AI GPUs force new power‑delivery designs
As GPUs push power density limits, the industry is shifting to shorter power paths, higher efficiency designs and vertical stacking — changes that trickle down into advanced packaging, substrates and data‑center power components. That technical trend will reshape demand for power delivery parts and thermal management solutions. (x.com)
NVIDIA’s Hopper H100 SXM5 modules are specified at up to 700 W per GPU, while NVIDIA’s Grace Hopper GH200 “superchip” configurations are documented with power envelopes up to about 1,000 W per node in vendor materials. ( ) Rack-scale designs shipping today support extreme totals—Supermicro’s GB200 NVL72 rack reference lists a 48U rack power budget of 132 kW and options for liquid cooling to service that load. (supermicro.com) Industry analysts and foundry reports show CoWoS and other wafer-level packaging lines are capacity-constrained because larger interposers and more HBM stacks reduce yields per wafer, pressuring lead times for AI GPU supply. (tomshardware.com) Market data and trade bodies put substrate demand and investment on the rise: the global advanced-substrate market was projected around $15.3 billion in 2024, and substrate makers invested more than $15 billion during 2021–22 to expand capacity for AI-driven packaging. ( ) Supply-chain reporting flags upstream material tightness—ABF and copper-clad laminate (CCL) lead times lengthened as CoWoS demand surged, with glass-cloth and prepreg cited as pinch points by substrate suppliers. (globalsmt.net) System vendors and infrastructure suppliers are shipping new power architectures and PDUs: Flex announced an 800 VDC power-shelf program to support NVIDIA rack platforms, and Schneider and Vertiv have rolled out PDU and liquid‑cooling products aimed at high-density AI racks. ( )
