Definition: Project Colossus

An xAI datacenter for AI training (xAI is Elon Musk's AI company). Colossus is a fitting name because at an estimated total cost of $4 billion, Musk claims it is the largest AI datacenter in the world. More advanced chips from NVIDIA will be used, and by 2030, Colossus is expected to operate with the equivalent of 50 million H100s GPUs (see H100).

Revitalizing an old Electrolux vacuum cleaner factory in Memphis, Tennessee, Project Colossus will eventually house more than a hundred thousand AI servers. This "Memphis Supercluster" trained xAI's Grok-3 chatbot (see xAI). See Project Stargate, GPU and AI datacenter.

230,000 Liquid Cooled GPUs

In July 2024, the datacenter went into operation with 100,000 NVIDIA H100 chips. By mid-2025, there were 150,000 H100, 50,000 H200 and 30,000 GB200 GPUs. SuperMicro's water cooling keeps the required temperature. See H100 and Blackwell. (Image courtesy of xAI and Supermicro Computer, Inc.)

A Massive Network

Networking all this equipment is no simple feat. An RDMA fabric ties all servers together in one huge Ethernet running at 400 Gbps, which is 400 times faster than common local network speeds (see RDMA and Terabit Ethernet). (Images courtesy of xAI and Supermicro Computer, Inc.)

Launched in Only Four Months

Phase I requires 150 MW (megawatts) of electricity. Combined with 8 MW from the local grid, 14 transportable VoltaGrid generators were installed to launch the datacenter in July 2024 with a third of the H100s. To keep the power uniform, all electricity flows into batteries and then to the equipment. (Image courtesy of VoltaGrid LLC.)