Google global fleet average annual PUE 1.09 in 2024, first sub-1.10 reading in six years, against a stated industry average of 1.56

SELF-REPORTED, and the report is unusually explicit about scope, which is why it is usable at all: Google states it (a) includes ALL data centers in the fleet, not just the newest or most efficient facilities; (b) measures PUE continuously through the year rather than in the coolest seasons; (c) includes all sources of overhead energy in the calculation; and (d) begins reporting a data center's PUE only once it reaches stable operations. The published figure is an average ANNUAL value; the report also maintains quarterly and trailing-12-month series (Figure 4, 2009-2024, plotted on a 1.08-1.20 axis). engine_id is ollama only as a schema placeholder: this is a fleet-facility metric with no engine.

REPRODUCIBLE: FALSE - it is an audited-report figure (Google states the underlying quarterly and T12M fleet PUE is publicly disclosed on its Data Centers: Efficiency site) but no machine-readable, independently re-computable meter data is published, and a third party cannot re-run the fleet. THE CONTRAST IS THE POINT: 1.09 versus a stated industry average of 1.56 means Google claims 84% less overhead energy than the industry average, i.e. 0.09 W of overhead per IT W against 0.56 W. THIS IS A CEILING, NOT A TYPICAL VALUE. A 1.09 PUE belongs to a fleet deliberately engineered for it (custom power distribution, on-site generation, years of retrofit) and is not what a colocation tenant or an enterprise closet will see - the LBNL all-equipment national average of 1.45 is the more honest planning number for most inference deployments. Google separately reports over six times more computing power per unit of electricity than five years earlier, attributing much of that to TPU deployment rather than to PUE. 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.