Everyone fixated on Nvidia’s GPU allocation this year missed the quieter squeeze happening one layer below the processor. In 2026, supply constraints have increasingly shown up in critical power-delivery parts, including power management ICs (PMICs), voltage regulator modules (VRMs), and silicon carbide (SiC) power components, as AI data centers push rack power densities far beyond traditional deployments. The companies that regulate the energy spikes inside those clusters are now sitting on backlogs that GPU makers would recognize.
A single NVIDIA GB200 NVL72 (Blackwell) rack is widely described as a ~120 kW-class system, and some analyses cite figures closer to ~140 kW depending on configuration. By contrast, NVIDIA’s DGX H100 SuperPOD reference architecture shows rack power exceeding 40 kW in an example configuration. That step-change in density is a primary reason many legacy data center environments can’t accommodate modern AI deployments without substantial electrical and cooling redesign. Every one of those redesigned racks needs a dense chain of power conversion hardware that most investors have never heard of.
Every AI server node requires a sophisticated array of step-down converters, multiphase buck regulators, and high-efficiency switches to convert facility power down to the sub-1V levels used by processor cores. Lead times can stretch into the 20-to-30-week range for certain AI-grade power components when demand spikes and supply is tight, particularly around higher-end PMIC and power-stage content.
Monolithic Power Systems is the clearest financial expression of this crunch. MPS has established itself as a top-tier supplier of power management for AI servers, including solutions that sit close to the processor and handle fast-changing load transients. In Q2 2026, MPS reported record quarterly revenue of $980.6 million. Enterprise Data revenue was $380.6 million, up 44.8% sequentially (about 45%) on broad-based demand. Management also raised its 2026 Enterprise Data growth outlook to 130% from 85%.
For NVIDIA’s next-generation Vera Rubin platform, analysts have argued MPS could capture as much as roughly 70% share of key power sockets on the platform, a dynamic some believe could add on the order of $100 million to 2026 estimates in the second half of 2026. That kind of socket dominance compounds because once a socket is secured, it tends to translate into follow-on orders over the life of the platform.
Infineon is the other name worth watching. Infineon has pointed to very strong demand for AI data center power supply solutions and has discussed expectations for revenue from AI data-center applications of roughly 1.5 billion euros in fiscal 2026 and approximately 2.5 billion euros in fiscal 2027. Infineon has also introduced dual-phase smart power stages aimed at AI accelerators, describing power density above 2 A per square millimeter in a compact package and support for up to 300 amps peak current.
The risk is real. Many power management components are produced on mature process technologies, where capacity additions can be slower and more conservative than the capital flowing into leading-edge compute nodes. If hyperscalers slow orders or the industry’s shift toward higher-voltage architectures (including 48V and beyond) stalls, backlogs can deflate fast. MPS itself has warned that AI server demand and customer ordering patterns could prove more volatile than current expectations.
The wealth takeaway is simple: the loudest AI trade this cycle ran through GPU makers. The quieter one, still underway, runs through the companies keeping those GPUs from overheating and frying themselves.
