Not Well Documented: Berkeley Lab's 2025 Update on United States Data Center Energy Use

Lawrence Berkeley National Laboratory published its "United States Data Center Energy Usage Report: 2025 Update" in June 2026. It updates the laboratory's report of December 2024, "leveraging the same methodology and many of the same assumptions while incorporating updated datasets for information technology (IT) equipment shipments and revised assumptions about equipment power consumption as of December 2025." The research had support from the Department of Energy's Industrial Technologies Office.

Its headline numbers are about energy over a year. The update estimates that data centers used 192 TWh of electricity in 2024, 4.7% of the national total. Its Reference Case forecast for 2030 is 649 TWh, which it puts at 11.8% of the electricity use forecast for that year in the North American Electric Reliability Corporation's 2025 Long-Term Reliability Assessment. Its Compounded Uncertainty range for 2030, a forecast too, runs from 521 TWh to 843 TWh.

A grid connection is measured another way. The report's glossary defines grid interconnection capacity as "The maximum amount of electrical power a data center requests from the utility or grid operator at its point of connection to the grid, typically established through a formal interconnection request." Parts of the update follow, not the whole.

How Much of the Connection

"Data centers are often discussed in terms of power capacity," the update says, and it calls converting "between estimates of annual electricity consumption and power capacity" challenging without data on how average demand, maximum demand, nameplate capacity and requested interconnection capacity relate. The next sentence gives an estimate. "Current utilization of interconnection capacity is not well documented but is estimated to be around 50% due to high redundancy requirements and maintenance needs."

The update adds that there is "a concerted effort in the data center industry to increase interconnection capacity utilization through innovations in power system design, energy balancing, and demand management." It also says that the capacity of power generation that must be added to the grid to serve new loads "is not inherently equal to the interconnection capacity requested," and that some data centers "may forgo grid interconnection altogether and build their own generation capacity with unique redundancy requirements and utilization rates." Those considerations, it says, are out of scope for its "consumption-focused analysis."

The update then converts its energy forecast into capacity by applying an average utilization rate. Assuming a 50% average utilization rate, its Reference Case forecast translates into 148 GW of interconnection capacity needed to serve data centers in 2030.

For one campus, that share is a question a lender can put in writing: how much of its connection will this load draw, and how often?

Waiting Servers, Working Servers

The update's estimate of what an AI server draws rests on stated assumptions. It assumes "that idling AI servers draw power at 20% of their rated capacity," and says the assumption is based on measurement data reported by Brookhaven National Laboratory.

Industry experts, the update reports, gave feedback that "in commercial AI operations, idle power may constitute a much higher fraction of rated power than that indicated by available measurement data." Power management settings, it says, "may be adjusted from factory settings to keep servers in a more active state to reduce latency." A hardware architect at a major semiconductor manufacturer confirmed the 20% ratio, "though it was not confirmed if this ratio may change in practice once the hardware is installed and managed by end users."

The 2024 report had assumed that AI servers running inference work operate at an average utilization of 40%. Feedback from industry experts now indicates that it is "likely much lower," because inference workloads are "largely user-facing and highly dependent on behavioral demand patterns, requiring systems to be provisioned for peak usage when large numbers of users simultaneously query models." The Reference Case assumes inference utilization reaching 20% by 2030, and training servers at 80%. The update says "there continues to be a lack of verifiable data to confirm typical inference utilization levels."

Provisioned for peak usage, in its words.

What a server draws while it waits is influenced, by the update's account, by its settings, and how busy an inference server is depends on how its users behave. Both sit on the computing side of the meter, and the power side's numbers are built on them.

More Energy and the Same Capacity, in One Scenario

The update tests its assumptions in four sensitivity scenarios. The one it says "has the greatest impact" is High Inference Energy, which raises the idle power of AI servers to 30% of rated power and raises the utilization of servers running inference. It forecasts 782 TWh for 2030, against the Reference Case forecast of 649 TWh.

The capacity needed would not move. "While electricity consumption is significantly higher under the High Inference Energy scenario, the installed equipment is the same as the Reference Case. Therefore, total capacity needs would be unchanged; the facilities in the High Inference Energy scenario would simply have higher capacity utilization rates."

In that scenario, more energy would pass through the same capacity.

In the update's arithmetic, a hall that draws more asks the grid for more energy and no more capacity. Which of the two a utility should plan for is the question this scenario leaves open.

What the Model Leaves Out

The update says it does not model "substantial departures from the current trajectory that could arise from technological breakthroughs, shifts in market conditions, or changes in external factors such as supply chains, regulatory environments, or grid power availability." Grid power availability is on that list.

Its conclusion asks for more work. Future research, it says, should "characterize the temporal variability of data center electricity demand to better understand how these loads can be served by the power grid and/or on-site electricity generation resources." And it says that "potential mechanisms for adjusting the timing of these loads (i.e., demand flexibility) should be explored and evaluated to increase the amount of computational demand that can be served with a given set of energy resources."