inference/research
verifiedhigh · 0.90updated 2026-10-05

US national average PUE across all data center equipment: 1.45 in 2024, down from 1.55 in 2018; 1.36 projected for 2030 in the Reference Case

1.45 dimensionless PUE ratio (total facility energy / IT equipment energy). Not an accuracy score and not a speedup: 1.0 means the facility overhead is nil.

Fields

id
benchmarks/pue-national-average-us-all-equipment-2024-1-45
Metric
dimensionless_ratio
Value
1.45
Unit
dimensionless PUE ratio (total facility energy / IT equipment energy). Not an accuracy score and not a speedup: 1.0 means the facility overhead is nil.
Methodology
MODELED, NOT METERED AT A SINGLE SITE. The 2025 Update report estimates total data center electricity by summing installed-base x unit-wattage per equipment category and then multiplying by modeled PUE values; the PUE model itself is inherited unchanged from the 2024 Report. The 1.55 (2018) and 1.45 (2024) figures describe the ALL-EQUIPMENT national average, i.e. the whole installed US stock, not AI facilities. The 2030 Reference Case value of 1.36 is a projection under unchanged methodology, and the report explicitly does NOT vary PUE across its scenarios - so scenario-to-scenario differences in its energy totals come entirely from AI server energy, not from efficiency. Attribution of the 2018-2024 improvement is given as 'the continued shift in server energy use into the types of data centers that have the lowest PUEs, namely those deploying liquid cooling systems for AI servers' - the national average improved partly because POWERFUL MACHINES MOVED to efficient buildings, which is a composition effect, not a retrofit effect. engine_id is ollama only as a schema placeholder: PUE is a facility metric and has no engine.
Measured by
third_party
Reproducible
yes
notes
USE THIS, NOT A GPU TDP NUMBER, FOR ANY WHOLE-FACILITY ENERGY ARGUMENT. A PUE of 1.45 means every 1 W of IT load costs 0.45 W of cooling and power distribution on top. The same report puts facility infrastructure at 31% of total US data center electricity in 2024, down from 36% in 2018. PUE IS NOT A QUALITY METRIC - it is filed under metric 'quality' purely because the benchmark metric enum has no energy or efficiency member and adding one requires a schemas/ edit that is out of scope for this task; see the joules-in-the-unit-string convention already used by the H200 power records. Confidence is high on the number and high on the caveat: the figure is a model output, so it inherits the 2024 Report's PUE uncertainty, which the 2025 Update says it no longer propagates explicitly. METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> dimensionless_ratio. PUE is a ratio by DEFINITION (total facility energy / IT equipment energy) and carries no faster/slower reading, so it is a dimensionless_ratio rather than a speedup_ratio. It was previously filed as `quality`, the catch-all that gotcha bench-metric-enum-hides-a-dimensionless-ratio names as the reason `quality` cannot be trusted as a semantic filter: that key was spanning accuracy deltas in percentage points, PUE ratios and power ratios simultaneously.

Notes

USE THIS, NOT A GPU TDP NUMBER, FOR ANY WHOLE-FACILITY ENERGY ARGUMENT. A PUE of 1.45 means every 1 W of IT load costs 0.45 W of cooling and power distribution on top. The same report puts facility infrastructure at 31% of total US data center electricity in 2024, down from 36% in 2018. PUE IS NOT A QUALITY METRIC - it is filed under metric 'quality' purely because the benchmark metric enum has no energy or efficiency member and adding one requires a schemas/ edit that is out of scope for this task; see the joules-in-the-unit-string convention already used by the H200 power records. Confidence is high on the number and high on the caveat: the figure is a model output, so it inherits the 2024 Report's PUE uncertainty, which the 2025 Update says it no longer propagates explicitly.

METRIC CORRECTION 2026-10-05 (schema authority pass, agent w5-schema). Value UNCHANGED; `metric` moved quality -> dimensionless_ratio. PUE is a ratio by DEFINITION (total facility energy / IT equipment energy) and carries no faster/slower reading, so it is a dimensionless_ratio rather than a speedup_ratio. It was previously filed as `quality`, the catch-all that gotcha bench-metric-enum-hides-a-dimensionless-ratio names as the reason `quality` cannot be trusted as a semantic filter: that key was spanning accuracy deltas in percentage points, PUE ratios and power ratios simultaneously.

Cross-references

Engine (1)