AI Compute’s Power Wall: Silicon Carbide, Gallium Nitride and HVDC architecture

AI Compute's Power Wall. Silicon Carbide and Gallium Nitride: $25,000 to $35,000 of new chip demand per megawatt

Rapid growth in AI training and inference workloads is driving server rack power from kilowatts toward megawatt levels, creating a “power wall.” At the same time, 800V EV adoption, smart-grid expansion, stricter energy-efficiency standards, and demand for compact high-performance systems are accelerating the shift toward High Voltage DC (HVDC) architectures.

Silicon Carbide (SiC) and Gallium Nitride (GaN) are the core technologies enabling the shift from traditional AC multi-stage power delivery to HVDC architectures in AI data centers. This "grid-to-rack" approach reduces conversion losses, lowers heat generation, cuts copper usage, and enables more compact, scalable designs. SiC handles high-voltage, high-power conversion with strong thermal performance, while GaN delivers ultra-high efficiency and power density at the rack level.

Future AI infrastructure deployments are projected to generate total GaN and SiC semiconductor of roughly $25,000-$35,000 per megawatt. GaN could contribute around $10,000-$15,000 per megawatt, while SiC could account for another $15,000-$20,000.

Key companies expanding in this area include Infineon Technologies, STMicroelectronics, onsemi, ROHM Semiconductor, Texas Instruments, TSMC, and auto companies such as Tesla, BYD, Hyundai, and Toyota.

Ionic View

AI data centers are quietly becoming one of the most compelling semiconductor demand drivers we've seen in years, as the industry's shift to high-voltage DC architectures creates a structural content uplift of $25,000–$35,000 per megawatt for SiC and GaN chips, a number that scales directly with every megawatt of new AI infrastructure deployed globally. This is increasingly more than a thematic story, it's showing up in design wins and order books across the broader power semiconductor industry, one of the ways investors may look to gain exposure to the AI infrastructure buildout beyond the usual hyperscaler and GPU names.