NVIDIA outlines 800 VDC power plan for AI factories

NVIDIA has outlined an 800 VDC power architecture for AI factories, aimed at supporting higher-density AI computing infrastructure.

The architecture centres on distributing electricity through 800-volt direct current rather than conventional alternating current systems with multiple conversion stages. The approach is reflected in NVIDIA's DSX reference designs, which are intended to guide operators from existing AC-based environments to hybrid systems and then to fully native 800 VDC facilities.

The move comes as operators of large AI data centres look for ways to increase rack density without overhauling entire sites. NVIDIA argued that the challenge is not only how much power can be delivered to a facility, but how efficiently that power can be moved from the grid to the compute equipment inside it.

Hybrid route

For existing sites, NVIDIA is preparing an MGX-compatible 800 VDC power rack due in the second half of 2026. The unit is designed to work with current AC infrastructure and deliver 800 VDC to compute racks within the row without requiring changes to a building's electrical system.

The hybrid approach targets operators that have already invested in land, power access and physical infrastructure for AI facilities. Instead of waiting for a full rebuild, they would be able to introduce higher-density compute within existing sites.

"800 VDC unlocks the compute performance and power density required for AI at scale," said Vladimir Troy, Vice President of Data Centre Infrastructure at NVIDIA.

Through the Open Compute Project, NVIDIA is working with more than 80 ecosystem companies to give AI factories a practical path forward, not just a future vision, Troy said.

Industry work

NVIDIA said it has been developing the 800 VDC architecture with Google and Microsoft through the Open Compute Project. The three companies published a joint white paper and a low-voltage direct-current solid-state transformer specification through the group, setting out a common framework for equipment vendors and infrastructure providers.

More than 80 equipment manufacturers and infrastructure companies are already building products to that specification, according to NVIDIA. The aim is to create common interfaces so hardware from different suppliers can operate within the same 800 VDC facility.

That matters for operators trying to avoid dependence on a single vendor for critical power infrastructure. The open specification also gives component makers a clearer basis for product development as demand for AI data centre build-outs rises.

Scaling path

NVIDIA described the new architecture as a staged roadmap rather than a one-step redesign. The first stage is the power rack for hybrid environments, which it sees as the likely entry point for many operators.

The second stage is a row power centre, a centralised power station for a full rack row that uses an overhead 800 VDC busway. NVIDIA expects that product to be available in 2027 and said it would support up to 2 megawatts per row.

Beyond that, NVIDIA outlined plans for a DC power block intended for new facilities. The unit would convert grid power directly to 800 VDC in a single step and is aimed at larger sites designed specifically around future AI demand.

The broader argument behind the architecture is that repeated AC-to-DC conversion adds overhead as rack power levels rise. By reducing the number of conversion stages, more incoming electricity can be delivered to accelerators and other compute systems while reducing infrastructure complexity.

Investment pressure

The announcement comes against the backdrop of rising capital spending expectations for AI and data infrastructure. NVIDIA cited a Wood Mackenzie projection of USD $9 trillion in global AI and data infrastructure investment through 2040.

That spending outlook is likely to intensify scrutiny of facility design, especially among operators trying to ensure that both new and retrofitted sites can support denser hardware over time. Power architecture has become a central issue as data centre operators try to align electrical systems with the demands of AI training and inference clusters.

NVIDIA's DSX reference designs are intended to serve as a system-level blueprint linking power distribution, rack-scale computing and facility infrastructure. The goal is to give operators a way to connect those elements as they expand AI factories across different stages of development.