\n\n
Key developments: Recent grid data show an unprecedented surge in power requests from AI‑focused data centres, with Oncor reporting 186 GW of pending interconnections in 2025 and the total queue swelling to 271 GW by early 2026. MIT Technology Review’s measurements illustrate how model size drives energy use—from 57 J per query for a modest 8‑billion‑parameter Llama model to 3.4 kJ for a 400‑billion‑parameter version, and up to a kilowatt‑hour for just five seconds of AI‑generated video. The International Energy Agency confirms that while per‑task efficiency is improving, AI‑center electricity consumption jumped 50 % in 2025, and U.S. projections suggest data centres could consume up to 15 % of national electricity by 2030 (Berkeley Lab). Additionally, tech giants are scaling up infrastructure: Microsoft plans multi‑gigawatt Maia 300 deployments from 2027, while utilities now scrutinise requests, approving only about 23 % as high‑confidence projects.
Why it matters: The rapid growth in AI workloads—10‑fold annual token increases since 2024—places massive strain on power grids, risking bottlenecks, higher electricity costs, and potential delays in AI deployments. As AI’s share of data‑centre power climbs, the mismatch between fast‑moving AI demand and slower grid upgrades could shape future regulatory and investment decisions.
What to watch next:
1. Grid‑expansion initiatives and fast‑deployed resources (e.g., gas turbines, renewables) that utilities and policymakers propose to meet AI demand.
2. Adoption of higher‑efficiency architectures such as Microsoft’s Maia 300 and 800 VDC designs, which could ease megawatt‑scale power pressure.
3. Regulatory actions on utility‑data‑centre contracts and rate‑payer protections as AI’s electricity share approaches double‑digit percentages.
4. Continued IEA and Berkeley Lab updates tracking AI‑related electricity trends and the effectiveness of efficiency gains versus overall demand growth. [[1, 2]]