xAI opens Colossus supercomputer in Memphis

xAI brought online Colossus, a training cluster of 100,000 Nvidia H100 GPUs housed in a converted Electrolux appliance factory in South Memphis, Tennessee. The company said it had taken roughly 122 days to build the cluster from an empty industrial shell to a training-ready facility — a pace well outside the years data centres of comparable scale typically require, achieved by compressing procurement, construction and networking work that other builders ran sequentially.

Colossus’s stated purpose was training xAI’s Grok family of chatbots, though the company said the cluster also supplied computing capacity to Musk’s other ventures, including X and SpaceX. Nvidia said the cluster used its Spectrum-X Ethernet networking platform to link the GPUs, which the company described as necessary to make a cluster of that size function as a single coherent training system rather than a collection of smaller ones.

The scale and speed of the build made Colossus a reference point in the broader compute race among frontier labs, illustrating how quickly a well-capitalised newcomer could assemble training infrastructure that had previously been the preserve of a handful of established hyperscalers. It also established a pattern of unconventional, fast-moving construction — including the use of on-site power generation to meet the cluster’s electricity demands — that drew scrutiny from local residents and environmental groups over air quality and grid impact in the surrounding Memphis community in the following year.

xAI continued expanding the site rapidly. It later built a second facility, Colossus 2, and by the end of 2025 was working toward a combined capacity approaching two gigawatts across the Memphis area, with industry trackers estimating several hundred thousand GPUs deployed across the expanded site.