Nvidia has unveiled new datacentre power-management systems aimed at squeezing more usable computing capacity out of the electricity supplied to AI facilities, since chip sales are increasingly constrained not by manufacturing but by how much power grids can actually provide. The move matters because ageing GPU generations remain in service for years, so demand for grid capacity keeps rising, and any unused "stranded" power in a datacentre represents GPUs Nvidia could otherwise have sold; by helping operators run facilities more efficiently, the company aims to keep growth from stalling on infrastructure limits.

The new DSX platform includes DSX MaxLPS, which links compute hardware with physical infrastructure such as power cabinets, batteries, coolant units and chillers so they can share data and adjust output dynamically rather than running at constant maximum capacity, addressing the fact that datacentres typically use only around 80 percent of their power budget for actual compute. A companion API called DSX Exchange lets this system pull in data from third-party equipment makers such as Vertiv and Schneider Electric. In a demonstration with cloud provider Lambda, the approach allowed 19 compute nodes to run within the power budget previously needed for 16, lifting cluster throughput and performance-per-watt by roughly 24 percent; separately, Nvidia worked with Emerald AI and utility Silicon Valley Power on its DSX Flex platform to free up datacentre power during periods of high grid demand.