The data center gold rush hinges in part on the idea that models will improve in line with the laws of scaling—so long as they’re trained on more data and compute. “There’s this basic story that scale is the thing, and that’s been true for a while, and it might still be true,” Jonathan Koomey, a visiting professor at UC Berkeley who studies computing and data center efficiency, told WIRED in September, talking about an OpenAI-Oracle deal to create three new data center sites. “That’s the bet that many of the US AI companies are making.”
Derek Thompson, a prominent economics reporter, noted in a recent post that the tech industry is projected to spend around $400 billion on AI infrastructure this year—while demand for AI services from US consumers stands at just about $12 billion a year, according to one study.
OpenAI had a strategic relationship with AMD prior to today’s announcement. At AMD’s Advancing AI event in Silicon Valley in June, Altman briefly joined Su on stage. Su said that AMD has been getting feedback from customers for a few years as the company has been designing the upcoming MI400 series of chips, and that OpenAI is one of those marquee customers.
Altman noted on stage that the industry’s move towards reasoning models has been putting pressure on AI model makers in terms of efficiency and long-context roll outs, and that OpenAI needs “tons of compute, tons of memory, and tons of CPUs,” in addition to the Nvidia GPUs that the generative AI industry is so reliant on.
Su, at that event, described Altman as a “great friend” and an “icon in AI.”
Lauren Goode contributed to this report.
