Metro Colo Advisory Research
Data Center Power Consumption
How Much Power Data Centers Use and What the Numbers Mean
US data centers used 192 TWh in 2024, 4.7% of the country's electricity. A single site can run from a few MW to several GW, and one MW figure can mean four different things. Here is how to read them, with a converter that turns any headline into usable IT load.
Send Us the Site DetailsUpdated November 2026.
Data center power consumption in the United States reached 192 terawatt-hours (TWh) in 2024, 4.7% of all US electricity and 14% more than in 2023, according to Lawrence Berkeley National Laboratory. That is about what 17.8 million American homes use in a year. How much power one data center uses depends on what it is: the median retail colocation building in our market data is rated about 6.6 MW, a wholesale building 33 MW and a leased hyperscale campus 168 MW. Announced AI campuses now reach 5,000 MW and more (Meta's Hyperion in Louisiana and the UAE-US AI Campus in Abu Dhabi are each announced at 5,000 MW). The largest running one, SpaceXAI's Colossus I and II in Memphis and Southaven, had about 1,000 MW of installed GPU power in SpaceX's June 2026 prospectus; company-wide, SpaceX's nameplate compute reached 1,700 MW by September 30, 2026.
Those megawatt figures do not all measure the same thing. A headline MW can be the utility connection, the whole facility's load, the critical IT load servers can use, or installed nameplate that counts spare equipment, and on one site the largest of these can run 50% or more above the smallest. This guide explains each one, how PUE, power factor and redundancy connect them, and gives you a converter that turns any figure into the others. When the figure sits in an offering memo or a press release you are about to act on, our data center consulting team checks it against the filings.
Sources: LBNL, United States Data Center Energy Usage Report: 2025 Update; EIA, average US home electricity use (homes figure is our calculation: 192 billion kWh divided by 10,791 kWh per home); Metro Colo Advisory market data, median of operator-stated MW across 281 verified buildings in Northern Virginia, Dallas-Fort Worth, Atlanta and Phoenix; Meta, Richland Parish data center; WAM via UAE Embassy; SpaceX prospectus, June 11, 2026 (Colossus I and II); SpaceX Form 10-Q, quarter ended September 30, 2026 (company-wide nameplate compute); ranked in Metro Colo Advisory's AI Campus Power Tracker.
Sources: LBNL 2025 Update; LBNL publication page (2030 range); Uptime Institute Global Data Center Survey 2026; Uptime Institute, The growing PUE advantage of larger data centers; NVIDIA DGX GB200 user guide; Supermicro GB200 NVL72 datasheet.
Data Center Energy Consumption: US and Global
US data centers used 192 TWh of electricity in 2024, and the Department of Energy's national lab projects 649 TWh by 2030 in its reference case, 11.8% of forecast US electricity use. Its full range for 2030 runs from 521 to 843 TWh, or 9.5% to 15.3%, and its reference case for 2028 is 464 TWh. In LBNL's December 2024 report, use climbed from 58 TWh in 2014 to 176 TWh in 2023. The 2025 Update revised recent years slightly lower and puts 2024 at 192 TWh, 14% above its own 2023 estimate. The 2023 and 2024 bars below come from different editions, so the gap between them is not the growth rate. The spread between the low and high 2030 cases, 322 TWh, is larger than all US data center use in 2024.
Sources: LBNL 2025 Update (2024 use, 2028 and 2030 reference cases, revision of prior years); LBNL publication page (2030 bounds of 521 to 843 TWh); DOE, December 20, 2024 release (2014 and 2023 use, 2028 range of 325 to 580 TWh).
Worldwide, data centers used about 485 TWh in 2025, up 17% in a year, and the IEA now expects about 950 TWh by 2030, roughly double. In 2024 the figure was about 415 TWh, about 1.5% of global electricity. In the IEA's April 2025 report, the United States and China account for nearly 80% of the growth to 2030, and the US had the highest data center use per person, around 540 kWh in 2024.
Sources: IEA, Key Questions on Energy and AI (2026) (2025 use, 17% growth, 2030 projection); IEA, Energy and AI: Energy demand from AI (April 2025; 2024 use and share, US and China share of growth, use per person).
US data center use in 2024 divided by the average American home's 10,791 kWh a year. A useful scale check, not a claim that homes and servers compete for the same supply.
Our calculation from LBNL and EIA.A gigawatt drawn every hour of the year is 8.76 billion kWh. Real loads run below their peak, so treat this as the ceiling for a 1 GW campus.
Our calculation from EIA, 10,791 kWh per home (2022).A third-party range for data centers' share of US electricity use by 2030, up from 4% to 5% today. EPRI's 2026 update lifted its projected 2030 demand range 60% above its earlier scenarios; LBNL's 11.8% reference case sits inside it.
Source: EPRI, Powering Intelligence 2026.How Big Is a Data Center? Capacity by Type
Most data centers are far smaller than the AI campuses in the news: the median retail colocation building in our market data is rated about 6.6 MW, a wholesale building 33 MW and a leased hyperscale campus 168 MW. The medians use each operator's own stated MW, which is usually critical IT or UPS capacity and sometimes a planned full build. Individual leases are smaller still: Avison Young classes retail colocation as leases under 250 kW and wholesale as leases of 250 kW to 4 MW. Definitions vary; we put wholesale colocation at 500 kW to 20 MW, with the dividing line usually around 500 kW.
Many tenants in cages and cabinets. Median size about 85,600 sq ft across the buildings with a published footprint. 34 buildings with a stated MW.
Metro Colo Advisory market data, operator-stated MW.Fewer, larger tenants leasing whole halls or suites, typically priced per kW per month. Median size about 139,000 sq ft. 45 buildings with a stated MW. See wholesale colocation.
Metro Colo Advisory market data, operator-stated MW.Multi-building campuses built for one or a few cloud or AI tenants. Median size about 929,000 sq ft. 44 campuses with a stated MW. Synergy Research counted 1,360 hyperscale data centers worldwide at the end of 2025.
Metro Colo Advisory market data; Synergy Research, April 7, 2026.Lease bands: Avison Young, Q2 2025 US data center update; wholesale pricing per kW: Inflect. Medians cover 281 verified buildings in Northern Virginia, Dallas-Fort Worth, Atlanta and Phoenix and count only rows with a stated MW figure.
One building, two very different MW figures
Operators often publish IT capacity and utility capacity side by side. Each panel is drawn on its own scale.
Centersquare DFW3
Dallas-Fort Worth colocation building, operator spec sheetStream DFW-C
Wilmer, Texas campus, planned full buildGalaxy Helios, Phase I
Texas AI campus leased to CoreWeaveSources: Centersquare DFW spec sheet; Stream Data Centers, DFW-C Wilmer; Galaxy, Phase I of Helios complete. Ratios are our arithmetic.
The largest AI data center campuses by power
From Metro Colo Advisory's ranking. Read the basis tag before comparing two rows: a nameplate figure, an IT figure and an unstated figure are not the same quantity.
Source: Metro Colo Advisory's AI Campus Power Tracker, Largest AI Campuses ranking, November 14, 2026, with sources per row: SpaceX prospectus, June 11, 2026 (Colossus I and II together, about 1,000 MW of compute power; installed GPU nameplate, not measured power draw; company-wide nameplate was 1 GW at March 31, 2026); Oracle Q1 FY2027 call, September 2026 (via Webull); Core Scientific Q3 2026 earnings deck (Denton about 260 MW, total billing as of November 5, 2026); Applied Digital; TeslaNorth on Tesla's Q3 2026 deck (Cortex 1 and 2 at more than 105 MW and 120 MW, our sum, a floor); Galaxy; TeraWulf Q3 2026 results; Meta (the same page also says "over two gigawatts"); WAM via UAE Embassy; Rio Grande do Sul government; Cowboy State Daily (2.7 GW; the same report says the project can scale to up to 10 GW); REBusiness Online; IURC Cause 46097 and About Amazon (St. Joseph County, "one of the Project Rainier sites"); The Energy Mag; Baxtel. Colossus is left out of the announced list: SpaceX's only forward figure is company-wide, nameplate compute "in the range of 2.5 to 3 gigawatts" expected by the end of the fourth quarter of 2026, with no site named (SpaceX Q3 2026 webcast, October 20, 2026). Company-wide nameplate was 1,700 MW at September 30, 2026 (SpaceX 10-Q).
The Four Kinds of Data Center MW
A data center megawatt figure can describe four different things: the grid connection, the total facility load, the critical IT load, or installed nameplate capacity that counts spare units, and on one site they can differ by 50% or more. They are measured at different points in the power chain, which is why two "100 MW" data centers can offer very different amounts of space to lease.
Grid or utility capacity
Measured at the utility meter or substation, often in MVAWhat the utility connection or service agreement is rated for. It caps everything downstream, and it is the figure in interconnection requests.
Example: Stream's DFW-C campus plans a 372 MW on-site substation for 240 MW of critical capacity.Total facility load
Measured at the facility's main meterEverything the building draws: IT plus cooling, lighting and losses in transformers and UPS systems. Cleanview defines it as "everything the building draws from the grid."
Example: Galaxy's Helios Phase I, about 200 MW of gross power.Critical IT load
Measured at the UPS outputPower delivered to servers, storage and network gear: "the power that reaches the servers." Colocation is leased against this number, and it is often what operators mean by MW.
Example: the same Helios Phase I includes 133 MW of critical IT load.Installed or nameplate capacity
Counted, not measuredEvery unit installed, including redundant ones, or every GPU multiplied by its rated power. Always larger than what can be used at once, and not a measure of what the site draws.
Example: SpaceX's 1.7 GW of nameplate compute at September 30, 2026.Sources: Stream Data Centers; Cleanview data center tracker (definitions); Galaxy; SpaceX 10-Q.
Our calculation: facility = IT x PUE; grid MVA = facility MW / power factor; installed = usable IT x 4/3 for three-plus-one N+1. This example uses the annual average PUE of 1.36, Uptime Institute's 2026 capacity-weighted average (Uptime Institute). The converter below sizes facility load and grid capacity on the design-day PUE instead (default 1.45, our assumption), so for 100 MW of IT its default answer is higher: 145 MW of facility load and about 152.6 MVA.
Worked example: SpaceX, 1,700 MW of nameplate
1,700 MWAt September 30, 2026, SpaceX counted 1,700 MW of nameplate compute across all of its data centers. It calculates the figure as "the number of GPUs installed in our data centers at the end of the period multiplied by their respective all-in power," and its filing says it does not represent actual power consumption or utilization. The filing adds that it leaves out power for supporting infrastructure such as cooling systems, power distribution losses, lighting, security systems and facility-level overhead. So 1,700 MW is an installed, IT-side number. If every GPU ran at its full rating at once, a 1.3 PUE would put the facility near 2,210 MW (our arithmetic), and the grid connection would need more still. Measured draw is lower whenever the GPUs run below their rating.
Source: SpaceX Form 10-Q, Key Business Metrics.Why the difference matters
Interconnection studies, substation transformers, backup generators and air permits are all sized on facility load, not IT load. Microsoft's draft air permit for its La Porte, Indiana campus lists 42 diesel generators rated 3,000 kWe and six rated 500 kWe, about 129 MW of standby (our arithmetic), and limits their combined NOx to 95 tons per twelve consecutive months. Cooling plants add a fifth unit: they are rated in thermal MW of heat removed, which is not the electricity they draw.
Source: IDEM draft air permit F091-50526-00197, Microsoft LPO Campus.Firms define MW differently
How research firms, trackers and operators state their megawatts, in their own words. When a basis is not stated, you cannot compare the figure to one that is.
| Who | What its MW measures | In its own words |
|---|---|---|
| Synergy ResearchHyperscale capacity counts | Critical IT | Capacity "measured by MW of critical IT load." |
| DC ByteMarket supply data | IT load | Reports IT load in MW "as opposed to MVA," the external utility feed. |
| Wood MackenzieData center dataset | IT Utility | Keeps them apart: existing IT capacity is "current power dedicated to IT equipment," while planned utility capacity is "Planned total utility power capacity in megawatts (MW) when the project reaches full build-out." |
| CleanviewData center tracker | Facility IT | Facility power is "everything the building draws from the grid"; IT capacity is "the power that reaches the servers." |
| Epoch AIAI data center research | Total facility | "All stated power capacities refer to total facility power, including power for GPUs, cooling, lighting, etc." |
| SpaceXCompany filing (parent of SpaceXAI, formerly xAI) | Nameplate | GPUs installed "multiplied by their respective all-in power." |
| SemiAnalysisIndustry model | IT, in use | Uses "critical IT power capacity" and "live IT load" without a public definition. |
| CBREMarket reports | Not stated | Tables "in Megawatts (MW)," with no IT or utility basis given on the public page. |
| JLLMarket reports | Not stated | Figures in GW with no IT or utility basis given on the public page. |
Sources: Synergy Research; DC Byte glossary (filed with Slough Borough Council); Wood Mackenzie dataset definitions; Cleanview; Epoch AI; SpaceX 10-Q; SemiAnalysis; CBRE press release; JLL H1 2026. Pages read November 2026.
Data center power terminology
The terms that show up in leases, offering memos and market reports, and what each one counts.
| Term | What it means |
|---|---|
| Critical IT load (MWc, kW IT) | Power available to IT equipment at the UPS output. Stream, for example, writes its campus capacity as "240 MWc." |
| UPS capacity | The rated output of the uninterruptible power supply systems, the usual ceiling on critical IT. Centersquare lists DFW3 at "UPS capacity: 4.8MW." |
| Utility capacity, MVA | What the utility feed or substation is rated for, often in MVA (apparent power) rather than MW. |
| Nameplate or installed | The sum of everything installed, including redundant units or the rated power of every GPU. Not a measure of use. |
| PUE | Power usage effectiveness: total facility energy divided by IT equipment energy. 1.0 would mean every watt reaches IT. |
| Power factor | Real power (W) divided by apparent power (VA). Converts MW to the MVA that grid gear is rated in. |
| N, N+1, 2N | Redundancy levels. N is the capacity needed at full IT load; N+1 adds one spare component; 2N mirrors the whole system. |
| Commissioned power | datacenterHawk: "Colocation power commissioned within the market, measured in MW, includes preleasing." |
| Live capacity | DC Byte: "Determined IT power that is operational whether it is let or not." |
| Committed (DC Byte) | A supply-certainty stage, not a leasing measure: "Committed supply does not mean sold space." |
| Net absorption | CBRE: "the amount of space leased compared to the amount vacated over a given period." |
Sources: Stream Data Centers; Centersquare DFW spec sheet; TechTarget on PUE; Schneider Electric; CoreSite; datacenterHawk; DC Byte glossary; CBRE.
What Is PUE? Power Usage Effectiveness Explained
PUE, or power usage effectiveness, is total facility energy divided by IT equipment energy: at the 2026 industry average of 1.52, a data center spends 0.52 kWh on cooling and power losses for every kWh that reaches its servers. The metric was created by members of The Green Grid, an industry group, and 1.0 is the floor, where every watt would reach IT. Uptime Institute calls the 1.52 average "a seven-year trend of relative stasis," but averages hide a wide spread: weighted by capacity, the figure falls to 1.36, because big new sites run better than small old ones.
Sources: TechTarget, What is power usage effectiveness (PUE)? ("PUE = Total facility power/IT equipment energy"); Uptime Institute Global Data Center Survey 2026. Worked examples are our arithmetic.
PUE calculation, step by step: take the energy the whole site drew over twelve months from the utility meters, take the energy the IT equipment drew at the UPS output over the same twelve months, and divide the first by the second. A site that drew 136 GWh in a year while its IT used 100 GWh has a PUE of 1.36. Use a full year: a single summer afternoon or winter night gives a different answer, which matters for capacity (see design PUE below).
Sources: Uptime Institute 2026 survey ("The 2026 industry average PUE is 1.52"); Uptime Institute Intelligence (1.36 capacity-weighted; large new sites "routinely report PUEs at 1.3 or lower"); Microsoft Azure blog, 2022 ("Our newest generation of datacenters have a design PUE of 1.12"); Microsoft, Measuring energy and water efficiency (global PUE 1.17 in FY25, for datacenters Microsoft fully owns and controls); Google Data Centers, Efficiency (1.09 fleet average, 2025).
Data center energy efficiency: what moves PUE
What lowers it
- Size and age. Uptime finds "Newer and larger data centers outperform the industry average, often by a significant margin."
- Cooling design. Most of the gap between a site's PUE and 1.0 is cooling and power conversion, so designs that cool with outside air or liquid shrink the top of the fraction.
- Running near full load. Fixed overheads spread across more IT when halls are full.
What raises it
- Hot, humid climates. Uptime lists climatic conditions among the factors that "all exert an influence on PUE."
- Resiliency choices. Redundant systems that idle on standby still draw power; Uptime names resiliency objectives as a driver.
- Partly empty halls. A new building leased at a fraction of its capacity carries full-size cooling for a small load.
Source: Uptime Institute Global Data Center Survey 2026 (quoted lines). Points without a quotation are our explanation of how the drivers work.
Design PUE vs annual PUE: why the hottest day sets capacity
Annual PUE averages a whole year, but cooling and power gear are sized for the hottest design day, so the IT load a site can support follows the higher design-day figure, not the average. Engineers calculate peak PUE "at the highest expected IT load and at the hottest design-day conditions," and "use extreme outdoor dry bulb and wet bulb temperature design set-points to size chillers." Microsoft said in 2022 that its actual operating PUE beat its designs in almost every region, which is the point: the design number is the conservative one, and it is the one that decides how many MW of IT a grid connection supports.
100 MVA at a 0.95 power factor is 95 MW of real power. Divided by the 1.36 annual average, it looks like room for 69.9 MW of IT.
Our arithmetic.Size on the hottest day, here an illustrative design PUE of 1.45, and the same connection supports 65.5 MW. The 4.3 MW gap is IT you cannot sell, even though the yearly average says it is there.
Our arithmetic. 1.45 is an assumption for illustration; ask the operator for the site's design-day figure.Sources: IES, Dynamic PUE modelling for data centres (quoted lines); Microsoft Azure blog, 2022 ("In almost every region, our actual operating PUE is more efficient than our designs"). The 69.9 and 65.5 MW figures and their 4.3 MW gap are our arithmetic from unrounded values.
MVA to MW: Power Factor Explained
To convert MVA to MW, multiply by the power factor: MW = MVA × power factor, so a 100 MVA connection delivers 95 MW of real power at a 0.95 power factor and 80 MW at 0.8. Watts are the real power equipment draws, volt-amperes are the apparent power that transformers and grid connections must carry, and "The ratio of the watts to VA is called the 'Power Factor.'" To go from MW to MVA, divide by the power factor instead.
| Power factor | MW from 100 MVA | MVA needed for 100 MW | Where you see it |
|---|---|---|---|
| 0.80 | 80 MW | 125.0 MVA | Rating of many legacy UPS systems |
| 0.90 | 90 MW | 111.1 MVA | For comparison |
| 0.95 | 95 MW | 105.3 MVA | Converter default, our working assumption |
| 0.99 | 99 MW | 101.0 MVA | Modern rectifiers often reach this or better |
Sources: Schneider Electric, VA vs watts; Fuji Electric, power factor in data center UPS selection ("Historically, many UPS systems were rated at 0.8 power factor"; modern rectifiers often achieve 0.99 or better). Table values are our arithmetic. We found no published source for one whole-site power factor, so the converter's 0.95 default is a working assumption.
Substations are often built so the site keeps running with one transformer out of service, written N-1. The remaining units must then carry the full facility load in MVA, so installed transformer capacity runs well above what the site draws. It is also why DC Byte reports IT load in MW "as opposed to MVA," the utility feed: the two are different quantities.
Redundancy: How N+1 and 2N Shrink Sellable Capacity
Redundant equipment protects uptime but adds no capacity you can lease: with four UPS modules in N+1, the load of three is shared across four, so a quarter of installed capacity is held in reserve and the installed total overstates leasable load by a third; 2N doubles it. CoreSite defines N as "the minimum capacity needed to power or cool a data center at full IT load." N+1 adds "a single component," which takes over when one unit is offline, and 2N "creates a mirror image of the original UPS, cooling system or generator arrangement." POWER Magazine's guide to Uptime's Tier III standard says the tier "ensures N+1 redundancy, allowing maintenance without downtime."
Sources: CoreSite, What is Data Center Redundancy? N, N+1, 2N, 2N+1 (definitions; its 2N example pairs four UPS units with "an additional four UPS units, for a total of eight systems"); POWER Magazine on Uptime's Tier III standard ("This tier ensures N+1 redundancy, allowing maintenance without downtime").
The practical question for any headline figure: does it count the spare modules? A 2N building quoted at its installed UPS total has half that to lease. The converter below lets you switch between N, N+1 and 2N and see the usable IT load move.
From Requested to Drawing: The Stages of a Megawatt
A data center megawatt moves through four stages: requested from the utility, signed under an agreement, energized and ready to serve, and actually drawn by running servers, and most announced megawatts are still in the first two. The gap is not small. Texas's grid operator told the state Senate on July 29, 2026 that about 474 GW of large loads were seeking to connect (counted as of June 2026), while the large loads it has approved to energize reached a combined monthly peak of about 4.1 GW in September 2026. That figure adds each load's own peak for the month, so it is not load running at one time.
Requested
A load request, study or letter of intent with the utility. It shows interest, not a commitment to serve. ERCOT's 474 GW queue is this stage.
Signed
An executed agreement with money behind it: a construction letter of agreement or an electric service agreement. Dominion Energy had 12.2 GW under electric service agreements as of September 2026.
Energized
The utility has connected the load and it may draw power. Approved to energize is a ceiling, not a reading: ERCOT had 9.5 GW approved as of September 2026.
Drawing
Servers running and metered. Customers ramp into their capacity over months or years, so this lags energized: a combined monthly peak of about 4.1 GW in ERCOT in September 2026, and a 4.8 GW data center billing peak at Dominion.
Sources: ERCOT Senate panel deck, July 29, 2026 ("approximately 474 GW of Large Loads seeking interconnection, of which ~90% are data centers," counted as of June 2026); ERCOT Monthly Operational Overview, September 2026, p. 9 ("Of the 9,456 MW that have received Approval to Energize," a "non-simultaneous monthly peak consumption of 4,132 MW in September 2026"). The same figures appear in Metro Colo Advisory's AI Power Reality Check.
Sources: Dominion Energy Q3 2026 earnings call slides (stage figures and definitions); billing peak from Metro Colo Advisory's AI Power Reality Check.
Ramp schedules: contracted in full, energized in blocks
Utility contracts expect customers to grow into their capacity over years, and they price that ramp. AEP Ohio's data center tariff, for loads of 25 MW or more, says "The load ramp period will not exceed four years," sets the contract term at the ramp plus eight years, and bills no less than the greater of 85% of the customer's highest monthly billing demand in the past 11 months or a minimum set from contract capacity (for a site above 75 MW, 57.5 MW plus all capacity above 75 MW, capped at 85% of contract capacity). On the supply side, Talen's 1,920 MW nuclear deal with Amazon says the delivery schedule "will ramp over time, expecting to achieve the full volume no later than 2032."
Sources: AEP Ohio, Data Center Tariff (PUCO docket 24-508-EL-ATA); Talen Energy release, June 11, 2025.
Phased builds: announced vs delivered
Large campuses are built and leased in blocks. What each company says is in service against its stated full build or, where none is given, the sum of its stated phases.
Sources: Oracle Q1 FY2027 call, September 2026 (via Webull) (618 MW, "accounting for 75% of the total capacity"); Crusoe states 1.2 GW of total power for Abilene, a different basis, so we do not divide one by the other; Applied Digital, October 2026; Core Scientific Q3 2026 earnings deck (Denton about 260 MW, total billing as of November 5, 2026) and Core Scientific 8-K, April 2026 (262 MW planned); TeraWulf Q3 2026 results; Galaxy, July 2026. Each bar uses the company's own basis, so compare shares, not MW, across rows.
The same pattern holds on the power plant side. Our AI Campus Power Tracker counts 98.3 GW of dedicated power planned for named US data center campuses, and 2.0 GW of it, about 2%, confirmed running. 640 MW of that 2.0 GW, Google's Quantum solar project in Texas, rests on its owner's statement alone; without it, 1.36 GW is confirmed running. Our AI Power Reality Check follows requested, signed and running data center power utility by utility, so you can see which stage a market's headline is at.
Source: Metro Colo Advisory's AI Campus Power Tracker, November 14, 2026 (142 of 197 projects count toward totals).
Data Center MW Converter: IT Load, Facility Load and MVA
Enter any megawatt figure, say what it measures, and the converter returns the rest: usable IT load, total facility load, grid capacity in MVA and installed nameplate. It applies the PUE formula, the power factor and the redundancy level you pick, and shows what a year-one ramp leaves you to use. The defaults are an annual PUE of 1.36 (Uptime's 2026 capacity-weighted average), a design-day PUE of 1.45 (our assumption, set higher because cooling is sized for the hottest day), and a 0.95 power factor (a working assumption).
Data Center MW Converter
Turn any megawatt figure into the other three: usable IT load, total facility load, grid capacity in MVA, and installed nameplate. Change the assumptions to match the site.
Sources for defaults: Uptime Institute, capacity-weighted PUE 1.36; Uptime Institute, average PUE 1.52; Fuji Electric on power factor. Results are our arithmetic and a starting point, not a site assessment.
kW per Rack: Typical vs AI Racks
The average of the most typical rack densities reported to Uptime's 2026 survey passed 11 kW for the first time, while an NVIDIA GB200 NVL72 rack draws about 120 to 130 kW, more than ten times as much in the same footprint. Leave out high-density sites and the survey average is 7.8 kW. 24% of survey respondents say they have some racks at 30 kW or more, up from 19% in 2025, and the majority of facilities still have none.
Sources: Uptime Institute Global Data Center Survey 2026; NVIDIA DGX GB200 user guide ("The rack power consumption is approximately 120kW"); Supermicro SuperCluster GB200 NVL72 datasheet ("Operating Power 125kW to 135kW," 72 GPUs, liquid cooled).
That gap is why AI tenants cannot simply take space in an ordinary colocation hall: a single 130 kW rack draws what a dozen typical racks do, and almost all of it leaves as heat, which is why the GB200 NVL72 is built for liquid cooling. Our high density data center guide covers the buildings and liquid cooling that can take these racks.
How Much Energy Does AI Use?
Per prompt, the companies report small numbers: Google puts its median Gemini text prompt at 0.24 Wh and OpenAI's CEO puts the average ChatGPT query at about 0.34 Wh. The totals are large: LBNL projects AI servers will use 55% of all US data center energy by 2030. Google's figure covers answering prompts (inference); OpenAI's CEO gave no method or scope. Both are the companies' own numbers. An older third-party estimate, a 2023 study cited by EPRI, was about 9 to 12 times higher.
Google says the per-prompt energy is "equivalent to watching TV for less than nine seconds."
Source: Google Cloud blog, Measuring the environmental impact of AI inference."About what an oven would use in a little over one second." Stated in a blog post without a published method.
Source: Sam Altman, The Gentle Singularity.Against 0.3 Wh for a Google search. An outside estimate from 2023, two years before either company figure.
Source: EPRI, Powering Intelligence (2024), citing de Vries, Joule (2023).LBNL's reference case has AI servers at 84% of server energy and 55% of total data center energy by 2030. This is a projection, not today's share.
Source: LBNL 2025 Update, Section 3.6.The IEA finds accelerated servers account for almost half of the net increase in global data center electricity use in its base case.
Source: IEA, Energy and AI.Questions to Ask of Any MW Figure
Before you compare two sites or underwrite one, pin down six things about every megawatt figure: where it was measured, at what PUE, whether redundancy is in or out, which stage it is at, whether it is demonstrated or design, and whether the power is firm or can be cut. Check how many of them the document in front of you answers; each one left open changes the number you should be using.
Where was it measured?
At the utility meter or substation, at the facility main, or at the UPS output. Each step down the chain is a smaller number.
Ask: is this utility, facility or critical IT MW?At what PUE?
Converting between facility and IT load needs a PUE, and the annual average flatters capacity. The hottest design day sets what the site can support.
Ask: what is the design-day PUE, and the measured annual PUE?Is redundancy in or out?
Installed totals that count N+1 or 2N spares overstate what can be leased by a third (four modules in N+1) to double (2N).
Ask: is the figure installed capacity or usable after redundancy?Which stage is it at?
Requested, signed, energized or drawing. A requested MW and a drawing MW can be years apart.
Ask: show the signed service agreement and the energization date.Demonstrated or design?
A design figure is what the plans allow. Demonstrated capacity has run at full IT load, at peak outdoor temperature, through commissioning.
Ask: has the site carried this load on a design day?Firm or non-firm?
Non-firm service can be curtailed when the grid is tight. FERC's December 2025 order for PJM directs separate Firm Contract Demand and Non-Firm Contract Demand transmission services for co-located loads.
Ask: how many MW are firm, and what triggers curtailment?Source for firm and non-firm services: Jones Day on FERC's December 18, 2025 PJM co-location order.
Have Us Check the MW Behind the Headline
We read the filings behind a site's power claims and tell you, in writing, what the megawatts are, what stage they are at, and what they support in usable IT load. Our fixed-fee consulting work covers single sites to portfolios. Engagements start at $45,000.
Developers and landowners
Turn a utility letter into a leasable IT figure a tenant will accept, and see where the redundancy, PUE and ramp assumptions put you against a powered shell or build to suit buyer's bar.
Buyers and tenants
Know whether the 20 MW on the proposal is critical IT you can load in year one or a utility figure that shrinks by a quarter to a third once cooling and losses come out.
Investors and lenders
Check an offering memo's MW against interconnection filings, service agreements and energization dates before the number goes into a model.
Data Center Power FAQ
How much power does a data center use?
It depends on the type. In our market data the median retail colocation building is rated about 6.6 MW, a wholesale building about 33 MW and a leased hyperscale campus about 168 MW, mostly stated as critical IT load. The largest AI campuses are far bigger: Meta's Hyperion is announced at 5,000 MW, and SpaceX counts 1,700 MW of installed GPU power across all of its data centers as of September 30, 2026; its June 2026 prospectus put Colossus I and II at about 1,000 MW. Altogether, US data centers used 192 TWh in 2024.
How much electricity do data centers use in the US?
US data centers used 192 TWh in 2024, 4.7% of US electricity and 14% more than in 2023, according to Lawrence Berkeley National Laboratory's 2025 Update. LBNL's reference case reaches 464 TWh in 2028 and 649 TWh, 11.8% of US electricity, by 2030, within a range of 521 to 843 TWh. Worldwide, the IEA puts data center use at about 485 TWh in 2025, up 17% in a year, and expects about 950 TWh by 2030.
How many homes can 1 MW power?
A 1 MW load running every hour of the day uses 24 MWh (24,000 kWh) a day and 8,760 MWh a year. The average US home used 10,791 kWh a year, according to the EIA, so 1 MW running all year matches about 812 homes, and 1 GW about 812,000. Real data center loads run below their peak, so treat this as a ceiling.
What is PUE in a data center?
PUE, or power usage effectiveness, is total facility energy divided by IT equipment energy. A PUE of 1.52, the 2026 industry average in Uptime Institute's survey, means the site uses 0.52 kWh for cooling and power losses for every kWh that reaches IT. Weighted by capacity the average is 1.36, and Google's fleet ran at 1.09 in 2025.
How do you calculate PUE?
Divide the energy the whole facility drew over a year by the energy the IT equipment drew at the UPS output over the same year. A site that used 136 GWh while its IT used 100 GWh has a PUE of 1.36. The same formula converts loads: facility MW equals IT MW times PUE, so 100 MW of IT at a PUE of 1.3 needs 130 MW at the meter, and 100 MW at the meter leaves about 77 MW for IT.
How do you convert MVA to MW?
Multiply MVA by the power factor: MW = MVA x power factor. A 100 MVA connection delivers 95 MW at a 0.95 power factor and 80 MW at 0.8. To go from MW to MVA, divide by the power factor, so 100 MW needs about 105.3 MVA at 0.95. Many legacy UPS systems were rated at 0.8, while modern rectifiers often reach 0.99.
How much energy does AI use?
Per prompt, the company figures are small: Google reports 0.24 Wh for a median Gemini text prompt and OpenAI's CEO about 0.34 Wh for an average ChatGPT query, both company-reported. An older third-party estimate, a 2023 study by Alex de Vries cited in EPRI's 2024 paper, put a ChatGPT request at 2.9 Wh. In total, LBNL projects AI servers will use 55% of all US data center energy by 2030, and the IEA finds accelerated servers drive almost half of global data center growth.
What is a typical kW per rack?
Uptime Institute's 2026 survey found the average of the most typical rack densities its respondents reported passed 11 kW for the first time, or 7.8 kW without high-density sites. 24% of survey respondents say they have some racks at 30 kW or more. An NVIDIA GB200 NVL72 AI rack draws about 120 to 130 kW (NVIDIA puts it at approximately 120 kW, Supermicro at 125 to 135 kW), more than ten times that average.
Cite this page
Metro Colo Advisory, Data Center Power Consumption, updated November 2026, https://metrocoloadvisory.com/data-center-power-consumption/
Free to use in news coverage and research under a Creative Commons Attribution 4.0 license, with credit to Metro Colo Advisory and a link to this page.
AI Power Reality CheckData center power requested, signed and running at 52 US utilities and 6 grid operators, quarter by quarter.
AI Campus Power TrackerThe power plants planned for named AI campuses, how much is running, and the largest campuses ranked by MW.
Data center consultingFixed-fee advisory for developers, owners, buyers and investors, from a single site to a portfolio.
How this page was built, and corrections
Our research team read every outside figure on this page from the source linked under it in November 2026 and quoted the wording where a definition matters. Building medians come from Metro Colo Advisory's verified building data for four US markets and use each operator's own stated MW, whose basis varies. Campus figures come from our AI Campus Power Tracker and keep each company's basis (IT, nameplate, facility or not stated). Homes equivalents, PUE and power factor conversions, redundancy shares and phased-build percentages are our arithmetic from the cited figures. The 1.45 design-day PUE and 0.95 power factor are stated assumptions, not survey results. To report an error, write to us through the contact page.