Think of electricity as a river. Coming out of the dam (the grid), it flows at very high pressure (high voltage). To make it useful inside a building, you need to step it down through a series of pressure-reducing valves until it reaches a faucet you can actually drink from. Every "valve" is a power semiconductor. Most are tiny silicon chips. The newest, most demanding ones — for EVs, AI servers, and grid infrastructure — use silicon carbide (SiC) for high-voltage stages or gallium nitride (GaN) for very fast switching.
For AI servers, this matters a lot. Every electron going to an AI chip is converted 4–5 times: transformer step-down → rectifier → SiC/GaN power conversion → board-level DC-DC → on-chip voltage regulator. NVIDIA's Rubin generation is moving the datacenter to an 800-volt high-voltage DC architecture — which requires SiC and GaN at every conversion stage. Navitas was named NVIDIA's lead GaN partner. Monolithic Power Systems owns an estimated 60–70% of the PMIC (power management IC) content on every Blackwell GPU.
Beneath all of that sits the grid : transformers and switchgear made by Eaton, Hubbell, and nVent — sold out 24+ months — needed to connect AI datacenters to the rest of the electrical world. You cannot have AI without power conversion at every layer.