SpaceX’s computing power rental operation is kind of like hoarding graphics cards—then when the “mining collapse” hits and things fall apart, you quickly sublease them to recoup cash back. The issue of a hardware performance gap is truly a long-standing headache in AI infrastructure.

CryptoWorld reports that SpaceX encountered delays and hardware discrepancies while developing the Grok model at the Memphis Colossus 1 data center, and therefore decided to lease all of Colossus 1’s computing power to Anthropic. The original plan was to train frontier models using a three-center cluster, but the connection spanning more than 10 miles and an aging network caused latency. Colossus 1 uses multiple generations of NVIDIA chips (Hopper/Blackwell and older accelerators), which do not match with Colossus 2/3, whose unified architecture is based on Blackwell.