Try Free

What AI Data Centers Really Do With All That Water

Into Silicon and AI Revolution November 22, 2026 13m 2,324 words 1 views
▶ Watch original video

About this transcript: This is a full AI-generated transcript of What AI Data Centers Really Do With All That Water from Into Silicon and AI Revolution, published November 22, 2026. The transcript contains 2,324 words with timestamps and was generated using Whisper AI.

"A big AI data center can go through millions of gallons of water in a single day. And you already know why. It's for cooling. But that answer has a hole in it. Your car's engine needs cooling too, and it gets by on the same few quarts of coolant for years. Nobody tops up their radiator every..."

[00:00:00] Speaker 1: A big AI data center can go through millions of gallons of water in a single day. And you already know why. It's for cooling. But that answer has a hole in it. Your car's engine needs cooling too, and it gets by on the same few quarts of coolant for years. Nobody tops up their radiator every afternoon. So if the water is just for cooling, why does a data center keep needing more? Where does it all go? It starts with heat, and AI makes a ridiculous amount of it. Pretty much every watt that goes into a chip comes back out as heat. So as far as physics is concerned, an AI chip is a space heater that happens to do math. One of NVIDIA's Blackwell chips pulls about 1,200 watts, squeezed into a slab of silicon smaller than a credit card. That's several times more heat per square inch than a stovetop burner on high. Now, pack 72 of those chips into one rack, the way NVIDIA's big AI systems do, and that single cabinet throws off around 120 kilowatts. That's roughly five home furnaces running flat out in a box the size of a fridge. And the next generation is expected to come close to doubling it. That heat doesn't wait around. When cooling fails, a data hall can get hot enough that the only safe move is to switch everything off. On July 19, 2022, Britain hit 104 degrees Fahrenheit for the first time on record. In London, the cooling at Google and Oracle data centers couldn't keep up. Google said several of its backup cooling systems failed at the same time, and it powered down machines before the heat could damage them. For hours, part of the cloud was knocked out by the weather. Then, late on Thanksgiving night in 2025, a chiller plant failed at a data center outside Chicago. That building runs the electronic trading system of CME Group, the biggest derivatives exchange in the world. Futures on oil, gold, currencies, and stock indexes froze around the globe for about 10 hours, while most of America was sleeping off dinner. So cooling isn't a side detail. It's what keeps the whole building alive. And the inside part is actually a solved problem. Liquid cooling is old. Back in 1985, the Cray-2 supercomputer had its circuit board sitting in a bath of liquid that doesn't conduct electricity. It costs 12 to 17 million dollars. Your phone does hundreds of times more math, and it's cooled by your pocket. Modern AI chips do it more neatly. Each one sits under a cold plate, a block of metal with tiny channels inside, and liquid runs through it, right on top of the chip. That liquid lives in a sealed loop and goes around and around for years, exactly like the coolant in your car. So the chips themselves drink basically nothing. The trouble is the last step. Every bit of that heat eventually has to be dumped into the outside air. And there are really only two ways to do that. You can do what your car does. Or you can do what you do. Your car pushes hot coolant through a radiator, and air blows across it. It works, but it has one hard limit. A radiator can never make the water colder than the air around it. On a 100 degree Fahrenheit afternoon, the water coming out of a radiator stays above 100 degrees Fahrenheit, no matter how big the fans are. And for decades, data centers cooled the air around their servers with water chilled to about 45 degrees Fahrenheit, or even lower. The only way to get water colder than the outside air is to pump heat uphill with a chiller, which is basically a giant air conditioner. And that burns a lot of electricity. Your body found a better trick a long time ago. When sweat evaporates off your skin, it takes heat with it. And the amount of heat is enormous. Here's a neat way to see it. Warming one pound of water by one degree Fahrenheit takes one British thermal unit. That's literally how the unit is defined. Yes, British. America sizes its air conditioners in a British unit and seems to have made its peace with that. Evaporating that same pound of water takes about 1,000 British thermal units. So turning a little water into vapor carries off as much heat as warming 1,000 times more water by a degree. One gallon of evaporating water takes about 2.5 kilowatt hours of heat with it. That's about what one of those AI chips puts out in two hours. That's also where a rule of thumb engineers use comes from. If a pound of evaporating water carries off 1,000 units of heat, then cooling 100 pounds of water by 10 degrees Fahrenheit costs about 1 pound of evaporation. So roughly 1% of the water leaves as vapor for every 10 degrees Fahrenheit of cooling. And evaporation has a second, sneakier advantage. It can cool water below the temperature of the air. That's why you feel cold stepping out of a pool on a hot day. A radiator can't do that. Sweat can. So that's what most big data centers built. It's called a cooling tower. Warm water from inside gets pumped up and sprayed over stacks of plastic sheets so it spreads out thin while giant fans pull air up through it. A small part of the water evaporates, and the rest falls into a basin around 10 degrees Fahrenheit cooler and goes back for another lap. Even data centers that use chillers usually get rid of the chillers' heat in a tower like this because sweating is so much cheaper than running bigger compressors. So a cooling tower is a building whose whole job is to sweat. That white plume you sometimes see rising from the roof isn't smoke. It's the water. That's where most of it goes. Up. The rest goes down the drain. Tap water has minerals dissolved in it, and when water evaporates, the minerals stay behind. With every lap, the water left in the tower gets a little saltier and crustier. Let that go too far, and it coats the pipes in scale, like the inside of an old kettle. So the system keeps dumping some of its most concentrated water and topping up with fresh. Engineers call it blowdown, which is a pretty dramatic name for opening a drain. So every drop that comes into a data center leaves one of two ways. It rises as part of a cloud, or it goes down the sewer, carrying everybody else's minerals. And you can check the scale yourself. A big AI campus might draw a gigawatt of power, and all of it ends up as heat. If you got rid of every bit of that heat by evaporating water, at 2.5 kilowatt hours per gallon, you'd go through about 9.5 million gallons a day. Shaolei Ren, a professor at UC Riverside who studies this, estimates that a 1 gigawatt AI data center can use up to 10 million gallons a day. The back of the envelope physics lands almost right on it. That's about 15 Olympic swimming pools, every day. Across the whole U.S., data centers used about 17 billion gallons in 2023, according to Lawrence Berkeley National Laboratory. And that could double, or even quadruple, by 2028. Now, evaporation works best when the air is dry. Engineers track this with something called the wet bulb temperature. Wrap a thermometer in a wet cloth, blow air over it, and the reading tells you how cold evaporation can get. In Phoenix, on a dry, 110-degree Fahrenheit afternoon, the wet bulb temperature is only around 70 degrees Fahrenheit. In Miami, at 90 degrees Fahrenheit and muggy, it's around 80. So a cooling tower in the desert can make colder water than one in Miami, on a day that's 20 degrees Fahrenheit hotter. Which means sweating works best in exactly the places with the least water to sweat with. Phoenix is one of the biggest data center hubs in America. And some Arizona cities have already capped or banned the use of drinking water for cooling data centers. So the obvious alternative is to stop sweating altogether and go where the air is always cold. Up there, a plane radiator works all year. Alaska is betting on exactly that. On November 10, 2026, the state approved a 50-year land lease for a huge data center campus on the North Slope, well above the Arctic Circle, cooled by nothing but Arctic air. The buildings would sit raised above a gravel pad so the cold can flow underneath them. The developer says that cuts water use by 90% or more compared with a sweating data center farther south. So why doesn't Alaska already have a single big data center? Because cooling turns out to be the cheap part. Google's data centers use only about 9% more electricity than their computers do. And that covers cooling, lights, everything. The expensive part, the one every AI company is fighting over, is the electricity itself. And Alaska's main power grid is an island. It isn't connected to the rest of the country. Its power costs well above the national average. And the gas fields that feed it are running low. So the North Slope campus would have to bring its own. The plan is a gas power plant of 1 to 3 gigawatts, burning natural gas that's sitting unused up there. The turbines for it are backordered for years. Construction is expected to take about 6 years, and no customer has been named yet. Alaska solved the cooling in one move. The power is the whole project. So how much of all this is you? In 2023, Wren's team estimated that a conversation of roughly 20 to 50 questions with GPT-3, an older OpenAI model, cost about a bottle of water. That number went everywhere. Then, in 2025, Google said a typical Gemini text prompt uses about five drops. Sam Altman put a chat GPT question at about a 15th of a teaspoon. That's a massive gap, and it mostly comes down to which drops you count. Google counted the water evaporating at its own buildings. Wren's estimate was for an older model on older hardware, and it also counted the water at the power plants making the electricity. And that turns out to be the bigger number by far. Because power plants sweat too. A coal or nuclear plant throws away roughly two units of heat for every one it turns into electricity. Most of America's electricity still comes from plants that make steam to spin a turbine. And that steam has to be cooled back down, often in the exact same kind of evaporating tower. Lawrence Berkeley's report estimated that the power plants feeding U.S. data centers used about 211 billion gallons of water in 2023. That's about 12 times what the data centers used on site. For every drop that evaporates off a data center roof, roughly 12 more evaporate at a power plant you'll never see in a story about AI. And that sets a trap. A data center can save water by running chillers instead of a cooling tower. But chillers burn more electricity. And more electricity means more water evaporating at the power plant. Squeeze the water out at one end, and some of it comes back out the other. So the real fix has to come from the chips. Remember the radiator's one limit. It can't make water colder than the air. But what if the water doesn't need to be cold? That's exactly where liquid-cooled AI chips are heading. A lot of new systems already run on water around 90 degrees Fahrenheit. At CES in January 2026, Jensen Huang said NVIDIA's NextRacks, called Vera Rubin, can be cooled with water at 113 degrees Fahrenheit. That sounds backwards. Cooling a chip with water hotter than a hot tub? But the chip runs hotter still, so heat keeps flowing into the water. And water that only has to get back down to 113 degrees Fahrenheit doesn't need sweat. Plain outside air can do it almost anywhere for most of the year. No tower, no giant air conditioner, and no need to move to the Arctic. In Huang's words, no water chillers are necessary. The companies that make chillers were listening too. The next day, shares of Carrier and Johnson Controls fell. Microsoft has already gone that way on the building side. Since August 2024, it's been designing data centers where the cooling water is sealed in a loop, filled once during construction, and never topped up again. The heat leaves through giant radiators with fans. Microsoft says each of these buildings avoids more than 33 million gallons of water a year. There's a catch, though, and it's that same hard limit. A radiator only works when the outside air is cooler than the water. Phoenix summer days can hit 115 degrees Fahrenheit, and on those days, you still need extra cooling. And Microsoft admits its waterless design uses a bit more electricity. Not much. Microsoft calls it a nominal increase. But by now, you know where extra electricity sends its water. So, in the end, AI doesn't really use water the way a farm or a city does. It uses evaporation, the cheapest way anyone has ever found to throw heat into the sky. The newest data centers are learning to stop sweating. The power plants feeding them haven't. As long as AI runs on electricity made with steam, every answer still sends a little cloud into the air. It just rises from a different tower. Subscribe and turn on notifications so you catch the next film. And if this one surprised you, leave a like. Thanks for watching!

Transcribe Any Video or Podcast — Free

Paste a URL and get a full AI-powered transcript in minutes. Try ScribeHawk →