NVIDIA published its 800 VDC (volts direct current) power architecture specifications for AI data centers in collaboration with Google and Microsoft through the Open Compute Project. The architecture shifts from traditional 54 VDC or 415-480 VAC distribution to direct 800 VDC delivery, reducing conversion stages, copper use by 45%, and improving end-to-end efficiency by up to 5%. An NVIDIA MGX-compatible 800 VDC power rack is arriving in H2 2026 to retrofit existing AC-based facilities, with a centralized row power center supporting up to 2 MW per row expected in 2027.
More than 80 equipment manufacturers and infrastructure companies have committed to building to the 800 VDC spec, including silicon providers (Analog Devices, Infineon, Innoscience, STMicroelectronics, Texas Instruments), power systems (Eaton, Vertiv, Schneider Electric), and partners (Delta, Flex Power). Texas Instruments demonstrated a complete reference architecture in March with an 800V-to-6V DC/DC converter delivering 97.6% peak efficiency—proving the supply chain can deliver the building blocks needed for megawatt-scale racks.
The architecture enables 1 MW IT racks and beyond, critical for next-generation accelerators like NVIDIA's Rubin Ultra and Kyber. Full-scale deployment begins in 2027; data centers planning new builds need 800 VDC readiness now to avoid stranded infrastructure. Wood Mackenzie projects $9 trillion in global AI and data infrastructure investment through 2040, with facilities resolved on power architecture before compute demand arrives.
For operators, this matters because 800 VDC maintenance costs drop up to 70% through fewer power supply units, total cost of ownership falls ~30%, and existing facilities have a hybrid path that doesn't strand AC infrastructure. This is no longer a future vision—it's a procurement cycle decision for new capacity in 2026-2027.