PUE answers an infrastructure question
PUE is total facility energy divided by IT energy over a stated period. FEMP explains that it characterizes supporting infrastructure rather than the productivity of the IT equipment. Cooling changes can affect both sides of the ratio, as the Vertiv/NVIDIA modeled study discusses. A lower PUE alone is not proof of less electricity used or more useful AI work delivered.
Consider purely hypothetical arithmetic: 100 kWh of support energy plus 300 kWh of IT energy gives PUE 1.33. Keeping that support demand while increasing IT energy to 600 kWh gives PUE 1.17, although total energy rises from 400 to 700 kWh. These are illustrative inputs, not a benchmark. Report the underlying energy totals and the completed workload alongside the ratio so the denominator remains visible.
Best Practices Guide for Energy-Efficient Data Center Design — DOE FEMP, July 2024 (external source) · Quantifying the impact on PUE and energy consumption when introducing liquid cooling — Vertiv/NVIDIA study summary (external source)Declare the meters and aggregation method
The DOE-hosted 2011 task-force guidance separates dedicated and mixed-use boundaries and describes different metering capabilities. The ESIF operator's current PUE page supplies a concrete example of metered support loads and disclosed allocations for shared equipment. These sources show why two identical ratio labels can conceal different measurement records; the older guidance is not represented here as a newly issued standard.
Specify the facility boundary, IT measurement point, all included energy sources and the treatment of shared loads. Align timestamps and intervals. An annual PUE should use the ratio of energy sums rather than an unweighted average of instantaneous ratios. For a fleet, aggregate numerator and denominator energy before forming a portfolio ratio, then publish site values separately. Identify estimated allocations, coverage gaps and changes in meter locations.
Recommendations for Measuring and Reporting Overall Data Center Efficiency, version 2 — DOE-hosted industry task force (external source) · HPC Data Center Power Usage Effectiveness — National Laboratory of the Rockies (external source)WUE needs a numerator definition
Site WUE expresses annual site water usage in liters per kWh of IT energy. FEMP distinguishes that site boundary from a source-based assessment that includes water associated with energy production. Specify whether the reporting convention uses supplied water or a consumptive accounting estimate, which end uses are included and whether alternative supplies are counted. A freshwater-only intensity is useful when clearly labeled, but it should not quietly replace total site WUE.
Publish absolute water volumes, freshwater source volumes and peak withdrawal alongside intensity. Keep the IT denominator aligned with the water reporting period. A change in IT load can change WUE even if water volume stays constant; that is arithmetic, not proof of a better cooling process. Avoid comparing annual observations with a modeled design point or a seasonal best result.
Best Practices Guide for Energy-Efficient Data Center Design — DOE FEMP, July 2024 (external source) · Cooling Water Efficiency Opportunities for Federal Data Centers — DOE FEMP (external source)GPU telemetry is one measurement layer
NVIDIA's NVML reference documents a GPU energy counter in millijoules since the driver was last reloaded, along with power-reading behavior and unsupported-feature errors. Such telemetry can help analyze supported devices, but its device boundary excludes the rest of the server and facility. A manufacturer's power limit is also a configuration limit, not accumulated workload energy.
Record the device identifier, driver, supported counter, reset behavior and sampling interval. Convert units explicitly and check counters across restarts. Reconcile device data with server or rack meters where the evaluation requires them. Label estimated integration of power samples separately from an energy-counter difference. Include CPUs, memory, networking and storage at the relevant system boundary rather than calling the GPU sum total compute electricity.
Device Queries: Total Energy Consumption and Power Usage — NVIDIA NVML API Reference (external source)Define useful tokens before dividing by kWh
Tokens per kWh can be a useful proposed workload metric, but it needs an explicit numerator. Specify tokenizer and model revision, input and output counts, prompt lengths, decoding settings and the task-quality acceptance condition. Count completed accepted work, and report errors, cancellations and retries. More generated text is not automatically more useful work, while cached inputs and repeated attempts may have different commercial accounting.
Keep separate device, server, cluster and allocated-facility energy results. In hypothetical arithmetic, one million accepted output tokens divided by 100 kWh gives 10,000 tokens per kWh. That value changes when the energy boundary changes; it is not an Aguasto result or a model comparison. For billable-compute reporting, disclose the billing unit and allocation rules as well as technical work, so revenue accounting cannot silently redefine physical efficiency.
Device Queries: Total Energy Consumption and Power Usage — NVIDIA NVML API Reference (external source) · Data Center Metering and Resource Guide — DOE Better Buildings (external source)Utilization and overhead need operating context
LBNL's United States Data Center Energy Usage Report: 2025 Update, published June 18, 2026, treats idle power, utilization and equipment lifetime as important modeling assumptions. This consumption-focused national analysis is not a benchmark of Aguasto or a guarantee about a future site's capacity. It nevertheless reinforces the need to make the operating assumptions visible rather than treating nameplate power as annual energy.
For a future service evaluation, include loaded idle periods, spare capacity kept for failover, warmup and actual customer demand. Separate technical utilization from occupancy and revenue utilization. A short saturated test may be valuable for configuration work, while a monthly facility-energy-per-billable-unit metric answers a different operating question. Do not convert one into the other without an explicit time series and allocation method.
Data Center Metering and Resource Guide — DOE Better Buildings (external source) · United States Data Center Energy Usage Report: 2025 Update — LBNL, published June 2026 (external source)Use a dashboard that preserves the trade-offs
Aguasto's proposed evaluation record would place total facility energy, IT energy, cooling electricity, water volumes, PUE, WUE and accepted workload together. Add configuration, site weather, measurement coverage and the reporting period. Keep embodied impacts and upstream electricity-related water or emissions in separate named assessments, since neither disappears into PUE.
Compare options only after aligning quality, latency, workload, reliability and environmental boundaries. Show modeled, manufacturer-reported and measured values as distinct evidence states. Publish uncertainty and retain the raw reconciliation record when measurements become available. Until then, formulas and conceptual examples can explain the method, but an empty results field is more accurate than an invented efficiency score, a projected SLA or a claim of superiority over another operator.
Best Practices Guide for Energy-Efficient Data Center Design — DOE FEMP, July 2024 (external source) · Data Center Metering and Resource Guide — DOE Better Buildings (external source)Sources & verification
The register preserves each source's publication and verification date. Manufacturer claims and documentation are attributed; Aguasto has not run the cited products in production.
- Best Practices Guide for Energy-Efficient Data Center Design — DOE FEMP, July 2024 ↗
PUE measures supporting infrastructure rather than IT productivity; Annual PUE and site/source WUE distinctions; Heat-reuse host and backup heat rejection. Publication: 2024-07-26. Verified: 2026-10-13.
- Quantifying the impact on PUE and energy consumption when introducing liquid cooling — Vertiv/NVIDIA study summary ↗
Cooling changes can affect both sides of the PUE ratio; Vendor-modeled configurations are not universal measured savings. Publication: 2023-02-15. Verified: 2026-10-13.
- Recommendations for Measuring and Reporting Overall Data Center Efficiency, version 2 — DOE-hosted industry task force ↗
Dedicated and mixed-use measurement boundaries; Metering categories and disclosed assumptions; Comparisons need climate and reliability context. Publication: 2011-05-17. Verified: 2026-10-13.
- HPC Data Center Power Usage Effectiveness — National Laboratory of the Rockies ↗
Example of metered support loads and attributed shared loads; Site-specific ESIF reporting; not an Aguasto benchmark. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Cooling Water Efficiency Opportunities for Federal Data Centers — DOE FEMP ↗
Closed equipment loop can reject to an evaporative tower; Annual site WUE definition; Treatment can add electricity and maintenance. Publication: 2019-01-09. Verified: 2026-10-13.
- Device Queries: Total Energy Consumption and Power Usage — NVIDIA NVML API Reference ↗
GPU energy counter in millijoules since driver reload; Power sampling behavior and unsupported-feature errors; Device telemetry does not meter all facility loads. Publication: date not stated in the reviewed source. Verified: 2026-10-13.
- Data Center Metering and Resource Guide — DOE Better Buildings ↗
Metering enables efficiency diagnosis and verification; Separate planning guidance from dated market statistics. Publication: 2016-07-28. Verified: 2026-10-13.
- United States Data Center Energy Usage Report: 2025 Update — LBNL, published June 2026 ↗
2030 reference case and sensitivity scenarios; Idle draw, utilization and lifetime assumptions; Consumption forecasts do not establish available grid capacity. Publication: 2026-06-18. Verified: 2026-10-13.