OpenAI Moves to Lessen Reliance on Some Nvidia Hardware

After breaking the news on Friday about how Microsoft and OpenAI are drawing up plans for an ambitious data center project costing as much as $100 billion, we were bombarded with questions from people across the artificial intelligence and cloud industries: What part of the country will it be located in? What chips would it use? How would the companies acquire the nearly-unfathomable amount of power needed to run it?

Our story includes answers to some of these questions—and we also make the point that nothing has been finalized. But it’s probably instructive to look at the more-concrete plans the companies have for a less-powerful data center (but still a multibillion-dollar one) expected to launch in Wisconsin as soon as 2026.

First off, Nvidia chips are likely to populate many of the data center’s server racks. No shocker there. More interesting is that it is not likely to include Nvidia’s networking cables, which stitch together the AI chip servers and help data flow in and out quickly. If you ever hear the word “fabric” in the context of data centers, that’s what it refers to. OpenAI has told Microsoft, which runs its servers, that it doesn’t want to use Nvidia’s InfiniBand networking gear anymore. Instead, it will likely use good ol’ Ethernet-based cables.