inference/research
verifiedhigh · 0.88updated 2026-10-05

US average PUE for facilities serving AI equipment: 1.145 in 2024, modelled down to 1.136 in 2030

1.145 dimensionless PUE ratio (total facility energy / IT equipment energy), 2024. Not an accuracy score and not a speedup: 1.145 means the facility draws 14.5% more than the IT equipment it houses. PUE is a ratio by DEFINITION and carries no faster/slower reading, which is why it belongs to dimensionless_ratio rather than speedup_ratio.

Fields

id
benchmarks/pue-us-facilities-serving-ai-equipment-2024-1-145
Metric
dimensionless_ratio
Value
1.145
Unit
dimensionless PUE ratio (total facility energy / IT equipment energy), 2024. Not an accuracy score and not a speedup: 1.145 means the facility draws 14.5% more than the IT equipment it houses. PUE is a ratio by DEFINITION and carries no faster/slower reading, which is why it belongs to dimensionless_ratio rather than speedup_ratio.
Methodology
MODELED. Same LBNL model as the all-equipment record: installed base x unit wattage summed per equipment category, multiplied by modeled PUE. The report states this figure is specifically for FACILITIES SERVING AI EQUIPMENT, which is a different population from the 1.45 all-equipment national average - AI-serving facilities are modeled as more efficient than the average stock because liquid cooling for dense AI racks is assumed to have better PUE than legacy air-cooled rooms. The report also states that despite evolving cooling system designs it does NOT expect meaningful industry-wide changes to AI-serving PUEs relative to its previous modeling results, so the 1.145 -> 1.136 decline is a modest trend continuation, not a step change from liquid cooling. engine_id is ollama only as a schema placeholder: this is a facility metric.
Measured by
third_party
Reproducible
yes
notes
THE MOST USEFUL SINGLE NUMBER HERE FOR AI DEPLOYMENT PLANNING, because it is the population that actually hosts accelerators. The gap between 1.145 (AI facilities) and 1.45 (all equipment) is roughly 0.30 of overhead W per IT W - about 21% lower overhead for AI-serving facilities. Consequence for cost-per-token work: multiplying an accelerator's joules/token by 1.145 is defensible for a modern AI facility, while the industry-average 1.56 figure Google quotes (see the Google record) is a blended number dominated by older, less efficient stock and will overstate overhead for a new AI build. Do not present 1.145 as a metered fleet average - it is a model output whose uncertainty the 2025 Update declines to carry forward explicitly. METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> speedup_ratio. Both members were live in this record's case and the enum forced a choice between two wrong answers: `quality` implies a score and this is a ratio, and there was no dimensionless member at all. This record is one of the five named in gotcha bench-metric-enum-hides-a-dimensionless-ratio, whose root_cause records that the `quality` catch-all was carrying accuracy deltas, PUE ratios and power ratios at once. METRIC CORRECTED AGAIN 2026-10-05: speedup_ratio -> dimensionless_ratio. The first pass moved this record's eight siblings to dimensionless_ratio and left this one on speedup_ratio, which made it the only PUE figure in the repo filed as a speedup. It was caught by the new regression tests rather than by review, which is the argument for having them.

Notes

THE MOST USEFUL SINGLE NUMBER HERE FOR AI DEPLOYMENT PLANNING, because it is the population that actually hosts accelerators. The gap between 1.145 (AI facilities) and 1.45 (all equipment) is roughly 0.30 of overhead W per IT W - about 21% lower overhead for AI-serving facilities. Consequence for cost-per-token work: multiplying an accelerator's joules/token by 1.145 is defensible for a modern AI facility, while the industry-average 1.56 figure Google quotes (see the Google record) is a blended number dominated by older, less efficient stock and will overstate overhead for a new AI build. Do not present 1.145 as a metered fleet average - it is a model output whose uncertainty the 2025 Update declines to carry forward explicitly.

METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> speedup_ratio. Both members were live in this record's case and the enum forced a choice between two wrong answers: `quality` implies a score and this is a ratio, and there was no dimensionless member at all. This record is one of the five named in gotcha bench-metric-enum-hides-a-dimensionless-ratio, whose root_cause records that the `quality` catch-all was carrying accuracy deltas, PUE ratios and power ratios at once.

METRIC CORRECTED AGAIN 2026-10-05: speedup_ratio -> dimensionless_ratio. The first pass moved this record's eight siblings to dimensionless_ratio and left this one on speedup_ratio, which made it the only PUE figure in the repo filed as a speedup. It was caught by the new regression tests rather than by review, which is the argument for having them.

Cross-references

Engine (1)