Colossus (Memphis, phase 1)
xAI brings up a roughly 100,000-GPU H100 training cluster in Memphis in 122 days, later doubled to about 200,000 GPUs.
A single-site cluster built in a converted Memphis factory in 122 days and used to train Grok 3. The build speed was the point, since it gave xAI frontier-scale compute within its first two years.
- Date
- 2024-09
- Lab
- xAI
- Kind
- infra
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Build time | 122 days to full operation | company |
| Initial GPUs | ~100,000 Nvidia H100 Wikipedia summary | third-party |
Day of month not confirmed in sources I opened (Wikipedia gives July to September 2024). 'access' field is not meaningful for infra; internal training cluster. xAI's Grok 3 post says it was trained on Colossus with plans to reach 200,000 GPUs.
Sources
- en.wikipedia.org/wiki/Colossus_(data_center)
- interestingengineering.com/ai-robotics/elon-musk-xai-gigawatt-scale-ai-training-cluster
- x.ai/news/grok-3
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check