Most forecasts of data center electricity use come from utilities seeking approval to build, consultants selling advice, or companies selling chips. The report that Lawrence Berkeley National Laboratory released in December, and summarized in a January 15 news release, comes from a federal laboratory with a long record of measuring the sector, and it was released by the Department of Energy. Its numbers have become the reference point for nearly every serious discussion of the issue since.
The core findings are simple. US data centers consumed about 176 TWh of electricity in 2023, roughly 4.4% of total US consumption. Depending on how quickly AI servers are deployed and how efficiently they run, data centers are projected to consume between 325 TWh and 580 TWh in 2028, or between 6.7% and 12% of total US electricity. In the space of five years, data centers could go from a mid-sized consumer to a share comparable to large industrial sectors.
A decade of flat, then a surge
The historical series is as important as the projection. Total data center electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023, a tripling. But the growth was not even. For much of the 2010s, efficiency gains in servers, storage and cooling, together with the shift of workloads from small corporate server rooms to large, efficient cloud facilities, kept consumption growing slowly despite enormous growth in computing. The laboratory finds that between 2017 and 2023, data center power demand more than doubled, largely because of the growth in AI servers built around graphics processing units.
That change in slope is the key to understanding the debate. The efficiency gains that held demand flat for years have not stopped, but they are no longer large enough to offset the growth in computing intensity that AI requires. Accelerated servers draw much more power per unit than conventional ones, and they are being installed in very large numbers.
Why the range is so wide
A projection that spans 325 to 580 TWh, a difference of almost 80%, can look like an admission that nobody knows. It is better read as an honest map of the key uncertainties. The first is the number of AI accelerators shipped and installed, which depends on chip supply, customer demand and capital spending by a small number of very large companies. The second is how hard those chips are run: average utilization rates for AI servers vary widely between training and inference workloads. The third is cooling and power distribution efficiency, measured as power usage effectiveness, which is far better at the newest hyperscale facilities than at older enterprise sites. The fourth is the rate of efficiency improvement in the chips themselves.
The low end of the range assumes slower deployment and continued efficiency gains. The high end assumes rapid deployment of AI servers and more modest efficiency improvement. Both are plausible. Planners who pick one end without saying why are making a bet, not a forecast.
What 580 TWh would mean
The scale of the high case deserves a closer look. An increase of roughly 400 TWh over five years would be larger than the total annual electricity consumption of most US states. Much of it would land in a handful of places: northern Virginia, central Ohio, Texas, Georgia, Arizona and the Pacific Northwest. Each of those regions would need new generation, new transmission and new substations at a pace utilities have not managed in decades.
The 4.4% national share in 2023 also understates the local impact. In regions where data centers cluster, their share of utility load is already far higher, and it is in those regions that capacity prices, transmission plans and rate cases have started to reflect data center demand.
The pace matters as much as the total. Generation and transmission projects typically take longer than five years from conception to operation once permitting, interconnection studies and equipment procurement are counted. Demand that doubles or triples inside that window has to be met largely with what already exists or is already under construction, which is why existing gas plants, delayed coal retirements and nuclear uprates have featured so heavily in utility responses.
What it does not say
The report is a demand estimate, not a supply plan. It does not say where the electricity will come from, how much it will cost, or what it will do to emissions. Those depend on decisions by utilities, grid operators and state regulators. It also reflects conditions as of late 2024. Breakthroughs in model efficiency or slower growth in AI investment would move consumption toward the low end of the range. A further acceleration in AI spending would push it higher.
There is also a definitional point. The report covers data centers, including conventional cloud and enterprise facilities as well as AI-focused ones, and it includes the energy used for cooling and power conversion, not only servers. Cryptocurrency mining, which is often lumped together with data centers in utility forecasts, is treated separately. Comparing these numbers with utility pipelines that combine crypto and data centers will overstate the gap.
Why it matters for policy
Two things follow from the report. First, it confirms that data center demand is not a statistical artifact or a bubble in utility marketing materials. It is real, it has already doubled in six years, and it is likely to double again. Policymakers who treated earlier warnings as hype will have to plan for growth.
Second, the width of the range is a policy problem in itself. If utilities build for the high case and demand comes in at the low end, ratepayers will pay for stranded assets. If they build for the low case and demand comes in high, reliability and prices will suffer. That asymmetry argues for tools that shift risk to data center developers, such as long-term contracts with minimum demand charges, collateral requirements and flexible interconnection that lets loads connect early if they agree to curtail in tight hours. The federal estimate tells planners how big the uncertainty is. It is up to regulators to decide who carries it.