A typical hyperscale data center draws 100 to 1,000 megawatts continuously, the equivalent of a small to mid-sized city. The number that matters for your community is how that load is paid for and whether residential rates change to fund new generation and transmission.
Annual consumption of a single 100MW hyperscale facility, equivalent to 80,000 to 100,000 homes
Data center share of US electricity, projected 2023 to 2028 if current build-out velocity continues
Share of Virginia electricity already consumed by data centers, the densest US cluster
A modern data center rack runs at 6 to 10 kilowatts continuously for traditional cloud workloads. AI training and inference at high server density push that to 30 to 80 kilowatts per rack and beyond. A 100-megawatt facility houses 1,500 to 5,000 racks depending on density, plus cooling, lighting, and ancillary load.
Unlike residential demand, which peaks in the evening, data center load is continuous. The grid sees a flat 100-megawatt baseload from a single industrial customer, 24 hours a day, 365 days a year.
The first source is the existing regional grid. Most large data center applications are now requiring new generation to be built explicitly to serve the project. Power purchase agreements (PPAs) with new wind, solar, and natural gas generation are the most common structure. Geothermal and small modular nuclear are emerging.
The structural problem: a new combined-cycle gas plant takes 3 to 5 years to build, a new transmission line 5 to 10 years, and new nuclear 10 years or more. Data center applications are arriving faster than new generation can be approved and built. This is why grid operators across MISO, PJM, ERCOT, and CAISO have flagged data center clustering as a reliability concern.
The honest answer: it depends on rate-class allocation. Three structural questions matter:
Public utility commissions in Virginia, Ohio, Indiana, and Georgia are actively considering rate-class separation specifically because of data center load growth. The political fight is about who pays for the grid the data centers need.
The single most useful question to ask at a public hearing: what is the projected continuous load of this facility, what new generation and transmission are required to serve it, who pays for that build-out, and is there a residential rate-protection mechanism?
A specific answer to that question, from the developer and the local utility, with a number and a paid-by name on the new infrastructure, is the difference between an informed council vote and a surprise ratepayer revolt three years later.
Independent analysis. Same numbers for every reader.
Data sourced from US Department of Energy, EIA, EPRI, regional transmission operators, and operator environmental reports. Methodology is public and reproducible.
Read the full methodologyQuestions we hear
Sourced from US DOE, EIA, EPRI, and operator environmental reports.
A typical hyperscale data center campus draws between 100 and 1,000 megawatts continuously. A 100-megawatt facility consumes roughly 876,000 megawatt-hours per year, the same as 80,000 to 100,000 average US homes. Smaller enterprise data centers draw 1 to 50 megawatts. The largest announced hyperscale and AI training campuses target 1 gigawatt or more.
AI training campuses are pushing power per rack from 6 to 10 kilowatts to 30 to 80 kilowatts and beyond. A 200-megawatt AI training cluster running at high utilization consumes roughly 1.7 terawatt-hours per year, comparable to a city of 150,000 people. The next generation of frontier-model training facilities is being designed at 500 megawatts to 1 gigawatt of continuous IT load.
A 100-megawatt facility uses 2,400 megawatt-hours per day at full operation. A 500-megawatt campus uses 12,000 megawatt-hours per day. The number is fairly stable across the year because cooling and IT loads run continuously, unlike residential demand which peaks evenings and weekends.
For a 100-megawatt facility running at typical utilization, annual consumption is roughly 700,000 to 876,000 megawatt-hours. US data center electricity consumption was about 4.4 percent of national demand in 2023 and is projected to reach 9 to 12 percent by 2028 if current build-out continues. Virginia is the densest cluster, with data centers consuming over 25 percent of state electricity.
Google reported 25.9 terawatt-hours of total electricity consumption across its data center fleet in its 2024 environmental report, equivalent to a country like Ireland. Per facility, average annual consumption is in the 700,000 to 1.2 million megawatt-hour range. Individual sites vary significantly with their IT density and cooling design.
It depends on whether the local utility is allowed to allocate the cost of new generation, transmission, and distribution to the rate base. Several state public utility commissions are now reviewing or approving rate-class separation, where data center load is billed at higher industrial rates and not pooled with residential. Without that separation, residential bills can rise to fund grid expansion that primarily serves a single industrial customer. The disclosure question to ask: which rate class does this facility fall under, and is there a cost-allocation mechanism that protects residential rates?
It comes from whatever generation mix is on the regional grid, plus any new generation built to serve the project. Most large new data centers are negotiating power purchase agreements that fund new wind, solar, geothermal, or natural gas generation. Some are co-locating with nuclear plants. The disclosure question for any new project: what new generation is being built, and is it carbon-aligned with the operator commitments?
Data centers consumed approximately 4.4 percent of US electricity in 2023 (about 176 terawatt-hours). Department of Energy and EPRI projections show this rising to 6.7 to 12 percent by 2028, depending on AI build-out velocity. Virginia, Texas, and Arizona have the highest data center share of state load, with Virginia exceeding 25 percent.
There is no single average. Enterprise data centers (under 5 megawatts) make up most facilities by count but a small share of total load. Hyperscale facilities (100 megawatts and up) make up a small share by count but dominate total load. The "average" depends on which population you measure. By total load, the typical megawatt of US data center capacity comes from a hyperscale facility.
The data center operator pays the utility directly for the electricity consumed. The complication is the cost of grid expansion, new transmission, and new generation needed to serve the project. Those costs are negotiated between the operator, the utility, and the public utility commission. They can be borne by the operator (developer-funded), spread across all rate classes (rate base), or split. Whose pocket the cost ultimately comes out of is the question communities should ask.
Hyperscale facilities, typically 100 megawatts of IT load and up, consume 876,000 megawatt-hours or more annually per site. Multi-building campuses can multiply that: a 4-building, 100-megawatt-each campus draws 3.5 terawatt-hours per year, comparable to a small US state.
New AI training campuses are designed at 200 to 500 megawatts of continuous IT load, with the leading edge pushing toward 1 gigawatt. A 500-megawatt facility consumes about 4 terawatt-hours per year. The latest AI training runs alone consume tens of gigawatt-hours over weeks. Inference workloads at scale (chatbot serving) are now matching the training scale in aggregate.
Data centers are large, constant, hard-to-curtail loads. Adding several gigawatts of new continuous load in one region tightens reserve margins, lengthens transmission queues, and can drive up wholesale power prices. Several Regional Transmission Organizations have flagged data center clustering as a reliability concern. New generation and transmission build-out lags behind data center applications by several years.
Direct blackouts are unlikely; grid operators have many tools to balance load. The realistic risk is rolling brownouts during peak summer or winter events when generation is constrained. The bigger risk for ratepayers is that the cost of preventing reliability problems (new transmission, new generation, capacity payments) gets passed through to residential bills. Specific market-by-market analysis is what matters.
Send this to someone who can act
Pre-written templates you can review, edit, and send. No signup, no tracking. Pick who you want to reach.
Email the developer or applicant
Hi, I am a community member with concerns about this proposed data center. The questions on my mind are specific: 1. What is the projected continuous load, and what new generation and transmission are required to serve it? 2. Who pays for the new infrastructure: the operator, or all rate-class customers? 3. Is there a residential rate-protection mechanism in place before this load arrives? I came across LSARS, an independent health and environmental data platform. The applicant pays for the analysis but the methodology is published (EPA AirToxScreen, California OEHHA), so the developer cannot change the numbers. The same report goes to the council, the community, and the applicant. Would you consider funding an independent LSARS analysis for this project? It would address the questions above transparently, before they become a public hearing problem. Reference page I read: https://www.lsars.com/data-center-electricity-use Developer overview: https://www.lsars.com/for/developers Thank you for considering it. [Your name]
The message opens in your default email app where you can edit before sending. Replace [Your name] and any bracketed placeholders. The reference link points back to this page so the recipient sees exactly what you read.
Project-specific continuous load, paid-by structure for new infrastructure, and residential rate-impact analysis on a public dashboard.
See your community page