The most repeated claim in AI infrastructure is that data centers are about to strain the electricity grid. The official US numbers behind it are real, but they are a mix of measured history, scenario ranges and assumptions, and the sources rarely say how much of the load is AI. This article reads four US government publications closely to separate those layers, so you can tell a measurement from a projection when you next see a data center electricity use figure.
What is measured: 2014 to 2023
The Department of Energy announced a Lawrence Berkeley National Laboratory report on 20 December 2024. The announcement says total data center electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023, that load growth has tripled over the past decade, and that data centers used about 4.4% of US electricity in 2023.[1]
Two points of care. First, this is an estimate built by a national laboratory, not a meter reading, and the announcement page we read does not describe the method or give error margins. We did not read the full report, so we cannot say how it handles, for example, small server rooms or colocation. Second, "data centers" here means all of them: web hosting, video streaming, enterprise IT, cloud and AI. The page names AI applications as one driver of rising demand, alongside data center expansion, domestic manufacturing growth and industrial electrification, but gives no AI share.[1]
What is measured at the grid edge
A second, independent view comes from utility sales data. EIA's analysis of 2019 to 2023 shows US commercial electricity sales were only 14 billion kWh, or 1%, higher in 2023 than in 2019. Underneath that flat total, the ten fastest-growing states added 42 billion kWh (10%) while the other forty fell by 28 billion kWh (3%).[2]
| State | Change in commercial sales, 2019 to 2023 |
|---|---|
| Virginia | +14 billion kWh, the largest increase |
| Texas | +13 billion kWh |
| North Dakota | +37% (2.6 billion kWh), the fastest relative growth |
EIA ties this growth to new computing facilities, and says Virginia had 94 new data center facilities connected since 2019, a count it takes from a Dominion Energy investor presentation.[2] It also attributes Texas's growth partly to low electricity and land costs that drew data centers and cryptocurrency mining. So the load is geographically concentrated, which matters more to a grid operator than the national percentage does. EIA is also careful to say data center developments are "evolving rapidly" and that it would re-evaluate forecasts as it learned more.[2]
What is projected: 2028
DOE's announcement gives the 2028 range: 325 to 580 TWh, or 6.7% to 12% of US electricity.[1] The range is wide because it has to be: it spans very different assumptions about how many accelerators are bought and how efficiently they are used.
A small piece of arithmetic of our own, which is an inference and not a figure from the source: 176 TWh at 4.4% implies total US consumption near 4,000 TWh, while 325 TWh at 6.7% and 580 TWh at 12% both imply a total near 4,850 TWh. If the 2028 shares are computed on a larger national total, then part of the headline share increase is the denominator moving, not only the data centers. We cannot confirm that from the announcement alone, so treat it as a question to put to the full report rather than a finding.
What is assumed: 2050
EIA's Annual Energy Outlook 2026, released 8 April 2026, is the clearest example of a projection that states its assumptions. EIA's May 2026 note says server electricity was about 7% of commercial-sector electricity in 2025 in all cases, and reaches 22% to 33% by 2050, or 446 to 818 billion kWh of server consumption.[3]
The gap between the cases is made of explicit choices:
- The High Electricity Demand case assumes no server efficiency improvement and a larger share of AI servers in the installed stock than the baseline.[3]
- The baseline assumes that, after 2040, average annual operational power draw per server falls 10% every three years, which EIA describes as beyond historical trends.[3]
- Servers are modelled with an essentially flat load shape, meaning demand is the same in every hour of the day.[3]
- Data center space needs up to 2.9 times the cooling of non-data-center floorspace, on average.[3]
EIA states plainly that these are long-term projections, not measured data, and that AEO2026 mostly considers only laws and regulations as of December 2025.[3] That framing is the useful one for any forecast you read: find the assumption that moves the answer. Here it is server efficiency.
What the near term looks like
EIA's September 2026 release expects US electricity sales of 4,135 billion kWh in 2026, up almost 2%, and 4,211 billion kWh in 2027, up about 2% more. It says demand growth is particularly strong in the West South Central region because of data center development, and that the region remains the largest regional contributor to sales growth "despite a pause in new data center projects in Texas." The release gives no data-center-specific figures.[4]
A reading guide
| Kind of claim | Example | How much weight |
|---|---|---|
| Measured or estimated history | 58 TWh in 2014 to 176 TWh in 2023 | High, with the method caveat above |
| Regional sales data | Virginia +14 billion kWh, 2019 to 2023 | High for sales, attribution to data centers is EIA's reading |
| Near-term forecast | 2026 and 2027 US sales | Moderate, revised regularly |
| Scenario range | 325 to 580 TWh by 2028 | Treat as bounds, not a prediction |
| Long-range case | 22% to 33% of commercial electricity by 2050 | Depends on stated assumptions |
| Claim about "AI's" share | Frequently asserted | Not isolated in the pages we read |
What this means for AI builders
- Cost and availability follow the grid. Power availability in a few regions is a plausible constraint on where capacity grows. That is our interpretation of the regional concentration above; none of the pages say it will limit any specific provider.
- Efficiency is the swing factor. Both the 2028 range and the 2050 cases turn on how much work each watt delivers. The serving techniques in Inference Serving Explained are part of that story, but we have no measurement of their aggregate effect.
- Be sceptical of per-query figures. If you see a watt-hour-per-prompt number, ask for the model, hardware, batch size and the boundary of what was counted; the same discipline as in How to Read an AI Benchmark Claim. We did not evaluate any per-query claim here.
- Cost per token is a separate question. Electricity is one input to the economics in Tokens, Context Windows and What an AI Request Really Costs; the sources here say nothing about prices.
Where to go next
The open questions are the ones the sources leave unanswered: how much of the 2023 load was AI-specific, how the 2028 shares were computed, and how fast efficiency gains are being offset by more demand. When a primary source answers them, this page will be updated. Until then, the honest summary is a measured tripling over a decade, a wide range ahead, and an AI share that is assumed more often than it is measured. For the policy side of AI regulation, see EU AI Act: General-Purpose AI Model Obligations Explained.
