How Submerged Data Centers Just Solved The AI Energy Crisis

[00:00:00] Speaker 1: Every time you ask Cloud a question, somewhere in the world a server heats up. Multiply that by 2.5 billion prompts a day and you start to understand the problem. AI is boiling the planet's data centers alive. Nearly half of a data center's total energy doesn't go toward computing. It goes toward keeping the machines from melting. Giant fans, cooling towers, millions of liters of fresh water evaporated into the sky. One Google hyperscaler in Oregon was consuming nearly a third of its city's entire water supply. So engineers ask a radical question. What if we stop fighting the heat and just sank the whole thing into the ocean? I know it sounds insane, but right now, off the coast of Shanghai, a 232 million dollar underwater data center is operational. Powered by offshore wind, cooled by the sea, using zero fresh water. And in Finland, Microsoft isn't just cooling its servers. It's piping the waste heat into homes, warming 100,000 households for free. This is the story of how AI Bigger's problem accidentally created two of the most elegant engineering solutions I've ever seen. Traditional data centers cool their servers the same way we've been doing it since the 1940s. Blowing air over hot machines. It works, but it's brutally inefficient. In many facilities, for every watt spent computing, you spent another watt just on cooling. And AI has made everything worse. The new GPU chips powering large language models generate far more heat than anything before them. Data centers are being packed denser, running hotter and demanding more water than ever. Ireland's data centers now consume a fifth of the country's entire electricity supply and residents are already paying higher bills because of it. The industry tried incremental fixes, direct-to-chip liquid cooling, immersion cooling in non-conductive oil. This helped, but what if the ocean itself was the cooling system? Microsoft was the first to try it. In 2018, they sank a sealed cylinder containing 864 servers onto the seafloor off Scotland's Orkney Islands. The capsule was filled with dry nitrogen. No oxygen, no humidity, no corrosion. The ocean did all the cooling. When they pulled it up two years later, the results were remarkable. The underwater servers had one-eighth of the failure rate of identical servers on land. The capsules was covered in algae and barnacles, but they were pristine on the inside. Microsoft eventually paused the project. Not because it didn't work, but because sealed capsules are hard to upgrade. And AI hardware evolves too fast for a five-year tomb. But others ran with the idea. A company called Subsea Cloud redesigned the concept with pressure-equalized pods filled with non-conductive liquid instead of sealed gas. Their pods can go as deep as 3,000 feet, connect directly to undersea fiber cables, and achieve a near-perfect power efficiency of 1.01, meaning 99% of electricity goes to computing, virtually nothing to cooling. And then China went big. In early 2026, Shanghai's High Cloud launched the world's first commercial-scale underwater data center, a $222 million facility sitting 10 meters below the surface, connected to offshore wind turbines. The system saves an estimated 122 million kilowatt-hours of electricity per year and eliminates the need for 105 tons of fresh water annually. They're now planning to scale to 500 megawatts. Now, here's where the story takes an unexpected turn. Others are asking, why waste it at all? In Finland, Microsoft is building what Fortum, the local energy utility, calls the world's largest data center waste heat recovery project. The servers generate heat. The heat gets captured, boosted by heat pumps, and then piped through an underground district heating network directly into homes, schools, and offices. When it's fully operational, it'll provide 40% of all the district heating for Finland's second-largest city, warming roughly 100,000 homes. Google is doing the same thing in Hamina in Finland, where its data center already supplies 80% of the town's heating demand. A Finnish engineer named Ari Kurvi now takes hot showers heated by a data center 5 kilometers away. Now, think about what's happening here. The same heat that costs millions to get rid of in Arizona and Oregon is being sold as a byproduct in Finland. AI's biggest liability, its waste heat, now becomes a city's heating infrastructure. The data center stops being a parasite on the grid and becomes a power plant. Neither solution is perfect. Underwater data centers are hard to maintain. When a GPU fails, you can just walk in and swap it. Sealed modules become outdated fast in an industry where hardware changes every two years. And we still don't fully understand the long-term impact of dumping heat into marine ecosystems, even at small scales. The heat recovery only works when district heating infrastructure already exists, mostly Nordic countries right now. And there's also a sovereignty question. Should a city's heating supply depend on a foreign tech company's business decisions? But the direction is clear. Germany's new energy-efficient act now requires data centers to reuse at least 10% of their waste heat starting July 2026. London just announced plans to heat 9,000 homes from local data centers. The era of simply venting heat into the sky is ending. The next time you send a prompt to chat GPT, that heat might be warming someone's home in Finland. And honestly, that might be the most useful thing AI has done yet. If you enjoyed this video, please subscribe for more. I'll see you in the next one.