AI Power Swings Cause Equipment Failures and Project Delays

Unprecedented power fluctuations at artificial intelligence data centers are forcing developers to delay major computing projects and re-engineer critical infrastructure, threatening the financial assumptions underlying billions of dollars in AI investments.

To guarantee the 99.999% power reliability demanded by Microsoft Corp., developers of a planned 2.67-gigawatt AI campus in West Texas have pushed back its operational target from 2027 to 2028. Chris James, chief executive officer of Joulent Inc., which is building the facility alongside Chevron Corp., said the extension was required to build in extra engineering time to withstand extreme electrical swings.

While data centers are modeled on the premise of 24-hour continuous uptime 365 days a year, unmanaged power swings are pushing actual availability down toward 80% at some sites, according to a person involved in project financing. With offline compute capacity costing operators anywhere from thousands to hundreds of thousands of dollars per minute in lost revenue, unexpected downtime presents a severe threat to project returns within the next 12 to 24 months.

“If essential equipment breaks down prematurely, the financial consequence is not primarily replacing a pump or a breaker or some power component — it’s the the value of that expensive compute capacity not generating revenue because it’s offline,” said Jason Hoffman, chief strategy officer at data-center builder and operator Switch.

The physical toll on power-generation and stabilization hardware has already manifested across multiple global facilities. Gas-fired turbines at xAI’s Colossus computing site in Memphis, Tennessee, developed internal cracks, prompting the installation of battery arrays to smooth power flows and reduce strain on spinning equipment. Parent company SpaceX did not respond to requests for comment. Similar turbine cracking has occurred at smaller sites in the UK, according to Andrew Cunningham, chief executive officer of GeoPura Ltd., which provides hydrogen fuel cells to stabilize data center power.

Equipment failures are multiplying across global supply chains, from broken crankshafts on small natural gas combustion engines to electrical arc flashes capable of damaging expensive AI chips, according to Jennifer Scanlon, chief executive officer of certification group UL Solutions Inc. At several facilities, backup batteries intended to absorb power shocks have required complete replacement within weeks or months of operation.

“AI does create very unusual power demand,” said Amber Villegas-Williamson, principal consultant at the Uptime Institute in the UK, which advises electricity suppliers and data centers on standards and reliability. “It’s like over-revving your car wears out the engine faster than keeping a constant speed.”

The root cause lies in how AI models are trained. When hundreds of thousands of graphics processing units (GPUs) execute workloads simultaneously, electricity usage can surge or plummet across hundreds of megawatts in milliseconds. Drew Baglino, former Tesla Inc. executive and founder of Heron Power Electronics Co., noted that power consumption during training spikes up to 50% above a site’s design capacity. “so a 1 gigawatt facility may use 1.5 gigawatts for a split second,” said Baglino, whose firm is building management hardware for Nvidia Corp.’s next-generation server architectures planned for 2027.

Standard electrical componentry struggles to manage such instantaneous shifts. Jon Parrella, chief executive officer of energy-storage developer Terraflow Energy, likened the dynamic to driving a high-performance vehicle and shifting instantly from sixth gear to first. “You can’t swing that fast,” he said.

The sheer scale of these spikes exacerbates the strain. A single 1-gigawatt facility draws as much power as the city of Boston, with half that capacity turning on or off in seconds, according to Shannon Miller, founder and president of Mainspring Energy Inc. Proposed AI campuses in Texas and the Midwest are planned at five times that capacity, approaching the average energy draw of New York City.

The volatility also poses broader systemic risks to regional electric grids. Power-quality expert Sreemant Roy, global offer manager at Schneider Electric in Nashville, Tennessee, warned that AI loads induce sub-synchronous oscillations that can travel back into the transmission system and damage external utility equipment. “These loads are extremely dynamic or fluctuating, which causes grid instability and can lead to, if not corrected, potential blackouts or power outages,” Roy said. “That has made utility companies globally very worried.”

Regulators have begun intervening. The North American Electric Reliability Corp. evaluated more than 33 gigawatts of operating US data centers and determined in a September report that roughly three-quarters relied on load models that failed to capture dynamic power shifts. NERC subsequently issued a rare level-three alert requiring big data centers to address these immediate risks and submit their responses by Aug. 3.

In response, operators have experimented with running continuous “dummy math” calculations to maintain flat power profiles during idle training periods, though the practice has drawn scrutiny for consuming vast amounts of excess electricity. Hardware manufacturers are also modifying designs. Nvidia started working more closely with power experts when it developed Blackwell GPUs, which were first released in 2024 and have become pervasive in data centers. Dion Harris, senior director of hyperscale infrastructure solutions at Nvidia, said “We’re building the chips and processors” but also using them in the company’s own data centers, working to make deployments smoother “both on the data-center build out, design and engineering phase, as well as on the power delivery.”

To address grid integration challenges, the US Department of Energy established a dedicated testing site last year at the National Laboratory of the Rockies near Denver, Colorado. Program manager Martha Symko-Davies said power suppliers are utilizing the facility’s on-site GPUs and generation assets to test batteries, control software, and grid-interface equipment designed to absorb harmful oscillations. “We have the opportunity right now to get it right,” said Symko-Davies.