Datacenter growth meets energy and geopolitics
The practical limit on scaling AI is increasingly power, sites and global supply chains, not just models or chips. Recent coverage highlights datacenter energy and siting as core constraints, pointing to power availability, interconnection timelines and regional risks — and Taiwanese chipmakers are even urging stockpiles of helium and LNG after recent geopolitical shocks. Those constraints turn AI capacity planning into an operational exercise tied to utilities and logistics, not only software selection. (youtube.com) (tomshardware.com)
The bottleneck in artificial intelligence is starting to look less like a chip problem and more like a utility problem. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024, or roughly 1.5% of global power use, and artificial intelligence runs mainly inside those buildings. (iea.org) A data center is basically a warehouse full of computers, and the computers are only part of the bill. The International Energy Agency says servers take about 60% of electricity in a modern facility, while cooling can range from about 7% in efficient hyperscale sites to more than 30% in less-efficient enterprise sites. (iea.org) That is why the first question for a new artificial intelligence campus is no longer “Which chips can we buy?” but “How fast can we get power?” Jones Lang LaSalle said on January 5, 2026 that power, not location or cost, is becoming the primary site-selection criterion because grid connections in major markets can take more than four years. (jll.com) The buildings are getting bigger at the same time the plugs are getting harder to secure. Jones Lang LaSalle expects nearly 100 gigawatts of new data center capacity between 2026 and 2030, which would roughly double global capacity, and says the sector could require up to $3 trillion of investment by 2030. (jll.com) Developers are reacting by chasing electricity the way factories once chased railroads. Bloom Energy’s 2026 Data Center Power Report says capital is concentrating in power-advantaged regions, interconnection timelines are widening, and on-site generation is moving from a temporary bridge to a long-term strategy. (bloomenergy.com) That shift reaches all the way down to the wiring inside the building. Jones Lang LaSalle says rack densities are approaching 100 kilowatts and liquid cooling is becoming a design requirement, which means the data center itself is being rebuilt around heavier power delivery and more aggressive heat removal. (jll.com) The grid side is getting touchier too. A Harvard Belfer Center policy brief published on February 10, 2026 says United States data center demand could rise from 176 terawatt-hours in 2023 to 325 to 580 terawatt-hours by 2028, and it cites a July 2024 voltage fluctuation in northern Virginia that disconnected 60 data centers and created a 1,500-megawatt power surplus. (belfercenter.org) Now add geopolitics, and the story stops being local. Taiwan Semiconductor Industry Association president Wu Chih-i said Taiwan should hold strategic reserves of helium and other critical gases, while Taiwan News reported on April 9, 2026 that Taiwan has about 11 days of liquefied natural gas supply and no strategic helium stockpile. (taiwannews.com.tw) Helium sounds obscure until you see where it sits in chipmaking. Taiwan News reported that chipmakers use helium for wafer cooling and leak detection, and that damage to Qatar’s Ras Laffan helium complex in March tightened supply and pushed buyers toward the United States and Canada. (taiwannews.com.tw) Liquefied natural gas matters for a different reason: it keeps the lights on. Taipei Times reported on March 18, 2026 that Taiwan had secured more than half of its liquefied natural gas needs for May after Middle East fighting threatened the fuel flows that support the island’s power system and semiconductor industry. (taipeitimes.com) So the new map of artificial intelligence is being drawn by transmission queues, gas tankers, industrial gas plants, and local substations. If you can get the latest model but not 500 megawatts, chilled water, and a reliable fuel chain, you do not have an artificial intelligence strategy yet; you have a purchase order.