With the AI boom, construction of new data centers has skyrocketed, and not without consequence — some communities that count these facilities as neighbors are now facing water shortages and strained power supplies. While tech's data center footprint has been growing for decades, generative AI has seemingly shifted the impacts of these operations toward the catastrophic. What exactly makes these new data centers such a burden on the environment and existing infrastructure, and is there anything we can do to fix it?
Chips
The industry believes AI will work its way into every corner of our lives, and so needs to build sufficient capacity to address that anticipated demand. But the hardware used to make AI work is so much more resource-intensive than standard cloud computing facilities that it requires a dramatic shift in how data centers are engineered.
Typically the most important part of a computer is its “brain,” the Central Processing Unit (CPU). It's designed to compute a wide variety of tasks, tackling them one at a time. Imagine a CPU as a one-lane motorway in which every vehicle, no matter the size, can get from A to B at extraordinary speed. What AI relies on instead are Graphics Processing Units (GPU), which are clusters of smaller, more specialized processors all running in parallel. In the example, a GPU is a thousand-lane motorway with a speed limit of just 30 mph. Both try to get a huge number of figurative vehicles to their destination in a short amount of time, but they take diametrically opposite approaches to solving that problem.
Advertisement Advertisement
Phil Burr is Head of Product at Lumai, a British company looking to replace traditional GPUs with optical processors. “In AI, you repeatedly perform similar operations,” he explained, “and you can do that in parallel across the data set.” This gives GPUs an advantage over CPUs in large but fundamentally repetitive tasks, like graphics, executing AI models and crypto mining. “You can process a large amount of data very quickly, but it’s doing the same amount of processing each time,” he said.
In the same way that thousand-lane highway would be pretty wasteful, the more powerful GPUs get, the more energy hungry they become. “In the past, as [CPUs evolved] you could get a lot more transistors on a device, but the overall power [consumption] remained about the same," Burr said. They're also equipped with “specialized units that do [specific] work faster so the chip can return to idle sooner.” By comparison, “every iteration of a GPU has more and more transistors, but the power jumps up every time because getting gains from those processes is hard.” Not only are they physically larger — which results in higher power demands — but they “generally activate all of the processing units at once,” Burr said.
In 2024, the Lawrence Berkeley National Laboratory published a congressionally mandated report into the energy consumption of data centers. The report identified a sharp increase in the amount of electricity data centers consumed as GPUs became more prevalent. Power use from 2014 to 2016 was stable at around 60 TWh, but started climbing in 2018, to 76 TWh, and leaping to 176 TWh by 2023. In just five years, data center energy use more than doubled from 1.9 percent of the US’ total, to nearly 4.4 percent — with that figure projected to reach up to 12 percent by the start of the 2030s.
Heat
Like a lightbulb filament, as electricity moves through the silicon of computer chips, it encounters resistance, generating heat. Extending that power efficiency metaphor from earlier, CPUs are closer to modern LEDs here, while GPUs, like old incandescent bulbs, lose a huge amount of their power to resistance. The newest generation of AI data centers are filled with rack after rack of them, depending on the owner’s needs and budget, each one kicking out what Burr described as “a massive amount of heat.”
... continue reading