As AI Racks Head Toward a Megawatt, Power Delivery Becomes a Chip Business

A single rack of AI servers used to draw tens of kilowatts. Nvidia now says a megawatt per rack is "on the horizon," and that number is reshaping the market for the power semiconductors that move electricity from the grid to the chip.

The change starts with physics. Nvidia's engineers wrote that moving from its Hopper to its Blackwell architecture raised the power of each GPU by 75%, but growing the NVLink domain to 72 GPUs in one system increased rack power density 3.4 times. Delivering that kind of power at the 54 volts of direct current used inside today's racks would require enormous currents, with high resistive losses and an "unsustainable volume of copper cabling."

Nvidia's answer, laid out in a technical blog and white paper presented at the Open Compute Project summit last year, is to distribute power through the data hall at 800 volts DC. Medium-voltage alternating current from the utility would be converted to 800 volts at the facility level, eliminating layers of AC switchgear, transformers and power-distribution units. At the same wire gauge, Nvidia said, 800 volts DC can carry 157% more power than 415 volts AC.

Inside the rack, the conversion steps change too. In Nvidia's planned Kyber rack, 800 volts is delivered to each compute node, where a single high-ratio converter steps it down 64 to 1, to 12 volts, right next to the GPU. Nvidia said that single stage is more efficient and takes 26% less board area than multi-stage designs.

Every one of those stages is built from power transistors, and different materials suit different jobs. Navitas Semiconductor, one of the suppliers on Nvidia's list, described the split in its own release: high-voltage silicon carbide devices for grid-tied and medium-voltage equipment, 650-volt gallium nitride parts for converters fed by the 800-volt line, and new 100-volt gallium nitride transistors for the lower-voltage stages on GPU power boards.

That division of labor is visible in the list of chip suppliers Nvidia named as collaborators on 800-volt designs: AOS, Analog Devices, Efficient Power Conversion, Infineon Technologies, Innoscience, MPS, Navitas, onsemi, Power Integrations, Renesas, Richtek, ROHM, STMicroelectronics and Texas Instruments. Power-system makers including Delta, Flex and LITEON, and data-center equipment suppliers including ABB, Eaton, GE Vernova, Schneider Electric, Siemens and Vertiv, are also on the list.

Power density is only half the problem. Nvidia also highlighted volatility: during AI training, thousands of GPUs compute and then exchange data in near unison, and a rack's draw can swing from about 30% to 100% of its rating and back in milliseconds. Across a data hall, that means hundreds of megawatts ramping in seconds. Nvidia's proposed fix is layered energy storage, with capacitors near the racks to absorb fast spikes and large battery systems at the grid connection for slower swings.

For investors, the shift turns power delivery from a commodity into a design decision tied to each new GPU generation. The open questions are timing and standards. Nvidia has said the move to a fully 800-volt architecture will happen in phases, and the industry still has to agree on voltage ranges, connectors and safety practices through bodies such as the Open Compute Project, whose 2026 summit runs in San Jose this week, Oct. 12 to 15.

The markers to watch are the first production deployments of Kyber-class racks, qualification announcements from the power-chip suppliers on Nvidia's list, and any new specifications for 800-volt equipment from the Open Compute Project.