Research

AI Data Centers Are Coming for the Power Grid

US power demand was flat for a decade and is rising again. What data centers actually consume, what the 2030 projections really claim, and how it reaches your bill.

By the Euphoria team · 2026-07-21 · 7 min read

The bill nobody itemized

For about fifteen years, American electricity demand did something strange: it stopped growing. Factories moved offshore, lighting went LED, appliances got stingier, and the total amount of power the country pulled off the grid drifted sideways even as the economy expanded. Utilities got comfortable planning for flat.

That era is over, and the reason is a building with no windows.

The Energy Information Administration now describes US electricity consumption as rising again after more than a decade of little change, and in January it forecast the strongest four-year growth in US electricity demand since 2000, naming data centers as the fuel. That is a genuinely big deal, and it is the kind of story that shows up in a household budget long before it shows up in the news.

• 4.4% of all US electricity went to data centers in 2023
• 176 TWh what they consumed that year, up from 58 TWh in 2014
• 11.8% Berkeley Lab's reference projection for 2030

What a data center actually does with power

Strip away the branding and a data center is a warehouse full of computers that are never turned off. Two things inside it consume electricity, and it is worth keeping them separate in your head.

The first is the computing itself. A chip doing arithmetic draws current. Training a large AI model means running enormous numbers of those operations in parallel, for weeks, across thousands of specialized processors. Answering your questions afterwards, which the industry calls inference, draws less per request but happens constantly and at enormous volume.

The second is cooling. Every watt a chip consumes comes back out as heat, and silicon fails when it cooks. So a meaningful share of a data center's power goes not into computing but into removing the warmth that computing produced. The industry tracks this with a ratio called power usage effectiveness: total facility power divided by the power actually reaching the computers. A perfect building would score 1.0. Real ones do not.

A data center is two appliances in a trench coat: a computer that never sleeps, and an air conditioner working to undo it.

The numbers, and why they are contested

Here is where honest reporting gets uncomfortable, because the most-quoted figures in this story are projections, and the projections disagree with each other by a factor of two.

What is measured is solid enough. A Department of Energy report prepared by Lawrence Berkeley National Laboratory put data center consumption at 58 terawatt-hours in 2014 and 176 terawatt-hours in 2023. A terawatt-hour is a billion kilowatt-hours, so 176 of them is roughly what forty million typical American homes use in a year. That 2023 figure worked out to about 4.4 percent of all electricity consumed in the United States.

The forward-looking number is the argued one. The same body of work put 2028 consumption somewhere between 325 and 580 terawatt-hours, and Berkeley Lab's 2025 update carried a reference case of 649 terawatt-hours by 2030, with a plausible range spanning 9.5 to 15.3 percent of national electricity use.

Notice what that range is really saying. The low end and the high end describe different worlds: one where chips keep getting dramatically more efficient per calculation and buildings keep getting better at cooling, and one where demand for computing grows faster than efficiency can absorb it. Nobody knows which one we are in yet. When you see a single confident number for 2030 in a headline, someone picked a scenario and dropped the word scenario.

Why a warehouse in Virginia touches your bill

An electricity bill has two halves that behave completely differently.

One half is the energy itself, the kilowatt-hours you actually consumed. The other half is the delivery system: the transmission lines, substations, and transformers that carry power to your street, plus the cost of keeping enough generating capacity standing by to cover the single worst hour of the year.

Data centers press hardest on the second half. A large one can request hundreds of megawatts at a single site, and it wants that power at a steady rate around the clock, which is a very different shape of demand from a neighborhood that peaks at dinnertime. Meeting it means new wires and new capacity, and those are long-lived investments that get recovered from everyone connected to the system through regulated rates.

Meanwhile retail prices have been climbing. The EIA has residential electricity going from about 12 cents per kilowatt-hour in 2013 to 16 cents in 2023, which over that stretch tracked inflation reasonably closely. The last few years have been less gentle.

Be careful with the causation here, because this is where commentary usually overreaches. Electricity prices rose for several reasons at once, including natural gas price swings, storm hardening, and wildfire mitigation spending. Data center growth is one pressure on the system, not the sole explanation for your bill. Anyone telling you otherwise is selling something.

The part that is genuinely unresolved

Three questions will decide how this lands, and none of them has an answer yet.

• Who pays for the interconnection. Regulators in several states are actively rewriting the rules for very large customers, so that the cost of new wires sits with the company that required them rather than being spread across residential ratepayers. How that shakes out matters more to your bill than the total terawatt-hours do.
• Whether efficiency keeps up. Computing has a long history of getting more done per watt. If that continues at the historical pace, the high-end projections never happen. If specialized AI hardware plateaus, they might.
• What gets built to supply it. Because the demand is around the clock, it does not match solar output well without storage, which is why nuclear, gas, and grid batteries all keep appearing in the same sentence as data centers. The EIA has separately noted that faster-than-expected data center growth would raise fossil generation, which is an awkward fact for everybody involved.

How to read the next headline about this

You are going to see this story dozens of times over the next few years, usually with a scary number in the title. Three habits will keep you oriented.

Ask whether the number is measured or projected. Measured figures come with a year that has already happened. Projections come with a scenario, and the scenario is the actual claim.

Ask what the denominator is. "Data centers will use as much power as Japan" and "data centers will use 12 percent of US electricity" can describe the same forecast while feeling completely different, and the percentage is usually the more useful framing.

Ask who pays. A cost that exists is not the same as a cost that lands on you, and the rules deciding which is which are being written right now in state utility commissions rather than in Congress.

Why this matters for you

This is a good case study in something that shows up everywhere in money: the most important part of a story is often the part with no number attached. The terawatt-hours are knowable. Who ends up paying for the wires is a policy choice, and policy choices are where household outcomes actually get decided.

Inside Euphoria you can practice exactly this move, pulling apart a claim into the measured part, the projected part, and the assumption doing the heavy lifting, on live market data and in interactive lessons, without a dollar at stake.