Anyone who specifies power devices reads a datasheet in a fixed order: blocking voltage first, then the curves that say what the part costs every time it switches. The EPFL work is strong on the first reading and openly incomplete on the second, and the second is where a design gets decided.
The diodes did the heavy lifting. Schottky barrier diodes grown on a relatively inexpensive silicon substrate [5] broke down above 3.9 kV with specific on-resistance as low as 4.7 milliohm cm2 [6]. They stayed above 3.3 kV at 125 C [10], and their dynamic on-resistance rose by less than 15 percent under repeated operation up to 3 kV [11].
The transistors are the parts that would actually sit in a converter, and they were characterised less far. They sustained 3.5 kV normally-on and 3.4 kV normally-off [8], but dynamic on-resistance was measured only up to 650 V, because the team's setup introduced parasitic effects at higher switching voltages [12]. That covers roughly 19 percent of the normally-off blocking figure [15].
That 650 V has a second meaning. Interesting Engineering's account describes 3.4 kV as more than five times the rating of many commercial GaN power devices [9], which puts the implied baseline at 680 V or below [16], roughly the class GaN already sells into. So the switching behaviour of this transistor has been shown at the top of the voltage range GaN already serves, and the blocking behaviour five times above it. The two kilovolt figures describe different properties measured at very different voltages: blocking voltage and switching-loss voltage, and they should not be read as a single rating.
The mechanism earns its attention. GaN's spontaneous and piezoelectric polarization already yields a sheet of mobile electrons, and the team added a layer that produces a matching sheet of mobile holes, tuning the cap thickness until the two nearly cancelled: about 1.03 x 10^13 electrons per square centimetre against 1.05 x 10^13 holes [2][3], a mismatch near 1.9 percent [18]. A control device with mismatched sheets broke down below 1 kV where the balanced one reached 3.9 kV [7], a factor of at least 3.9 attributable to the match itself [17]. Because the balance comes from layer thickness rather than implanted dopants [4], it sidesteps what EPFL PhD student Yuan Zong calls the high temperature sensitivity of doping-based charge balancing in GaN [4].
For the person specifying parts, this stays a lab result rather than a catalog item. The account reports device-level measurements and gives no wafer size, yield, qualification status or availability [19]. What is usable is a two-column table per candidate device. Column one: the voltage at which off-state blocking was demonstrated, and the temperature it was demonstrated at. Column two: the voltage at which repeated switching loss was demonstrated, and its temperature. On this evidence, the diodes fill both columns out to about 3 kV; the transistor fills the first to 3.4 kV and the second to 650 V. Where column two is a fraction of column one, the loss line in the thermal model belongs to whoever builds the model, not to the device, and column two is the thing worth asking a supplier to date.
