xAI's Colossus Supercomputer to Reach 1.44 Million GPUs by End of 2026
Summary
- • xAI adding 660,000 Nvidia GB300 GPUs this year, reaching 1.44M total by late December
- • Colossus 2 uses only Blackwell GPUs; Colossus 1 leased to Anthropic for inference
- • xAI building 1.2-GW dedicated power plant amid lawsuits over gas turbines
- • Long-term goal: 7x data center capacity by 2027, 50M H100-equiv GPUs by 2030
Details
Total GPU target: 1.44 million units by late December 2026
Three tranches of 220,000 Nvidia GB300 GPUs added ~weekly, in November, and late December, joining 780,000 existing GPUs across Colossus 1 and 2.
Colossus 2: 110K GB200 + 440K GB300 (all-Blackwell architecture)
Exclusively Blackwell-generation GPUs eliminate mixed-architecture bottlenecks, making Colossus 2 the primary cluster for Grok model training.
Colossus 1: 150K H100 + 50K H200 + 30K GB200 (mixed generation)
Mixed Hopper-Blackwell architecture is inefficient for Grok training; xAI has leased this cluster to Anthropic for inference workloads instead.
1.2-GW dedicated power plant under construction at Colossus site
Power — not GPUs — is the binding constraint. xAI ran unpermitted gas turbines to bridge the gap, resulting in a neighborhood lawsuit; turbines are being removed over a one-year period.
7x data center capacity growth targeted by 2027; 50M H100-equiv by 2030
Musk's roadmap extends far beyond the 1.44M GPU milestone, with plans for sevenfold capacity expansion by 2027 and a 50 million H100-equivalent GPU target by 2030.
SpaceX planning Orbital Data Center System with 1 million satellites
A long-term orbital compute vision is in early planning; Nvidia CEO Jensen Huang described the concept as a 'dream' for now.
xAI is only 3 years old vs. Anthropic (6 years) and OpenAI (10 years)
Despite its shorter history, xAI has assembled one of the world's largest GPU clusters, closing the gap with rivals that had a multi-year head start.
Infrastructure build-out details and strategic roadmap for xAI's Colossus supercomputer expansion.
What This Means
xAI's Colossus expansion to 1.44 million GPUs by year-end positions it among the largest AI compute operators in the world despite being only three years old. The pragmatic decision to lease the architecturally mismatched Colossus 1 to rival Anthropic reveals that the economics of AI infrastructure have matured to the point where idle compute has a market regardless of competitive relationships. Power — not silicon — is now the defining bottleneck for frontier AI labs, and xAI's construction of a dedicated 1.2-GW plant underscores how the AI infrastructure race has expanded from chip supply chains into utilities, permitting, and grid policy.