100K+ GPU Training Supercluster (Colossus-class) — AI datacenter / managed cluster

100K+ GPU Training Supercluster (Colossus-class) — AI datacenter / managed cluster

Companies committing: spacex, xai, anthropic, nvidia.

Opportunity: Per verified facts: Colossus reached 100K+ NVIDIA Hopper GPUs in Memphis with ~150MW on-site gas turbines; Anthropic became the first external customer ~May 6, 2026; HN comment claims a Colossus 2 with 2× the GPUs. Frontier-lab compute is now a salable utility, not a self-hosted asset.

The 100K-GPU cluster has become a productized SKU — Anthropic buying capacity from xAI/SpaceX is the proof point. The $1.75-2T IPO partly capitalizes the next Colossus. NVIDIA is the silent L2 winner. Constraint named in the JDs and facts: power (~150MW) and on-site generation, not silicon.