your data drinks more water than you do 🥵 #data #nvidia #datacenters

By 2030, data centers will consume as much electricity as the entire country of Japan. Let's talk about what's going on with them in the AI era. Data centers are the physical buildings behind every cloud service, streaming platform, and AI tooling used today. Inside are rows of servers that process data, storage systems that hold it, and networking equipment that moves it, all running 24/7. There are around 12,000 operational data centers worldwide and close to half of them are in the United States. The US also holds half of the world's hyper-scale facilities. The massive campuses operated by companies like Amazon, Microsoft, Google, and Meta. Following the US is Europe with about 25% of the total, then the Asia-Pacific region, and finally the rest of the world with only about 10 to 15% combined. Running these centers requires two things at a scale that is starting to strain both energy and water. In 2024, data centers consumed 415 terawatt hours globally, but 1.5% of all electricity used on Earth. The IAA projects that reaches 945 terawatt hours by 2030. That's more than double and roughly equivalent to Japan's entire annual electricity consumption. This energy usage contributes to a constant generation of heat that has to be cooled down, usually by means of water or air. To understand the scale at which these resources are being used, you can think of one AI rack today, which draws about 100 kilowatts, as roughly equivalent to 10 or more general purpose rocks from a decade ago, which drew around 5 to 15 kilowatts. And these pressures don't stay inside the facility. A third of all US data centers are concentrated in just three states. For January with around 660, Texas around 400, and California around 300. In the community surrounding these hubs, the strain shows up on electricity bills. Utilities making expensive grid upgrades to handle data center demand pass those costs onto households. In Northern Virginia, connecting a new large skill project to the grid can now take up to seven years because of this congestion. One of the industry's responses to these strains has been planning new capacity to be built outside of these saturated hubs, following available power rather than proximity to users. The other responses are technological. The major cloud providers are signing nuclear power agreements. Unlike solar, wind, nuclear runs continuously at the scale's data centers need, without drawing from the public grid. Notably, a Microsoft signed a 20-year deal to restart three-mile island. The industry has also shifted toward liquid cooling, where water or coolant runs directly to the chip, or servers are submerged in non-conductive fluid. Because liquid cooling recirculates the same fluid in a closed loop, rather than continuously evaporating it, facilities using it consume a fraction of the water than typical evaporative systems require. And at the furthest edge, companies plan to put data centers in orbit, with startups like Star Cloud and partners like Nvidia, relying on solar power and the natural cooling of space to sidestep terrestrial resource constraints entirely.