AI power bottleneck
Hyperscalers and customers are running up against real limits on deliverable power for AI, not just chip supply: AWS added 3.9GW of capacity in 2025 but still reports “capacity constraints,” and some large customers reportedly tried to buy all available instance capacity for 2026. That squeeze is pushing data‑centre operators toward fast, dispatchable generation like gas turbines and is changing utility behaviour and permitting debates around new AI loads. (networkworld.com) (turbomachinerymag.com)
Artificial intelligence builders are running into an older limit than chips: electricity they can actually get to a data center on time. Amazon Web Services said it added 3.9 gigawatts of power capacity in 2025 and still had “capacity constraints” that left demand unserved. (aboutamazon.com) Amazon Chief Executive Andy Jassy said in his 2025 shareholder letter, published April 9, 2026, that two large customers asked to buy all available 2026 capacity for Amazon Web Services’ Graviton instances. He said Amazon Web Services refused because other customers also needed the capacity. (aboutamazon.com) (networkworld.com) A data center is a warehouse full of servers, and servers need both chips and steady power to run. Amazon said it expects to double its total power capacity by the end of 2027, a sign that the bottleneck now includes substations, transmission lines, and generation as well as semiconductors. (aboutamazon.com) Federal energy regulators have already moved to write new rules for these giant new loads. In December 2025, the Federal Energy Regulatory Commission ordered PJM Interconnection to develop rules for colocating data centers and other large loads at power plants while protecting grid reliability and allocating costs. (ferc.gov) (utilitydive.com) The load forecasts behind that fight are large. The Department of Energy said United States data center electricity use rose from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023 and could reach 325 to 580 terawatt-hours by 2028. (energy.gov) The Electric Power Research Institute now projects data centers could consume 9% to 17% of United States electricity by 2030, up from about 4% to 5% today. Its 2026 scenarios say the heaviest impacts will be concentrated in a small number of states and utility territories. (powering-intelligence.epri.com) That timing problem is why gas turbines are back in the conversation. Turbomachinery Magazine reported that data center developers are turning to aeroderivative gas turbines because they can ramp quickly and provide dispatchable power while grid connections and longer-lead projects lag. (turbomachinerymag.com) The scramble is now affecting the power industry’s own queue. Bloomberg reported on April 1 that turbine prices are rising as electricity demand from data centers surges, making it harder and more expensive to secure on-site generation for new campuses. (bloomberg.com) Utilities are also pressing regulators to speed up approvals tied to new demand. In Wisconsin, a coalition including Xcel Energy and American Transmission Company asked federal regulators this month to pause some competitive bidding rules, saying delays threaten power infrastructure needed for data centers and artificial intelligence growth. (wisconsinwatch.org) The result is that the artificial intelligence buildout is now constrained by the same things that limit factories and refineries: land, wires, permits, turbines, and megawatts. Amazon’s 2026 message to investors was not that demand for computing is cooling, but that power is becoming the gating item. (aboutamazon.com)
