25% normalized	13,888 GPUs 14 MW
XTX Markets Cluster	—	1.29× 26% normalized	20,000 GPUs 15 MW
EuroHPC LUMI	FI	1.29× 26% normalized	11,912 GPUs 12 MW
OpenAI/Microsoft Mt Pleasant, Wisconsin Phase 2	US	1.31× 27% normalized	700,000 GPUs 1,500 MW
Lawrence Livermore NL Tuolumne	US	1.34× 29% normalized	4,608 GPUs 6 MW
Samsung SSC-21 Scalable Module	KR	1.34× 29% normalized	144 GPUs 0.2 MW
G42 Artemis	AE	1.34× 29% normalized	1,296 GPUs 1 MW
Intel Stability Gaudi 2	US	1.34× 29% normalized	4,000 GPUs 5 MW
Sandia NL El Dorado	US	1.35× 29% normalized	1,520 GPUs 2 MW
Sesterce Southern France 250MW	FR	1.35× 30% normalized	120,000 GPUs 250 MW
Petrobras Fênix Phase 2	BR	1.38× 32% normalized	720 GPUs 0.7 MW
Nscale Loughton	GB	1.39× 32% normalized	23,040 GPUs 90 MW
Tesla Cortex Phase 3	US	1.41× 34% normalized	120,000 GPUs 140 MW
Fluidstack France Gigawatt Campus	FR	1.41× 34% normalized	500,000 GPUs 1,000 MW
Nebius New Jersey	US	1.41× 34% normalized	150,000 GPUs 300 MW
OpenAI/Microsoft Mt Pleasant, Wisconsin Phase 1	US	1.41× 34% normalized	150,000 GPUs 300 MW
Sesterce Grand Est France A	FR	1.41× 34% normalized	150,000 GPUs 300 MW
Sesterce Grand Est France B	FR	1.41× 34% normalized	150,000 GPUs 300 MW
NVIDIA Taipei-1	TW	1.42× 35% normalized	768 GPUs 0.9 MW
SURF Snellius Phase 3	NL	1.43× 35% normalized	640 GPUs 0.7 MW
AGH Cyfronet Helios	PL	1.44× 36% normalized	464 GPUs 0.5 MW
Center for Advanced Intelligence Project, RIKEN, RAIDEN	JP	1.46× 37% normalized	416 GPUs 0.2 MW
Petrobras Gaia	BR	1.48× 39% normalized	704 GPUs 1 MW
Alps Supercomputer Phase 2	CH	1.48× 39% normalized	10,752 GPUs 12 MW
Oak Ridge NL Summit	US	1.50× 40% normalized	27,648 GPUs 13 MW
Los Alamos NL Venado	US	1.51× 41% normalized	2,560 GPUs 3 MW
CEA EXA1-HE Phase 2	FR	1.55× 43% normalized	1,908 GPUs 2 MW
Gemini 1.0 Ultra training cluster A	US	1.57× 44% normalized	28,672 GPUs 20 MW
Microsoft Ares/Maia	US	1.58× 45% normalized	2,048 GPUs 2 MW
Google Oklahoma TPU v4 Pods	US	1.61× 47% normalized	32,768 GPUs 24 MW
Paper on PaLM	US	1.61× 47% normalized	6,144 GPUs 4 MW
Google TPU v4 Pod	US	1.62× 47% normalized	4,096 GPUs 3 MW
SK Group AWS Uslan Phase 2	KR	1.64× 49% normalized	60,000 GPUs 103 MW
Google TPUv3 POD Generic	—	1.64× 49% normalized	1,024 GPUs 1 MW
Google MLPerf 0.7 Submission	US	1.67× 50% normalized	4,096 GPUs 4 MW
Argonne NL Aurora	US	1.75× 54% normalized	63,744 GPUs 60 MW
University of Oxford Jade2	GB	1.75× 55% normalized	504 GPUs 0.2 MW
LUMI Supercomputer (CSC Finland)	FI	1.90× 62% normalized	10,240 GPUs 7 MW
Eclairion Bruyères-le-Châtel (Mistral Compute)	FR	1.94× 64% normalized	13,800 GPUs 44 MW
Oak Ridge NL Titan	US	1.94× 64% normalized	18,688 GPUs 16 MW
Tesla Dojo 1 Phase 1	US	2.05× 69% normalized	3,000 GPUs 2 MW
Lawrence Livermore NL El Capitan Phase 2	US	2.09× 71% normalized	44,544 GPUs 35 MW
CSCS Piz Daint Phase 2	CH	2.14× 73% normalized	5,704 GPUs 4 MW
Jupiter, Jülich	DE	2.15× 73% normalized	23,536 GPUs 18 MW
ORNL Frontier (Exascale HPC)	US	2.19× 75% normalized	37,888 GPUs 23 MW
GSIC TSUBAME 3.0	JP	2.32× 80% normalized	2,156 GPUs 2 MW
US Government Supercomputer 1	US	2.37× 81% normalized	4,160 GPUs 3 MW
US Government Supercomputer 2	US	2.37× 81% normalized	4,160 GPUs 3 MW
Meta 2017 P100 Cluster	US	2.42× 83% normalized	992 GPUs 0.7 MW
Google TensorFlow Research Cloud	US	2.50× 86% normalized	4,096 GPUs 3 MW
Jean Zay Supercomputer Phase 4	FR	2.52× 86% normalized	3,704 GPUs 2 MW
GSIC TSUBAME 2.5	JP	2.58× 88% normalized	4,224 GPUs 3 MW
Preferred Networks MN-1b	JP	2.69× 91% normalized	1,536 GPUs 0.9 MW
Moscow State University Lomonosov 2	RU	2.81× 95% normalized	1,472 GPUs 0.9 MW
Eni HPC2	IT	3.49× 111% normalized	3,000 GPUs 1 MW
Aramco Groq Inference Cluster	SA	3.82× 117% normalized	19,725 GPUs 8 MW
MIT LLSC TX-GAIA	US	4.08× 121% normalized	896 GPUs 3 MW
Universitaet Frankfurt Goethe-NHR	DE	4.35× 125% normalized	880 GPUs 0.3 MW
JAMSTEC ZettaScaler-2.2 Gyoukou	JP	6.32× 145% normalized	10,000 GPUs 3 MW
Paper on AlphaZero	US	9.33× 161% normalized	5,000 GPUs 0.9 MW

What this ranking means

Tight bounds (ratio ≤ 1.5×): two independent inputs reconcile under the methodology defaults. A facility whose inputs are both wrong in the same direction, both nameplate-only or both phase-1, reconciles just as tightly.
Loose bounds (ratio > 3×) usually mark a gap in the record: a power figure disclosed with no GPU model (the H100 SXM5 default then stands in for the deployed hardware), a nameplate capacity larger than the draw, or a phase-1 GPU count set against a full-build substation. The loose rows are the investigation queue.
The cost path is excluded here because the records hold no per-facility investment figure; organization-level capex is annualised and pooled across facilities. The ranking is chip count against power, and the estimator above runs all three paths on inputs a reader supplies.

285 of 314 reconciled sites resolve a specific chip from the per-facility hardware inventory; the remaining 29 run the H100 SXM5 default, because no recognisable model is recorded. Throughput cancels between the two paths and TDP does not, so a divergence computed on the fallback chip is conditional on that default; those 29 rows are the ones a chip model would move.

Model archaeology

The infrastructure behind seven frontier models

Every claim links to its source. Layers 01 to 03 are evidence about the model itself; 04 to 07 are industry context.

DeepSeek · December 2024

DeepSeek-V3

Keys 1–7 for model · ↑↓ for layers · Esc for training cluster

Training hardware: NVIDIA H800; Reported · Model run; Eight GPUs per node.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Run allocation: 2,048; Reported · Model run; GPUs in the reported training cluster.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Compute cost: ~$5.576M; Estimate · Model run; GPU rental estimate at $2/hour; official training only.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Pretraining duration: Under 2 months; Reported · Model run; Context extension and post-training are separate stages.; DeepSeek technical report§1, Table 1; §3.1; §4.2

02 / Training cluster

NVLink within nodes; InfiniBand between them

Node layout: 8 H800 GPUs; Reported · Model run; NVLink and NVSwitch connect the GPUs within each node.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Official training: 2.788M H800 GPU-hours; Reported · Model run; 2.664M pretraining; 119K context extension; 5K post-training.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Elapsed time: Pretraining under 2 months; Reported · Model run; The report gives this bound and no exact full-training duration.; DeepSeek technical report§1, Table 1; §3.1; §4.2
Run power: —; Not established · Model run; No measured whole-facility power figure is established by this report.; DeepSeek technical report§1, Table 1; §3.1; §4.2

Continue

Thresholds runs the same estimate against the EU AI Act’s 10²⁵ FLOP trigger and its counterparts elsewhere: the days of sustained operation each site needs to accumulate that total, and how many clear it inside a month.

Capacity

Estimate daily compute

Chip-count estimates of daily compute exceed power-based estimates at 279 of 314 sites

Published hardware and compute for frontier training runs

Shared assumptions

Compare estimates by facility

The infrastructure behind seven frontier models

DeepSeek-V3

671B parameters; 37B active per token

NVLink within nodes; InfiniBand between them

The physical training site is not identified here

Packaging connects compute and memory

Foundries manufacture the accelerator dies

Manufacturing equipment sits upstream of the chip

Wafers, chemicals and gases enter during manufacturing

DeepSeek-V3

671B parameters; 37B active per token

NVLink within nodes; InfiniBand between them

The physical training site is not identified here

Packaging connects compute and memory

Foundries manufacture the accelerator dies

Manufacturing equipment sits upstream of the chip

Wafers, chemicals and gases enter during manufacturing