AI strain hits infrastructure
Enterprise AI demand is outpacing available compute and electricity, forcing providers to ration access and raise prices as outages and disruptions spread across the sector. Several reports say buyers are locking capacity years ahead and that this scarcity is shifting the primary constraint from models to hardware and energy (enterpriseai.economictimes.indiatimes.com).
Artificial intelligence demand is running into a harder limit than software: enough chips, data centers and electricity to keep the systems online. (iea.org) A data center is a warehouse of servers, networking gear and cooling equipment, and the International Energy Agency said data centers used about 415 terawatt-hours of electricity in 2024, or roughly 1.5% of global power demand. Servers account for about 60% of that load in modern facilities, with cooling taking 7% to more than 30% depending on the site. (iea.org) The spending race has moved from model labs to concrete and transformers. Microsoft said on January 3, 2025 that it was on track to invest about $80 billion in fiscal 2025 on AI-enabled data centers, and Meta said on January 28, 2026 that its 2025 capital expenditures reached $72.22 billion. (blogs.microsoft.com) (investor.atmeta.com) Buyers are locking up capacity years ahead instead of waiting for spot access. CoreWeave said on March 10, 2025 that it signed a deal worth up to $11.9 billion to provide OpenAI with dedicated compute capacity for model training and services. (coreweave.com) The bottleneck is now visible in the power queue. Dominion Energy told Virginia regulators in a February 2, 2026 filing that data centers had requested about 70,000 megawatts of power, nearly three times the utility’s record peak load of 24,678 megawatts on January 23, 2025. (virginiabusiness.com) That queue keeps growing. Dominion said it receives about 10 new large-load requests a month totaling 2,000 to 3,000 megawatts, after reporting 451 Virginia data centers using 3,583 megawatts in 2024. (virginiabusiness.com) Cloud companies are now shopping for electricity the way they used to shop for servers. Oracle agreed this week to buy as much as 2.8 gigawatts of Bloom Energy fuel-cell capacity for artificial intelligence data centers. (bloomberg.com) Power sourcing has widened beyond the grid. Reuters reported on April 9, 2026 that large technology companies were backing next-generation nuclear projects to secure future electricity for AI data centers. (reuters.com) The buildout is also running into local resistance. Reuters reported on April 6, 2026 that Amazon, Microsoft and Google had each recently abandoned multibillion-dollar data center projects after community opposition over water, power use and pollution. (reuters.com) Alphabet’s investor site shows it is due to report first-quarter 2026 results after the market closes on April 23, 2026, giving the market another read on how fast the infrastructure bill is still climbing. For now, the contest is less about who has the smartest model than who can actually plug the next cluster in. (abc.xyz)
