Research Note

Research Note: Data Center Resource Flow and Metric Dictionary

Electricity enters a data center through grid or on-site supply. Power-conversion and distribution equipment conditions it for servers, storage, and network equipment. Co

Aug 4, 20264 min readBy Dalton Anderson
In this article

Research Note: Data Center Resource Flow and Metric Dictionary

Physical model

Electricity enters a data center through grid or on-site supply. Power-conversion and distribution equipment conditions it for servers, storage, and network equipment. Computing turns almost all of that electrical work into heat. Fans, pumps, chillers, cooling towers, dry coolers, or liquid loops move the heat to an outdoor sink. Lighting, controls, security, fire protection, and other facility systems add smaller loads.

Water can appear at several boundaries. A facility may consume water directly through evaporative heat rejection, use water without consuming all of it, receive reclaimed rather than potable supply, discharge water to a sewer, or use little on-site water while drawing electricity from generation that has its own water footprint. Direct facility water and electricity-related source water are different quantities.

AI training and inference both run on computing equipment. Their resource use depends on model architecture, hardware, precision, batch size, utilization, data movement, cooling, duration, and the number of requests or training runs. A universal multiplier against one web search hides more than it reveals.

Metric dictionary

TermMeaningWhat it does not establish
MWPower or capacity at a momentEnergy consumed over a month or year
MWhEnergy equal to one megawatt sustained for one hourPeak demand or contracted capacity
PUEAnnual total facility energy divided by annual IT-equipment energyCarbon, total demand, useful computing work, or local grid effect
Site WUEAnnual on-site water use divided by IT-equipment energy, usually liters per kilowatt-hourSource water used to produce electricity
Source WUEWater associated with both site operations and energy supply under a stated methodLocal scarcity, water quality, or every supply-chain use
WithdrawalWater removed from groundwater or surface water for useThe portion no longer available locally
Consumptive useWater not returned to the immediate local water environmentTotal water delivered or withdrawn
Potable waterWater treated to drinking-water standardsThe only water source a data center can use
Reclaimed waterWastewater or another source treated and repurposed for a defined useZero treatment, pumping, infrastructure, or health requirements

The US Energy Information Administration explains the power and energy distinction. One megawatt sustained for one hour is one megawatt-hour. A 300 MW request is not the same statement as 300 MWh of annual consumption.

The Department of Energy design guide defines PUE as total annual facility energy divided by annual IT-equipment energy. It defines WUE as annual site water use divided by IT energy and discusses source WUE as an expanded boundary.

The US Geological Survey calls water removed from ground or surface sources a withdrawal. Consumptive use is the portion that does not return to local water bodies or groundwater, including evaporation and water incorporated into products.

The EPA water-reuse overview defines water reuse as treating and repurposing wastewater, stormwater, or another source for an appropriate beneficial use. Treatment requirements depend on the source and intended use.

Current national values

The 2024 LBNL report estimated 176 terawatt-hours of US data-center electricity use in 2023 and described facility infrastructure as about 30 percent of total data-center energy in that model year. It estimated 66 billion liters of direct site water use in 2023, with hyperscale and colocation facilities accounting for most of that total.

The report's modeled national average site WUE sat just above 0.36 liters per kilowatt-hour through 2023 and rose in its scenarios as facility mix and cooling changed. These are national modeled values. They cannot be assigned to a named facility without its own measurement record.

Editorial use

Every numerical claim should carry geography, time period, unit, denominator, operating or utilization assumption, facility boundary, water type, measured or modeled status, and source date.

PUE and WUE should be written as partial indicators. A lower PUE can coexist with higher total electricity when IT load grows. Lower on-site water use can coexist with higher energy demand. Reclaimed water can reduce pressure on potable supply while still requiring treatment, infrastructure, and a local allocation decision.

Sources

Follow the evidence.

  1. eta-publications.lbl.gov: united states data center energy 2025eta-publications.lbl.gov
  2. emp.lbl.gov: electricity rate designs large loadsemp.lbl.gov
  3. epa.gov: caapsepa.gov
  4. energy.gov: best practice guide data center design 0energy.gov
  5. energy.gov: data centers tribal economic development frequently asked questionsenergy.gov
  6. fairfaxcounty.gov: board supervisors approve new data center zoning ordinance amendmentfairfaxcounty.gov
  7. aepohio.com: data center tariffaepohio.com
  8. jlarc.virginia.gov: Rpt598 2jlarc.virginia.gov
  9. epa.gov: basic information about water reuseepa.gov
  10. energy.gov: doe releases new report evaluating increase electricity demand data centersenergy.gov
  11. iea.org: executive summaryiea.org
  12. youtu.be: uViM0ExISB0youtu.be
  13. eta.lbl.gov: revisiting relationship betweeneta.lbl.gov
  14. energy.gov: cooling water efficiency opportunities federal data centersenergy.gov
  15. apps.oregonlegislature.gov: HB3546apps.oregonlegislature.gov
  16. iea.org: energy supply for aiiea.org
  17. epa.gov: clean air act resources data centersepa.gov
  18. open.spotify.com: 43HYAwxKL3bFrE8qrR32BGopen.spotify.com
  19. eia.gov: measuring electricityeia.gov
  20. usgs.gov: water use united statesusgs.gov
  21. daltonanderson.ghost.io: the hidden costs of ai data centers power water straindaltonanderson.ghost.io
  22. energyanalysis.lbl.gov: 2024 lbnl data center energy usage reportenergyanalysis.lbl.gov

From this episode

Two useful next steps.

Research Note · 1 min

Research Note: Public Data Center Project Review

The civic guide should help a resident use limited time before a hearing or comment deadline. It must identify records and decision makers without pretending that every j

Research Note · 1 min

Research Note: E075 Demand Outlook and Publication Boundary

The July 2025 recording used dramatic comparisons to make AI infrastructure feel physical. The revised Episode Story needs to preserve Dalton's concern while replacing ar

Return to the episode
Venture Step