Power Semiconductor Industry Targets Data Center Efficiency - THE ELEC, Korea Electronics Industry Media

The power curve of AI data centers is breaking legacy power architectures. Once single-GPU TDP crosses 1000W, 800V HVDC shifts from optional to existential—rewriting value allocation across the power semiconductor stack. SiC MOSFETs will capture 40%+ of AC-DC primary conversion within 18 months, while GaN locks down auxiliary and DC-DC stages, forming a rigid high-voltage SiC plus low-voltage GaN pairing. The real bottleneck sits upstream: 6-inch SiC substrates. Over 60% of global capacity concentrates in a handful of players, with critical epitaxy nodes held by firms in Japan and Taiwan, China. A geopolitical shock here would hit harder than chip-level export controls—a subtler chokepoint most supply-chain risk models still underweight. On competition, Infineon and onsemi are bundling die, module, and thermal solutions to lock in hyperscalers, while Navitas and Power Integrations weaponize the 1%-efficiency-equals-measurable-OPEX-savings narrative. Korean players' window: anchor to domestic AI infrastructure buildout and sell compliance-driven efficiency, sidestepping head-on price wars with European and American incumbents. The 12-to-24-month tail: procurement logic pivots from buying compute to buying cost-per-watt. Power semiconductors migrate from a BOM line item to the number-one variable in data-center TCO. Whoever defines the 800V device spec owns the pricing power of the next generation of data centers.