US hyperscalers to guzzle 22% more grid juice by end of 2025 (www.theregister.com)

451 Research (S&P Global) warns U.S. hyperscale datacenters will need about 22% more grid power by the end of 2025 than a year earlier, jumping to 61.8 GW this year (+11.3 GW), then projected to 75.8 GW in 2026, 108 GW in 2028 and 134.4 GW by 2030. The surge is driven primarily by new machine‑learning workloads that require dense, GPU‑packed server clusters and heavy cooling — a build‑new‑rather‑than‑retrofit economics — and excludes enterprise-owned sites. Massive hyperscaler capex underpins the boom (Amazon ~$100B, Microsoft ~$80B, Google ~$85B, Meta $66–72B), and Virginia and Texas are already the largest demand centers (Virginia ~12.1 GW in 2025 from 9.3 GW; Texas ~9.7 GW). That scale-up has immediate infrastructure implications for the AI/ML community: utilities face strained interconnection capacity, uneven regional availability, and new tariff regimes that can chill growth. AEP Ohio’s rule forcing datacenters to be liable for 85% of subscribed energy cut interconnection requests dramatically (from >30 GW to 13 GW), highlighting overprovisioning and jurisdictional duplicate requests. Operators are therefore scouting alternative sites (Idaho, Louisiana, Oklahoma, West Texas), and increasingly adopting on‑site generation — retired nuclear, gas, batteries, fuel cells, renewables or hybrids — as stopgaps. The result: faster access to GPU compute in some markets but potential delays, higher costs, and shifting sustainability tradeoffs that will affect model training cadence, deployment economics, and infrastructure planning.