NVIDIA argues AI scale will be limited by power delivery, pushing an 800 VDC data center architecture

NVIDIA says the next wave of AI compute will run into a bottleneck that is not measured in GPU performance alone, but in how electricity gets from the grid to the hardware inside data centers. As rack density rises and inefficiencies compound at higher power levels, the company argues that traditional alternating-current (AC) power delivery and multiple conversion stages can leave less usable power for the accelerators themselves.

In NVIDIA’s view, the core issue is not simply wattage capacity at the facility, but the path the electricity takes. Under conventional designs, power is supplied from the grid as AC and then converted multiple times through various stages before it reaches racks. NVIDIA says each conversion adds overhead and complexity, and that small losses can add up quickly once the power levels demanded by next-generation AI hardware grow.

To address that, NVIDIA is promoting an 800 VDC architecture, short for “800 volt direct current.” The company’s central claim is that distributing power at higher voltage using direct current reduces the number of conversions between the grid and the compute accelerators. In practical terms, NVIDIA says this approach can increase the share of available power that reaches the computing equipment.

NVIDIA also frames 800 VDC as an evolution path for operators, rather than a clean-slate rebuild. The company says its DSX reference designs are meant to guide AI factories through a transition that starts with today’s AC infrastructure, moves through hybrid models, and ultimately reaches “fully native” 800 VDC facilities. The goal, as described by NVIDIA, is to give data center owners options that align with how capital is already deployed in buildings, land, and electrical systems.

A key element of that on-ramp, according to NVIDIA, is a rack-level product. NVIDIA says the NVIDIA MGX-compatible 800 VDC power rack, expected to arrive in the second half of 2026, is designed to fit into existing AC infrastructure while still delivering 800 VDC to compute racks within a row. The company says operators would not need to change the building’s electrical system for that step, positioning it as a near-term upgrade for higher-density compute without a facility-wide retrofit.

Beyond rack deployments, NVIDIA describes additional building blocks at larger scales. It points to a row power center, described as a centralized power station that uses an overhead 800 VDC busway to scale distribution across multiple rack rows, with support for up to 2 megawatts per row. Separately, for new facilities planned for the decade ahead, NVIDIA points to a DC power block that it describes as converting grid power directly to 800 VDC in a single step through medium-voltage conversion.

NVIDIA links the effort to broader industry standardization. It says NVIDIA, Google, and Microsoft have been developing the 800 VDC architecture together through the Open Compute Project (OCP), and that the trio published a joint white paper in March 2026 along with the “LVDC Solid-State Transformer Specification v0.3” in July 2026. NVIDIA adds that more than 80 equipment manufacturers and infrastructure companies are already building products to the specification, and it characterizes this as the start of a supply chain forming around an open standard.

The company’s motivation, in its own words, is that power architecture decisions today can either enable or constrain AI growth. Vladimir Troy, vice president of data center infrastructure at NVIDIA, is quoted saying 800 VDC unlocks the compute performance and power density required for AI at scale, and that through OCP NVIDIA is working with more than 80 ecosystem companies to provide a practical path forward instead of a distant vision.

NVIDIA also places its push in the context of a larger investment cycle. It cites Wood Mackenzie’s projection of $9 trillion in global AI and data infrastructure investment through 2040, arguing that the facilities best positioned to absorb that spend will have resolved their power architecture before compute demand outran what infrastructure could deliver. The company stops short of publishing specific performance benchmarks or cost comparisons in the blog post, and it does not detail pricing, availability constraints, or how quickly operators can migrate from AC-designed environments to full 800 VDC. Operators considering the approach will likely want to examine integration details with their existing power equipment, utility interconnect constraints, and the timeline for rack, row, and facility-scale components.