AI’s own power surge is starting to bite

Rapid fluctuations in power demands caused by AI data centers are straining essential equipment, leading to malfunctions and accelerated wear. These issues raise concerns about added costs and reliability problems, coming at a time when investors and lenders are already wary of hyperscalers' massive spending on AI facilities. The problems also pose a potential threat to the stability of power grids already struggling to meet increasing demand.

Amber Villegas-Williamson, a consultant at the Uptime Institute in the UK, explains that AI computing demands are unlike any seen before, akin to rapidly revving a car's engine which wears it out faster than maintaining a steady pace. Data centers designed for AI processing have unique needs, as their power demands can surge dramatically and change rapidly, causing repeated shocks to connected equipment.

These fluctuations can be so extreme that a gigawatt data center can switch on and off like a city the size of Boston every few seconds, according to Shannon Miller, founder of Mainspring Energy Inc. Some planned AI campuses in Texas and the Midwest are even larger, consuming power similar to New York City.

The strain on power supply becomes particularly evident during the training of new AI models, a process that engages all graphics processing units (GPUs) simultaneously. This can cause power usage to spike up to 50% above the facility's design capacity, meaning a 1 gigawatt facility may temporarily use 1.5 gigawatts.

Most equipment isn't designed to handle such rapid power swings. Jon Parrella, CEO of energy-storage developer Terraflow Energy, compares it to suddenly shifting gears from sixth to first in a Ferrari, which isn't feasible. Batteries installed to smooth out power swings are often replaced within months due to the high strain.

Several power experts interviewed for this story reported visible physical stresses on the facilities, including broken cranks on natural gas engines, cracked gas-fired turbines, and premature wear on various components. These issues lead to electrical arc flashes, which can damage AI chips. While some stabilizing equipment exists, such as batteries, capacitors, transformers, and flywheels, there is a shortage of these technologies in new data centers, according to Villegas-Williamson.