AI’s Rushed Power Swings Are Bricking Its Own Data Centers Before They Turn a Profit

(SeaPRwire) –   By: Ethan Gallagher
I’ve consulted on more than 40 hyperscale data center builds over the last 15 years. The current AI boom is creating a crisis no one planned for. Standard power infrastructure playbooks don’t account for millisecond-scale load swings. It’s like building a highway for passenger cars, then letting 18-wheelers zip back and forth at 200 mph. Operators are rushing to meet demand, but they’re ignoring basic equipment design limits. That’s already leading to costly, avoidable breakdowns across the industry.
The core physical damage is well-documented. More than 36 industry experts confirmed the strain. Cranks on small natural gas engines have broken off at multiple sites. At xAI’s Colossus facility in Memphis, gas turbines developed cracks. UK data centers have reported similar turbine failures, per GeoPura’s Andrew Cunningham. A 1-gigawatt AI data center uses power equal to a city the size of Boston. Half of that load can flicker on and off every few seconds, per Mainspring Energy’s Shannon Miller. Some planned campuses are five times larger, using nearly as much power as New York City on average. AI workloads can see power usage spike 50% above design capacity. A 1-gigawatt facility may use 1.5 gigawatts for a split second, per Heron Power’s Drew Baglino. Standard data center equipment is built for steady, predictable load. It can’t absorb the rapid swings AI workloads create. Amber Villegas-Williamson of the Uptime Institute compared the stress to overrevving a car’s engine. The industry is skipping stabilization gear to build faster. Batteries installed to smooth swings often need replacement in months, not their rated years.
The financial stakes are staggering. Even a few minutes of downtime can cost operators thousands to hundreds of thousands of dollars in lost revenue. The planned 2.67-gigawatt AI campus in West Texas was delayed a full year to 2028. The delay fixes power stability issues for Microsoft’s required 99.999% uptime. Some facilities are already seeing uptime drop to just 80%. That’s far below the 365-day-a-year assumption built into financing models. Investors are already jittery about hundreds of billions in hyperscaler spending. Faster equipment depreciation means ROI timelines are blowing out. These issues also threaten wider power grid stability. Grids already struggle with aging equipment and rising demand.
The data center power supply chain hasn’t caught up to AI’s unique demands. Right now, no one is willing to hit the brakes on building until the first wave of projects tanks completely.
Author bio: Ethan Gallagher, Silicon Valley Hardware Architect and Infrastructure Strategist with 15+ years designing hyperscale data centers.