That Datacenter Outside Your Town: What's Actually in There?

There's a decent chance one of these has shown up near you recently: a windowless building the size of several football fields, ringed by fencing, humming day and night. No sign of workers. A vague name on the gate like "Project XYZ LLC." And around it, a fight — neighbors angry about noise, water, power bills, farmland, and the fact that nobody asked them. These are valid concerns.

Let me say upfront: I'm not here to tell you to fear this building. I'm here to tell you what it is, because you can't push back on — or make peace with — something you can't see into. So I want to do two things in this post. First, tell you what's physically inside, because "the cloud" is a marketing term and the fog around it is doing real damage to the public conversation. Second, and this might surprise you coming from someone who works in tech, tell you that most of your objections are more legitimate than the industry wants to admit.

The cloud is a building

Every photo you've backed up, every email, every digital paper, every Netflix stream, every ChatGPT answer lives somewhere physical. That somewhere is a datacenter: a warehouse full of racks of computers (servers), stacked floor to ceiling, running 24/7.

For decades these buildings were boring. They ran websites and stored spreadsheets, sipped power at a predictable rate, and nobody protested them because there was really mostly nothing to protest.

AI broke that. Training and running modern AI models requires specialized chips (GPUs) that are extraordinarily power-hungry and packed extraordinarily densely. A rack of AI hardware can draw five to ten times what a traditional server rack does, and all that electricity turns into heat, which then requires industrial-scale cooling, which is where the water comes in. So the new generation of facilities isn't just "more datacenters." It's a different beast wearing the same name.

The scale-up has been explosive. According to the Department of Energy's 2024 report, produced by Lawrence Berkeley National Laboratory, US datacenters used 58 terawatt-hours of electricity in 2014, 176 TWh in 2023 — about 4.4% of all US electricity — and are projected to hit somewhere between 6.7% and 12% of the national total by 2028. Read that range again. The people who study this most closely can't tell you within a factor of two how big this gets. That uncertainty is itself part of the problem: utilities are being asked to build for a demand curve nobody can actually see.

Globally, if datacenters were a country, their electricity appetite would rank among the largest on Earth — Brookings puts it in the neighborhood of Japan and Russia by 2026, on a growth curve four times faster than overall electricity use.

Yes, it can show up on your power bill

For roughly twenty years, US electricity demand was flat. Efficiency gains canceled out growth, and your utility planned accordingly. Datacenters ended that era almost overnight, and here's the mechanical problem: a datacenter can go from groundbreaking to switched-on in about 18 months. A power plant takes five to ten years. The gap between those two numbers gets paid for by somebody.

In the PJM region — the grid operator covering 13 states from Virginia to northern Illinois — capacity auction prices jumped more than tenfold in about two years, driven substantially by datacenter demand. Those costs flow into the monthly bills of tens of millions of people. So when your neighbor says "my electric bill went up because of AI," they are not being a crank. The causal chain is real, documented, and largely uncompensated.

The water question: where both sides get it wrong

Water is where the discourse gets sloppiest, so let me be precise, because precision here cuts against both the industry and some of the activists.

The direct numbers first: large AI facilities can draw millions of gallons per day for cooling. Contracts of two million gallons a day are now routine, and the biggest sites run up to five million. That's the municipal demand of a small city, frequently drawn from the same drinking-water system you use, often in places that are already water-stressed, because dry air happens to be good for cooling.

Now the nuance: the same Berkeley Lab research found that roughly 92% of datacenter-related water use doesn't happen at the datacenter at all — it's consumed generating the electricity the datacenter buys. Which means when a company says "our facility barely uses water," they can be technically truthful and deeply misleading in the same sentence, because the water bill was just moved to the power plant. And when an activist says "this building will drain the aquifer," the honest answer is: maybe. It depends on the cooling design and the local grid, and the company usually won't tell you, because water usage is routinely treated as a trade secret.

That secrecy, not the water itself, is the real problem. You cannot run a town's water planning around a facility that won't disclose its consumption.

About those jobs

The sales pitch to your county board goes: jobs, investment, tax revenue. Here's the honest version. Construction genuinely does employ hundreds or thousands of people — for eighteen months. A finished hyperscale facility typically runs with somewhere between a few dozen and a couple hundred permanent staff. That's a decent employer for a rural county; it is nothing like the factory it replaced in the pitch deck. And it frequently arrives with decades of tax abatements — many states run datacenter incentive programs — so the "revenue" was negotiated away before you heard about the project.

None of this is illegal. Much of it, in my opinion, is weak governance: deals negotiated under NDAs, utilities and municipalities outmatched by trillion-dollar companies, residents finding out after the zoning is done. But notice what kind of problem that is. It's not a technology problem. It's a rules problem — and rules problems have solutions.

The good news: none of this is inevitable

Here's the part the loudest voices on both sides tend to skip. The solutions to almost everything above already exist, and some facilities already use them.

On water: closed-loop cooling systems recirculate the same water instead of constantly drawing fresh supply. Immersion and air-cooled designs cut water draw dramatically — some sites use almost none. Others run on recycled wastewater instead of drinking water.

On energy: the best-run facilities are startlingly efficient. Google's fleet wastes only about 9% of its power on overhead like cooling, versus more than 50% at a typical facility. Efficiency gains like these are literally why US datacenter demand stayed flat for two decades before the AI boom.

Even the transparency problem has a working example: some operators now publish facility-by-facility water numbers in their environmental reports. It can be done. It is being done.

Which sharpens the real question. If the good version of this building exists, why is your town being offered the mystery version? The gap between the best facility and the average one is the whole argument for standards: not to stop these buildings, but to make the good version the required version. That's not anti-technology. That's how we've handled every other piece of industrial infrastructure, from power plants to factories.

So why do these buildings exist at all?

Fairness requires the other side. Everything digital you rely on — your bank, your hospital's records, your kid's school portal, emergency dispatch, and yes, the AI tools I use in cybersecurity work — runs on this infrastructure. There is no version of modern life without datacenters, and there was no groundswell against them until AI supersized the footprint. The demand is real, the usefulness is real (that's the next post), and a modern facility run by a company that discloses its numbers, pays its grid costs, and cools efficiently is a legitimate neighbor.

The problem is that nothing currently requires any of that. Disclosure is voluntary. Cost allocation is contested. Siting fights happen town by town, each community negotiating alone against companies that do this every week. You're not anti-technology for noticing the asymmetry. You're paying attention.

Next in the series: the case for this technology — the things AI is genuinely, verifiably good at, including some that might save your life. I've spent two posts earning the right to make that argument. Time to make it.