15 kV on the Label, 4,516 Volts Short of a 13.8 kV Peak

4,516 volts.

That is the gap between a chip rated 15,000 volts and the peak on a feeder rated 13,800, which the October 21 announcement never computes. University at Albany said a team led by Woongje Sung had demonstrated a 15-kilovolt silicon-carbide insulated-gate bipolar transistor, funded by the Army's DEVCOM research laboratory, and called it the second such result in the country and the first inside a university (UAlbany, October 21, 2026; Newswise). A 13.8 kV line is ordinary North American medium voltage, listed by Schneider Electric beside 4.16 kV and 34.5 kV and named, in the same company's distribution note, as the native voltage of a 60-hertz generator plant at a large data center (Schneider Electric White Paper 61; Schneider, July 24, 2020). Peak line-to-line voltage is √2 times the RMS rating: 1.4142 × 13,800 = 19,516 volts, so 19,516 minus 15,000 leaves 4,516, about 23 percent of the peak, with nowhere to sit if one chip is the whole plan.

One chip does not cover this feeder.

Nameplate, average, and the voltage that actually sits there

An unsmoothed six-pulse rectifier, feeding an inductive load with no large capacitor, averages (3 × √2 / π) times RMS line voltage, and 1.3505 × 13,800 is 18,637 volts (the six-pulse average, written as 1.350 × V_LL). Add the capacitor an inverter or solid-state transformer actually uses and the bus climbs toward the peak, less two diode drops that barely register near twenty kilovolts (capacitive case: √2 × V_LL). Even that kinder 18,637 leaves the new chip 3,637 volts short, and a 10 percent load sag, inside the range the rectifier note describes, only pulls it to about 16,773, still above 15,000. Covering the average with one 15 kV nameplate would take a 19.5 percent sag. Schneider also says measured voltage can run 10 percent hot, which lifts this peak from 19,516 volts to 21,468.

Infineon's FZ750R65KE3, a 750-amp module at the top of the silicon catalog, is rated 6,500 volts blocking and 3,800 volts of DC stability at 25°C under a 100 FIT condition (FZ750R65KE3 datasheet, revision 1.10; Infineon's 6.5 kV family). Three of those modules sum to 19,500 volts of nameplate against a 19,516-volt peak, a miss of 16 volts, so the nameplate count is four, while five times the 3,800-volt continuous figure is 19,000 and still 516 volts short, which makes the continuous count six.

Borrow the ratio and label it a test, not a measurement of this die: 3,800 / 6,500 equals 0.5846, which rounds to 58.5 percent, and 0.5846 × 15,000 is 8,769 volts. Two chips at that imagined rating sum to 17,538 and miss the peak by 1,978 volts, so the derated count is three, whereas a true continuous 15 kV rating needs only two, because 30,000 clears both the nominal peak and the 21,468-volt high line. UAlbany says stacks of "three or more" silicon parts could give way to a single 15 kV carbide device. Three is what you get by dividing the unsmoothed average, 18,637, by 6,500 and rounding up. It is neither the peak nor Infineon's continuous rating, and on this feeder one chip replaces the stack only as a sentence.

Devices in series to hold the peak line-to-line voltage (√2 times nominal RMS) at three North American medium voltages from Schneider's table. Counts are ours. The derated carbide column applies Infineon's 3,800/6,500 ratio and is not a published rating of the Albany part.
Feeder	Peak	Silicon, 6.5 kV nameplate	Silicon, 3,800 V DC stability	15 kV SiC, nameplate	15 kV SiC, 58.5% derate
4.16 kV	5,883 V	1	2	1	1
13.8 kV	19,516 V	4	6	2	3
34.5 kV	48,790 V	8	13	4	6

At 4.16 kV the new label buys nothing one 6.5 kV module does not already block on nameplate, while at 34.5 kV the cut is real and still a stack: 13 continuously rated silicon modules become four carbide chips if the label holds, or six if the derating sticks. Down in the hall, a Berkeley Lab note puts server parts at 1.35 to 12 volts and the internal bus an AC supply already builds near 380, and a 480-volt service peaks at only 679 (Lawrence Berkeley National Laboratory). NoMIS, the spinout packaging this device, says its carbide span runs from 400 volts to 20 kV (NoMIS services). Racks do not care.

Why build an IGBT when the MOSFET already blocked 15 kV

October's part is an IGBT, a bipolar transistor with a MOSFET gate, and this same group already blocked 15 kV with a MOSFET in 2021. Sung's 2024 workshop slides, citing that paper, put the split-gate device at 281 milliohm-square-centimeters of specific on-resistance at a 15 kV block, against 273 for the conventional layout on the same wafer (Sung, Sandia workshop, 2024; Lynch, Yun, Sung et al., WiPDA 2021). Cree, now Wolfspeed, published a 4H-SiC n-channel IGBT in 2013 with a differential specific on-resistance of 25.6 in those units while blocking 17 kV, on 0.32 square centimeters of active area (Ryu et al., 2013). Divide 281 by 25.6 and the quotient is 10.98, so at equal current and equal area the MOSFET burns about eleven times more in the resistive term alone.

Eleven is a ceiling, not this month's wattage, because differential resistance ignores the knee and Albany published no on-state voltage, current or die size. The release says the IGBT shares about 90 percent of that MOSFET process and is supposed to carry higher current, without saying higher than what, and the voltage itself dates to Cree: a 2012 p-channel IGBT blocked 15 kV at 24 mΩ·cm² while the n-channel sibling stopped at 12.5 kV, a 2013 test switched a 15 kV, 20-amp n-IGBT up to 11 kV, and a 2015 comparison put the IGBT ahead only below about 5 kHz as MOSFET specific resistance rose from 204 to 570 between 25°C and 150°C (Ryu et al., 2012; Kadavelugu et al., ECCE 2013; Ryu et al., 2015). Grid gear sits under that line, so a faster data-center supply can still prefer the MOSFET, and UAlbany's claim to be second in the country and first in a university never names the first institution. Sung had already put the crossover near 7 kHz in a 2009 design study (Sung, Wang, Huang and Baliga, ISPSD 2009).

Who holds the wrench is the new fact. UAlbany says the chain stayed domestic, its lab on design, GE Aerospace on most of the fabrication, a fab the release identifies only as DoW on the last steps, and NoMIS on the package, which Sung calls a "strategic domestic advantage." Sixty days earlier NoMIS had announced a 6.5 kV MOSFET above 8 kV blocking, 90 milliohms and 55 amps, with production aimed at the fourth quarter of 2026 and with 10 kV and 20 kV IGBTs still "in development" (NoMIS, August 22, 2026). That page did not mention a 15 kV IGBT. The university now says NoMIS is commercializing the class, and that China and Japan already make it, but I found no public datasheet for a 15 kV part anyone can order.

What You Can Do

If a medium-voltage drawing is about to change because of the October 21 release, wait for three numbers it does not contain: continuous voltage at a stated reliability condition, current at a stated temperature, and switching energy at the frequency you actually switch. Size from the table until those arrive, skip the new part on a 4.16 kV feeder, and treat 34.5 kV as a shorter stack rather than a single switch. Buyers should chase NoMIS's 6.5 kV MOSFET, the device with samples and a production quarter, and demand curve-tracer plots before the IGBT reaches a roadmap. People who run servers can ignore the yard: a 480-volt service peaks under 700 volts, and the carbide that belongs on a rack starts near 1.2 kV, already inside this company's published range.

Limitations

Nothing here proves the Albany transistor blocks 15 kV beyond the university's account: no current, on-state voltage, switching loss, yield or price came with the news. The 3,800-volt anchor is one Infineon silicon module, so 58.5 percent is a sensitivity test, and this die's cosmic-ray curve could be kinder or worse. Taps, twelve-pulse rectifiers and multilevel cells move the bus; I used a plain six-pulse peak and average. Schneider's plus or minus 10 percent is a tolerance, not one campus, and I did not inventory factories in China or Japan. The factor 10.98 pairs a 2021 MOSFET with a 2013 IGBT from another firm and skips knee voltage, so it is not this press release's loss.

The Strongest Case Against This Analysis

For the jobs the release names first, a continuous-voltage ruler is the wrong tool. Microwave pulses, fusion switches and rock-cracking discharges last microseconds to milliseconds, not the years a traction module sits biased, and Infineon's 3,800 volts is a 100 FIT reliability budget for that long sit, not the voltage where silicon breaks. On nameplate alone, two 6.5 kV modules sum to 13,000 and miss 15,000 by 2,000, so three modules are the smallest stack that matches the label, which is the release's "three," and one 15 kV chip really does replace them, resistors and matched gate drives included. Cree showed the class in 2012 and never catalogued it, yet a chain through GE Aerospace into a spinout already selling 3.3 kV parts is a factory fact a conference paper was not, and cutting six silicon modules to two carbide chips is a real win if the rating holds. Both readings fit one missing datasheet: one chip for a pulse, two for this feeder.

The Bottom Line

A domestic 15 kV silicon-carbide IGBT, from a group on this problem since 2009, is news about who can build the part. On a 13.8 kV feeder the peak is 19,516 volts: one chip rated 15,000 falls 4,516 short, two cover it if the label is continuous, and three cover it under Infineon's derating. The stack shrinks, and only a curve, not a sentence, makes it disappear.