Who makes GaN and SiC for AI data centers

Gallium nitride and silicon carbide switch faster and lose less energy per switching cycle than the silicon MOSFETs they replace. On a board delivering hundreds of amps to an accelerator, that buys two things: higher switching frequency, which shrinks the inductors and capacitors around the converter, and lower loss, which means less heat to remove from a place where there is no room to remove it. The two materials do different jobs. GaN is generally the choice at lower voltages and very high frequency, which suits the point-of-load and intermediate-bus stages sitting close to the processor; SiC holds up better at higher voltage and is more at home further upstream. On an accelerator board this part is mostly the GaN case, and the SiC case belongs to the power shelf and the electrical room. AI boards are the demand story here because the current a single accelerator draws has climbed faster than the board area available to deliver it, which is what forces the conversion stage closer to the die and higher in frequency. What to look for is whether a supplier makes its own devices or buys wafers, since substrate supply has been the binding constraint in wide-bandgap before, and whether it sells discrete devices or integrated power stages — the second is a design-in that lasts a platform generation.