Nvidia's 800-Volt Racks Aim to Push 157% More Power Through the Same Wire

Blackwell's power draw jumped 75% over Hopper

Nvidia's own account of the shift, published October 13, 2025, lays out the problem its 800-volt direct current design is meant to solve: GPU thermal design power rose 75% moving from the Hopper generation to Blackwell, and a 72-GPU rack now carries 3.4 times the power density of its predecessor, part of an overall 50-times jump in performance across the two generations. Today's data centers distribute power at 415 or 480 volts AC, a standard built for a much lower load. That load growth is also why Australian datacenter operator Firmus is seeking a $5 billion IPO to fund capacity after signing an OpenAI deal, the kind of buildout the voltage question sits underneath.

800 volts carries 157% more power through the same copper

Nvidia's design converts 13.8-kilovolt AC grid power directly to 800 volts DC at the perimeter of the building, then distributes it on row-level busways, cutting the multiple AC/DC conversion steps that can drop end-to-end efficiency below 90%. At 800 volts, Nvidia says the same wire gauge carries 157% more power than it would at 415 volts AC, while needing less copper to do it. The company's Kyber rack architecture pairs the new voltage with a 64-to-1 LLC converter that Nvidia says occupies 26% less area than a traditional multi-stage design, and sits alongside the existing MGX rack standard rather than replacing it outright.

Texas Instruments built the chips the ecosystem needs

Nvidia does not manufacture the power components itself, and its GTC event on March 16 through 19, 2026, gave the partners doing that work a stage. Texas Instruments unveiled a complete 800-volt reference design built around Nvidia's architecture, spanning an 800-volt hot-swap controller, an 800-to-6-volt DC/DC bus converter rated at 97.6% peak efficiency and more than 2,000 watts per cubic inch, a 6-volt-to-under-1-volt multiphase buck converter for the GPU core itself, a 30-kilowatt 800-volt AC/DC power supply, and 800-volt capacitor banks rated at 40 watts per cubic inch. "The exponential growth of AI computing demands a fundamental rethinking of how we deliver power in data centers," said Kannan Soundarapandian, TI's vice president and general manager of high-voltage power.

TI 800V component	Peak efficiency or density
800V-to-6V DC/DC bus converter	97.6% efficiency, >2,000 W/in³
800V AC/DC power supply unit	30 kW capacity
800V capacitor bank unit	40 W/in³

1. GPU power draw1.75 x multiple
2. Rack power density, 72-GPU system3.4 x multiple

Source: Nvidia developer blog, "Building the 800 VDC Ecosystem for Efficient, Scalable AI Factories," accessed 2026-09-14

The load itself swings by 70 percentage points in milliseconds

Nvidia's blog post flags a problem that has nothing to do with voltage: GPU power draw swings from around 30% to 100% utilization within milliseconds as inference workloads start and stop, a volatility that infrastructure like AWS's SageMaker HyperPod is built to smooth over at the software layer, not the electrical one. That swing is what makes AI power delivery different from a steady industrial load: the grid and the rack both have to absorb spikes an older datacenter never saw, and a conversion stage that is 97.6% efficient at steady state is not the same claim as 97.6% efficient during a spike.

Why 415-volt AC ran out of headroom

The voltage a datacenter distributes on has stayed close to 415 or 480 volts AC for decades because IT loads stayed low enough that the copper and conversion losses at that voltage were tolerable. Nvidia's own figures describe what changed the arithmetic: a 3.4-times increase in power density for a 72-GPU rack, layered on top of the 75% jump in per-GPU draw from Hopper to Blackwell. Moving that much more power through a rack at the same voltage means thicker copper, more conversion stages, and more heat shed at every stage before it reaches a chip. Nvidia's 157% figure for 800-volt DC is a statement about that copper problem specifically, not about the GPUs themselves.

Two conversion stages, not one

The TI design does not take a rack from 800 volts straight to the sub-1-volt level a GPU core actually runs at. It converts 800 volts to 6 volts first, at the 97.6% efficiency figure, then steps 6 volts down to under 1 volt with a separate multiphase buck converter closer to the chip. Each stage has its own loss, so the headline efficiency number describes one link in the chain rather than the full path from the grid to the silicon, a distinction Nvidia and TI's own materials do not collapse into a single end-to-end figure.

What Nvidia has not said

Nvidia's October 2025 post describes the transition as happening "in phases" without naming a date when a rack running fully at 800 volts actually ships in volume, and its efficiency and density figures are Nvidia's own, not independently measured at an operating facility. Until a hyperscaler discloses real power-usage-effectiveness numbers from an 800-volt deployment, the case for the redesign rests on the vendor doing the redesigning, and on component suppliers like TI whose own numbers describe a converter on a test bench rather than a rack under production load.