Data Centers and the Grid

Data Centers and the Grid

LEARN / REINDUSTRIALIZATION

Data Centers and the Grid

Data centers consumed 4.4 percent of American electricity in 2023 and are projected to reach between 6.7 and 12 percent by 2028. The widely quoted gigawatt figure derived from that projection rests on an assumed utilisation rate stated in the report and dropped almost everywhere it is cited.

Key points

What is actually consumed

The Energy Act of 2020 directed the Department of Energy to update an earlier Berkeley Lab study of data center electricity use. The resulting 2024 report is the closest thing to an authoritative American baseline, and its historical findings are more useful than its projections because they describe what happened.

Data58 → 176 TWh
U.S. data center electricity consumption, 2014 to 2023, reaching 4.4 percent of total national electricity use. Consumption sat near 60 TWh through 2016, reached about 76 TWh and 1.9 percent by 2018, then accelerated. Source: Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report.

Series and provenance
Agency
Lawrence Berkeley National Laboratory for the U.S. Department of Energy
Program
2024 United States Data Center Energy Usage Report
Series
Bottom-up estimate mandated by the Energy Act of 2020
Measure
U.S. data center electricity consumption and projection
Units
Terawatt-hours per year; percent of national consumption
Adjustment
Bottom-up from equipment shipments and installed base
Period
History 2014 to 2023; projection to 2028
Latest
176 TWh and 4.4 percent in 2023; 325 to 580 TWh and 6.7 to 12.0 percent projected for 2028
Source tier
Primary
Retrieved
August 9, 2026

The report converts its energy projection into 74 to 132 GW of power demand by assuming an average capacity utilisation rate of 50 percent, an assumption stated in the report and omitted from most citations. The authors also note that direct metered energy data for the sector does not exist, which limits a bottom-up estimate in a segment changing this quickly.

Open the source

The shape of that curve matters more than its endpoint. Compound annual growth ran near 7 percent between 2014 and 2018 and 18 percent between 2018 and 2023. The inflection sits at 2017, when the installed base of servers began growing again and graphics processing units for artificial intelligence became a significant share of the stock. What changed was not that computing grew, but that a more power-dense form of it did.

The projection, and what it assumes

Berkeley Lab projects 325 to 580 TWh by 2028, equivalent to 6.7 to 12.0 percent of forecast national consumption, with compound growth between 13 and 27 percent. The range is wide because the sector is changing faster than the evidence base describing it.

Measurement noteThe report converts that energy range into a power figure of 74 to 132 GW by assuming an average capacity utilisation rate of 50 percent. That assumption is stated plainly in the report and is dropped in most citations, which present 74 to 132 GW as though it were measured. Utilisation is the single lever: a facility fleet running at 60 percent rather than 50 percent would need proportionally less installed power for the same energy. Any gigawatt figure derived from an energy projection carries a utilisation assumption, and a reader should ask what it was.

The authors also state the constraint on their own work. The estimate is built bottom-up from equipment shipments and installed base rather than from metered consumption, because direct energy data for the sector does not exist, and they note that this limits the analysis in a segment changing as quickly as this one.

Institute analysisA projection produced by a national laboratory under a congressional mandate, which states its own utilisation assumption and its own data limitation, is more useful than a confident number from a source that states neither. The quality of a forecast is largely visible in how much of its own uncertainty it publishes, and this is the rare case where the primary source is more candid than the commentary built on it.

How data centers behave as load

A large data center draws close to its maximum continuously, which gives it a high load factor. That makes it a heavy consumer of energy relative to its peak, and a comparatively easy load to plan for, because it does not concentrate demand into the hours the system already finds hardest.

The complication is that this load arrives in very large individual increments at single points on the network. A single campus can request more power than a mid-sized city uses, at one substation, on a schedule set by a construction programme rather than by the utility planning cycle. The problem is less the total than the concentration and the speed.

Flexibility changes the calculation substantially. NERC notes that because new large loads can be curtailed during energy emergencies, their effect on planning reserve margins is smaller than their raw size implies, which is why curtailment terms have become a central negotiating point rather than a technicality.

Why siting concentrates

Data centers cluster, and the clustering is not accidental. Fibre routes, land, tax treatment, water availability and speed of grid connection all pull in the same direction, which produces regional concentrations that stress particular systems rather than the national one. Dallas County alone is projected to hold roughly 10 GW of large load by 2030, and ERCOT has identified billions of dollars of transmission projects in response.

For industrial capacity this concentration has a direct consequence. A manufacturer siting in a region where data center load is already queued is competing for the same interconnection studies, the same transformers and the same construction labour, and is generally competing against a counterparty with a larger balance sheet and a shorter decision cycle.

Common misconceptions

That the gigawatt figures are measurements. The commonly quoted 74 to 132 GW is derived from an energy projection using an assumed 50 percent utilisation rate. It describes a modelled implication rather than an observation.

That efficiency gains will resolve it. Efficiency per computation has improved throughout the period in which total consumption tripled. Whether efficiency reduces total demand depends on whether it also reduces the amount of computing purchased.

That data center load competes with industrial load nationally. The competition is regional and specific: the same substation, the same queue, the same long-lead equipment. National totals conceal where it actually binds.

What the evidence says, and where it is contested

ContestedHow much of the projected load will materialise is genuinely disputed. One position holds that the compute demand behind it is real and capitalised, that firms filing requests have the balance sheets to build, and that the constraint will be power availability rather than willingness to pay. A second holds that the projections extrapolate an inflection that may not persist, that model efficiency has improved faster than expected in specific cases, and that the gap between requested and energised load is already documented at individual utilities. Both cite real evidence. Berkeley Lab’s own range, from 325 to 580 TWh, encodes almost the whole disagreement inside a single official projection.

Related Institute research

Why Electricity Demand Is Growing Again
Queue figures, utility forecasts and committed load.

How Electricity Demand Is Measured
Why energy and peak are not interchangeable.

Powering the Buildout
The Institute’s projections across data centers, manufacturing and defense.

Sources

Reference entry maintained by the Institute for American Manufacturing & Technology. Figures are drawn from primary sources and cited above. Where the Institute states a position rather than a finding, it is marked as such.

Private Professional Network

Institute for American Manufacturing & Technology

Join the IAMT Network. Connect with researchers, policymakers, and industry leaders working on American manufacturing, energy, and technology policy.

Apply for Membership →