Google, Microsoft and Nvidia are pushing to make 800‑volt direct current (800VDC) the open standard for powering the next generation of high‑density AI data centers. Working through the Open Compute Project (OCP), the three companies are defining common requirements for 800VDC power distribution, including power conversion, power quality, system interfaces and safety.cio+2

The move is a direct response to AI workloads driving power consumption sharply higher. GPUs and AI accelerators now demand so much energy that traditional low‑voltage AC distribution is becoming difficult to scale in dense rack configurations.cio+2

Why 800VDC?

Today’s data centers mostly run on alternating current (AC) with multiple conversion stages (AC→480V→48V→~1V at the chip). Each stage adds losses, heat and extra hardware.hw+1

An 800VDC backbone simplifies this:

• Fewer conversion stages, improving end‑to‑end efficiency by up to ~5% in some estimates.adikumar

• Less copper and less heat: DC uses two conductors instead of four, and higher voltage means lower current for the same power. Reports cite 40–50% lower copper usage versus traditional 415VAC systems, with some designs claiming up to 80% reduction in specific scenarios.

• Lower costs: Operators can see 8–12% annual energy‑related OpEx savings and $4–8M CapEx savings per 10 MW build for AI‑first facilities. At 1 GW scale, that’s millions of pounds of copper saved.

What the OCP effort covers

The OCP work stream focuses on making 800VDC interoperable and safe across vendors and sites. Key areas include:cio+2

• Power conversion: Specifications for solid‑state transformers (SSTs) and rectifiers that convert medium‑voltage AC directly to 800VDC.diarioti

• Power quality and interfaces: Requirements for the 800VDC bus (voltage tolerance, ripple, communication protocols) and coordination with energy storage.diarioti

• Safety: DC‑specific protections such as high‑sensitivity arc detectors and fast DC breakers to handle sustained arcing risks.

Two closely related architectures are emerging in parallel:

• OCP “Mount Diablo”: a bipolar ±400V (800V pole‑to‑pole) LVDC spec for rack‑level DC power, targeting 100 kW–1 MW per rack.adikumar+1

• NVIDIA’s 800VDC reference: a unipolar 0–800V rail used in Blackwell/Rubin‑class AI factories, with sidecar power racks that separate 800V→48V conversion from the IT rack.adikumar

The fact that both the open hyperscaler track (OCP) and the proprietary chip track (NVIDIA) have converged on the same 800V class makes the standard far more “bankable” for 2027‑era equipment capex.adikumar+1

Why this matters

AI racks are moving from tens of kilowatts to hundreds of kilowatts and beyond 1 MW. At those densities, traditional low‑voltage AC distribution becomes bulky, lossy and expensive to scale. An 800VDC backbone lets operators feed ultra‑dense AI racks with fewer cables, less switchgear and higher efficiency, while standardizing on a common voltage class for multi‑vendor interoperability.adikumar

For now, the effort is focused on defining common requirements and reference designs (including SSTs and sidecar power racks). If adoption follows current momentum, 800VDC could become the default power architecture for new AI‑first data centers within the next few years.