In the span of a decade and a half, from 2010 to the end of 2024, the number of data centers in the US quadrupled. The trend is similar worldwide: more data centers, bigger, now or soon. The number of the construction projects of centers over 100 megawatts announced over the last four years total 377, according to data center certification and research agency Uptime Institute.
But before we allow Big Tech’s feverish race toward more compute, which environmentalists would not like us to allow, let us pause and consider another option: making do with what we have. Can we retrofit our current data centers to match the needs of our newest technology? Perhaps the building frenzy is not merited; perhaps we have all the facilities we need. A few upgrades here, some fresh servers over there, a new lick of paint, and voilà — an AI data center built from the shell of a legacy one.
“Most of the time what it’s going to mean is bulldozing the building and starting over from scratch.”
I took this idea to data center experts, who told me, in so many words, that no, our current data centers cannot readily be retrofitted to become AI superhouses. The problem is as physical as the ground you’re standing on: Legacy data centers cannot bear the weight of the latest AI technology. The racks that house computer chips or AI chips are simply too damn heavy for the floors, and they would crack under the weight.
Chris Brown, chief technical officer at Uptime Institute, summarized the situation: “We can retrofit the old ones to an extent, but not to the extent that a lot of these AI factories need.” Small sections of small data centers can accommodate small AI-focused workloads for a single Fortune 500 company, for example, he said. “But most of the time what it’s going to mean is bulldozing the building and starting over from scratch,” Brown said.
AI racks, the metal cabinets that house stacks of metal boxes called servers, which house the chips that do the computer or generative AI processing, have a weight problem. Thirty years ago, at the start of Brown’s career in data centers, racks averaged around 400 to 600 pounds. Think of the weight of a home refrigerator up to a baby grand piano. Now, it’s normal for racks to weigh 1,250 pounds to up to 2,500 pounds, falling in the range of a grizzly bear up to a Toyota Prius. But racks specialized with AI equipment fall to the upper end of the spectrum and beyond — Brown said the projected milestone weight of an AI rack is 5,000 pounds.
The extra weight, Brown said, is due to the amount of electronics crammed into the metal racks. Gaps between GPUs slow data transmission, which slows AI model training, which wastes precious compute power and, ultimately, money. The latest high-density racks come packed with memory chips (leading to the decline of the global supply of RAM) and hundreds up to 1,000 GPUs. Whereas traditional computer chip workloads of a decade ago averaged around 10 kilowatts per rack, AI workloads are now 35 times that, up to 350 kilowatts per rack. “They’re packing as much as they can into each rack and putting racks as close together as humanly possible to maximize that capability,” he said.
More power generates more heat that needs to dissipate before a fire breaks out or the chips melt. Air blown over chips has been replaced or supplemented with cooling plates full of liquid, often a watery mixture of toxic coolants. Water weighs a little over 8 pounds per gallon. And don’t forget about the cables. There are often 10 to 35 racks lined up to form a single row in the bowels of a data center. In order to deliver enough power, the diameter of the cables, or a cable-like copper plate called a busway, needs to increase. (Imagine putting out a house fire with a kitchen sink faucet; better to spray water from a fire hose with its wide diameter.) Brown said that a modern busway weighs 37 pounds per linear foot.
“It’s all those things — it’s the weight of all the processors, all the memory, all of the chips that you need to be able to run the IT devices, all of the cooling hardware that you need inside of there,” he said. The structure of legacy data centers is not up to the task, Brown said. Many have raised floors, which top out at around 1,250 pounds per square foot for a static load, he noted. Dynamic loads, he said, such as a rack pushed across the floor, require higher weight bearing.
Even if you did reinforce the floor of an old center, other geometric problems persist, Chris McLean, president of data center construction firm Critical Facility Group, told The Verge. He has been designing data centers for nearly two decades, and the height of racks has grown by 3 feet during that time, from 6 feet to 9 feet. (The footprint has only increased from 2 by 2 feet to 2 by 3 feet.) The new height stands taller than industrial doorframes from a few years ago. Freight elevators, too, cannot withstand the weight of the gigantic racks, the apparatus on which they rest during moving, and the weight of the humans pushing the thing: “All of a sudden, you’re getting into a pretty beefy elevator for a multi-story,” McLean said.
“What’s caused the huge growth in the last two years is just the fact that artificial intelligence is gobbling up everything.”
Big Tech companies are obviously building new data centers to accommodate their increasing push for AI dominance. And when OpenAI, Microsoft, or others run out of space in their own AI-built data center complexes, they rent space at colocation facilities owned by companies like CoreWeave or Digital Realty or Compass, which are, in turn, building new AI-focused data centers. “What’s caused the huge growth in the last two years is just the fact that artificial intelligence is gobbling up everything,” Brown at Uptime said.
Despite the hype, generative AI is not the only type of computation, lest we forget that typical computer workloads still exist. Actually, non-AI data workloads are actually increasing, Brown said. Traditional data centers, therefore, are as important as ever before. Universities, hospitals, midsize companies, and municipalities will all need to continue to store their non-AI data files, just as you still have your blurry photos housed on some cloud provider’s server, McLean said. “All those people still need that legacy data center environment,” he said. “It’s never going to go away.”
