Signal & Supply
← Archive
November 16, 2026 POWER SEMICONDUCTORS

Carmakers Overbuilt Silicon-Carbide Factories for an EV Boom That Slowed. Nvidia's New 800-Volt Racks Just Offered to Buy the Leftovers.

On October 16, Infineon locked in a new multi-year deal to buy silicon-carbide wafers from Wolfspeed — one of several such deals signed this year as chipmakers scramble to feed a market nobody originally built this capacity for. Automakers spent the early 2020s wiring up silicon-carbide fabs for electric vehicles. EV sales grew slower than planned. Now Nvidia's new 800-volt AI racks need the very same chips to survive currents that would otherwise melt their own internal wiring — and the leftover EV capacity just found a second customer.

Key takeaway Nvidia says a 1-megawatt AI rack wired the old way needs up to 200 kilograms of copper busbar just to carry its own current without overheating. Move the rack's internal power bus from roughly 54 volts to 800 volts DC, and the same power flows through about half the current — Nvidia's own estimate is over 150% more power through the same copper cross-section. The catch: switching and rectifying cleanly at 800 volts requires 1,200-volt silicon-carbide (SiC) transistors, not ordinary silicon — and the factories that make them were built, and overbuilt, for electric cars.

Same 500 kW rack, three bus voltages. Drag it.

Simplified model of one AI rack's internal power bus. Power delivered is held fixed at 500 kW (a representative rack load; real racks today span roughly 50–1,000 kW) and conductor current density is held fixed, so only voltage, current, copper and switch type change. Drag the slider, tap a preset, or press Run.

GRID RACK COPPER BUSBAR — 100% (baseline) ORDINARY SILICON SWITCHES OK
BUS VOLTAGE 48 V
CURRENT 10,417 A
COPPER NEEDED 100%
SWITCH CHIP Silicon
This is roughly today's legacy 54-volt-class rack design. A 500 kW load at 48V already draws over 10,000 amps — which is why the busbar has to be this thick, and why ordinary silicon switches are still fine at this voltage.

The plain version

Picture plugging a hair dryer into a wall socket, then picture pushing the same kind of electricity into a server rack that draws as much power as 150 houses. The newest AI racks already pull 200–600 kilowatts, and Nvidia's next generation is designed for a full megawatt. The old way of wiring that up — relatively low-voltage power pushed through thick copper bars bolted onto the back of the rack — runs into a hard limit: at low voltage, moving a lot of power means pushing a lot of current, and current is what makes wires heat up. Keep the voltage low, and you just keep adding copper to stop the bars from overheating. Nvidia says a single 1-megawatt rack, wired the old way, needs up to 200 kilograms of copper busbar — more metal than an adult rhinoceros weighs.

The fix, which Nvidia is pushing as a new industry standard, is to raise the voltage way up, to 800 volts — more than triple what's in a typical electric car's battery pack. Higher voltage moves the same power with less current, the same way raising water pressure lets you use a thinner pipe. The catch: ordinary silicon switches start leaking and overheating past a few hundred volts. The chips that handle 800 volts safely are made from silicon carbide (SiC), a tougher crystal than standard silicon.

Here's the twist. Nobody built silicon-carbide factories for data centers — they built them for electric cars, which also need high voltage to charge fast and drive far. EV sales grew more slowly than those factories expected, leaving spare capacity sitting idle. Chipmakers like Wolfspeed and Infineon are now redirecting that leftover capacity toward Nvidia's racks, signing new supply deals as recently as mid-October. Whether AI alone can soak up the leftover EV chips fast enough to matter is still an open question.

The expert version

Rack-level power draw in frontier AI deployments has climbed from under 50 kW a few years ago to a current range of roughly 200–600 kW, with Nvidia and McKinsey both citing 1 MW as the near-term design target for platforms like Nvidia's next-generation "Kyber" rack, expected around 2027. At that density, delivering power through conventional 480/415 VAC three-phase distribution — standard in today's data centers — would require more than 700 amps per phase, per McKinsey's analysis. Because resistive loss in a conductor scales with I²R, and copper cross-section has to grow with current to hold current density (and heat) roughly constant, that current level forces enormous conductor mass: Nvidia states a 54-volt busbar architecture needs up to 200 kg of copper per 1 MW rack.

The industry's answer is to move the rack's internal distribution bus to 800 volts DC. Since power equals current times voltage, raising bus voltage for the same power lowers current proportionally, which — for a fixed current density — shrinks the required copper cross-section by roughly the same factor. McKinsey estimates a 40–50% cut in copper mass for solid-state-transformer-based 800VDC designs; Nvidia claims its 800VDC architecture pushes over 150% more power through an equivalent copper cross-section versus a 54VDC design. (Both figures come from parties selling the new architecture, not independent audits.)

The tradeoff shows up in semiconductors, not metal. Switching and rectifying at 800 volts with acceptable efficiency needs power transistors rated near 1,200 volts, and ordinary silicon devices at that blocking voltage carry high conduction losses and poor thermal behavior. Wolfspeed — a silicon-carbide supplier with an obvious stake in the outcome — claims 1,200V SiC MOSFETs cut AC-DC and DC-DC conversion losses 25–40% versus silicon; gallium nitride, which tops out at lower blocking voltages, instead handles the higher-frequency conversion stages closer to the GPU. The supply irony: 150mm/200mm SiC wafer capacity at Wolfspeed, Infineon, onsemi, STMicro and ROHM was built chasing electric-vehicle demand that arrived more slowly than forecast, leaving fabs underused. On October 16, Infineon extended a multi-year 200mm wafer deal with Wolfspeed explicitly to secure AI-rack volume — redeploying automotive-era SiC capacity into a market analysts doubt AI alone can fully absorb before 2027–28.

Why it matters for tech + supply chain: AI's power-delivery bottleneck just got wired into the same silicon-carbide supply chain electric cars depend on — so a shortage, or a glut, in one market now ripples straight into the other.

Why it matters for tech + supply chain: 800VDC racks tie AI capex directly to wide-bandgap semiconductor capacity planned around automotive demand cycles, so EV demand swings now double as a leading indicator for AI power-delivery lead times.

Get it in your inbox every morning