AI data centers are using electricity at a pace so intense that it's causing equipment to break down faster than expected. Drew Baglino, CEO of Heron Power Electronics, explains that in a single moment, these facilities can surge to 1.5 gigawatts of power, which is 50% higher than their typical capacity. Such extreme power fluctuations are putting massive strain on components that weren't designed to handle such sharp changes.
At xAI's data center in Memphis, Tennessee, gas-fired turbines began to crack under the pressure, according to a report from one of the people interviewed. The company added batteries to help manage the power swings and protect the turbines, but even these solutions are suffering. In some cases, the batteries have had to be replaced in just weeks, far shorter than their expected lifespan. The rapid degradation is a sign of the stress caused by AI’s fluctuating energy needs.
How AI Strains Power Systems
AI demands power in sudden and dramatic ways, especially during the training of new models. When training, hundreds of thousands of graphics processing units (GPUs) all activate or shut off almost instantly. This causes massive shifts in power use, creating repeated shocks to the system. Unlike traditional data centers, which operate with more predictable electricity consumption, AI facilities experience extreme, fast-moving surges that stress their infrastructure.
Shannon Miller, founder of Mainspring Energy, compared the power demands to a city the size of Boston suddenly turning half its lights on and off every few seconds. These rapid changes are overwhelming the equipment. For example, some upcoming AI campus projects in Texas and the Midwest are so large that their average energy use equals that of New York City. These extreme energy needs are pushing existing power systems to their limits.
The Hidden Costs of AI Expansion
The physical stresses on data center equipment are becoming hard to ignore. Smaller components like gas-powered combustion engines, which help generate power, are breaking down more frequently under the strain. Jennifer Scanlon, CEO of UL Solutions, warned that these stresses can result in electrical arc flashes—dangerous surges of electricity that jump between conductors. These events can harm crucial AI chips and lead to system failures.
The financial costs are also significant. Investors and lenders are closely tracking how these failures impact the bottom line. Replacing batteries and generators quickly adds billions in unexpected expenses to the already enormous capital outlays of hyperscale companies. Jon Parrella, CEO of Terraflow Energy, compared the situation to shifting a high-performance sports car from the highest gear directly to the lowest one. It's a jarring move that stresses the engine far beyond its normal limits, leading to damage or even breakdowns.
These issues are emerging at a time when power grids across the world are already struggling to meet demand. AI facilities, with their extreme power consumption and unpredictable patterns, are adding another layer of instability. The result is a growing worry among grid operators and energy providers that the infrastructure may not be able to keep up with the rapid expansion of AI computing.
The strain is not limited to a few sites. Reports from over 30 power experts and industry professionals in the U.S. and Europe indicate that the physical stresses are widespread. Cranks on small natural gas combustion engines used in power generation at data centers have broken off, as have turbines at facilities like xAI’s Memphis operation. Similar issues are being reported at much smaller data centers in the UK, where even newer hydrogen-powered fuel cell technologies are being tested to stabilize the power flow.
Despite the availability of technologies like batteries, capacitors, transformers, and flywheels to smooth out power fluctuations, many AI data centers are not deploying these solutions in time to keep up with the pace of expansion. The lack of these stabilizing components means the risk of power failures and equipment damage is higher, especially as new, more energy-intensive servers from companies like Nvidia are set to arrive in 2027.