Why does OpenAI need six giant data centers? (arstechnica.com)

OpenAI, Oracle and SoftBank announced five new U.S. sites for their joint Stargate AI infrastructure project, bringing planned capacity to nearly 7 gigawatts and pushing investment past $400 billion over the next three years on a path to the full $500 billion / 10‑gigawatt commitment by end of 2025. Three Oracle‑developed sites (Shackelford County, TX; Doña Ana County, NM; and an undisclosed Midwest location) plus a 600‑MW expansion near the Abilene flagship add roughly 5.5 GW of potential draw; two SoftBank sites (Lordstown, OH, and Milam County, TX) could scale to ~1.5 GW over 18 months. The Abilene campus is already operational with Oracle delivering Nvidia GB200 racks and OpenAI running both training and ChatGPT inference there. Technically, the buildout underscores the enormous compute and power requirements of modern LLM training and high‑volume inference (ChatGPT reports ~700 million weekly users), with on‑site power draw measured in gigawatts and tens of thousands of expected jobs created. The push accelerates vertical integration of hardware, datacenter capacity and energy procurement, but critics warn the scale and capital structure may be hard to sustain economically. For the AI/ML community, Stargate means more colocated high‑density GPU capacity for next‑generation model development — and a growing focus on energy, supply chains and deployment economics as limiting factors for future model scale.