Last time: a data center is a building full of computers, shaped by the workload. The AI hall is different because the job is different — one long factory job for training, millions of short ones for inference, and an ordinary web storefront in front. Those racks run hot.
Water. Water is where a lot of the scare stories live, and most of them confuse using water with using it up. A big AI hall turns almost every watt it draws into heat, and that heat has to go somewhere. Older, cheaper data centers often dump it with cooling towers: water evaporates on purpose, which is why you see those huge annual gallon figures.
The new high-density AI data centers are different. The chips themselves are liquid-cooled, and that coolant runs in a closed loop. This is just like in your car — the water runs in a closed loop from the engine to the radiator, where the heat is dumped before the coolant goes back. You fill it once — you don't need a fire hose providing water every time you start the engine. Heat leaves the building through big outdoor radiators — dry cooling. Any additional water is just used to top off the radiator.
Related: Data Center Joe Friday: Just the Facts (Part One)
The scare story is that AI data centers drink rivers. Part of that is adding in water used by power plants. Some old data centers still use evaporative cooling; new designs use closed-loop cooling, which cuts the ongoing demand for makeup water, like that old car. It's nothing like the evaporative cooling model, but it does still require some water to top off the cooling system. Not a firehose, but some.
Power. These data centers do use a lot of power, and the plants have to be built somehow.
Dominion Energy Virginia is Virginia's largest investor-owned utility. Dominion serves Loudoun County, home to the densest data center cluster in the country. The load grew faster than capacity, and Dominion has a duty to serve the new load. So it builds generation and transmission ahead of that load, and under the old allocation rules, a share of those costs is spread out to households that didn't cause it. That share is what shows up as extra line items on a Virginia bill — and why the State Corporation Commission finally carved the biggest users into their own class.
Related: The AI 'Pause' Might Not Be What You Think it Is
In Georgia, they took a different approach. In January 2025, the Georgia Public Service Commission put new customers over 100 megawatts on special terms: longer contracts, minimum bills, and requirements that they pay for upstream generation, transmission, and distribution as the hall goes up. Georgia Power is now advertising that this book of large-load contracts will mean lower residential bills — on the order of $180 a year for a typical house — starting in 2029. The difference is timing. Georgia is pushing those costs onto the data centers while they are being built, not playing catch-up after the iron is in the ground.
Land. Another scare story is that these AI data centers will take up immense amounts of land that would otherwise be used for farming or parkland. The whole data center campus may be 500 to 1,000 acres, but that's not just the data centers — that includes land for the various supporting structures, like the power substation and cooling yard.
A useful comparison is a solar farm. A solar panel field is 5,000–10,000 acres per GW, and that isn't 24/7 power — that's several times the size of a data center campus.
So, the data centers aren't paving the state, but some of them are going to be neighbors. The data centers for training can be far from people. The middle piece, inference, still wants to be close enough that the answers come back fast. So some data centers are competing for land near homes and businesses, and that means there will be more ZIP codes with data centers and more zoning fights.
Related: The Little Red Hen: Why the Data Center Fight Is a Culture Fight
