The AI Spotlight Follow
AI Data Centers Are Failing the Kitchen-Table Test. The Real Fight Isn't About Water, It's Your Power Bill
Energy & Water

AI Data Centers Are Failing the Kitchen-Table Test. The Real Fight Isn't About Water, It's Your Power Bill

Data centers went from invisible warehouses to the most fought-over buildings in America in about three years. The backlash is not really about AI, or even water. It is about a building that uses a city's worth of electricity for people who live somewhere else, and who ends up paying for it.

Advertisement

Contents

The most valuable acre in town

An acre of land is 66 feet wide and 660 feet long. Farm it for corn and it might earn about $1,200 a year. Raise cattle and maybe $3,000 in a good year. Lease it for a wind turbine and perhaps $5,000.

Put an AI data center on it and, by some industry estimates, the economic value it generates can run into the millions of dollars per acre per year, the kind of number once reserved for prime real estate in Manhattan, Paris or Tokyo.

In about three years, data centers went from buildings nobody thought about to some of the most contested construction projects in America. Ask people today whether they want one next door, and opposition rivals what nuclear plants faced in the 1970s, during the era of duck-and-cover drills.

How did a windowless box full of computers become less popular than a nuclear reactor? The answer runs from a parking garage in Virginia to the way your electricity bill is calculated.

Advertisement

How a garage became the internet

In the early days of the internet, the long-distance backbone linking cities was government-funded and reserved for universities and research labs. Commercial traffic was not allowed. You could email a professor. You could not email a customer.

When the first commercial internet providers appeared, they had a problem. If a customer of one wanted to email a customer of another, the only connecting wire was the one they could not use. So they physically connected their networks.

A few engineers ran lines from each company into a plain office suite in Tysons Corner, Virginia. Within months, several providers were linked there. As more joined, the exchange moved into a small room off a parking garage nearby. Nothing was computed there. It was simply a meeting point where networks touched.

That kind of exchange, later known as MAE-East, became one of the most important hubs of the early commercial internet, carrying a huge share of global traffic. It was the seed of what lies about 30 minutes away today: Data Center Alley in Loudoun County, the largest concentration of data centers on Earth.

Why everyone crowded into Virginia

Once networks met in Northern Virginia, everyone wanted to be nearby. Distance equals delay, and on the internet, milliseconds are money.

Consider online advertising. Every time you load a page with ads, an auction runs. Bidders get a profile of who you might be and roughly 100 milliseconds to decide what to bid. If your servers are in San Francisco and the auction runs in Virginia, about 35 milliseconds go to receiving the request and another 35 to sending your answer back. Seventy percent of your time is gone before you have even analyzed anything.

So ad companies placed servers not just near major exchange points, but in the same buildings. That auction machinery, running billions of times a day, is how much of the free internet is funded.

Amazon Web Services launched its first region, us-east-1, in Northern Virginia in 2006. It became the default for countless apps, which pulled even more infrastructure into the area.

Loudoun County is often described as the richest county in America. Three things are true at once: its residents are wealthy mainly because it is a Washington suburb; data centers are there because the internet's hub is there; and the county government collects enormous tax revenue from them, which local officials say has helped fund schools and cut residential property tax rates.

For decades, nobody protested. Data centers stored your photos, streamed your movies and carried your messages, and drew as much attention as a warehouse full of lawn chairs. Then a different kind of computer moved in.

What a watt actually measures

To understand the fight, start with a watt.

A phone charger draws about 10 watts. A TV about 100. A microwave about 1,000. The wattage label on appliances tells electricians how much power a device might demand, so they can size wires and breakers. It is why big appliances get bigger outlets, and why plugging a space heater into the wrong room trips the breaker.

An average American home uses about 1,200 watts around the clock, a little more than a microwave running constantly.

Streaming is efficient. Netflix compresses each movie once, copies it to servers close to viewers, and sends the file when you press play. By one illustrative estimate, the servers needed for every household in a city the size of New Orleans to stream at once would draw roughly the power of a handful of homes.

Why AI needs so much more power

A movie already exists before you hit play. When you ask AI for something, nothing exists yet. The system has to create it from scratch.

Video makes this concrete. A two-second clip at 24 frames per second means generating 48 images, each with around two million pixels, and each frame must stay consistent with the ones around it.

That work runs on a different chip. A CPU, like the one in your laptop, has a handful of very fast cores: think eight excellent chefs each cooking their own dish. A GPU has thousands of simpler cores, on the order of 17,000 in high-end models, all working on the same task at once: thousands of cooks on one dish.

The result: a single rack of modern AI servers can draw on the order of 100,000+ watts, roughly the continuous demand of about 100 homes. Multiply by thousands of racks and you get campuses that consume as much power as a city.

Which leads to the awkward question neighbors ask: what do I get for all that electricity, land and money? Too often the honest answer sounds like "something cool, it can write emails," followed by a request to build another city-sized box across the street.

The kitchen-table test

Planners have a name for projects nobody wants nearby: LULUs, locally unwanted land uses. Landfills, highways, power lines, prisons.

Consider a prison. North Dakota's state penitentiary dates to the 1880s Dakota Territory era. Reporting on the facility has described cramped cells, poor sight lines and problems meeting modern safety standards for staff and inmates. Replacing such facilities can take decades, because every proposed site triggers protest.

But a prison passes what you might call the kitchen-table test. A family can sit down and say: we do not love it, but it serves us. People who break the law in our state are housed there. The benefit stays local.

A data center flips that logic. When roughly a million people around the world send a request to a giant AI campus every couple of seconds, almost none of them live in the town hosting it. Some of those towns have a few thousand residents, many of whom rarely use AI.

A landfill is "our mess, so our dump." A data center is matched not to its users but to four inputs: power, water, fiber and a deal. Around the kitchen table, people focus on two things.

Is it really about water?

Number two on the list is water.

Many data centers cool chips the way humans cool themselves: by evaporation. Water absorbs heat and evaporates away. Some large facilities using this approach can consume around a million gallons on hot days. In a dry county with a small utility, neighbors will ask questions.

What angers people most is often secrecy. In The Dalles, Oregon, records eventually showed Google's data centers used 355 million gallons in 2021, about 29% of the city's water. The city had initially gone to court to keep the figure confidential, calling it a trade secret, before settling and agreeing to disclose it.

Context matters. Agriculture is by far the largest user of water in the US. Crops like almonds and pistachios, along with golf courses, use far more than data centers. Documented cases of residents losing water because of a data center are relatively few, and the same examples circulate repeatedly. One widely shared case in Newton County, Georgia, involved several private wells; local officials have said the water problems predated the nearby Meta construction, though residents disputed that.

The industry has noticed the polls. Many new facilities are designed for closed-loop cooling, where water is filled once and recirculated for years, or for air-based and waterless designs. A million-gallon headline costs more than the savings.

The number one issue, though, shows up at the top of poll after poll across political parties: will this raise my electricity bill?

How big is a gigawatt?

Big AI announcements are measured in gigawatts. OpenAI's Stargate partners have talked about multi-gigawatt campuses, and Nvidia has discussed investing up to $100 billion in OpenAI tied to deploying around 10 gigawatts of systems.

Here is the scale:

  • 1,000 watts (a microwave) = 1 kilowatt
  • 1,000 kilowatts = 1 megawatt
  • 1,000 megawatts = 1 gigawatt, about a million microwaves

Imagine every home in a large city running its microwave at the same moment. That is roughly the territory of a gigawatt.

Add up all the proposed data centers, and projections show the US surplus of generating capacity shrinking over the next decade, especially in certain regions. Which raises the obvious question: where will the power come from?

The Electric States of America

People imagine the power grid as one national highway system. It is not. It is a patchwork of regions that operate very differently. Think of them as the Electric States of America.

Electricity does not come from the nearest plant. It comes from whatever plants are running in your region at that moment. Over a century, small local utilities merged into larger systems for reliability and lower cost, much like fuel refining consolidated. That is why national totals matter less than whether each region has enough.

Three broad models dominate:

1. Organized markets with capacity auctions

Much of the population, including the region hosting the world's largest data center cluster, lives in areas where different companies own power plants, transmission lines and local wires, and a regional operator coordinates them.

Prices come from auctions. In the hourly energy market, plants bid the lowest price they will accept. The operator stacks bids until demand is covered. The last bid needed sets the clearing price, and every accepted plant gets it. If you bid $5 and I bid $10 and both are needed, we both get $10.

There is also a capacity auction, held well in advance, that pays plants to be available on the hottest day years ahead. It works like insurance, and its cost is spread across everyone's bills.

2. Vertically integrated utilities

In much of the Southeast and West, one company owns the plants, wires and billing. In Georgia, for example, the utility forecasts its needs, asks state regulators to approve rates, and customers pay the approved rate. One company, one hearing, one rate.

3. Texas

Texas runs its own grid largely within state lines, which keeps it mostly outside federal interstate rules. Its market is energy-only: plants get paid when they generate. Most hours, power is cheap. During extreme heat or storms, prices can spike to the market cap of $5,000 per megawatt-hour.

Three systems, three different ways the same data center boom shows up on a bill.

Why bills are rising

Regional projections show comfortable surpluses today but tightening margins by 2030, with some of the same markets already reporting bill increases expected to get tighter by 2035.

Not all of those increases come from data centers. Some come from grid maintenance, aging equipment and forecasting errors. In capacity markets, if operators expect lots of new supply that never shows up, prices can jump, and those costs land on household bills.

Still, the core worry is reasonable. If a company asks for a city's worth of electricity, and utilities must spend billions on new substations, lines and plants to serve it, who pays? If the answer is ordinary customers, the world's AI ambitions have become your monthly bill.

The case for building anyway

There is another way to see the same request.

One reaction to a company wanting a gigawatt is: where will we find it? The other is: someone just offered to buy a gigawatt, all day, every day, for years.

For decades, US electricity demand grew slowly. AI changed that. Regions that can build generation quickly, without burdening existing customers, may become places with abundant, reliable power, which also attracts factories and advanced manufacturing.

The money is staggering. Alphabet, Amazon, Meta and Microsoft are on track for roughly $700 billion or more in combined capital spending in 2026, much of it for AI infrastructure. For comparison, the entire Interstate Highway System cost about $500 billion in today's dollars, spread over decades.

The strategic pitch: compute may become an export as important as oil. When someone in London, São Paulo or Jakarta asks an AI a question, the answer is often generated in the US. Like Northern Virginia, an early center of gravity attracts companies, suppliers and talent, and then more building. Do what Loudoun did, but nationally: build here, generate power here, collect taxes here.

That is not a ridiculous bet. But the person at the kitchen table is not the United States. They are a household with a bill.

What nuclear and highways teach us

A nuclear plant makes electricity for people nearby. A data center uses electricity for people mostly far away. Same category of unwanted building, opposite direction of benefit.

Protests did not kill American nuclear power. Costs did. After the 1979 Three Mile Island accident, plants under construction had to be redesigned mid-build, and costs soared. In Washington state, a public power system that launched five plants defaulted on billions in bonds, one of the largest municipal defaults in US history. Around 100 planned plants were canceled, and ordinary ratepayers were often left holding the bill. For decades, almost nobody started a new reactor.

The interstates went differently, not because they were popular, but because the federal government paid 90% of construction costs. Fights over routes were bitter and some neighborhoods were saved, but the fights were about where, not who pays.

Data centers today face both questions at once. And everyone is watching the bill.

How states are making data centers pay

That is where policy is heading. The headlines sound like bans, but most are really demands that data centers prove they can cover their costs.

  • New York: Governor Kathy Hochul imposed a one-year moratorium on new hyperscale data centers of 50 megawatts or more in July 2026 while the state develops higher standards.
  • Texas: Governor Greg Abbott directed state environmental regulators to pause data center permits in September 2026 pending a grid-connection audit, expected to report in December.
  • Ohio: Regulators ordered AEP Ohio to create a data center tariff requiring large new customers to pay for at least 85% of the capacity they reserve, even if they use less, for up to 12 years.
  • Virginia: Regulators have moved large data centers into their own customer class with separate rates.
  • Louisiana: Entergy won approval for three gas plants totaling about 2.26 gigawatts to serve Meta's Hyperion campus, with Meta covering the costs under its agreement.
  • Pennsylvania: Constellation is restarting a Three Mile Island reactor, now the Crane Clean Energy Center, under a 20-year power deal with Microsoft. It is targeting the second half of 2027.

That last one is poetic. The site whose accident helped end the last nuclear buildout is being revived for a building even less popular than a reactor. The difference: this time, the customer can afford the bill.

Why this matters

The data center debate is often framed as pro-AI versus anti-AI. That misses the point. The real question is cost allocation: who pays for the grid that AI needs.

If tech companies pay their own way, communities may gain tax revenue and new power capacity. If costs leak onto household bills, backlash will grow and could slow the entire buildout, just as cost overruns stalled nuclear power for a generation.

Practical takeaways

  1. Ask who pays. When a data center is proposed locally, ask whether there is a dedicated tariff or cost-recovery agreement.
  2. Ask about cooling. Closed-loop or air-cooled designs use far less water than evaporative systems.
  3. Demand transparency. Public water and power figures prevent the secrecy that fuels distrust.
  4. Know your grid. Whether you live in a capacity market, a regulated utility area or Texas changes how costs reach you.
  5. Watch tax deals. Abatements can erase the local revenue that makes hosting worthwhile.

Counterpoints and risks

  • Forecasts may be wrong. Some proposed data centers will never be built, so projected shortages may be overstated, or capacity built for them could go unused.
  • Local benefits are real. Counties like Loudoun have used data center taxes to fund services and cut tax rates.
  • Efficiency improves. Chips and cooling get more efficient each generation, though total demand has kept rising.
  • Pausing has costs. Moratoriums may push projects to other states or countries.
  • New generation takes time. Even with willing buyers, plants and transmission lines take years to build.

FAQ

Why do people oppose AI data centers?

Mostly over electricity costs, then water use, along with noise, land use and a lack of transparency. Polls consistently show bills as the top concern.

Do data centers use a lot of water?

Some do, especially those using evaporative cooling, which can use around a million gallons on hot days. Nationally, agriculture uses far more, and many new data centers use closed-loop or waterless cooling.

Do data centers raise electricity bills?

They can, depending on how a region allocates grid costs. Several states now require data centers to pay for the capacity and infrastructure they need.

How much power does an AI data center use?

Large AI campuses can require hundreds of megawatts to multiple gigawatts, comparable to the demand of a large city.

Is Three Mile Island really reopening for AI?

Yes. Constellation is restarting a reactor at the site, renamed the Crane Clean Energy Center, to supply Microsoft under a 20-year agreement, targeting the second half of 2027.

The bottom line

Data centers became unpopular not because people fear AI, but because they fail a simple test: they use a city's worth of power for people who live somewhere else, and nobody has clearly answered who pays.

The places that win the AI buildout will be the ones that answer that question honestly, making the newcomers cover their own costs. Get that right, and the kitchen table might just say yes.

Sources

Independent Evaluators Say AI Is Now Speeding Up Its Own Research. Fable 5.2 Shows What That Looks LikeUp next · 14 min readIndependent Evaluators Say AI Is Now Speeding Up Its Own Research. Fable 5.2 Shows What That Looks LikeAn independent evaluator embedded at Anthropic says progress is running roughly twice its old pace, and most of the speed-up comes from AI helping build AI. Fable 5.2 is the clearest window yet into that loop, from a 41% freelance-job score to a model that knows when it is being tested.Keep reading →