Why people are throwing AI data centers into the ocean

[00:00:00] Speaker 1: - What happens when you ask ChatGPT to reword an email? Your prompt goes to a data center and triggers a wave of computations on hundreds of servers. As those servers get hit with requests from thousands of users at once, they start to heat up. To keep them from overheating, data centers typically cool them with water, fans, and air conditioners. OpenAI doesn't publicly share exactly how much energy or water its models consume, but one researcher we spoke to put together some estimates. For the simplest tasks, like rewording your email, their popular GPT-5 model uses as much electricity as it takes to power a 3-watt LED light bulb for around 20 minutes, and a little less than a teaspoon of water. For one person in one prompt, it may not seem like much. OpenAI says it receives 2.5 billion queries a day. Add those up, and at the very least, you're looking at enough water to meet the daily drinking needs of all the people, in Monterrey, Mexico. And enough energy to binge watch Netflix for more than 3,000 years. The number of AI users is expected to roughly double over the next half decade. And as AI usage is increasing, so too is the number of data centers needed to support it. That's why researchers are racing to find new and better ways to keep server racks cool, like sinking data centers deep into the ocean and even launching them into space. Cooling has always been a core part of how data centers work. Even one of the first computers in the 1940s had vacuum tubes which were cooled by fans and ventilation. despite being literally trillions of times more powerful, the data centers of today still look kind of similar. The big difference is they're using a lot more energy. Nearly 50% of their entire energy use can go towards keeping the CPUs that process your prompts, also known as microchips, from overheating. This Google data center in the U.S. state of Oregon is large enough to run computations needed for AI applications. A typical data center covers an area of about 9,000 square meters. This one is 13 times bigger. It's called a hyperscale data center. [00:02:13] Speaker 2: The scale of the AI data centers, which we call hyperscalers, is so much larger than the data centers that we have been living with pretty much without even noticing them for the past decade plus. [00:02:33] Speaker 1: That's Professor Scanlon from the University of Wisconsin, who wrote this paper about the hidden environmental cost of AI data centers. Packing in more servers means a lot more cooling. Google's hyperscaler in the Dales, Oregon, used nearly a third of its city's water supply in 2021, according to public records obtained by the state's largest newspaper. Relying on water for cooling is a common approach. That's why when you look closely at this map, you see big clusters of data centers here, close to fresh water. Nearly a tenth of the world's data centers are in this part of the U.S. The region is called the Great Lakes and is home to one-fifth of the world's surface fresh water, much of it naturally cold. Using it lets data centers spend far less electricity on cooling. The water gets sprayed into rooftop cooling towers, allowed to cool naturally through evaporation, and then used to keep servers from overheating. Fewer fans and less electricity, the method is called evaporative cooling. The trade-off? A lot more water. [00:03:34] Speaker 2: Less than one percent of the water in the Great Lakes is annually renewed. And many ecosystems, city drinking water, industries rely on a relatively stable level of the lakes. So the water that can be taken out for use by humans, it has to fit within that water budget. [00:03:59] Speaker 1: Now let's look at the other side of this trade-off. Google's data center in Mesa, Arizona, sits in one of the driest regions in the United States. So they went with air cooling. Giant fans pushing air across the servers. No water, but a lot more electricity. [00:04:18] Speaker 3: The data centers in Ireland are consuming a fifth of all of our electricity. [00:04:22] Speaker 1: Rosie Leonard is an environmental activist from Ireland. [00:04:27] Speaker 3: So it's really, really high. So they are directly causing a strain on the energy system. And the public is paying for this through increased bills. [00:04:35] Speaker 1: Research found that using evaporative rather than air cooling can reduce the energy needed to cool a data center by between 30 and 60%. But as new AI models drive data centers to pack in more powerful chips and denser servers, they're generating more heat than ever, meaning cooling needs to work even harder to keep up. Or it can work smarter. One solution is to directly cool the microchips where most heat is generated, instead of the whole server room. Direct-to-chip cooling uses a water-based liquid coolant. This circulates across metal plates, typically made from copper or aluminum, stuck onto the chips. Companies say using liquid to absorb heat from chips in a recirculating closed-loop system cools more effectively than chilling the air. [00:05:18] Speaker 4: Liquids move heat faster than air does. Air is very light and liquids are very dense. So the same volume moving through the system, liquid can remove way more of the heat. [00:05:31] Speaker 1: That's Maxi Reynolds from the data center company Subsea Cloud. We'll hear more from her later. Research found that direct-to-chip cooling can reduce the power used by server fans by up to 80% compared with air-based methods. And the liquid itself remains sealed in a closed-loop. So no on-site water consumption is necessary to make it happen. It looks like this method is becoming the new norm for large-scale data center providers. Microsoft is building what it calls its most powerful AI data center in Wisconsin, in that same Great Lakes region. They say this facility will house hundreds of thousands of processors, and they're using direct-to-chip cooling to save on water. Studies by McKinsey and Future Market Insights estimate that by the end of 2030, direct-to-chip cooling build-outs will represent about a third of the data center cooling sector. But what if you could take that idea even further and dunk an entire server into cooling liquid? I know what you're thinking. Tossing an AI server rack into a cold water bath is a sure way to fry the system. So for immersion cooling to work, data center providers use special oily liquids that won't conduct electricity or cause metal to rust. Gigabyte, a data center services provider in Taiwan, says cooling its servers in this bubbling liquid allows them to be stacked more densely. With the direct-to-chip cooling method we explored earlier, some fans are still needed. With immersion cooling, there's no air around the server, removing the need for fans altogether. One operator in Athens, Greece, has reported that over 90% of the energy that the data center uses goes strictly to crunching computations instead of cooling, a big jump from the roughly 50% seen in many traditional air-cooled data centers. Industry researchers found that the immersion-cooled data center sector was already valued at US$1.7 billion in 2025. That represents a third of the liquid-cooled data center market as a whole. In the frosty harbors of Denmark, one operator is taking cooling to a whole new level by plunging data centers directly into the frigid waters of the North Sea. [00:07:48] Speaker 4: By placing them subsea, we eliminate the electrically driven cooling. So we see about a 40% reduction in the power that's consumed and then about 40% decrease in the carbon emissions. There's Maxi again. Her company's shipping container-like capsules weigh 46 tons. [00:08:05] Speaker 1: Inside, high-density server racks hold over 1,100 processors, suspended in — you guessed it — a non-conductive, oily coolant goo. The heat this coolant absorbs from the computing components quickly emanates outward into the frosty waters just outside. The modules can be connected, for example, to offshore wind turbines to power the computations, though no electricity is needed for cooling. No fans, no lights, and no pumps for circulating the gooey coolant. This means that 99% of electricity goes directly to computing power, according to the company. There are problems with underwater data centers, though. [00:08:43] Speaker 5: Every day there's some GPU that fails due to plenty of different reasons. [00:08:48] Speaker 1: Pengfei Li is a researcher focused on AI's environmental impact. [00:08:52] Speaker 5: Pengfei Li: If it's in the ground, we can easily replace them or fix the errors on these GPUs or some other hardware. But if it's undersea, then it's kind of hard to maintain. [00:09:07] Speaker 1: Surveys show that over half of facilities have experienced at least one outage in the past three years. Though most of these are minor hiccups, it shows that problems can come from all angles, like memory errors, network glitches, power issues, or even natural disasters. But there's one more data center frontier, and this one goes far beyond the limits of Earth itself. Operators are exploring the idea of shooting data centers into space to orbit around our planet. The idea is they can rely on solar power for their computing, while the cold of space would help to naturally keep the servers cool. Pengfei Li: By staying close to Earth's surface, data could travel back and forth using optical lasers without too much lag time. In 2026, Elon Musk's rocket launch company SpaceX merged with his artificial intelligence company XAI with the goal of launching AI data centers into space. [00:10:02] Speaker 6: But researchers are skeptical data centers require a lot of cooling. That's actually really hard to do in space. [00:10:11] Speaker 1: Samantha Lawler is an astronomer who studies space junk. [00:10:14] Speaker 6: It seems like, oh, space is cold, infinite heat sink capacity, but it's actually really hard to get the heat to leave. Right when the sun is shining, all that heat builds up. [00:10:27] Speaker 1: Objects in space exposed to direct sunlight can get as hot as 120 degrees Celsius. And because there's no air, that heat can't easily go anywhere. Lawler also points out that despite its name, space is getting crowded, with around 11,000 satellites already in low Earth orbit. [00:10:44] Speaker 6: We do need to think more carefully about prioritizing which satellites are most beneficial to the most people. There are satellites that are useful for saving human lives, for planning crop rotations, for watching climate change unfold and helping us to mitigate that. [00:11:05] Speaker 1: Some of these new cooling technologies are more feasible than others, but they all have the same goal: making AI use less water and energy. But that might not be enough. It can also make sense to reduce the need for cooling in the first place. Something as simple as shifting workloads to cooler hours or water-secure regions can significantly reduce environmental impact. There's also a school of thought in AI called Small Data, Big Tasks. This approach relies on changing the way we train AI models altogether, by teaching it in small, deliberate steps, rather than brute force programming them with millions of possible scenarios. This reduces the mass of computations that keep data centers running, hot and thirsty, around the clock. And then we, the users of AI, can also play our part. You can opt for less water-intensive LLMs or select which model they use when entering prompts. Another solution, of course, is choosing not to use a chat bot to draft every email or automate every moment of friction in our lives. Because behind every quick AI response is something very physical: water, energy, and communities already stretched thin. What about you? How often do you use AI tools like CLAWD or ChatGPT, and are you concerned about how much water and energy they use? Let us know down in the comments and visit DW.com for more.