AI chip shelves are very heavy

In a decade and a half, from 2010 to the end of 2024, the number of data centers in the United States increased. quadrupled. the The trend is similar around the world: More data centers, bigger, now or soon. Number of construction projects Of the 100-megawatt-plus centers announced over the past four years a total of 377, according to the Uptime Institute for Data Center Certification and Research.

But before we let big tech companies Hectic race about more accountany Environmentalists don’t want us to allow thatLet’s stop and consider another option: making do with what we have. Can we modernize our existing data centers to match the needs of our latest technology? Maybe the building frenzy isn’t worth it; We probably have all the facilities we need. A few upgrades here, some new servers there, a fresh lick of paint, and there you have it, an AI data center built on top of an old core chassis.

I took this idea to data center experts, who told me, in so many words, that no, our existing data centers cannot be easily retrofitted into giant AI centers. The problem is as physical as the ground you stand on: legacy data centers cannot support the weight of the latest AI technology. Racks containing computer chips or AI chips are simply too heavy for the flooring, and will crack under the weight.

Chris Brown, chief technical officer at the Uptime Institute, summed up the situation by saying: “We can modernize older factories to some extent, but not to the extent that a lot of these AI factories need.” Small sections of small data centers could accommodate small, AI-focused workloads for a single Fortune 500 company, for example, he said. “But most of the time what that means is tearing down the building and starting from scratch,” Brown said.

AI racks, which are metal cabinets that hold groups of metal boxes called servers, which contain the chips that do the computer or generative AI processing, suffer from a weight problem. Thirty years ago, at the beginning of Brown’s career in data centers, racks averaged 400 to 600 pounds. Think of the weight of a home refrigerator up to the size of a grand piano. Now, it’s normal for racks to weigh anywhere from 1,250 pounds to 2,500 pounds, which falls in the range of a grizzly bear to even a Toyota Prius. But racks that specialize in AI equipment are at the upper end of the spectrum and beyond — Brown said the expected weight limit for an AI rack is 5,000 pounds.

Brown said the extra weight is due to the amount of electronic devices crammed into the metal shelves. Gaps between GPUs slow data transfer, which slows down AI model training, wasting precious computing power and, ultimately, money. The latest high-density racks come loaded with memory chips (resulting in… Decreased global supply of RAM) and favorable Up to 1000 GPU. Whereas a decade ago traditional computer chip workloads averaged about 10 kilowatts per rack, AI workloads are now 35 times that, up to 350 kilowatts per rack. “They pack as much as possible into each rack and place the racks as close together as possible to maximize that capacity,” he said.

More power generates more heat that must dissipate before a fire can start or the chips can melt. The air blown over the chips has been replaced or supplemented by cooling panels filled with liquid, often a water-based mixture of Toxic refrigerants. Water weighs just over 8 pounds per gallon. And don’t forget 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 provide sufficient power, the diameter of the cables, or a cable-like copper plate called a bus path, must be increased. (Imagine putting out a house fire with a kitchen sink faucet; it’s better to spray water from a wide-diameter fire hose.) A modern busway weighs 37 pounds per linear foot, Brown said.

“It’s all those things — it’s the weight of all the processors, all the memory, all the chips you need to be able to run the IT hardware, all the cooling you need inside of it,” he said. The structure of legacy data centers is not up to the task, Brown said. Many have raised floors that go for around 1,250 pounds per square foot fixed He pointed to the pregnancy. Dynamic loads, such as a shelf being pushed across the floor, require more weight to be carried, he said.

Even if you reinforce the floor of an old center, other engineering problems remain, said Chris McLean, president of data center builder Critical Facility Group. Edge. He’s been designing data centers for nearly two decades, and has increased rack heights by 3 feet over that time, from 6 feet to 9 feet. (The space has only increased from 2 x 2 feet to 2 x 3 feet.) The new height is taller than the industrial door frames of a few years ago. Freight elevators also cannot support the weight of giant racks, the equipment they are resting on while in motion, and the weight of humans pushing the object. “All of a sudden, you’re in a very large, multi-floor elevator,” McLean said.

It’s clear that big tech companies are building new data centers to accommodate their increasing quest for AI dominance. When OpenAI, Microsoft, or others run out of space in their AI data center campuses, they lease space in distribution facilities owned by companies like CoreWeave, Digital Realty, or Compass, which in turn build new AI-focused data centers. “The reason for the tremendous growth in the last couple of years is the fact that AI is eating everything,” Uptime’s Brown said.

Despite the hype, generative AI is not the only type of computing, lest we forget that typical computer workloads still exist. Non-AI data workloads are actually increasing, Brown said. Therefore, traditional data centers are as important as ever. Universities, hospitals, mid-sized companies and municipalities will all need to continue storing non-AI data files, just as your blurry photos still reside on some cloud provider’s server, McLean said. “All these people still need this legacy data center environment,” he said. “It will never go away.”