How millisecond demand swings are quietly destroying data center hardware

AI data centers are exposing a growing infrastructure challenge because their electricity demand changes far more rapidly than that of traditional data centers, Bloomberg reports. 

During AI model training, hundreds of thousands of GPUs can power up and down within milliseconds, creating repeated swings in power consumption that strain batteries, generators, turbines, transformers and other critical power equipment. Experts report that these fluctuations are causing components to wear out or fail much sooner than expected, with some batteries requiring replacement within weeks or months and turbines and generators developing cracks under the stress. 

These reliability issues increase maintenance costs and can lead to expensive downtime, as interruptions can cost operators anywhere from thousands to hundreds of thousands of dollars per minute in lost revenue, depending on the facility. Some developers have delayed AI data center projects to improve power system reliability, while investors are becoming increasingly concerned that these facilities and their equipment may depreciate faster than expected. 

Beyond the financial impact on operators, the rapid and unpredictable power demands of AI facilities also pose risks to the broader electric grid. 

Large AI campuses can consume as much electricity as major cities, and their sudden shifts in demand can destabilize power networks already under pressure from rising electricity use, aging infrastructure and the increasing share of intermittent renewable energy. U.S. grid regulators have identified large AI data centers as a significant emerging risk to grid stability and are requiring operators to address these issues. 

In response, technology companies, utilities, equipment manufacturers and government researchers are developing batteries, power-conditioning systems, improved facility designs and other technologies to smooth power fluctuations and reduce stress on both data center equipment and the electric grid. These efforts aim to support continued AI infrastructure growth while minimizing reliability problems and grid disruptions.

Bloomberg has the full story.