xAI's Colossus 2 will be the first gigawatt-scale datacenter in the world, and it is being built faster than anyone thought possible. The project started March 7, 2025, when xAI acquired a 1 million square foot warehouse in Memphis plus 100 adjacent acres. By August 22, 2025, 119 air-cooled chillers were on site, representing roughly 200MW of cooling capacity, enough for approximately 110,000 GB200 NVL72 GPUs. That is six months of construction for what took Oracle, Crusoe, and OpenAI 15 months to achieve.
Colossus 1, built in 122 days with roughly 200,000 H100/H200s and 30,000 GB200 NVL72 units at 300MW, was already the largest single-coherent training cluster on earth. Colossus 2 makes it look modest. SemiAnalysis proprietary datacenter model, the same model that called Oracle's compute deals months before announcement, projects xAI's total single-cluster capacity will surpass both Meta Superintelligence and Anthropic by Q3 2025. Nvidia GPU allocations are secured. Capital raise is the remaining variable.
The full report covers two distinct topics worth reading in sequence. The first half details the construction mechanics and infrastructure specifics of Colossus 2. The second half covers Grok model trajectory and, more importantly, a unique reinforcement learning method xAI is deploying that SemiAnalysis believes could position xAI ahead of OpenAI, Anthropic, and Google on model capability. The RL methodology section alone justifies reading the original.
[READ ORIGINAL →]