NVIDIA, Google and Microsoft Advance 800 VDC Power for AI Factories
NVIDIA, Google and Microsoft advance an open 800 VDC architecture, with hybrid racks due in 2026 and 2 MW row systems expected in 2027 for AI factories.
Summary
On August 11, 2026, NVIDIA detailed an 800 VDC architecture intended to deliver more grid power to GPUs by replacing multiple alternating current conversions with fewer direct current stages. NVIDIA DSX reference designs guide AI factories from current AC systems through hybrid deployments to native 800 VDC facilities. NVIDIA, Google and Microsoft developed the architecture through the Open Compute Project, publishing a joint white paper in March 2026 and LVDC Solid-State Transformer Specification v0.3 in July 2026. More than 80 equipment and infrastructure companies are building products around its common, interoperable interfaces.
The NVIDIA MGX-compatible 800 VDC power rack, due in the second half of 2026, will fit existing AC facilities and supply 800 VDC within each row without building-level electrical changes, preserving land, power rights and infrastructure investments. A centralized row power center with an overhead 800 VDC busway is expected in 2027, supporting multiple rack rows and up to 2 megawatts per row. For new facilities, a DC power block will convert medium-voltage grid power directly to 800 VDC in one step, targeting infrastructure planned for the coming decade. NVIDIA data center infrastructure vice president Vladimir Troy called the ecosystem-backed hybrid approach a practical migration path. Wood Mackenzie projects $9 trillion in global AI and data infrastructure investment through 2040, increasing pressure to upgrade power delivery before compute demand exceeds facility capacity.
Positives
- More than 80 equipment and infrastructure companies are developing products for the open 800 VDC specification.
- The MGX-compatible power rack will bring 800 VDC to existing AC facilities without building-level electrical changes.
- Up to 2 megawatts per row will be supported by the row power center expected in 2027.
- NVIDIA DSX reference designs provide blueprints spanning power architecture, rack-scale computing and facility infrastructure.
- Common Open Compute Project interfaces will allow compatible power hardware from different vendors to operate within one facility.
Risks & concerns
- Multiple AC conversion stages add overhead and complexity as AI rack density and power requirements increase.
- Most existing AI factories use AC distribution, requiring hybrid equipment before they can adopt native 800 VDC infrastructure.
- The row power center is not expected until 2027, delaying broader multirow deployment.
- NVIDIA provided no availability date for the facility-scale DC power block intended for new construction.
- Compute demand could exceed facility power-delivery capacity unless operators resolve architecture constraints before absorbing projected investment.