Google, Microsoft, and Nvidia are working with the Open Compute Project to develop an open 800-volt direct-current power standard for future AI data centers. The initiative aims to make high-density computing facilities more efficient, easier to scale, and less expensive to build and operate.developer.nvidia

AI infrastructure is changing the basic design assumptions behind data-center power systems. GPU clusters are becoming larger, racks are drawing far more electricity than traditional server installations, and the old approach of distributing lower-voltage AC power is becoming increasingly difficult to scale.

Why 800V DC Matters

Most data centers currently receive medium-voltage AC power, step it down, condition it through UPS equipment, distribute it through switchgear and PDUs, and then convert it to DC again inside every server rack. Each conversion stage introduces hardware, energy loss, heat, cost, and potential failure points.

The proposed model would convert AC power to 800V DC at the facility level and distribute DC directly across the data hall. Compute racks would then reduce the voltage near the hardware, where it is needed by GPUs and other components. This simplifies the power path and removes several conversion stages.developer.nvidia

At higher voltage, the same amount of power can be delivered with lower current. Lower current means less resistive heat loss and fewer or smaller cables. Nvidia says that 800V DC can carry 157% more power than a 415V AC system using the same wire gauge, while requiring a simpler three-wire configuration rather than the four conductors commonly used for AC distribution.developer.nvidia

Lower Copper Use and Operating Costs

The potential savings are significant, especially for new AI-focused facilities:

• Copper consumption could fall by 50% to 80% because the system needs fewer conductors and less parallel cabling.
• Annual energy-related operating costs could decline by 8% to 12% through lower conversion and distribution losses.
• A 10 MW AI data-center build could save approximately $4 million to $8 million in capital expenditure by reducing upstream AC infrastructure.networkworld

These figures should be treated as estimates, not guarantees. The cited study was conducted by Enteligent, a company developing DC-power infrastructure, so organizations should independently assess the economics for their own projects. Still, the core engineering argument is straightforward: fewer power-conversion stages and lower current can reduce waste heat, cable volume, and equipment complexity.networkworld

AI Changes the Power Profile

The problem is not only total consumption. AI training workloads can cause rapid and synchronized changes in power demand. Thousands of GPUs may move together between intensive computation and data-exchange phases, causing rack power to swing from around 30% utilization to full load within milliseconds. Across a large facility, those swings can reach hundreds of megawatts and create challenges for electrical equipment and the utility grid.developer.nvidia

Nvidia’s proposed architecture pairs 800V DC distribution with multi-layered energy storage:

• Capacitors or supercapacitors near the racks could absorb extremely short spikes and drops in demand.
• Larger battery energy-storage systems at the facility level could handle slower power changes, protect grid stability, and support transitions to backup generation.

In this model, storage becomes an active component of the AI data center rather than a system used only during an outage.

Open Standards Are Essential

The Open Compute Project provides a forum for companies to align on voltage ranges, connectors, power-conversion systems, safety procedures, and interoperability rules. OCP began as an initiative to create more efficient and open data-center hardware designs, and it has since expanded to cover servers, storage, networking, racks, and power systems.networkworld

An open 800V DC standard could help avoid a market where every vendor develops incompatible power systems. That is especially important because the transition will involve equipment makers, data-center operators, chip vendors, electrical suppliers, UPS manufacturers, and cloud providers.

The shift will likely begin with new “greenfield” AI facilities rather than mass upgrades of older data centers. Retrofitting an existing AC-based site can be complex and expensive, while a new project can be designed around high-voltage DC from the beginning.networkworld

For the AI industry, the message is clear: future performance will depend not only on faster chips, but also on the ability to deliver, stabilize, and cool enormous amounts of power. 800V DC may become one of the key building blocks of that next generation of AI infrastructure.