Anthropic Leases xAI’s Colossus 1 Supercomputer for Inference Workloads

Anthropic is leasing xAI's Colossus 1 supercomputer for inference workloads, according to reports. The cluster, built by SpaceX, features a mixed GPU architecture that proved inefficient for training.

Mixed GPU architecture limits training efficiency

Colossus 1 houses over 220,000 GPUs, including roughly 150,000 H100s, 50,000 H200s, and 20,000 GB200s, drawing 300 megawatts of power. xAI's real-world GPU utilization reportedly sat at just 11 percent, making the system unsuitable for training its Grok model.

xAI is building Colossus 2, a homogeneous Blackwell-based cluster, for frontier training and a potential IPO. The Colossus 1 deal could generate $5 to $6 billion in annual revenue for xAI, according to Mirae Asset Securities, offsetting losses from the underutilized hardware.

Anthropic may generate $15 billion in incremental annual recurring revenue from the inference capacity, based on a 3x revenue multiplier cited by Anthropic CEO Dario Amodei. Financial details of the Colossus deal are not publicly available.