The Hydrogen Center for Artificial Intelligence

NewsletterNo. 8

The power problem behind AI

AI now runs into a limit that no software update can fix: electricity. A long look at how much power data centers need, who pays for it, and why it became a voting issue.

By Yonathan Shanko, Founding Editor

The short version

The biggest constraint on AI right now is not chips or talent. It is electricity.

Data centers are using more power every year, and the grid was not built for demand to rise this fast. Someone has to pay for the new power plants and transmission lines. Right now, part of that bill is landing on ordinary households, and voters have noticed. This piece explains how the system works and where the fight stands.

How much power are we talking about?

The research firm Gartner projects that global data center electricity use will reach 565 terawatt hours in 2026, up 26% from 2025. Gartner says power availability, not demand, is now what limits how fast AI capacity can grow.

In the US, the Electric Power Research Institute, a nonprofit research group funded by the utility industry, projects that data centers could use 9% to 17% of all US electricity by 2030, up from 4% to 5% today. Its range is 60% higher than its own earlier forecast, which shows how quickly expectations have moved.

A single large AI data center can draw as much power as a small city. That kind of demand does not fit easily into a grid planned years in advance.

Why your bill is involved

Most of the eastern US, including Maryland, is served by a regional grid operator called PJM. It covers parts of 13 states and Washington, D.C.

PJM has to make sure enough power plants are available for the hottest and coldest days of the year. It pays for that through a capacity auction. Plant owners bid to promise they will be available years ahead, and the price is set by the auction. That cost shows up in the supply portion of every customer's bill.

When forecasts say demand will jump, the grid needs more of those promises, and the price rises. According to the Maryland Office of People's Counsel, the state's consumer advocate, prices in a recent PJM auction rose 800%, an increase of $12.5 billion across the region. One estimate cited by the office attributes $9.3 billion of that to data centers, most of them outside Maryland.

The office is careful to add context. It says the biggest drivers of higher Maryland bills are rising distribution rates and extreme weather. An auction error that left two power plants near Baltimore out of the bidding also added cost. The office has filed a complaint with federal regulators and says customers could get refunds if it succeeds.

The fight in Washington

The policy question is simple to state: should the companies that create the new demand pay for the new infrastructure it requires?

In September, the House passed the Ratepayer Protection Act 417 to 3. It encourages state regulators to bill very large power users, like data centers, for the new generation and grid upgrades built to serve them.

The Senate blocked it. On September 30, the vote to move forward fell short at 57 to 43, Roll Call reported, below the 60 votes needed. The dispute was over strength, not direction. The bill asks states to consider a standard but does not require them to adopt one. Senate Minority Leader Chuck Schumer said Democrats want data centers to cover their costs by law. Majority Leader John Thune said Democrats were denying Republicans a win in an election year.

Meanwhile, the White House has taken a voluntary route. President Trump has urged data center developers and tech firms to sign a pledge to build or buy their own power supply, and the White House says hundreds of utilities and dozens of data center operators have signed on, CBS News reported. At the September 29 White House meeting, Trump also said AI executives agreed to address local opposition by funding schools or reducing energy costs, according to the Associated Press.

The fight at home

Public opinion has moved faster than Congress. In a Gallup poll conducted in March, 71% of Americans said they would oppose an AI data center in their area. That is more than the 53% who said they would oppose a local nuclear plant. Opposition crossed party lines, though it was somewhat lower among Republicans.

The reasons are practical. Opponents most often cited water and energy use, followed by quality of life and economic worries such as higher utility bills. Supporters most often cited jobs.

Some communities are writing that opposition into local law. Maine's legislature passed what would have been the nation's first statewide moratorium on data center development in April, but Governor Janet Mills vetoed it, the Bangor Daily News reported. Since then, at least 21 Maine towns have adopted their own limits, and eight more put bans or temporary moratoriums before voters on November 3. All three candidates to succeed Mills said they would have signed the statewide bill.

The trade-off, stated fairly

Both sides have a real point.

Data centers bring construction jobs, tax revenue, and the computing power that every AI product depends on. Communities that host them can see real benefits, and the US government sees AI capacity as a matter of national competitiveness.

But data centers also bring a large, fast increase in demand that the grid has to absorb. If the cost of that buildout is spread across every customer, households end up subsidizing an industry's growth, whether or not they benefit from it.

The question is not whether to build. It is who pays, and whether the answer is a law or a promise.

What to watch

  • Whether the next Congress takes up a mandatory version of the ratepayer bill.
  • Whether Maryland's complaint against the PJM auction leads to refunds.
  • Whether Maine's legislature brings back its statewide moratorium in the session that begins in December.
  • Whether the companies that signed the White House power pledge publish what they have actually built or bought.

Sources

If you work in energy, utilities, or local government and see this differently, send us a pitch. We publish people who disagree with us.