A 1 MW rack on today’s dominant 54 V in-rack distribution bus implies a bus current of about 1,000,000 / 54 = 18,519 A. That is not a problem you solve by making the copper busbar a little thicker. DC busbars, connectors and breakers rated for nearly 18.5 kA, together with the cooling and protection hardware they demand, consume the rack. NVIDIA puts another number on it in its 800 VDC architecture description: staying with 54 V distribution means roughly 200 kg of copper busbar in a single 1 MW rack, and about 200,000 kg of rack busbars in a 1 GW data center. On GB200 / GB300 NVL72 systems, eight power shelves already take a large share of the rack; at megawatt scale the same approach would grow to 64 U of power shelves on Kyber, leaving no room for compute. So 1 MW is not a matter of building a bigger power supply. The distribution voltage has to change first.
This article examines the current, copper loss and semiconductor stress boundaries that appear when a 1 MW class AI rack moves to an 800 V DC bus, and it carries the device selection arithmetic through to the values published in the official HXY Electronics datasheet. The HXYS165N120MPI is an N-channel 1200 V / 165 A (catalog rating) / 13 mOhm SiC MOSFET in a TO-247H-4L Kelvin-source package. At VGS = 18 V the datasheet specifies ID = 167 A at TC = 25 C and 118 A at TC = 100 C, RDS(on) rising from 13 mOhm at 25 C to 20 mOhm at 175 C, a total gate charge Qg of 216 nC, and a junction-to-case thermal resistance Rth(J-C) of 0.27 typically and 0.37 maximum C/W. Its 1200 V rating covers the roughly 1100 V peak produced when steady-state overvoltage on an 800 V bus is combined with turn-off overshoot, and its 13 mOhm on-resistance cuts conduction loss by about 48 percent versus a 25 mOhm device at the same current. Companion options include the SiC MOSFET HXYS125N120Q (1200 V / 125 A / 15 mOhm / SOT-227) and the SiC diode HC4D40120H (1200 V / 40 A / TO-247-2L) for the same 800 V bus power stage. One clarification: the applications listed in the datasheet are solar inverters, switch-mode supplies, high-voltage DC-DC converters and motor drives. An AI rack HVDC sidecar is an extension of the same high-voltage DC-DC topology family and requires device-level validation under the actual operating conditions.
The first wall is the square of the current. Resistive loss in a busbar follows P = I squared times R. If power doubles at a fixed voltage, current doubles and loss quadruples. An engineering example presented at OCP 2026 is representative: a 240 kW rack at 54 V draws about 4,444 A, while a 1 MW rack at the same voltage draws 18,519 A. Going from 40 kW to 1 MW scales current by roughly 25 times and resistive loss by roughly 625 times, with megawatt-class busbar loss estimated above 50 kW. That is no longer a few points of efficiency. Those 50 kW become heat that the liquid cooling system has to remove from inside the rack, equivalent to adding the thermal load of another full rack.
The second wall is space and weight. Conductor cross-section, connector current rating and DC interrupting capability all scale with 18.5 kA, and the weight of the copper, the structural load and the bend radius grow with them. The power shelves themselves occupy U space: under a 54 V architecture a megawatt rack needs around 64 U of power shelves, leaving little compute space. Moving the supplies into a dedicated rack solves the height problem but halves rack-position density, because each compute rack now needs its own power rack.
The third wall is the number of conversion stages. In a conventional chain, power arrives at 13.8 kV medium voltage and passes through transformers, switchgear, UPS, PDUs, rack PSUs, a 48 V or 54 V tray bus, an intermediate bus converter to 12 V, and multiphase VRMs down to the sub-1 V GPU core. Every stage adds loss and a failure point, plus fans, maintenance and floor space. Below 200 kW those costs are thin enough to absorb. At 1 MW they are amplified by the same physical constraints all at once.
One more item is regularly left out of the calculation. GPU loads step in milliseconds, and the di/dt between idle and full load is severe. The L-sigma times di/dt spike generated by busbar stray inductance therefore scales with the absolute current change. A low-voltage, high-current design has to carry both the steady-state copper loss and a larger transient current swing, so it is squeezed from both ends.
Raising the voltage releases all three constraints at once. Since I = P / V, 1 MW at 800 V is about 1,250 A, roughly 15 times lower than the 18,519 A at 54 V; for the same conductor cross-section, resistive loss falls by a factor of about 220. NVIDIA reports that moving from 415 VAC distribution to an 800 VDC busway lets the same conductor carry about 85 percent more power, and that reducing AC/DC conversion stages can improve end-to-end efficiency by up to 5 percent, cut maintenance cost by up to 70 percent and lower total cost of ownership by up to 30 percent. Those figures come from the architecture sponsor’s public material and should be rechecked against each project’s own distribution layout and load profile, but the direction is unambiguous: lift the voltage first, and 1 MW becomes implementable.
This is already past the concept stage. NVIDIA has stated that 800 VDC is required beginning with the Rubin Ultra and Kyber rack architectures in 2027, with Kyber racks spanning 600 kW to 1 MW. On the standards side, OCP contributors are converging on 800 VDC and plus/minus 400 V as an open specification rather than a single-vendor design; version 0.3 of the LVDC solid-state transformer specification was published in July 2026 with more than 80 equipment makers participating. Deployment is phased. A rack-adjacent power rack or sidecar first rectifies existing 415 VAC or 480 VAC feeds locally into 800 VDC without changing upstream AC distribution. Conversion then migrates to row-level Power Centers and data-hall-scale 4.8 MW DC power blocks, with medium-voltage solid-state transformers as the long-term form. TI’s 30 kW 800 V AC/DC PSU and 800 V capacitor bank unit shown at GTC 2026 are concrete modules in that chain. It is worth noting that 800 V is not delivered directly to the processor. Inside the rack it is stepped down by isolated DC-DC to 48 V, 54 V or even 6 V, then through IBCs and multiphase VRMs to the GPU. Raising the bus voltage redesigns the distribution segment; it does not discard everything downstream.
So why stop at 800 V instead of going higher? There are three hard boundaries.
The first is the semiconductor voltage class. An 800 V nominal bus does not mean the device only blocks 800 V. Steady-state overvoltage caused by sensing and protection delay can reach about 900 V, and stray inductance adds roughly 200 V of turn-off overshoot, for a combined peak near 1100 V. That sits inside the rating of mature 1200 V SiC devices with roughly 9 percent static margin, a practical balance of device cost and reliability. Push the nominal to 900 V and the corresponding steady-state overvoltage reaches about 1030 V, plus 200 V of overshoot, or about 1230 V: past the 1200 V rating. A move to 1700 V parts becomes necessary, and published estimates put 1700 V SiC at 1.8 to 2.5 times the price of 1200 V, with gate drive, insulation and creepage distances redesigned as well.
The second is the regulatory line. DC voltage above 1000 V falls under high-voltage distribution rules, which changes operator qualifications, insulation classes, grounding and fire protection, and installation practice. At 800 V, even a transient rise to 900 V stays below that 1000 V threshold. The third is the supply chain. A 700 V bus carries roughly 14 percent more current than 800 V at the same power, which hurts copper loss and makes the downstream conversion ratio less favorable. More importantly, there is no standardized 700 V ecosystem: rectifier modules, PDUs and server power supplies would all be custom. 800 V sits at the intersection of device availability, compliance boundaries and a standards-based supply chain.
The tradeoff is equally clear. Once the voltage goes up, the stress map for power semiconductors is redrawn. Where the 48 V / 54 V side cared about the on-resistance and current sharing of 100 V devices, the front-end rectifier, the primary side of isolated DC-DC, DC protection and snubber positions must now be selected against high-voltage stress from scratch.
One point is frequently written incorrectly: 1200 V is not the only answer for every switch position on an 800 V bus. Three-level topologies such as three-phase Vienna, ANPC and flying-capacitor FC-3L split the 800 V bus stress in half, so a single device sees roughly 400 V of static stress and a derating ratio in the comfortable 33 to 42 percent band. A published 1 MW sidecar design uses a three-phase Vienna rectifier with 750 V main switches, and Infineon’s 11 kW bidirectional PFC reference design likewise notes that 3L-ANPC can implement an 800 V bus with 600 V to 750 V devices. In the other direction, some industry voices argue that two-level, full-bus hard-switching positions on an 800 V bus should use 1400 V or 1700 V devices, because 1200 V parts operating continuously at about 66 percent of rating can see degraded terrestrial-neutron single-event-burnout FIT rates once turn-off overshoot is added.
Put both views together and the conclusion becomes sharper. The right use of 1200 V SiC is in positions that must cover full-bus overvoltage and turn-off overshoot, in three-level positions where extra reliability margin is worth more than the lowest possible device voltage class, and in SST isolation stages, DC blocking and snubber circuits where stress cannot conveniently be split by topology. It goes together with topology-based voltage division, low-inductance busbars and clamping to keep the spike under control. It is not a part to place everywhere, and it is not a part to be replaced by 1700 V at the first sign of concern. For a 1 MW sidecar that runs continuously and still has to hit a power-density target, this voltage class offers a calculable landing point between cost, maturity and reliability.
The HXYS165N120MPI from HXY Electronics is a representative device in that class. Stopping at the three catalog numbers of 1200 V / 165 A / 13 mOhm leaves both the thermal and the gate-drive budgets unresolved, so the datasheet values come first and the design consequences follow.
| Parameter | Datasheet value | Test condition | Relevance to an 800 V sidecar |
|---|---|---|---|
| Drain-source breakdown V(BR)DSS | 1200 V (Min) | VGS = 0 V, ID = 100 uA | Covers the roughly 1100 V peak with about 9 percent static margin |
| Gate-source voltage VGS | -5/+18 V recommended; -10/+22 V absolute maximum | — | Sets the window for positive drive and negative turn-off bias |
| Threshold VGS(th) | 2.9 V at 25 C / 2.1 V at 175 C (Typ) | VDS = VGS, ID = 20 mA | Threshold falls 0.8 V when hot, so parasitic turn-on risk grows with junction temperature |
| Continuous drain current ID | 167 A at TC = 25 C / 118 A at TC = 100 C | VGS = 18 V | At 100 C case temperature the margin must be recalculated from 118 A |
| On-resistance RDS(on) | 13 mOhm at 25 C / 20 mOhm at 175 C (Typ) | VGS = 18 V, ID = 75 A | A 1.54 times temperature coefficient; use 20 mOhm for thermal design |
| Input / output / reverse capacitance | Ciss 3815 pF / Coss 239 pF / Crss 29 pF (Typ) | VGS = 0 V, VDS = 1200 V, f = 100 kHz | Only 29 pF of Crss keeps Miller coupling contained, but clamping is still required |
| Gate charge Qgs / Qgd / Qg | 46 / 100 / 216 nC (Typ) | VDS = 800 V, VGS = -5/18 V, ID = 75 A | Directly sets drive power and dv/dt |
| Body diode forward voltage VSD | 4.7 V at 25 C / 4.1 V at 175 C (Typ) | VGS = -5 V, ISD = 37.5 A | High dead-time freewheel loss; a parallel SiC Schottky diode is advisable |
| Junction-to-case thermal resistance Rth(J-C) | 0.27 (Typ) / 0.37 (Max) C/W | — | Junction-to-case rise can be calculated directly |
| Power dissipation and junction temperature | PD 555 W at TC = 25 C; Tvj -55 to 175 C | TJ = 175 C | Upper boundary for the cold-plate thermal budget |
Several thermal and capacitive figures for the companion devices are worth recording from the same datasheet library:
| Part number | Additional datasheet values |
|---|---|
| HXYS125N120Q (SOT-227) | ID 125 A at 25 C / 90 A at 100 C, 250 A pulsed, Ptot 681 W, Rth(J-C) 0.22 C/W maximum, VGS -4/+18 V recommended |
| HXYS87N120T6 (TO-263-7L) | ID 87.5 A at 25 C / 64.2 A at 100 C, 140 A pulsed, Ptot 536 W at 25 C / 232 W at 110 C, Rth(J-C) 0.28 C/W maximum |
| HC4D40120H (TO-247-2L) | VF 1.5 V (Typ) / 1.8 V (Max) at 40 A and 25 C, 2.2 V / 3.0 V at 175 C; QC 167 nC at 800 V, EC 36 uJ at 800 V, RthJC 0.225 C/W; a majority-carrier device with no reverse-recovery charge |
| HC5D50120H (TO-247-2L) | VF 1.43 V (Typ) / 1.7 V (Max) at 50 A and 25 C, 2.0 V / 2.5 V at 175 C; QC 268 nC at 800 V, EC 138 uJ at 800 V, RthJC 0.23 C/W |
Its parameters convert into design benefits as follows:
• 1200 V blocking addresses the roughly 1100 V worst-case peak. The datasheet specifies V(BR)DSS at 1200 V minimum with VGS = 0 V and ID = 100 uA, so 900 V of steady-state overvoltage plus 200 V of turn-off overshoot remains inside the rating. In a three-level half-bus position at 400 V, derating falls to about 33 percent and long-term stress is more relaxed. The gate side has its own boundary: VGS is -5/+18 V recommended and -10/+22 V absolute maximum, so drive overshoot must stay inside that window. In a two-level full-bus hard-switching position at 800 V, derating is about 66 percent, so the decision should be revisited against measured overshoot, MOV clamping thresholds and reliability derating requirements, with a higher voltage class considered where needed.
• 13 mOhm on-resistance has to be read together with 20 mOhm at 175 C. For a 30 kW module on an 800 V bus, module current is about 37.5 A, and single-device conduction loss is 37.5 squared times 0.013, or roughly 18.3 W. At the 25 mOhm class available in the same catalog, the same current gives about 35.2 W: roughly 17 W more per device, a reduction of about 48 percent. The datasheet also specifies 20 mOhm at 175 C, which raises that same 37.5 A loss to 28.1 W, about 54 percent above the 25 C figure; the 25 mOhm reference device is specified at 41 mOhm at 150 C, or about 57.7 W at the same current. Across the thirty-plus modules and several dozen main switches of a 1 MW system, the kilowatt-scale difference therefore has to be converted into cold-plate load using hot values rather than room-temperature typicals. These are static estimates from datasheet typical values; RDS(on) is specified at ID = 75 A, so the real 37.5 A case will be slightly lower, and switching loss is not included.
• 167 A is the 25 C number; 118 A is the rating that matters in a hot cabinet. The catalog lists 165 A, while the datasheet gives 167 A at TC = 25 C and 118 A at TC = 100 C for VGS = 18 V, with PD limited to 555 W. Under the common rule of keeping continuous current below half of ID, 37.5 A against the 25 C half value of 83.5 A leaves about 2.2 times margin; at 100 C case temperature the half value is only 59 A and the margin narrows to about 1.6 times. More important, a 125 kW unit on an 800 V bus draws about 156 A, which exceeds the 118 A hot rating and cannot be carried by a single device: it needs paralleling or a lower unit power. Sizing current margin from the room-temperature rating is an easy trap in high-voltage sidecar design. N+1 paralleling, idle-to-full-load GPU transients and current-sharing imbalance all have to sit above that derating curve.
• The TO-247H-4L Kelvin source and Qg = 216 nC jointly set the switching-frequency ceiling. Power-loop di/dt no longer couples into the gate drive loop through common-source inductance, so gate ringing and parasitic turn-on risk fall and dv/dt is easier to hold at the designed value. The datasheet values Qg = 216 nC, Qgd = 100 nC and Crss = 29 pF against Ciss = 3815 pF quantify the benefit: Miller charge accounts for about 46 percent of total gate charge, so the gate plateau is not short during high dv/dt commutation and negative turn-off bias plus an active Miller clamp are not optional. The roughly 1:132 ratio of Crss to Ciss, combined with the Kelvin source, is what keeps gate ringing manageable on hard-switching legs above 40 kHz. One more item is easy to miss: the datasheet gives VSD = 4.7 V at 25 C with ISD = 37.5 A and VGS = -5 V. That body-diode drop is why a parallel SiC Schottky diode, rather than the MOSFET body diode, should carry the dead-time freewheel current.
Four calculations that can be reproduced. Take a common example of a 30 kW / 800 V module at 37.5 A and fsw = 40 kHz; every input appears in the two tables above.
• Conduction: 37.5 squared times 0.013 = 18.3 W at 25 C; 37.5 squared times 0.020 = 28.1 W at 175 C.
• Gate drive: the energy per charge cycle is approximately Qg times the gate swing, or 216 nC times (18 V minus -5 V) = 4.97 uJ. That is about 0.20 W at 40 kHz, 0.32 W at 65 kHz and 0.50 W at 100 kHz. Small for one device, but dozens of channels in a 1 MW sidecar add up in the auxiliary supply and driver-board thermal budget, and driver internal losses are excluded.
• Dead-time freewheel: assuming 100 ns of dead time and two commutations per cycle, VSD = 4.7 V gives about 4.7 V times 37.5 A times 2 times 100 ns times 40 kHz = 1.41 W. With a parallel HC4D40120H at VF = 1.5 V the figure is about 0.45 W, saving roughly 0.96 W per device. Its QC of 167 nC and EC of 36 uJ, both at 800 V, are capacitive charge and stored energy rather than reverse-recovery tail current, so no such spike is transferred to the MOSFET turning on.
• Junction temperature: with Rth(J-C) of 0.27 typical and 0.37 maximum C/W, 18.3 W produces a junction-to-case rise of 4.9 / 6.8 C and 28.1 W produces 7.6 / 10.4 C. That is only the junction-to-case segment. Case-to-cold-plate and cold-plate-to-liquid resistances complete the budget, and cold-plate capability should be sized against the 0.37 C/W maximum with operating margin.
A boundary has to be stated plainly. The datasheet still marks Eoss, EON / EOFF, td(on) / tr / td(off) / tf, internal gate resistance RG(int), body-diode IS / Qrr / trr / Irrm and pulsed current ID(pulse) as TBD. Eoss and switching energy are provided only as curves, the latter measured at VDD = 800 V, ID = 75 A, RG(ext) = 2.5 ohm, VGS = -5/+18 V and L = 150 uH, with no numeric table. Those are precisely the inputs needed for hard-switching loss above 40 kHz, dead-time tuning and desaturation blanking time. This article does not infer them; obtain the full characteristics from an FAE or close the gap on a double-pulse bench.
Mapping the available 1200 V devices by position makes selection more concrete:
| Circuit position | Part number | Key parameters (product catalog) | Selection rationale |
|---|---|---|---|
| 800 V bus rectifier leg, isolated DC-DC primary high-frequency position | HXYS165N120MPI | 1200 V / 165 A / 13 mOhm / TO-247H-4L | Low on-resistance with a Kelvin source for high-frequency switching |
| Paralleled modular power units, cold-plate mounting | HXYS125N120Q | 1200 V / 125 A / 15 mOhm / SOT-227 | Module-style package simplifies paralleling and isolated thermal design |
| Medium-power 800 V to 48 V isolated conversion | HXYS80N120MPI | 1200 V / 80 A / 25 mOhm / TO-247-4L | Balanced current class and loss, fewer gate drive channels |
| High-density board-level power stage | HXYS87N120T6 | 1200 V / 87.5 A / 39 mOhm / TO-263-7L | Surface-mount package for dense layouts |
| Boost, freewheel and DC blocking diode positions | HC4D40120H | 1200 V / 40 A / VF 1.5 V / TO-247-2L | SiC zero reverse recovery for hard-switched freewheeling above 40 kHz |
| High-current diode position | HC5D50120H | 1200 V / 50 A / VF 1.43 V / TO-247-2L | Lower forward voltage reduces conduction loss |
The parameters in this position table come from the product catalog and serve as a quick map from circuit function to part number; datasheet values for the devices that have them are listed in the two tables above. Zero reverse recovery is shared across the SiC diode product line, and the datasheet explicitly notes that a majority-carrier device has no reverse-recovery charge. Its value in a hard-switched freewheeling position is quantified by the dead-time calculation above: there is no reverse-recovery current spike of the kind a silicon fast-recovery diode produces, so that loss is not transferred to the MOSFET turning on.
This section walks the circuit in words and can serve as a topology review draft. Structurally it extends the totem-pole PFC concept into a three-phase HVDC sidecar: a front-end three-phase rectifier / PFC establishes the 800 V bus, or plus/minus 400 V with a midpoint, and an isolated DC-DC stage steps 800 V down to 48 V or 54 V before the tray-level IBCs and VRMs take over.
The main power loop runs: 415 VAC / 480 VAC three-phase input, input filter and breaker, boost inductors, three-level or two-level bridge legs, bus film capacitors with clamping and bleed branches, 800 VDC output through a DC breaker and connector into the IT rack, the primary half-bridge of an isolated LLC or DAB converter, a high-frequency transformer, secondary synchronous rectification, and finally the 48 V tray bus. In a sidecar configuration, rectification and protection move out of the compute rack into a separate adjacent power cabinet, and only the DC-DC stage remains in the compute rack. That move is what gives the space back to the GPUs.
The control signal chain splits into four linked stages:
• Control stage. A DSP or dedicated digital controller runs an outer voltage loop regulating the 800 V bus and an inner current loop for sinusoidal input current and power factor correction. Three-level topologies add a midpoint-balance loop or a flying-capacitor voltage-balancing loop, and paralleled units add interleaved carrier phase shifting to cancel bus ripple current. Switching frequency is typically in the 40 kHz to 100 kHz range: a published 1 MW three-phase Vienna example uses 40 kHz with about 90 microhenry per phase, while a three-level ANPC reference design uses 65 kHz. Insulation monitoring, arc-fault detection, overvoltage protection and bus-pump suppression during load dump all belong in the same protection logic.
• Gate drive stage. Isolated gate drivers amplify the PWM to the voltage SiC requires. The HXYS165N120MPI datasheet gives -5/+18 V recommended and -10/+22 V absolute maximum VGS, so -5 V turn-off bias and +18 V turn-on have a documented basis, with drive overshoot held inside -10 V / +22 V. The more consequential detail is the falling threshold: VGS(th) drops from 2.9 V at 25 C to 2.1 V at 175 C, so parasitic turn-on margin shrinks in hot, high dv/dt operation, while Qgd = 100 nC accounts for nearly half of Qg = 216 nC. An active Miller clamp and a separate Rg_off are therefore not optional. CMTI has to cover the tens of kV per microsecond of common-mode dv/dt produced by midpoint transitions in a three-level leg. Propagation delays between upper and lower devices must match, and while SiC dead times can reach the hundreds of nanoseconds, three-level commutation paths are numerous, so dead time has to be designed together with the clamping logic. Desaturation protection is a sensible default, with blanking time aligned to the measured short-circuit withstand window.
• Power stage. The HXYS165N120MPI sits in the rectifier leg or on the primary half-bridge of the isolated DC-DC. Qg = 216 nC, Qgd = 100 nC, Ciss = 3815 pF and Crss = 29 pF set drive power and dv/dt; at 40 kHz the per-device gate charge and discharge power is about 0.20 W, and separate turn-on and turn-off gate resistors are recommended. The switching-energy curves in the datasheet use RG(ext) = 2.5 ohm at VDD = 800 V, ID = 75 A, VGS = -5/+18 V and L = 150 uH, which is a reasonable starting point for Rg before converging on measured double-pulse ringing rather than swallowing loss with a large resistor. The TO-247H-4L Kelvin source decouples the power and gate loops. Third-quadrant conduction needs its own decision: VSD = 4.7 V at 25 C with ISD = 37.5 A produces about 1.41 W of freewheel loss per device at 100 ns dead time and 40 kHz, so a SiC Schottky diode such as the HC4D40120H (VF 1.5 V typical, QC 167 nC at 800 V) should be paralleled for boost and freewheel duty, bringing that figure to about 0.45 W and removing reverse-recovery current. Coss is specified at 239 pF and 1200 V, but Eoss remains TBD, so the capacitive turn-on loss of a hard-switched leg still has to be measured.
• Load stage. Bus film capacitors absorb GPU current steps, and the 800 V to 48 V isolated stage tracks the load in closed loop. Load dump deserves particular attention: when GPUs drop from full load to idle, returned energy pumps the bus up, and that is the main source of the 900 V steady-state overvoltage condition. Overvoltage thresholds, bleed branches and clamping devices must be coordinated for this case rather than sized only for a nominal 800 V.
The position table above maps devices to those functions, and static conduction loss, gate-drive power and junction-to-case rise can now be calculated from published datasheet values. Three quantities still require bench work: peak turn-off overshoot, which sets the voltage class and the MOV clamping threshold; power-loop stray inductance, which sets the spike and EMI; and the EON / EOFF and Eoss values marked TBD, which set total switching loss above 40 kHz. Junction-to-case thermal resistance is no longer an unknown at 0.27 / 0.37 C/W; what remains unknown are the case-to-cold-plate and cold-plate-to-liquid segments, which belong to mechanical design and have to be measured on a prototype.
• Measure the worst-case stress before choosing a voltage class, not after. On paper, 900 V of overvoltage plus 200 V of overshoot is 1100 V, but the real spike is set by busbar stray inductance and di/dt, and a two-level full-bus position adds a long-term reliability derating question. Use a high-voltage differential probe to measure VDS peaks under steady full load, load dump and power-on inrush; compare 1.1 times the measured peak against 1200 V, and verify that the MOV clamping threshold stays below the device rating.
• Low-inductance busbar layout is the first layout constraint in an 800 V system. Tightly laminated positive and negative layers, a small power-loop area, film decoupling capacitors placed close to the bridge pins, and a Kelvin-source trace kept under 10 mm do more than a snubber added later. Run a double-pulse test to measure the spike before fixing Rg, rather than swallowing the loss with a large gate resistor.
• Do not omit negative gate bias, the Miller clamp or desaturation protection, and give the body diode an alternative path. VGS(th) falls from 2.9 V at 25 C to 2.1 V at 175 C and Qgd = 100 nC is nearly half of Qg, so parasitic turn-on margin is smaller when the device is hot. Set the drive to the recommended -5/+18 V without crossing the -10/+22 V absolute maximum, use separate Rg_on and Rg_off, add an active Miller clamp, specify an isolated driver above 100 kV per microsecond CMTI, and check desaturation blanking time against the measured short-circuit withstand window. Do not leave the body diode carrying freewheel current either: VSD = 4.7 V at 25 C adds roughly 1.4 W per device at 100 ns dead time and 40 kHz, so a parallel SiC Schottky diode is the safer choice.
• Give three-level voltage balancing and circulating current their own control budget. Flying-capacitor voltage must stay at half the bus, and startup, light load and load dump are the conditions where it tends to drift; plan for precharge, a midpoint-balance loop and a suitable modulation strategy. With interleaved paralleled units, circulating current needs common-mode inductors, isolated DC sources or a dedicated suppression loop. A single-unit waveform is not sufficient evidence.
• Calculate thermal design and DC safety together. Do not extrapolate from the room-temperature typical: the datasheet specifies RDS(on) = 20 mOhm at 175 C, 1.54 times the 25 C value, which raises 37.5 A single-device conduction loss from 18.3 W to 28.1 W. At the 0.37 C/W maximum Rth(J-C) that is about 10.4 C of junction-to-case rise, and case-to-cold-plate plus cold-plate-to-liquid resistances complete the budget; PD of 555 W is a TC = 25 C boundary, not usable dissipation in a cabinet. Power devices in a 1 MW sidecar should be cold-plate mounted. On the 800 VDC side, plan high-resistance grounding, continuous insulation monitoring, active arc-fault detection and connector interlocks that prevent disconnect under load. Recheck creepage and clearance against the applicable standard, and ensure that no transient overvoltage crosses the 1000 VDC regulatory boundary.
Moving from 54 V to 800 V is fundamentally a trade of voltage for current: roughly 15 times less current, about two orders of magnitude less busbar copper loss, and power shelves relocated out of the compute rack. That is what turns a 1 MW rack from a headline number into a deliverable cabinet. Once the voltage rises, 1200 V SiC takes on the job of covering overvoltage and turn-off overshoot while holding losses down at high switching frequency. Choosing the right positions, controlling the spike and closing the thermal and safety budgets gets closer to an engineering answer than simply specifying a higher voltage class everywhere.
HXY Electronics was founded in 2016 in Shenzhen and focuses on three product lines: high-performance MOSFETs, SiC MOSFETs and SiC diodes. The company position is that of a specialist in advanced power semiconductors, supplying devices and application support for data centers, renewable energy and high-end industrial systems. For the 800 V bus positions discussed here, the practical next step is a sample request, followed by double-pulse and high-voltage bench work to close out the three calculations of VDS overshoot, switching energy and temperature rise before moving to online ordering, volume quotations, lead-time checks or forward orders. Drive voltage ranges, threshold, gate charge and junction-to-case thermal resistance are listed here from the official datasheets; Eoss, switching energy, body-diode reverse-recovery figures and surge capability are still marked TBD, so request the full characteristics from an FAE or complete them on the bench, and take every value from the latest datasheet. FAE engineers can also help map topology stress onto specific part numbers.
This article is provided for reference only and does not constitute a purchasing or design commitment. Some architecture figures are drawn from third-party public material and device parameters from the HXY Electronics product catalog and official datasheets; loss and temperature-rise examples are static estimates under nominal conditions, and all data is subject to update. For design selection, refer to the latest official HXY Electronics datasheet. For application-specific questions, contact HXY Electronics technical support at sales@hxymos.com.