When PUE Rises but Total GPU-Data-Center Energy Falls

Start with what PUE divides

The DOE FEMP/NREL guide revised July 2024 defines annual PUE as total facility energy divided by the energy drawn by all IT equipment. Section 8.1 distinguishes this infrastructure measure from overall data-center efficiency.

Match the numerator and denominator to the same facility boundary and interval. GPU-only telemetry cannot stand in for all IT-equipment energy. Dividing one year's facility energy by another year's IT energy creates a different, mismatched calculation.

References: [D1]

The denominator changes the meaning of an improvement

For the following algebra, call IT energy I and support energy S. Support is the remainder of facility energy after IT energy within the chosen boundary. Total facility energy F is I + S. For positive I, PUE = (I + S) / I = 1 + S / I. The energy units cancel.

The part above 1 therefore expresses support energy relative to IT energy. A PUE of 1.20 means 0.20 units of support energy per unit of IT energy; it does not mean support is 20% of total facility energy.

Reducing S while holding I fixed lowers both PUE and total energy. Reducing I while holding positive S fixed lowers total energy but raises S / I. That second case explains how a higher PUE can accompany lower consumption. The ratio records a relationship between two energy quantities, with no term for whether the IT activity produced a useful result.

A fictional annual comparison: 1.20 versus 1.25

Cases A and B are wholly fictional annual comparisons at the same facility boundary. Every energy figure is invented. For this example only, both deliver the same 1,000 accepted units of identical required work, with identical acceptance requirements. Support energy is stipulated to remain 20 MWh in both cases. That fixed amount is an assumption for the arithmetic, not a prediction of how a real optimization would affect supporting equipment.

A uses 100 MWh of IT energy plus 20 MWh of support energy. Total facility energy is 120 MWh and PUE is 120 / 100 = 1.20.

B uses 80 MWh of IT energy plus the same stipulated 20 MWh of support energy. Total facility energy is 100 MWh and PUE is 100 / 80 = 1.25.

Converting MWh to kWh, A's facility energy per accepted unit is 120,000 kWh / 1,000 = 120 kWh. B's is 100,000 kWh / 1,000 = 100 kWh. B uses 20 MWh less facility energy and 20 kWh less per accepted unit, despite its higher PUE.

The support-energy numerator in S / I stayed at 20 MWh while the IT-energy denominator fell from 100 to 80 MWh. Its share relative to IT energy rose from 20% to 25%. Nothing in those stipulated numbers says supporting equipment deteriorated. No particular GPU, software change or cooling intervention produced these fictional results.

The reverse arithmetic exposes a misleading incentive

With positive support energy fixed, extra wasted IT consumption can lower PUE without improving useful output. Add an energy amount X greater than zero to I, while holding S and accepted work unchanged. PUE becomes 1 + S / (I + X), which is smaller than 1 + S / I. Yet total facility energy rises by X.

If Q is the unchanged positive number of accepted units, facility energy per accepted unit also rises by X / Q. The dashboard ratio improves while the energy needed for the same useful result worsens.

This is an explanation of the ratio's behavior, never a reason to add load or change cooling or software to improve a dashboard number. A lower PUE alone cannot demonstrate a successful workload-efficiency change.

Choose the efficiency result the buyer needs

For infrastructure overhead, compare PUE alongside its facility and IT energy totals. For absolute consumption, compare facility energy over the stated period. For energy needed to deliver the same useful result, compare facility energy per accepted unit under equivalent work and acceptance requirements. In the fictional comparison, B has higher relative overhead, lower total consumption and lower energy per accepted unit. Those statements are compatible.

The equal-work condition does essential work here. If B simply completed less work, its lower annual energy total would not establish that it used less energy for the same result. If it produced more output that failed the required quality, counting that output as accepted would change the question.

Define the accepted unit before calculating the comparison. For a GPU workload, retain the task, required output quality and applicable response-time or completion requirements. Energy spent on unsuccessful attempts still belongs in the energy total if it falls within the declared boundary, even though those attempts add no accepted units. The workload-benchmark guide linked below develops comparable performance evidence; PUE cannot supply that missing evidence.

Keep the result within its measured boundary

An annual campus PUE does not become a tenant's measured energy for an individual GPU job. Multiplying a job's IT energy by campus annual PUE produces an allocation estimate under assumptions. It does not establish that the campus overhead during that job was measured and assigned to it. The operator's scope, interval and allocation method remain material.

Energy also differs from capacity: MW describes power, while MWh records energy over time. Do not divide a site's MW headline by annual PUE to approve racks or infer an available tenant allocation. The power-capacity guide owns that separate decision.

Neither the PUE ratio nor the fictional energy reduction establishes water use or carbon emissions. Those quantities need their own evidence. Renewable matching concerns the relationship between consumption and qualifying generation under a stated time and location method; it cannot be read out of PUE. The annual-versus-hourly renewable-matching guide addresses that distinct claim.

References: [E1]

Carry confirmed physical requirements into installation

Keep the efficiency conclusion with the operator and workload evidence that supports it. If the resulting equipment choice creates a physical installation requirement, carry the exact equipment, confirmed facility allocation and released physical plans into that discussion.

Leviathan's colocation infrastructure installation service covers agreed customer-side physical IT work in authorized third-party space, including equipment mounting, physical IT cabling, labeling and defined handoff records. Energy assessment and workload acceptance remain with the appointed operator and platform reviewers.

Use the linked colocation project inquiry to describe the nonconfidential site location, equipment population, requested physical tasks, available plans and intended work window. The scope discussion concerns those installation requirements, with availability and facility conditions confirmed for the engagement.