Who makes power smoothing / supercapacitors for AI data centers

A large training job running across many GPUs can start, checkpoint, or pause in near-unison, producing power swings sharp enough to stress voltage regulation on the equipment upstream — UPS (7.10), switchgear (7.05), and potentially the grid interconnection itself. Supercapacitors respond on a timescale that battery chemistry cannot match, discharging and absorbing energy fast enough to flatten these transients before they propagate further up the distribution chain, typically working alongside rather than instead of the UPS in 7.10. This is complemented by firmware-side smoothing built into the GPUs and cluster management software themselves, which can stagger workload transitions or throttle power draw to reduce the swing at the source rather than only absorbing it downstream. Because this is a problem largely specific to large, synchronized AI training clusters rather than conventional enterprise IT load, it is a newer product category with a small number of suppliers and no long incumbency to displace, and utilities are increasingly interested in it as a way to let data centers connect to the grid without the swings they would otherwise cause. For investors, this is a young, fast-evolving niche: watch which supercapacitor and power-electronics suppliers get qualified into major GPU cluster reference designs, and how much of the smoothing job software and firmware end up doing instead, which would reduce the addressable hardware market.