Severe Power Fluctuations in AI Data Centers Are Destroying Their Own Infrastructure
Miles Bennett
AI data centers' extreme power swings are wrecking their own equipment — batteries dying in weeks, generator crankshafts snapping, gas turbines cracking. This means → the real cost of the AI buildout isn't just electricity bills; it's the facilities themselves wearing out far too fast.
Why does AI power demand behave so differently from a normal data center?
Traditional data centers draw power at a relatively steady rate. AI training is different: hundreds of thousands of GPUs start and stop in sync within milliseconds, spiking power demand up to 50% above rated capacity.
In plain terms = a facility designed for 1 gigawatt can suddenly pull 1.5 gigawatts — equivalent to a city the size of Boston surging all at once.
Planned AI campuses in Texas and the US Midwest already exceed 5 gigawatts, with average consumption approaching New York City's. This signals a structural industry-wide problem, not a handful of edge cases.
What equipment is already breaking?
Small natural-gas generators at multiple data centers have suffered snapped crankshafts. At Elon Musk's xAI Colossus facility in Memphis, gas turbines cracked, forcing the site to retrofit batteries to smooth out power swings.
GeoPura CEO Andrew Cunningham confirmed that smaller UK data centers are seeing the same turbine cracking — This means → the damage is not limited to hyperscale sites; mid-size facilities are hit too.
Some batteries installed specifically to buffer these swings are failing under the stress and need replacement within weeks or months. UL Solutions CEO Jennifer Scanlon warned that cracks and wear can trigger arc flash — electricity jumping between conductors — potentially destroying the AI chips themselves.
How do insiders describe the stress?
Uptime Institute principal consultant Amber Villegas-Williamson's analogy: "Constant high-RPM driving wears out an engine faster than cruising at a steady speed."
Drew Baglino, former Tesla executive and founder of Heron Power Electronics, put it more bluntly: the power cycling is like "driving a Ferrari and slamming from sixth gear straight into first."
Baglino's company is building power-swing management hardware for Nvidia's next-generation, higher-wattage servers expected in 2027. This reflects the industry's recognition that the problem will worsen as compute scales up.
Why weren't these safeguards installed from the start?
Batteries, capacitors, transformers, and flywheels can all stabilize power flow in theory. But Bloomberg's interviews with more than 30 power specialists across the US and Europe found that new AI data centers routinely skip or under-deploy these technologies.
In plain terms = everyone is racing to bring capacity online; power-smoothing gear gets filed under "we'll add it later" — but the equipment is already breaking.
The gap drives up operating costs, and even a few minutes of downtime can directly hit a developer's revenue.
What does this mean for investors?
Investors and lenders are already wary of the hundreds of billions of dollars being poured into hyperscale data centers. Premature equipment failure deepens concerns that these facilities may depreciate far faster than projected.
This means → asset-life assumptions in data-center valuations may need rewriting, directly affecting financial models and lending risk assessments.
Violent power swings also pose a potential shock to already-strained public grids — making this not just a corporate cost issue but a public infrastructure risk.
Content is for reference only, not financial advice.