saef 3.0.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- saef/__init__.py +27 -0
- saef/__main__.py +3 -0
- saef/_calc.py +350 -0
- saef/_core.py +165 -0
- saef/cli.py +73 -0
- saef/data/cloud_region_factors.csv +144 -0
- saef/data/factors_v3_lifecycle.csv +103 -0
- saef/data/grid_intensity_all.csv +241 -0
- saef/data/model_class_mapping.csv +188 -0
- saef/data/model_factors.csv +33 -0
- saef/data/region_grid_crosswalk.csv +18 -0
- saef/data/region_water_factors.csv +18 -0
- saef/data/reserved_capacity_factors.csv +4 -0
- saef/data/token_basis_constants.csv +8 -0
- saef-3.0.0.dist-info/METADATA +115 -0
- saef-3.0.0.dist-info/RECORD +20 -0
- saef-3.0.0.dist-info/WHEEL +4 -0
- saef-3.0.0.dist-info/entry_points.txt +2 -0
- saef-3.0.0.dist-info/licenses/LICENSE +396 -0
- saef-3.0.0.dist-info/licenses/LICENSE-CODE +24 -0
|
@@ -0,0 +1,103 @@
|
|
|
1
|
+
model_class,class_value_basis,scenario,gpu,n_class_members,class_members,gpu_avg_power_w_total_over_gpus_mean_of_members,h_host_uplift,i_idle_uplift,pue,utilisation,gpu_j_per_output_token,gpu_j_per_output_token_lowest_member,gpu_j_per_output_token_highest_member,gpu_j_per_output_token_geometric_mean,gpu_j_per_output_token_median,gpu_j_per_output_token_named_proxy,gpu_wh_per_1k_effective_tokens,server_wh_per_1k_effective_tokens,facility_wh_per_1k_effective_tokens,region_id,region_name,grid_kg_co2e_per_kwh,grid_source,grid_unscaled_egrid_kg_per_kwh,electricity_kg_co2e_per_m_effective_tokens,embodied_kg_co2e_per_m_effective_tokens,construction_kg_co2e_per_m_effective_tokens,inference_footprint_kg_co2e_per_m_effective_tokens,inference_footprint_named_proxy_kg_co2e_per_m_effective_tokens,training_kg_co2e_per_m_billed_tokens,default_mix_effective_over_billed_tokens,factor_kg_co2e_per_m_billed_tokens_default_mix,four_term_kg_co2e_per_m_effective_equiv,development_line_kg_co2e_per_m_effective_tokens,development_inclusive_total_kg_co2e_per_m_effective_tokens,electricity_high_case_kg_co2e_per_m_effective_tokens,embodied_high_case_kg_co2e_per_m_effective_tokens,construction_high_case_kg_co2e_per_m_effective_tokens,inference_footprint_high_case_kg_co2e_per_m_effective_tokens,factor_high_case_kg_co2e_per_m_billed_tokens_default_mix,wue_l_per_kwh,wue_source,ewif_l_per_kwh,ewif_source,ewif_excl_hydro_l_per_kwh,water_wue_basis_as_published,water_onsite_consumption_basis_ml_per_1k_effective_tokens,water_onsite_as_published_ml_per_1k_effective_tokens,water_offsite_ml_per_1k_effective_tokens,water_offsite_excl_hydro_ml_per_1k_effective_tokens,water_total_ml_per_1k_effective_tokens,water_total_excl_hydro_ml_per_1k_effective_tokens,aware_factor,water_onsite_scarcity_weighted_ml_world_eq_per_1k_effective_tokens,water_onsite_share_of_total_pct,water_onsite_share_of_total_excl_hydro_pct,water_total_onsite_as_published_ml_per_1k_effective_tokens,water_total_excl_hydro_onsite_as_published_ml_per_1k_effective_tokens,ewif_high_hydro_l_per_kwh,water_total_high_hydro_ml_per_1k_effective_tokens,water_high_hydro_ratio_to_excl_hydro,water_confidence,water_wue_denominator,water_wue_provider,water_wue_site,water_wue_year,water_v3final_wue_l_per_kwh,water_v3final_total_ml_per_1k_tokens,water_v3final_total_excl_hydro_ml_per_1k_tokens,energy_source,lifecycle_source
|
|
2
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.00277,0.00089,0.00046,0.00411,0.00216,0.00060,0.3025,0.00185,0.00472,0.00377,0.00788,0.00457,0.00111,0.00063,0.00631,0.00251,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0008,0.0012,0.4332,0.0997,0.4340,0.1005,2.17,0.0017,0.18,0.80,0.4344,0.1009,103.306,7.4360,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,0.1426,0.0342,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
3
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.00383,0.00089,0.00046,0.00518,0.00273,0.00060,0.3025,0.00217,0.00578,0.00523,0.01041,0.00633,0.00111,0.00063,0.00807,0.00304,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0020,0.0030,0.2642,0.1483,0.2662,0.1503,4.99,0.0100,0.75,1.33,0.2672,0.1513,32.196,2.3192,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.0871,0.0494,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
4
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.00585,0.00089,0.00046,0.00720,0.00381,0.00060,0.3025,0.00278,0.00780,0.00798,0.01518,0.00967,0.00111,0.00063,0.01141,0.00405,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0006,0.0006,0.3269,0.1232,0.3275,0.1238,2.17,0.0013,0.18,0.48,0.3275,0.1238,50.222,3.6152,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,0.1065,0.0402,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
5
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.00969,0.00089,0.00046,0.01104,0.00587,0.00060,0.3025,0.00394,0.01164,0.01322,0.02426,0.01601,0.00111,0.00063,0.01775,0.00597,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0020,0.0030,0.1880,0.1580,0.1900,0.1600,3.52,0.0070,1.05,1.25,0.1910,0.1610,7.340,0.5303,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.0623,0.0525,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
6
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.01699,0.00089,0.00046,0.01833,0.00979,0.00060,0.3025,0.00615,0.01894,0.02316,0.04149,0.02806,0.00111,0.00063,0.02980,0.00962,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0020,0.0030,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
7
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.01696,0.00089,0.00046,0.01830,0.00977,0.00060,0.3025,0.00614,0.01891,0.02312,0.04143,0.02801,0.00111,0.00063,0.02975,0.00960,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0056,0.0084,0.1689,0.1332,0.1745,0.1388,2.29,0.0128,3.21,4.03,0.1773,0.1416,8.220,0.5972,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,0.0559,0.0443,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
8
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.01930,0.00089,0.00046,0.02065,0.01103,0.00060,0.3025,0.00685,0.02125,0.02632,0.04697,0.03189,0.00111,0.00063,0.03362,0.01077,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0017,0.0026,0.2479,0.2479,0.2497,0.2497,1.89,0.0032,0.68,0.68,0.2505,0.2505,3.690,0.2673,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,0.0818,0.0818,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
9
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.02033,0.00089,0.00046,0.02168,0.01158,0.00060,0.3025,0.00716,0.02228,0.02773,0.04940,0.03359,0.00111,0.00063,0.03533,0.01129,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0008,0.0012,0.1062,0.0815,0.1070,0.0823,1.17,0.0009,0.75,0.97,0.1074,0.0827,10.442,0.7524,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.0349,0.0269,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
10
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.02617,0.00089,0.00046,0.02752,0.01471,0.00060,0.3025,0.00893,0.02812,0.03568,0.06320,0.04323,0.00111,0.00063,0.04497,0.01421,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0068,0.0102,0.1389,0.1210,0.1457,0.1278,2.09,0.0142,4.67,5.32,0.1491,0.1311,11.726,0.8507,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.0467,0.0409,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
11
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.02187,0.00089,0.00046,0.02321,0.01240,0.00060,0.3025,0.00762,0.02382,0.02982,0.05303,0.03612,0.00111,0.00063,0.03786,0.01206,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0024,0.0036,0.1717,0.1426,0.1741,0.1450,1.62,0.0039,1.38,1.65,0.1753,0.1462,5.656,0.4095,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,0.0585,0.0491,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
12
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.02330,0.00089,0.00046,0.02464,0.01317,0.00060,0.3025,0.00806,0.02525,0.03177,0.05641,0.03849,0.00111,0.00063,0.04022,0.01277,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0048,0.0072,0.6837,0.0416,0.6885,0.0464,3.91,0.0188,0.70,10.33,0.6909,0.0488,116.219,8.3694,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,0.2408,0.0321,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
13
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.02701,0.00089,0.00046,0.02836,0.01516,0.00060,0.3025,0.00918,0.02896,0.03683,0.06519,0.04462,0.00111,0.00063,0.04636,0.01463,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.0140,0.0210,0.0926,0.0873,0.1066,0.1013,13.56,0.1897,13.12,13.81,0.1136,0.1083,1.948,0.1542,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.0350,0.0333,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
14
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.03102,0.00089,0.00046,0.03237,0.01731,0.00060,0.3025,0.01039,0.03297,0.04230,0.07467,0.05125,0.00111,0.00063,0.05299,0.01663,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0059,0.0089,0.1662,0.1255,0.1721,0.1314,1.78,0.0105,3.44,4.50,0.1751,0.1343,6.643,0.4840,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,0.0552,0.0420,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
15
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.03398,0.00089,0.00046,0.03532,0.01890,0.00060,0.3025,0.01129,0.03593,0.04633,0.08166,0.05613,0.00111,0.00063,0.05787,0.01811,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0059,0.0088,0.2234,0.1209,0.2293,0.1268,3.41,0.0201,2.57,4.64,0.2322,0.1298,13.930,1.0085,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,0.0754,0.0420,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
16
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.03605,0.00089,0.00046,0.03740,0.02001,0.00060,0.3025,0.01191,0.03800,0.04916,0.08655,0.05956,0.00111,0.00063,0.06129,0.01914,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.0628,0.0942,,,,,0.867,0.0544,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
17
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.03545,0.00089,0.00046,0.03680,0.01969,0.00060,0.3025,0.01173,0.03740,0.04835,0.08515,0.05857,0.00111,0.00063,0.06031,0.01885,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.0488,0.0732,0.1662,0.1075,0.2150,0.1563,0.637,0.0311,22.69,31.22,0.2394,0.1807,22.074,1.6375,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,0.0579,0.0388,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
18
|
+
A,mean of the measured class members (spec 11.2),H100 (central),H100,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),696.8,1.5,1.1,1.2,0.75,0.1309,0.0782,0.1927,0.1249,0.1282,0.0702,0.03635,0.05998,0.07197,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.03478,0.00089,0.00046,0.03612,0.01933,0.00060,0.3025,0.01153,0.03673,0.04742,0.08355,0.05746,0.00111,0.00063,0.05919,0.01851,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.0330,0.0330,0.2261,0.1131,0.2591,0.1461,,,12.73,22.58,0.2591,0.1461,17.497,1.2923,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,0.0842,0.0475,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
19
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.00239,0.00089,0.00040,0.00368,0.00217,0.00060,0.3025,0.00172,0.00428,0.00327,0.00695,0.00396,0.00111,0.00054,0.00561,0.00230,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0007,0.0010,0.3752,0.0864,0.3759,0.0871,2.17,0.0015,0.18,0.80,0.3762,0.0874,103.306,6.4402,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,0.1432,0.0343,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
20
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.00332,0.00089,0.00040,0.00460,0.00274,0.00060,0.3025,0.00200,0.00521,0.00453,0.00913,0.00548,0.00111,0.00054,0.00714,0.00276,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0017,0.0026,0.2288,0.1285,0.2306,0.1302,4.99,0.0086,0.75,1.33,0.2314,0.1311,32.196,2.0086,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.0874,0.0496,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
21
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.00507,0.00089,0.00040,0.00635,0.00383,0.00060,0.3025,0.00253,0.00696,0.00691,0.01327,0.00838,0.00111,0.00054,0.01003,0.00364,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0005,0.0005,0.2831,0.1067,0.2836,0.1072,2.17,0.0011,0.18,0.48,0.2836,0.1072,50.222,3.1310,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,0.1069,0.0404,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
22
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.00839,0.00089,0.00040,0.00968,0.00590,0.00060,0.3025,0.00353,0.01028,0.01145,0.02113,0.01387,0.00111,0.00054,0.01552,0.00530,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0017,0.0026,0.1628,0.1368,0.1645,0.1385,3.52,0.0061,1.05,1.25,0.1654,0.1394,7.340,0.4593,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.0625,0.0527,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
23
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.01471,0.00089,0.00040,0.01600,0.00982,0.00060,0.3025,0.00544,0.01660,0.02006,0.03606,0.02430,0.00111,0.00054,0.02596,0.00846,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0017,0.0026,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
24
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.01469,0.00089,0.00040,0.01597,0.00981,0.00060,0.3025,0.00543,0.01657,0.02003,0.03600,0.02426,0.00111,0.00054,0.02592,0.00844,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0048,0.0073,0.1463,0.1153,0.1511,0.1202,2.29,0.0111,3.21,4.03,0.1536,0.1226,8.220,0.5172,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,0.0561,0.0444,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
25
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.01672,0.00089,0.00040,0.01800,0.01107,0.00060,0.3025,0.00605,0.01861,0.02279,0.04080,0.02762,0.00111,0.00054,0.02927,0.00946,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0015,0.0022,0.2147,0.2147,0.2162,0.2162,1.89,0.0028,0.68,0.68,0.2170,0.2170,3.690,0.2315,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,0.0821,0.0821,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
26
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.01761,0.00089,0.00040,0.01889,0.01163,0.00060,0.3025,0.00632,0.01950,0.02401,0.04291,0.02909,0.00111,0.00054,0.03075,0.00990,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0007,0.0010,0.0920,0.0706,0.0927,0.0713,1.17,0.0008,0.75,0.97,0.0930,0.0716,10.442,0.6516,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.0351,0.0270,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
27
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.02266,0.00089,0.00040,0.02395,0.01477,0.00060,0.3025,0.00785,0.02455,0.03091,0.05485,0.03744,0.00111,0.00054,0.03910,0.01243,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0059,0.0088,0.1203,0.1048,0.1262,0.1106,2.09,0.0123,4.67,5.32,0.1291,0.1136,11.726,0.7368,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.0469,0.0410,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
28
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.01894,0.00089,0.00040,0.02022,0.01245,0.00060,0.3025,0.00672,0.02083,0.02582,0.04605,0.03129,0.00111,0.00054,0.03294,0.01057,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0021,0.0031,0.1487,0.1235,0.1507,0.1255,1.62,0.0034,1.38,1.65,0.1518,0.1266,5.656,0.3546,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,0.0588,0.0493,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
29
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.02018,0.00089,0.00040,0.02146,0.01322,0.00060,0.3025,0.00710,0.02206,0.02751,0.04897,0.03333,0.00111,0.00054,0.03499,0.01119,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0042,0.0062,0.5922,0.0361,0.5963,0.0402,3.91,0.0162,0.70,10.33,0.5984,0.0423,116.219,7.2485,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,0.2418,0.0322,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
30
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.02339,0.00089,0.00040,0.02468,0.01522,0.00060,0.3025,0.00807,0.02528,0.03190,0.05658,0.03865,0.00111,0.00054,0.04030,0.01279,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.0121,0.0182,0.0802,0.0756,0.0923,0.0877,13.56,0.1643,13.12,13.81,0.0984,0.0938,1.948,0.1335,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.0351,0.0334,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
31
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.02687,0.00089,0.00040,0.02815,0.01738,0.00060,0.3025,0.00912,0.02876,0.03664,0.06479,0.04439,0.00111,0.00054,0.04604,0.01453,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0051,0.0077,0.1439,0.1087,0.1491,0.1138,1.78,0.0091,3.44,4.50,0.1516,0.1163,6.643,0.4192,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,0.0554,0.0421,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
32
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.02943,0.00089,0.00040,0.03071,0.01898,0.00060,0.3025,0.00989,0.03132,0.04013,0.07084,0.04862,0.00111,0.00054,0.05027,0.01581,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0051,0.0076,0.1935,0.1047,0.1986,0.1098,3.41,0.0174,2.57,4.64,0.2011,0.1124,13.930,0.8734,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,0.0757,0.0422,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
33
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.03122,0.00089,0.00040,0.03251,0.02009,0.00060,0.3025,0.01044,0.03311,0.04257,0.07508,0.05158,0.00111,0.00054,0.05323,0.01671,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.0544,0.0816,,,,,0.867,0.0471,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
34
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.03071,0.00089,0.00040,0.03199,0.01977,0.00060,0.3025,0.01028,0.03259,0.04187,0.07386,0.05073,0.00111,0.00054,0.05238,0.01645,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.0422,0.0634,0.1439,0.0931,0.1862,0.1353,0.637,0.0269,22.69,31.22,0.2073,0.1565,22.074,1.4182,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,0.0582,0.0390,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
35
|
+
A,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,7,Llama-3.1-8B-Instruct (bfloat16); Qwen3-8B (bfloat16); NVIDIA-Nemotron-Nano-9B-v2 (bfloat16); gemma-3-12b-it (bfloat16); NVIDIA-Nemotron-Nano-12B-v2 (bfloat16); Qwen3-14B (bfloat16); Qwen3-30B-A3B-Instruct-2507 (bfloat16),788.7,1.5,1.1,1.2,0.75,0.1133,0.0750,0.1690,0.1081,0.0990,0.0705,0.03148,0.05195,0.06233,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.03012,0.00089,0.00040,0.03140,0.01941,0.00060,0.3025,0.01010,0.03201,0.04107,0.07248,0.04976,0.00111,0.00054,0.05141,0.01615,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.0286,0.0286,0.1959,0.0980,0.2244,0.1265,,,12.73,22.58,0.2244,0.1265,17.497,1.1192,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,0.0846,0.0477,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
36
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.00906,0.00334,0.00150,0.01390,0.01522,0.00293,0.3025,0.00714,0.01684,0.01235,0.02625,0.01496,0.00418,0.00206,0.02120,0.00935,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0026,0.0039,1.4186,0.3266,1.4213,0.3292,2.17,0.0057,0.18,0.80,1.4226,0.3306,103.306,24.3511,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,0.9730,0.2332,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
37
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.01255,0.00334,0.00150,0.01740,0.01913,0.00293,0.3025,0.00820,0.02034,0.01712,0.03452,0.02074,0.00418,0.00206,0.02698,0.01109,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0065,0.0098,0.8652,0.4858,0.8718,0.4923,4.99,0.0327,0.75,1.33,0.8751,0.4956,32.196,7.5949,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.5941,0.3370,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
38
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.01917,0.00334,0.00150,0.02402,0.02652,0.00293,0.3025,0.01020,0.02695,0.02614,0.05016,0.03167,0.00418,0.00206,0.03791,0.01440,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0020,0.0020,1.0705,0.4034,1.0725,0.4054,2.17,0.0043,0.18,0.48,1.0725,0.4054,50.222,11.8389,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,0.7265,0.2746,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
39
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.03174,0.00334,0.00150,0.03659,0.04058,0.00293,0.3025,0.01400,0.03952,0.04328,0.07987,0.05244,0.00418,0.00206,0.05868,0.02068,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0065,0.0098,0.6156,0.5173,0.6222,0.5238,3.52,0.0230,1.05,1.25,0.6255,0.5271,7.340,1.7365,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.4250,0.3584,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
40
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.05562,0.00334,0.00150,0.06047,0.06727,0.00293,0.3025,0.02122,0.06341,0.07585,0.13632,0.09190,0.00418,0.00206,0.09813,0.03262,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0065,0.0098,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
41
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.05553,0.00334,0.00150,0.06038,0.06717,0.00293,0.3025,0.02120,0.06331,0.07572,0.13610,0.09174,0.00418,0.00206,0.09798,0.03257,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0183,0.0275,0.5532,0.4361,0.5715,0.4544,2.29,0.0420,3.21,4.03,0.5807,0.4636,8.220,1.9558,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,0.3814,0.3021,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
42
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.06321,0.00334,0.00150,0.06805,0.07575,0.00293,0.3025,0.02352,0.07099,0.08619,0.15424,0.10442,0.00418,0.00206,0.11066,0.03641,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0056,0.0084,0.8120,0.8120,0.8176,0.8176,1.89,0.0106,0.68,0.68,0.8204,0.8204,3.690,0.8753,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,0.5580,0.5580,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
43
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.06658,0.00334,0.00150,0.07143,0.07952,0.00293,0.3025,0.02454,0.07437,0.09080,0.16223,0.11000,0.00418,0.00206,0.11624,0.03809,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0026,0.0039,0.3479,0.2669,0.3505,0.2695,1.17,0.0031,0.75,0.97,0.3518,0.2708,10.442,2.4638,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.2383,0.1835,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
44
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.08570,0.00334,0.00150,0.09054,0.10088,0.00293,0.3025,0.03032,0.09348,0.11686,0.20740,0.14158,0.00418,0.00206,0.14782,0.04764,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0223,0.0334,0.4549,0.3961,0.4771,0.4183,2.09,0.0465,4.67,5.32,0.4883,0.4295,11.726,2.7859,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.3188,0.2790,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
45
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.07160,0.00334,0.00150,0.07645,0.08513,0.00293,0.3025,0.02606,0.07939,0.09764,0.17409,0.11830,0.00418,0.00206,0.12454,0.04060,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0079,0.0118,0.5621,0.4669,0.5700,0.4747,1.62,0.0127,1.38,1.65,0.5739,0.4786,5.656,1.3409,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,0.3994,0.3349,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
46
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.07629,0.00334,0.00150,0.08114,0.09037,0.00293,0.3025,0.02747,0.08407,0.10403,0.18516,0.12604,0.00418,0.00206,0.13227,0.04294,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0157,0.0236,2.2391,0.1364,2.2548,0.1521,3.91,0.0614,0.70,10.33,2.2627,0.1599,116.219,27.4077,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,1.6432,0.2188,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
47
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.08845,0.00334,0.00150,0.09330,0.10396,0.00293,0.3025,0.03115,0.09623,0.12062,0.21391,0.14613,0.00418,0.00206,0.15237,0.04902,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.0458,0.0687,0.3033,0.2859,0.3491,0.3317,13.56,0.6211,13.12,13.81,0.3720,0.3546,1.948,0.5049,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.2388,0.2269,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
48
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.10159,0.00334,0.00150,0.10644,0.11865,0.00293,0.3025,0.03513,0.10937,0.13853,0.24497,0.16784,0.00418,0.00206,0.17407,0.05558,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0194,0.0291,0.5442,0.4109,0.5636,0.4302,1.78,0.0345,3.44,4.50,0.5733,0.4399,6.643,1.5851,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,0.3766,0.2863,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
49
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.11127,0.00334,0.00150,0.11611,0.12947,0.00293,0.3025,0.03805,0.11905,0.15173,0.26784,0.18383,0.00418,0.00206,0.19006,0.06042,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0193,0.0289,0.7316,0.3960,0.7509,0.4153,3.41,0.0657,2.57,4.64,0.7605,0.4249,13.930,3.3025,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,0.5142,0.2869,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
50
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.11805,0.00334,0.00150,0.12290,0.13705,0.00293,0.3025,0.04011,0.12583,0.16098,0.28388,0.19504,0.00418,0.00206,0.20127,0.06381,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.2056,0.3084,,,,,0.867,0.1782,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
51
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.11610,0.00334,0.00150,0.12095,0.13487,0.00293,0.3025,0.03952,0.12389,0.15832,0.27927,0.19182,0.00418,0.00206,0.19805,0.06284,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.1597,0.2396,0.5442,0.3520,0.7040,0.5118,0.637,0.1018,22.69,31.22,0.7838,0.5916,22.074,5.3623,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,0.3953,0.2651,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
52
|
+
B,mean of the measured class members (spec 11.2),H100 (central),H100,5,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16); Llama-4-Scout-17B-16E-Instruct (bfloat16),2090.1,1.5,1.1,1.2,0.75,0.4285,0.2642,0.5301,0.4131,0.4992,0.4790,0.11904,0.19641,0.23569,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.11388,0.00334,0.00150,0.11873,0.13239,0.00293,0.3025,0.03885,0.12167,0.15530,0.27403,0.18815,0.00418,0.00206,0.19439,0.06173,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.1080,0.1080,0.7405,0.3704,0.8486,0.4784,,,12.73,22.58,0.8486,0.4784,17.497,4.2319,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,0.5748,0.3241,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
53
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.00603,0.00334,0.00100,0.01037,0.01232,0.00293,0.3025,0.00607,0.01331,0.00822,0.01859,0.00996,0.00418,0.00137,0.01551,0.00763,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0017,0.0026,0.9444,0.2174,0.9462,0.2192,2.17,0.0038,0.18,0.80,0.9470,0.2201,103.306,16.2113,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,0.7342,0.1760,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
54
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.00836,0.00334,0.00100,0.01270,0.01527,0.00293,0.3025,0.00678,0.01564,0.01140,0.02410,0.01381,0.00418,0.00137,0.01936,0.00879,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0044,0.0065,0.5760,0.3234,0.5804,0.3277,4.99,0.0217,0.75,1.33,0.5825,0.3299,32.196,5.0562,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.4483,0.2543,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
55
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.01276,0.00334,0.00100,0.01711,0.02085,0.00293,0.3025,0.00811,0.02004,0.01740,0.03451,0.02108,0.00418,0.00137,0.02663,0.01099,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0013,0.0013,0.7127,0.2686,0.7140,0.2699,2.17,0.0028,0.18,0.48,0.7140,0.2699,50.222,7.8815,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,0.5482,0.2072,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
56
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.02113,0.00334,0.00100,0.02548,0.03146,0.00293,0.3025,0.01064,0.02841,0.02881,0.05429,0.03491,0.00418,0.00137,0.04046,0.01517,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0044,0.0065,0.4098,0.3444,0.4142,0.3487,3.52,0.0153,1.05,1.25,0.4164,0.3509,7.340,1.1560,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.3207,0.2704,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
57
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.03703,0.00334,0.00100,0.04138,0.05160,0.00293,0.3025,0.01545,0.04431,0.05050,0.09187,0.06118,0.00418,0.00137,0.06673,0.02312,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0044,0.0065,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
58
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.03697,0.00334,0.00100,0.04131,0.05152,0.00293,0.3025,0.01543,0.04425,0.05041,0.09172,0.06108,0.00418,0.00137,0.06662,0.02309,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0122,0.0183,0.3683,0.2903,0.3805,0.3025,2.29,0.0279,3.21,4.03,0.3866,0.3086,8.220,1.3020,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,0.2878,0.2279,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
59
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.04208,0.00334,0.00100,0.04642,0.05799,0.00293,0.3025,0.01698,0.04936,0.05738,0.10380,0.06952,0.00418,0.00137,0.07507,0.02564,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0037,0.0056,0.5405,0.5405,0.5443,0.5443,1.89,0.0070,0.68,0.68,0.5461,0.5461,3.690,0.5827,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,0.4211,0.4211,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
60
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.04433,0.00334,0.00100,0.04867,0.06084,0.00293,0.3025,0.01766,0.05161,0.06045,0.10912,0.07323,0.00418,0.00137,0.07878,0.02676,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0017,0.0026,0.2316,0.1777,0.2333,0.1794,1.17,0.0020,0.75,0.97,0.2342,0.1803,10.442,1.6402,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.1798,0.1384,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
61
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.05705,0.00334,0.00100,0.06140,0.07696,0.00293,0.3025,0.02150,0.06433,0.07780,0.13919,0.09425,0.00418,0.00137,0.09980,0.03312,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0148,0.0222,0.3028,0.2637,0.3177,0.2785,2.09,0.0310,4.67,5.32,0.3251,0.2859,11.726,1.8547,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.2406,0.2105,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
62
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.04767,0.00334,0.00100,0.05201,0.06508,0.00293,0.3025,0.01867,0.05495,0.06500,0.11702,0.07875,0.00418,0.00137,0.08430,0.02843,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0052,0.0078,0.3742,0.3108,0.3795,0.3160,1.62,0.0085,1.38,1.65,0.3821,0.3187,5.656,0.8927,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,0.3014,0.2527,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
63
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.05079,0.00334,0.00100,0.05513,0.06903,0.00293,0.3025,0.01961,0.05807,0.06925,0.12439,0.08391,0.00418,0.00137,0.08945,0.02999,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0105,0.0157,1.4906,0.0908,1.5011,0.1012,3.91,0.0409,0.70,10.33,1.5063,0.1065,116.219,18.2461,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,1.2399,0.1651,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
64
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.05888,0.00334,0.00100,0.06323,0.07929,0.00293,0.3025,0.02206,0.06616,0.08030,0.14353,0.09728,0.00418,0.00137,0.10283,0.03404,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.0305,0.0457,0.2019,0.1903,0.2324,0.2208,13.56,0.4135,13.12,13.81,0.2477,0.2361,1.948,0.3362,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.1802,0.1712,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
65
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.06763,0.00334,0.00100,0.07198,0.09037,0.00293,0.3025,0.02470,0.07491,0.09222,0.16420,0.11173,0.00418,0.00137,0.11728,0.03841,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0129,0.0193,0.3623,0.2735,0.3752,0.2864,1.78,0.0230,3.44,4.50,0.3816,0.2929,6.643,1.0553,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,0.2842,0.2160,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
66
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.07407,0.00334,0.00100,0.07842,0.09853,0.00293,0.3025,0.02665,0.08135,0.10101,0.17943,0.12238,0.00418,0.00137,0.12793,0.04163,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0128,0.0193,0.4870,0.2636,0.4999,0.2765,3.41,0.0438,2.57,4.64,0.5063,0.2829,13.930,2.1986,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,0.3880,0.2165,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
67
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.07859,0.00334,0.00100,0.08294,0.10425,0.00293,0.3025,0.02802,0.08587,0.10717,0.19011,0.12984,0.00418,0.00137,0.13539,0.04389,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.1369,0.2053,,,,,0.867,0.1187,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
68
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.07729,0.00334,0.00100,0.08164,0.10261,0.00293,0.3025,0.02763,0.08457,0.10540,0.18704,0.12770,0.00418,0.00137,0.13325,0.04324,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.1063,0.1595,0.3623,0.2343,0.4687,0.3407,0.637,0.0677,22.69,31.22,0.5218,0.3939,22.074,3.5699,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,0.2983,0.2000,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
69
|
+
B,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,4,Llama-3.1-70B-Instruct (bfloat16); Llama-3.3-70B-Instruct (bfloat16); Qwen3-32B (bfloat16); gemma-3-27b-it (bfloat16),1457.3,1.5,1.1,1.2,0.75,0.2853,0.1764,0.3796,0.2689,0.2926,0.3614,0.07925,0.13076,0.15691,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.07582,0.00334,0.00100,0.08016,0.10074,0.00293,0.3025,0.02718,0.08310,0.10339,0.18355,0.12526,0.00418,0.00137,0.13081,0.04250,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.0719,0.0719,0.4930,0.2466,0.5649,0.3185,,,12.73,22.58,0.5649,0.3185,17.497,2.8173,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,0.4338,0.2446,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
70
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.03372,0.01284,0.00560,0.05217,0.04941,0.01284,0.3025,0.02862,0.06501,0.04598,0.09815,0.05571,0.01605,0.00766,0.07942,0.03687,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0098,0.0146,5.2827,1.2163,5.2925,1.2260,2.17,0.0212,0.18,0.80,5.2973,1.2309,103.306,90.6782,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,3.0190,0.7237,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
71
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.04674,0.01284,0.00560,0.06519,0.06154,0.01284,0.3025,0.03256,0.07803,0.06374,0.12893,0.07723,0.01605,0.00766,0.10094,0.04338,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0244,0.0366,3.2219,1.8088,3.2463,1.8332,4.99,0.1217,0.75,1.33,3.2585,1.8454,32.196,28.2818,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,1.8434,1.0458,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
72
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.07138,0.01284,0.00560,0.08983,0.08449,0.01284,0.3025,0.04001,0.10267,0.09734,0.18716,0.11793,0.01605,0.00766,0.14164,0.05569,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0073,0.0073,3.9864,1.5022,3.9937,1.5095,2.17,0.0159,0.18,0.48,3.9937,1.5095,50.222,44.0853,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,2.2543,0.8521,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
73
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.11820,0.01284,0.00560,0.13664,0.12809,0.01284,0.3025,0.05417,0.14949,0.16118,0.29782,0.19527,0.01605,0.00766,0.21899,0.07908,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0244,0.0366,2.2925,1.9263,2.3169,1.9507,3.52,0.0858,1.05,1.25,2.3290,1.9629,7.340,6.4663,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,1.3188,1.1121,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
74
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.20713,0.01284,0.00560,0.22558,0.21092,0.01284,0.3025,0.08107,0.23842,0.28245,0.50803,0.34220,0.01605,0.00766,0.36592,0.12352,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0244,0.0366,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
75
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.20678,0.01284,0.00560,0.22522,0.21059,0.01284,0.3025,0.08097,0.23807,0.28197,0.50720,0.34162,0.01605,0.00766,0.36534,0.12335,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0683,0.1024,2.0599,1.6239,2.1282,1.6921,2.29,0.1563,3.21,4.03,2.1623,1.7263,8.220,7.2829,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,1.1834,0.9373,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
76
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.23536,0.01284,0.00560,0.25381,0.23722,0.01284,0.3025,0.08961,0.26665,0.32095,0.57476,0.38885,0.01605,0.00766,0.41256,0.13763,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0208,0.0312,3.0236,3.0236,3.0444,3.0444,1.89,0.0394,0.68,0.68,3.0548,3.0548,3.690,3.2594,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,1.7315,1.7315,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
77
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.24794,0.01284,0.00560,0.26639,0.24893,0.01284,0.3025,0.09342,0.27923,0.33810,0.60449,0.40963,0.01605,0.00766,0.43334,0.14392,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0098,0.0146,1.2954,0.9939,1.3052,1.0036,1.17,0.0114,0.75,0.97,1.3101,1.0085,10.442,9.1747,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.7395,0.5693,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
78
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.31911,0.01284,0.00560,0.33756,0.31522,0.01284,0.3025,0.11494,0.35040,0.43515,0.77271,0.52721,0.01605,0.00766,0.55093,0.17948,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0829,0.1243,1.6939,1.4749,1.7768,1.5578,2.09,0.1732,4.67,5.32,1.8182,1.5993,11.726,10.3740,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.9892,0.8656,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
79
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.26664,0.01284,0.00560,0.28508,0.26634,0.01284,0.3025,0.09907,0.29793,0.36360,0.64868,0.44052,0.01605,0.00766,0.46423,0.15326,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0293,0.0439,2.0932,1.7385,2.1225,1.7677,1.62,0.0474,1.38,1.65,2.1371,1.7824,5.656,4.9932,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,1.2394,1.0391,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
80
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.28408,0.01284,0.00560,0.30252,0.28258,0.01284,0.3025,0.10435,0.31537,0.38738,0.68990,0.46933,0.01605,0.00766,0.49304,0.16197,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0585,0.0878,8.3379,0.5078,8.3964,0.5663,3.91,0.2288,0.70,10.33,8.4256,0.5956,116.219,102.0601,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,5.0986,0.6788,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
81
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.32937,0.01284,0.00560,0.34782,0.32477,0.01284,0.3025,0.11805,0.36066,0.44914,0.79696,0.54416,0.01605,0.00766,0.56788,0.18461,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.1706,0.2558,1.1296,1.0647,1.3001,1.2352,13.56,2.3128,13.12,13.81,1.3854,1.3205,1.948,1.8803,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.7408,0.7042,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
82
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.37829,0.01284,0.00560,0.39674,0.37033,0.01284,0.3025,0.13284,0.40958,0.51585,0.91259,0.62498,0.01605,0.00766,0.64870,0.20905,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0721,0.1082,2.0265,1.5300,2.0987,1.6021,1.78,0.1284,3.44,4.50,2.1348,1.6382,6.643,5.9027,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,1.1687,0.8884,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
83
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.41433,0.01284,0.00560,0.43278,0.40390,0.01284,0.3025,0.14374,0.44562,0.56500,0.99777,0.68453,0.01605,0.00766,0.70824,0.22706,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0718,0.1077,2.7243,1.4745,2.7961,1.5463,3.41,0.2448,2.57,4.64,2.8320,1.5822,13.930,12.2978,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,1.5956,0.8901,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
84
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.43960,0.01284,0.00560,0.45804,0.42743,0.01284,0.3025,0.15139,0.47089,0.59945,1.05750,0.72627,0.01605,0.00766,0.74998,0.23969,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.7655,1.1483,,,,,0.867,0.6637,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
85
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.43234,0.01284,0.00560,0.45079,0.42067,0.01284,0.3025,0.14919,0.46363,0.58956,1.04034,0.71428,0.01605,0.00766,0.73799,0.23606,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.5949,0.8923,2.0265,1.3108,2.6214,1.9057,0.637,0.3789,22.69,31.22,2.9188,2.2031,22.074,19.9681,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,1.2265,0.8225,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
86
|
+
C,mean of the measured class members (spec 11.2),H100 (central),H100,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct-FP8 (fp8); Llama-4-Maverick-17B-128E-Instruct-FP8 (fp8),3990.4,1.5,1.1,1.2,0.75,1.5958,0.5294,2.4728,1.3270,1.7851,1.4862,0.44327,0.73139,0.87767,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.42408,0.01284,0.00560,0.44253,0.41298,0.01284,0.3025,0.14669,0.45537,0.57829,1.02082,0.70063,0.01605,0.00766,0.72435,0.23193,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.4023,0.4023,2.7576,1.3794,3.1599,1.7816,,,12.73,22.58,3.1599,1.7816,17.497,15.7585,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,1.7837,1.0057,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
87
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,eu-north-1,EU North - Sweden,0.0384,Ember 2023 SWE 38.42,,0.03481,0.01284,0.00578,0.05343,0.03594,0.01284,0.3025,0.02901,0.06628,0.04747,0.10090,0.05751,0.01605,0.00791,0.08147,0.03749,0.0200,AWS; AWS Region Europe (Stockholm) eu-north-1; 2025,6.019,Reig Sweden 1.59 gal,1.386,withdrawal,0.0101,0.0151,5.4533,1.2555,5.4633,1.2656,2.17,0.0218,0.18,0.80,5.4684,1.2707,103.306,93.6062,73.961,published-site,IT load,AWS,AWS Region Europe (Stockholm) eu-north-1,2025,0.090,1.9089,0.4576,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
88
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,eu-west-3,EU West - France,0.0533,Ember 2023 FRA 53.26,,0.04825,0.01284,0.00578,0.06688,0.04361,0.01284,0.3025,0.03307,0.07972,0.06580,0.13268,0.07972,0.01605,0.00791,0.10368,0.04420,0.0500,AWS; AWS Europe aggregate; 2025,3.671,Reig App.2 France 0.97 gal,2.061,withdrawal,0.0252,0.0378,3.3260,1.8672,3.3511,1.8924,4.99,0.1256,0.75,1.33,3.3637,1.9050,32.196,29.1950,15.427,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,1.1656,0.6612,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
89
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,europe-north1,EU North - Finland,0.0813,Ember 2023 FIN 81.33,,0.07369,0.01284,0.00578,0.09231,0.05812,0.01284,0.3025,0.04077,0.10516,0.10048,0.19279,0.12174,0.01605,0.00791,0.14570,0.05691,0.0100,"Li et al. 2025 (Microsoft); Finland* (projected, site under construction as of July 2023); as of July 2023",4.542,Reig Finland 1.20 gal,1.712,consumption,0.0076,0.0076,4.1151,1.5507,4.1226,1.5583,2.17,0.0164,0.18,0.48,4.1226,1.5583,50.222,45.5088,29.205,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),"Finland* (projected, site under construction as of July 2023)",as of July 2023,0.010,1.4253,0.5388,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
90
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,europe-west1,EU West - Belgium,0.1347,Ember 2023 BEL 134.67,,0.12201,0.01284,0.00578,0.14064,0.08569,0.01284,0.3025,0.05538,0.15348,0.16638,0.30702,0.20158,0.01605,0.00791,0.22554,0.08106,0.0500,AWS; AWS Europe aggregate; 2025,2.612,Reig App.2 Belgium 0.69 gal,2.195,withdrawal,0.0252,0.0378,2.3665,1.9885,2.3917,2.0137,3.52,0.0886,1.05,1.25,2.4042,2.0262,7.340,6.6751,3.315,continental,IT load,AWS,AWS Europe aggregate,2025,0.060,0.8338,0.7032,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
91
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,eu-27,EU-27 average,0.2360,Ember 2023 EU 236.0,,0.21382,0.01284,0.00578,0.23244,0.13806,0.01284,0.3025,0.08315,0.24529,0.29157,0.52401,0.35325,0.01605,0.00791,0.37721,0.12694,0.0500,AWS; AWS Europe aggregate; 2025,,no source (Reig has no EU-27 aggregate),,withdrawal,0.0252,0.0378,,,,,,,,,,,,,,continental,IT load,AWS,AWS Europe aggregate,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
92
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,europe-west2,EU West - UK (London),0.2356,Ember 2023 GBR 235.60,,0.21346,0.01284,0.00578,0.23208,0.13786,0.01284,0.3025,0.08304,0.24493,0.29108,0.52316,0.35266,0.01605,0.00791,0.37662,0.12676,0.1400,AWS; AWS Region Europe (London) eu-west-2; 2025,2.347,Reig App.2 Great Britain 0.62 gal,1.850,withdrawal,0.0705,0.1057,2.1264,1.6763,2.1969,1.7468,2.29,0.1614,3.21,4.03,2.2321,1.7820,8.220,7.5181,4.304,published-site,IT load,AWS,AWS Region Europe (London) eu-west-2,2025,0.050,0.7482,0.5926,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
93
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,europe-west4,Netherlands,0.2682,Ember 2023 NLD 268.17,,0.24296,0.01284,0.00578,0.26159,0.15469,0.01284,0.3025,0.09197,0.27444,0.33132,0.59291,0.40141,0.01605,0.00791,0.42537,0.14150,0.0356,"Microsoft; Hollands Kroon, Netherlands (site); FY25",3.445,Reig Netherlands 0.91 gal,3.445,withdrawal,0.0215,0.0323,3.1212,3.1212,3.1427,3.1427,1.89,0.0406,0.68,0.68,3.1535,3.1535,3.690,3.3647,1.071,derived-site,total site electricity,Microsoft,"Hollands Kroon, Netherlands (site)",FY25,0.060,1.0948,1.0948,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
94
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,eu-west-1,EU West - Ireland,0.2825,Ember 2023 IRL 282.50,,0.25595,0.01284,0.00578,0.27457,0.16210,0.01284,0.3025,0.09589,0.28742,0.34902,0.62359,0.42286,0.01605,0.00791,0.44682,0.14799,0.0200,AWS; AWS Region Europe (Ireland) eu-west-1; 2025,1.476,Reig Ireland 0.39 gal,1.132,withdrawal,0.0101,0.0151,1.3373,1.0259,1.3473,1.0360,1.17,0.0118,0.75,0.97,1.3524,1.0410,10.442,9.4710,9.142,published-site,IT load,AWS,AWS Region Europe (Ireland) eu-west-1,2025,0.020,0.4676,0.3599,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
95
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,eu-central-1,EU Central - Germany,0.3636,Ember 2023 DEU 363.59,,0.32942,0.01284,0.00578,0.34804,0.20401,0.01284,0.3025,0.11811,0.36089,0.44920,0.79725,0.54424,0.01605,0.00791,0.56820,0.18470,0.1700,AWS; AWS Region Europe (Frankfurt) eu-central-1; 2025,1.930,Reig App.2 Germany 0.51 gal,1.681,withdrawal,0.0856,0.1284,1.7486,1.5226,1.8342,1.6081,2.09,0.1788,4.67,5.32,1.8770,1.6509,11.726,10.7090,6.659,published-site,IT load,AWS,AWS Region Europe (Frankfurt) eu-central-1,2025,0.080,0.6254,0.5473,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
96
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,us-east-1,US East - N. Virginia,0.3038,eGRID2023 SRVC 596.3 lb/MWh x 1.1232,0.2705,0.27525,0.01284,0.00578,0.29387,0.17311,0.01284,0.3025,0.10173,0.30672,0.37534,0.66921,0.45474,0.01605,0.00791,0.47870,0.15763,0.0600,AWS; AWS Region U.S. East (Northern Virginia) us-east-1; 2025,2.385,Reig SRVC 0.63 gal,1.981,withdrawal,0.0302,0.0453,2.1608,1.7946,2.1910,1.8248,1.62,0.0489,1.38,1.65,2.2061,1.8399,5.656,5.1544,2.825,published-site,IT load,AWS,AWS Region U.S. East (Northern Virginia) us-east-1,2025,0.140,0.7836,0.6570,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
97
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,us-west-2,US West - Oregon,0.3237,eGRID2023 NWPP 635.3 lb/MWh x 1.1232,0.2882,0.29325,0.01284,0.00578,0.31187,0.18337,0.01284,0.3025,0.10718,0.32472,0.39988,0.71176,0.48448,0.01605,0.00791,0.50844,0.16663,0.1200,AWS; AWS Region U.S. West (Oregon) us-west-2; 2025,9.500,Reig App.1 NWPP 2.51 gal,0.579,withdrawal,0.0604,0.0906,8.6071,0.5242,8.6675,0.5846,3.91,0.2362,0.70,10.33,8.6977,0.6148,116.219,105.3556,180.224,published-site,IT load,AWS,AWS Region U.S. West (Oregon) us-west-2,2025,0.950,3.2238,0.4292,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
98
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,us-south-central,US South Central - Texas,0.3753,eGRID2023 ERCT 736.6 lb/MWh x 1.1232,0.3341,0.34001,0.01284,0.00578,0.35863,0.21005,0.01284,0.3025,0.12132,0.37148,0.46365,0.82228,0.56173,0.01605,0.00791,0.58569,0.19000,0.2915,Microsoft; San Antonio (TX) (site); FY25,1.287,Reig ERCT 0.34 gal,1.213,withdrawal,0.1761,0.2641,1.1660,1.0990,1.3421,1.2751,13.56,2.3875,13.12,13.81,1.4301,1.3631,1.948,1.9410,1.522,derived-site,total site electricity,Microsoft,San Antonio (TX) (site),FY25,0.250,0.4684,0.4452,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
99
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,us-south,US South - Southeast,0.4310,eGRID2023 SRSO 846.0 lb/MWh x 1.1232,0.3837,0.39051,0.01284,0.00578,0.40913,0.23886,0.01284,0.3025,0.13659,0.42198,0.53251,0.94164,0.64516,0.01605,0.00791,0.66912,0.21523,0.1233,Meta; Stanton Springs (GA) (site); 2024,2.309,Reig App.1 SRSO 0.61 gal,1.743,withdrawal,0.0745,0.1117,2.0920,1.5794,2.1665,1.6538,1.78,0.1326,3.44,4.50,2.2037,1.6911,6.643,6.0933,3.684,derived-site,total site electricity,Meta,Stanton Springs (GA) (site),2024,0.060,0.7389,0.5617,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
100
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,us-central1,US Central - Iowa,0.4721,eGRID2023 MROW 926.6 lb/MWh x 1.1232,0.4203,0.42771,0.01284,0.00578,0.44634,0.26008,0.01284,0.3025,0.14785,0.45918,0.58324,1.02958,0.70663,0.01605,0.00791,0.73059,0.23382,0.1227,Microsoft; Des Moines (IA) (site); FY25,3.104,Reig MROW 0.82 gal,1.680,withdrawal,0.0741,0.1112,2.8123,1.5222,2.8864,1.5963,3.41,0.2527,2.57,4.64,2.9234,1.6333,13.930,12.6949,7.953,derived-site,total site electricity,Microsoft,Des Moines (IA) (site),FY25,0.140,1.0088,0.5628,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
101
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,ap-southeast-1,Asia Pacific - Singapore,0.5009,Ember 2023 SGP 500.87,,0.45379,0.01284,0.00578,0.47242,0.27496,0.01284,0.3025,0.15573,0.48526,0.61881,1.09123,0.74972,0.01605,0.00791,0.77368,0.24686,1.5700,AWS; AWS Region Asia-Pacific (Singapore) ap-southeast-1; 2025,,no source (Reig has no Singapore),,withdrawal,0.7902,1.1854,,,,,0.867,0.6851,,,,,,,,published-site,IT load,AWS,AWS Region Asia-Pacific (Singapore) ap-southeast-1,2025,,,,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
102
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,ap-northeast-1,Asia Pacific - Japan,0.4926,Ember 2023 JPN 492.6,,0.44630,0.01284,0.00578,0.46493,0.27069,0.01284,0.3025,0.15347,0.47777,0.60859,1.07352,0.73734,0.01605,0.00791,0.76130,0.24311,1.2200,AWS; AWS Region Asia-Pacific (Tokyo) ap-northeast-1; 2025,2.309,Reig App.2 Japan 0.61 gal,1.494,withdrawal,0.6141,0.9211,2.0920,1.3531,2.7061,1.9672,0.637,0.3912,22.69,31.22,3.0131,2.2743,22.074,20.6129,10.478,published-site,IT load,AWS,AWS Region Asia-Pacific (Tokyo) ap-northeast-1,2025,0.200,0.7755,0.5200,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|
|
103
|
+
C,mean of the measured class members (spec 11.2),B200 (hardware sensitivity),B200,3,Qwen3-235B-A22B-Instruct-2507 (bfloat16); Llama-3.1-405B-Instruct (bfloat16); DeepSeek-V3.1 (fp8),6218.6,1.5,1.1,1.2,0.75,1.6473,1.1319,2.2436,1.5845,1.5664,0.9397,0.45758,0.75501,0.90601,global-average,Global average,0.4832,Ember 2023 World 483.19,,0.43778,0.01284,0.00578,0.45640,0.26582,0.01284,0.3025,0.15089,0.46925,0.59697,1.05337,0.72326,0.01605,0.00791,0.74722,0.23885,0.5500,Li et al. 2025 (Microsoft); U.S. Average; as of July 2023,3.142,Li T1 US average (proxy; Reig US 0.83 gal),1.572,consumption,0.4153,0.4153,2.8467,1.4239,3.2619,1.8391,,,12.73,22.58,3.2619,1.8391,17.497,16.2674,8.845,2023-site,IT load (server energy),Li et al. 2025 (Microsoft),U.S. Average,as of July 2023,0.550,1.1278,0.6359,ML.ENERGY v3 lm-arena-chat (2026-02-16); per member: min J/token over all num_gpus with median ITL <= 50 ms; class value = mean of the members,"embodied: mean over the class members of (3,300 + 672 kg) / (5 y x 75% x the member's measured 8-GPU H100 throughput); construction 0.00766 kg per server (IT) kWh; training: Meta Llama 3.1 / (12 months x 40 T/21 d)"
|