NVIDIA, Google, and Microsoft Unite on 800 VDC Power Architecture for AI Factories

The defining bottleneck of the generative AI era is no longer just how much raw electricity we can pull from the grid, but how efficiently we can deliver that power to the silicon. As individual compute racks demand hundreds of kilowatts, traditional alternating current (AC) distribution systems are hitting a thermal and physical wall. To keep the next generation of AI factories from melting their own infrastructure, the industry is forcing a fundamental shift in physics: bypassing legacy AC-to-DC conversions entirely at the rack level.

The Physics of the 800 VDC Power Architecture

For decades, data centers have relied on a standard power delivery pipeline: high-voltage AC from the utility grid is stepped down, converted to direct current (DC), converted back to AC for distribution, and finally converted back to low-voltage DC at the server motherboard. Each conversion step acts as a tax, throwing off waste heat and eating away at precious megawatt efficiency. At the scale of modern AI workloads, these compounded losses are no longer acceptable.

To eliminate this overhead, NVIDIA, Google, and Microsoft have co-developed an open-standard 800 VDC power architecture under the auspices of the Open Compute Project (OCP). By distributing power at 800 volts of direct current (VDC) directly to the compute racks, the architecture eliminates multiple conversion stages. Higher voltage means lower current for the same wattage, which significantly reduces resistive line losses (the I²R effect) and allows for thinner, less expensive copper busbars inside the cabinets.

"By distributing power at higher voltage through a direct current (DC), fewer conversion stages stand between the grid and the accelerator — which means more of the available power reaches the compute."

NVIDIA Architecture Group

NVIDIA MGX and the Path to Native DC Facilities

Rebuilding the world's data centers to be fully native DC facilities cannot happen overnight. Recognizing this friction, NVIDIA is introducing a modular transition path. The company's upcoming NVIDIA MGX-compatible 800 VDC power rack is scheduled to ship in the second half of 2026. This hardware allows operators of existing AC-powered facilities to deploy ultra-dense, high-voltage DC compute racks without initiating a complete, multi-million-dollar overhaul of their entire electrical infrastructure.

This hybrid approach allows data center operators to run highly efficient DC loops specifically for their power-hungry AI clusters while keeping legacy storage and cooling infrastructure on standard AC loops. It provides an immediate efficiency injection where it matters most: at the accelerator level.

A Rapidly Consolidating Ecosystem

Standards are only as good as their adoption, and the industry is moving quickly to align behind this new power paradigm. Currently, over 80 infrastructure and equipment manufacturers are actively building products designed around this new 800 VDC specification. This includes everything from specialized power shelf connectors and busbars to high-efficiency power supply units (PSUs) and liquid-cooling pumps designed to operate natively on DC loops.

• Reduced thermal overhead: Fewer conversions mean less waste heat generated by the power delivery system itself, lowering the cooling burden.
• Lower capital expenditure: Thinner copper busbars and fewer components reduce material costs at a time when raw metals are at a premium.
• Accelerated deployment: Modular standards like the NVIDIA DSX reference designs provide a blueprint that secondary manufacturers can build to immediately.

Why This Matters for the Future of Compute

The transition to 800 VDC represents a quiet but crucial phase change in AI hardware scaling. While algorithmic optimizations and node shrinks occupy the headlines, the ultimate limiting factor of deep learning is thermodynamics. The companies that build the most efficient power pipelines will be the ones capable of running the largest training clusters at the lowest marginal cost. By standardizing the electrical spine of the AI factory, hyperscalers are clearing the runway for the multi-hundred-megawatt clusters of the next decade.