Working with Nvidia hardware has become a status symbol—a sign that one is serious about A.I. Talking with engineers about the equipment, I was reminded of the time I saw a snaking line of young men standing in the cold to buy sneakers from the streetwear brand Supreme.
Earlier this year, CoreWeave went public. Venturo and his co-founders are now billionaires. The company owns several hundred thousand G.P.U.s, and its platform trains models for Meta and other leading labs, in addition to OpenAI.
This summer, I visited a CoreWeave facility on the outskirts of Las Vegas. The building, a large warehouse, was surrounded by a thick fence and dotted at regular intervals with security cameras. I went through a turnstile, where I was greeted by a security guard wearing a bulletproof vest and a holstered Taser. After surrendering my phone, I took two lime-green earplugs from a dispenser and entered the facility.
I was joined by three CoreWeave engineers, geeks who had adapted to hyper-scale capitalism as Darwin’s finches had to the Galápagos Islands. Jacob Yundt, from corporate, was lean and eloquent, with a swooping part in his hair. Christopher Conley, an enthusiastic explainer with sunglasses and a beard, oversaw the hardware. Sean Anderson, a seven-foot-tall former college-basketball center, wore a shirt that read “MOAR NODES.”
The nodes in question were shallow trays of computing equipment, each weighing around seventy pounds and holding four water-cooled G.P.U.s along with an array of additional gear. Eighteen of these trays are stacked, then connected with cables to a control unit, to form the Nvidia GB300 computing rack, which is a little taller than a refrigerator and costs a few million dollars. In a busy year, a typical rack will use more electricity than a hundred homes. Dozens of them stretched into the distance.
CoreWeave keeps its racks in white metal cabinets, to help them stay cool and to dampen noise. Conley unlatched a door to show me a rack in action, and I was buffeted with air. The noise was unholy, as if I’d opened a broom closet and found an active jet engine inside. I watched the blinking lights and the spinning of the fans. “Tinnitus is an occupational hazard,” Conley shouted at me.
I looked around. There were hundreds of identical cabinets in the facility. Above us was a metal catwalk, lined with power distributors for the computing equipment. I thought of monks in cloisters, soldiers in barracks, prisoners in cells. What type of person voluntarily worked in such a place, I wondered. “I was told by H.R. that I can’t ask this kind of question anymore, but I like to hire people that can endure a lot of pain,” Yundt later said. “Endurance athletes, that sort of thing.”
CoreWeave wouldn’t tell me which customer was using its technology that day, although Yundt suggested that the training run we were witnessing was a modest one. He began to detail the configuration of the rack. Unable to hear what he was saying, I nodded sagely, as if in a conversation at a night club. Even with the plugs in, my ears were starting to ring, and I was developing a headache. Yundt turned to me. “Sometimes a customer will tie up this entire place for weeks at a time,” he shouted. His parted hair began to flap in the fan exhaust. “We call those ‘hero runs.’ ”
CoreWeave’s hardware can train an A.I. from scratch to completion. Software developers, typically at a workstation in Silicon Valley, upload to the data center a file of numbers known as “weights” and a vast array of training data, which might be text or images or medical records or, really, anything at all. In their initial configuration, the weights are random, and the A.I. has no capabilities.
The A.I. is then exposed to a slice of the training data, and asked to offer a prediction about what should ensue—the next few letters in a sentence, say. An untrained A.I. will invariably get this prediction wrong, but at least it will learn what not to do. The weights must be modified to absorb this new piece of information. The math is unwieldy, and is especially dependent on an operation known as matrix multiplication.
