Case Studies
The real buildouts, decoded.
Four of the most consequential AI campuses on Earth — each one an expensive bet on a different philosophy of power, silicon, and geography. These pages separate public facts from marketing, cite sources, and extract the engineering and political lessons.
Colossus 1 & 2 — xAI's Memphis Gigawatt Campus
Memphis, Tennessee / Southaven, Mississippi
The fastest gigawatt ever built — 100k H100s online in 122 days. Controversial gas turbines, MLGW substations built in 97 days, and the cost of speed.
- Power
- 250 MW live → 2 GW target by end-2026
- Silicon
- 200K H100/H200 live · targeting 1M+ GPUs
- Milestone
- July 2024 (122-day build)
Stargate Abilene — OpenAI × Oracle × SoftBank's Flagship
Abilene, Texas (Lancium Clean Campus)
$500B committed. 10 GW target. The Trump-announced moonshot — now capped at 1.2 GW in Abilene after grid-availability wall.
- Power
- 1.2 GW (capped March 2026)
- Silicon
- 450,000 GB200 Blackwell under 15-year Oracle lease
- Milestone
- First 2 buildings September 2025
Project Rainier — Anthropic's 2.2 GW Trainium Campus
New Carlisle, Indiana (1,200 acres, near Lake Michigan)
The world's largest non-NVIDIA AI cluster. Why Anthropic bet on AWS silicon, and how UltraServers replace NVLink.
- Power
- Current 7 buildings → 30 planned · 2.2 GW ultimate
- Silicon
- ~500K Trainium2 live · targeting 1M+ Trainium2/3
- Milestone
- October 2025 (7 months from groundbreaking)
Hyperion — Meta's 5 GW Louisiana Campus
Richland Parish, Louisiana (2,250 acres)
$27B investment. 10 gas plants approved. Bigger than most utilities. When a single data center becomes 15% of a state's electricity demand.
- Power
- 2 GW initial → scaling to 5 GW
- Silicon
- Primarily NVIDIA for Llama training
- Milestone
- Under construction · operational target 2030
Editorial policy
Every capacity figure, silicon count, and cost number is sourced to a public press release, 10-K/S-1 filing, permit document, or analyst report (SemiAnalysis, The Next Platform, DCD). We flag speculation vs confirmed data.