Where AI Actually Lives

← Back to Newsletters

Go Deeper Guide  •  Companion to Issue #22

The numbers behind the headlines: power, water, and who pays

In this week's newsletter I gave you the short version. Here's the longer one, for anyone who wants to understand what's really going on before the next scary headline (or the next rosy one) comes along.

Same approach as before: I'm not taking sides. I just want you to have good numbers, where they came from, and where the experts disagree.

How Many Are There?

Nobody keeps one official count, so the number depends on who's counting. Estimates for the US run from about 4,000 to 5,400. The safe way to say it: more than 4,000, far more than any other country.

Most of them are modest buildings that have quietly run our email, banking and streaming for years. The ones making news are the giants. Only about 90 US data centers draw more than 100 megawatts today, roughly the power for tens of thousands of homes. If the projects already planned get built, that could climb to around 280 by 2030.

California's share

California has more than 200 active data centers, behind only Texas and Virginia. But they use only about 2% of the state grid's peak demand. Our high electricity prices for industry mean California is seeing less of the new AI building boom than many other states.

How Much Electricity?

This is where the growth shows up. In 2014, data centers used about 2% of all US electricity. By 2023 it was about 4.4%. The forecasts for 2030 are where it gets interesting, because the experts don't agree.

Why such a wide range for 2030? Nobody knows how many planned projects will actually get built, or how much more efficient the chips will get. Treat any single 2030 number with caution.

Worldwide, the International Energy Agency expects data center electricity use to roughly double between 2025 and 2030, with AI growing from about one-third of that to about half. In the US, data centers drove about half of all the growth in electricity demand last year.

What About Water?

Water is the most confusing part of this whole topic, and the source of a lot of bad numbers. Here's what's solid.

Many data centers use water to keep their computers cool. In 2023, US data centers used about 17 billion gallons directly for cooling, according to Lawrence Berkeley National Laboratory. That sounds enormous, but it's only about 0.14% of all the water the country uses.

There's a second, bigger number. Power plants also use water to make the electricity data centers run on. Count that, and the total climbs to more than 200 billion gallons a year.

What really matters

Location. A data center in a rainy region barely registers. The same data center in a dry desert town can compete with homes and farms for water. That's why you'll hear the loudest concerns from dry parts of the country, and why California's new laws require data centers to report their water use.

What Does One Question Cost?

AI toolElectricity per typical text questionRoughly the same as
Google Gemini0.24 watt-hours, about 5 drops of waterWatching TV for 9 seconds
ChatGPT0.34 watt-hoursRunning an oven for about 1 second

Two honest caveats. These numbers come from the companies themselves, and they're for plain text questions. Making images or video uses a lot more. Critics also point out that Google's water figure leaves out the water used at the power plant.

Either way, your own questions are a tiny slice. It's the billions of questions a day, from everyone, that add up.

How the Tech Companies Are Getting Their Power

The big AI companies need more electricity than the grid can supply quickly, so they're lining up their own. Here are the four main approaches.

ApproachExampleThe trade-off
Restart old nuclear plantsMicrosoft's 20-year, roughly $16 billion deal to restart Three Mile Island Unit 1 (now renamed the Crane Clean Energy Center). Target: late 2027, pending final approval from federal nuclear regulators.Clean, but only a handful of closed plants can be restarted.
Fund new small reactorsGoogle (500 megawatts with Kairos Power), Amazon (320 megawatts in Washington state), Meta (up to 6.6 gigawatts by 2035).Clean, but most won't be running until the 2030s.
Build their own gas plantsMore than 46 data centers are pursuing on-site gas. Meta is funding seven gas plants in Louisiana. About 80% of on-site power projects use gas.Fast, but not clean, and it locks in fossil fuel for decades.
Promise to pay their own wayThe White House Ratepayer Protection Pledge (March 2026), signed by Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI. Expanded in July to include utilities, developers and states.Voluntary, with no enforcement.

The "Who Pays" Fight

Here's the part that touches your wallet. When a giant data center moves in, the local utility often has to build new power lines, substations and power plants to serve it. The question is who pays for all that: the tech company, or everyone else's electric bill.

About 7 in 10 Americans worry data centers will raise their bills. In some places, they already have:

  • The eastern US grid (13 states plus Washington, DC): the grid's independent monitor estimates data center demand added about $29 billion in costs to 67 million customers over roughly two years.
  • Maryland: one state estimate puts the added cost at $168 to $216 a year per household.
Here in California

Seven new laws, signed in September

Supporters call them the most comprehensive data center laws in the country. Here's what they do:

  • Pay their own way: by January 2028, data centers must cover their own grid, transmission and power costs through special electricity rates set by the state's utility regulators.
  • No walking away: a data center that pulls out early pays a penalty, so the rest of us aren't stuck paying for equipment built just for it.
  • Report the numbers: data centers must report how much water and energy they use.
  • Full environmental review for new projects.

For Californians, the practical upshot: the rules are now on the books to keep data center costs off your electric bill. How well they work will depend on how regulators carry them out over the next two years.

And nationally?

In September, the House passed the Ratepayer Protection Act by a lopsided 417 to 3. It would require state regulators to consider making large data centers pay for the grid upgrades built for them.

Then it stalled in the Senate. On September 30 it got 57 votes, short of the 60 needed to move forward. Supporters said it would protect families from higher bills. Opponents said it didn't go far enough, because it only asks states to consider these rules rather than requiring them.

The debate isn't over. A larger bipartisan energy bill now before the Senate includes its own version: data centers above a certain size would pay the full extra cost of the grid built to serve them. And with a new Congress taking office in January, more proposals are expected.

The bottom line: California has rules on the books. Nationally, it's still promises and proposals.

The Good Side

It would be easy to come away from all this thinking data centers are simply a problem. They're not. A few things worth keeping in mind:

  • They already run your everyday life. Email, online banking, video calls with the grandkids, your medical portal and the weather app all depend on them.
  • They can speed up grid upgrades. Much of America's power grid is old. When tech companies pay for new lines and plants, everyone can benefit from the improvements.
  • They're pushing new clean power. The tech industry's demand is a big reason old nuclear plants are being restarted and new reactor designs are getting funded.

How to Check a Scary Stat

This topic is full of outdated and exaggerated numbers. You may have seen claims like "the US has 3 million data centers." That one's way off. When a number sounds alarming (or too good to be true), ask your AI:

"I read that [paste the claim]. Is that accurate? What's the original source, how recent is it, and do other experts agree?"

Then ask it for links so you can look for yourself. The higher the stakes, the more you verify.

My Closing Thought

Every time you ask an AI a question, a building somewhere hums a little harder. That's not a reason to stop using AI. Your share is tiny. But it is a reason to pay attention to how these buildings get powered and who picks up the bill.

AI holds real promise, in medicine, in science, and in everyday help for people like us. Getting there means building more of these places. The challenge ahead is striking the right balance: making room for what AI can do, while keeping power clean, water safe and bills fair.

Those decisions are being made right now, in Sacramento and in Washington. Now you know enough to follow along.

Facts checked November 16, 2026. Sources include Lawrence Berkeley National Laboratory, the International Energy Agency, the Electric Power Research Institute, the California Energy Commission, Google, OpenAI, Reuters and the US Nuclear Regulatory Commission. This topic moves fast, so details may differ by the time you read this.
AI Unlocked • aiunlocked.us • Your trusted guide on the AI journey

Next
Next

ChatGPT Can Talk, and Listen, Too!