ROOT / COMPUTE // AI TRAINING CLUSTERS
// Compute · decoded · deep-research pass
AI Training Clusters, decoded.
The gigawatt machines training frontier models, by GPU count and power draw.
SEMI-LIVE◊ confidence: Medium
~1 GW
first crossed (Anthropic)
// the signal · what the data says
Frontier clusters are now measured in gigawatts, not servers. In March 2026 an Anthropic facility became the first to cross 1 GW, beating OpenAI's Stargate; xAI's Colossus in Memphis pioneered the ~12-month gigawatt buildout. Stargate targets 10 GW across five+ US sites. Cluster size is the clearest proxy for who can train the largest models.
// comparison · every entry, benchmarked & annotated
| Name | Maker | GPUs (k) | Power (MW) | Interconnect | Online (yr) |
|---|
| Colossus / Colossus 2 | xAI | 200 | 300 | NVLink+RoCE | 2025 |
| ◊ Memphis; fastest gigawatt-scale buildout on record |
| Rainier | Amazon/Anthropic | 160 | 300 | EFA/Trainium2 | 2026 |
| ◊ Anthropic training on Trainium2; ~1 GW class |
| Fairwater | Microsoft | 150 | 210 | NVLink | 2026 |
| ◊ Wisconsin; OpenAI + Microsoft |
| Stargate Abilene | OpenAI/Oracle | 100 | 300 | NVLink | 2026 |
| ◊ Most advanced Stargate site, ~0.3 GW live, scaling |
| TPU v7 pods | Google | 120 | 160 | Optical OCS | 2025 |
| ◊ Ironwood pods, 9,216 chips each |
| Meta Hyperion | Meta | 130 | 190 | RoCE | 2026 |
| ◊ Louisiana; multi-GW plan on gas |
// leaderboard · GPUs (k)
01Colossus / Colossus 2200
02Rainier160
03Fairwater150
04Meta Hyperion130
05TPU v7 pods120
06Stargate Abilene100
// sources & where to verify
// full index, per-field sourcing & CSV export [ AUTHENTICATE → ]
◊ Deep-research pass, mid-2026. Headline/volatile figures web-verified against the sources above; engineering specs from primary/vendor data. Green = best value in column · confidence: Medium · verify volatile figures (prices, counts, live feeds) before publishing.