AI data centers’ energy demand is rising fast because training and running AI models requires racks of power-hungry chips running around the clock. In the US, data centers used about 4.4% of electricity in 2023, and a Department of Energy-backed study projects 6.7% to 12% by 2028. That new demand can push up power prices when grids must pay more to secure enough generating capacity, and those costs can flow through to household bills.

For most of the last two decades, US electricity demand barely grew. Efficiency gains offset population growth and new gadgets. Then AI arrived, and utilities started receiving requests to connect single facilities that want as much power as a mid-sized city.

The effect is already visible. PJM, the grid operator serving 13 states and Washington, D.C., has seen its capacity auctions clear at record prices in consecutive years, with data centers driving most of the forecast increase in peak demand. Those auction costs show up in what utilities charge customers.

Key takeaways

  • AI data centers draw far more power per rack than traditional data centers because of GPUs and specialized AI chips.
  • US data centers used about 176 TWh in 2023, roughly 4.4% of national electricity, per Lawrence Berkeley National Laboratory.
  • Rising demand tightens grid capacity, and capacity costs can be passed on to households and businesses.
  • Tech companies are signing deals for nuclear, gas and renewable power, but new supply takes years to build.
  • The impact on your bill depends heavily on your region and how your state regulates utilities.

What is an AI data center?

An AI data center is a facility packed with specialized processors, mainly GPUs and custom accelerators, designed to train and run artificial intelligence models at scale. It looks similar to a regular data center from the outside: a large, windowless building with heavy power and cooling infrastructure. Inside, the difference is density.

A traditional server rack running websites or email might draw a handful of kilowatts. Racks filled with modern AI accelerators can draw many times that, which requires heavier electrical systems and, increasingly, liquid cooling instead of air. Multiply that across tens of thousands of chips and a single campus can need hundreds of megawatts. The largest planned projects talk about gigawatts.

If you want the background on what these chips are actually computing, our explainer on how large language models work covers training and inference in plain English.

How much energy do AI data centers use?

The most-cited US figures come from the 2024 Department of Energy report produced by Lawrence Berkeley National Laboratory:

  • Data centers consumed about 176 terawatt-hours in 2023, around 4.4% of total US electricity.
  • By 2028, consumption could reach 325 to 580 TWh, or roughly 6.7% to 12% of US electricity.
  • The wide range reflects uncertainty about how fast AI chips are deployed and how efficient they become.

Globally, the International Energy Agency’s Energy and AI report projects data center electricity use roughly doubling by 2030, with AI the main driver. Put simply, AI data centers’ energy needs are the fastest-growing new source of power demand in many advanced economies.

Training vs inference

Training a frontier model is a huge one-off energy event. But inference, the work of answering billions of user queries every day, runs continuously, and it grows as more people and businesses use AI. Over time, inference is expected to account for a large share of AI’s total power use.

Why are data centers driving up electricity prices?

Electricity prices reflect three main costs: generating the power, delivering it over transmission and distribution lines, and paying for capacity, meaning the guarantee that enough power plants will be available at peak demand. Data centers touch all three.

  1. Capacity gets scarce. In PJM’s auction for the 2027/2028 delivery year, prices cleared at the cap of about $333 per megawatt-day, and PJM said the region fell short of its reliability target, according to Utility Dive’s report. PJM attributed most of the forecast peak-load increase to data centers.
  2. New lines and substations. Connecting a large campus can require new transmission. Depending on how regulators allocate costs, some of that can land on all ratepayers.
  3. Higher wholesale prices. When demand grows faster than supply, more expensive plants run more often, setting a higher price for everyone.

Not every rate increase is about AI. Fuel costs, storm repairs, aging infrastructure and inflation all play a role. With oil and gas markets disrupted in 2026, energy costs have risen for many reasons at once. But data centers are now a meaningful factor in regions where they cluster, such as Northern Virginia, Texas, Ohio and parts of Georgia and Arizona.

Where will the extra power come from?

Source Advantages for data centers Drawbacks
Natural gas Dispatchable, quick to build relative to nuclear Emissions, fuel price swings, turbine supply backlogs
Nuclear (existing and restarts) Carbon-free, runs around the clock Limited supply; new reactors take many years
Small modular reactors Potentially sited near campuses Mostly not yet commercially operating at scale
Solar and wind plus batteries Cheap energy, fast to build Intermittent; needs storage or backup
On-site generation Avoids grid connection waits Costly; may face local permitting opposition

Microsoft, Amazon, Google and Meta have all announced long-term deals with nuclear operators or developers. Those agreements help, but they rarely add new power quickly. The near-term reality in many regions is more gas generation and older plants staying online longer.

What does this mean for investors?

The AI build-out has turned utilities, power equipment makers and grid suppliers into growth stories. It’s also why investors are scrutinizing big tech spending. The late-October 2026 slide in tech stocks was driven partly by worries that capital spending on data centers is eating into free cash flow; our explainer on the Nasdaq correction goes into that. Higher energy costs are part of that math: a data center that costs more to power earns a lower return.

What this means for you

Our view: AI data centers are a real, though regionally uneven, contributor to rising electricity costs, and that pressure is unlikely to fade soon because new power plants and transmission lines take years to build. If you live in a data center hotspot, expect this to come up in utility rate cases and state politics.

Practical steps:

  • Read your bill. Look for capacity, transmission or “supply” line items that have changed.
  • Follow your state utility commission. Rate cases are public and many accept consumer comments. Several states are debating whether large data centers should pay a bigger share of grid upgrades.
  • Cut peak usage. Time-of-use plans, efficient heat pumps and smart thermostats reward shifting demand away from peak hours.
  • Budget for it. If utilities are a growing slice of your spending, revisit your plan using the 50/30/20 rule so higher bills don’t crowd out savings.

The US Energy Information Administration at eia.gov publishes state-level electricity price data if you want to see how your area compares.

Frequently asked questions

How much electricity does one AI query use?

Estimates vary widely by model, hardware and query length, and companies disclose little. Individual queries are small, but billions of them a day add up. Treat any single precise number you see online with caution.

Are AI data centers bad for the environment?

It depends on the power source. Data centers running on gas or coal add emissions, while those supplied by nuclear or renewables have a much smaller footprint. Water use for cooling is a separate concern in drought-prone areas.

Will my electricity bill go up because of data centers?

In some regions it already has, mainly through higher capacity and transmission costs. How much depends on your grid operator, your utility and how your state regulators allocate costs between large customers and households.

Why don’t tech companies just build their own power plants?

Some are, through on-site gas turbines or partnerships with nuclear developers. But power plants take years to permit and build, equipment like gas turbines has long waiting lists, and most campuses still rely on the grid.

Will AI chips become more efficient?

Yes, each generation does more computation per watt. History suggests efficiency gains tend to spur more usage rather than less total demand, which is why forecasts still show consumption rising.