XAI likely first AI company with Gigawatt plus campus (semianalysis.com)

xAI has rapidly followed Colossus 1 with Colossus 2, a buildout that went from land purchase (Mar 7, 2025) to ~200 MW of installed cooling capacity in six months (119 air‑cooled chillers by Aug 22). Colossus 1 — erected in 122 days — still holds the title of the largest single coherent training cluster (~200k H100/H200s + ~30k GB200 NVL72; ~300 MW). Colossus 2’s 200 MW of cooling could host ~110k GB200 NVL72, and xAI has vertically expanded power by acquiring a former Duke Energy plant in Southaven, MS, deploying Tesla Megapacks and gas turbines (seven 35 MW units operating). Through a JV with Solaris Energy Infrastructure (Solaris’ 1,700 MW orderbook, ~1,140 MW effectively tied to Musk’s projects), xAI aims to scale to >1.1 GW of generation by Q2 2027 and potentially >1.5 GW overall — enough to create the world’s largest single datacenter cluster once GPUs are installed. Nvidia has allocations ready; the bottleneck is tens of billions in CapEx. Significance: this cements the compute arms race—xAI’s speedy, hybrid approach (warehouse conversion + leased turbines + rapid turbine rentals) pressures OpenAI, Meta and Anthropic on time‑to‑market and single‑site scale. If realized, xAI could outsize Meta/Anthropic for single‑cluster capacity by Q3 2025 per industry modeling, changing training economics and model scale strategies. Key risks remain: massive fundraising needs (rumored $10sB raise, Middle East sovereign interest), regulatory/permit workarounds for power, and whether capital and GPU installation timelines match the infrastructure build.