Why New Power Architectures Are Needed for Scaling AI Computing Performance

As AI computing generations advance, efficiency and scalability in power distribution become as critical as performance and rack density. Existing AC-based infrastructure undergoes multiple conversions before power reaches the GPU, and even small losses accumulate significantly in high-density racks.

800 VDC is an architecture that distributes higher-voltage direct current to reduce power conversion stages and improve efficiency from the grid to the accelerator. Through the DSX reference design, NVIDIA has presented a path to transition from existing AC infrastructure to native 800 VDC facilities via a hybrid structure.

NVIDIA is promoting 800 VDC standardization through the Open Compute Project (OCP) alongside Google and Microsoft. The three companies released a joint white paper in March 2026 and the LVDC Solid-State Transformer Specification v0.3 in July 2026, with over 80 equipment manufacturers and infrastructure companies developing related products.

Existing facilities can also be transitioned without a complete overhaul of the electrical system. The NVIDIA MGX-compatible 800 VDC power rack, scheduled for release in the second half of 2026, offers a hybrid structure that connects to existing AC infrastructure to supply 800 VDC to computing equipment within rack rows.

The configuration by scaling stage is as follows:

• Power Rack: A short-term path for transitioning from existing facilities to high-density computing
• Row Power Center: Supplies power to multiple rack rows via overhead 800 VDC busways, with support for up to 2MW per rack row planned from 2027
• DC Power Block: A facility-scale configuration that converts medium voltage from the grid to 800 VDC in a single step, serving as the foundation for next-generation AI infrastructure

Open standards defining common interfaces allow power equipment from multiple vendors to interoperate within the same 800 VDC facility. NVIDIA is materializing this as a system design connecting power architecture, rack-scale computing, and facility infrastructure, while Wood Mackenzie forecasts global AI and data infrastructure investment to reach $9 trillion by 2040.