800-Volt Direct-Current Power Architecture: Nvidia, Google and Microsoft Face Computing Bottlenecks in AI Factories

The latest generations of accelerating computing face unprecedented demands on the core infrastructure of data centers, where the main obstacle is not only the amount of power consumed in megawatts but also how that power is delivered from the electrical grid to graphics processing units. In traditional distribution systems, electricity moves from the grid as alternating current, then undergoes multiple conversion stages that add operational burdens and increasing complexity as server density on the racks rises. With the massive levels required by modern AI systems, these modest losses accumulate into significant barriers to scaling. In this context, an 800-volt direct-current power architecture offers a direct solution: it relies on delivering higher-voltage electricity via direct current, reducing the conversion stages between the grid and accelerators and ensuring that the greatest possible share of available power reaches the compute units.

The reference designs for Nvidia DSX systems aim at guiding AI factories through a gradual transition, starting with traditional alternating-current infrastructure, moving through hybrid solutions, and reaching native facilities that operate entirely on an 800-volt direct-current architecture.This step results from a joint collaboration among Nvidia, Google and Microsoft within the Open Compute Project, where the three companies released a joint white paper in March 2026, followed by the launch of low-voltage DC solid-state transformer specifications in July 2026. Today, more than 80 equipment and infrastructure manufacturers are developing products that comply with these open specifications, establishing an integrated supply chain based on standardized standards that allow devices to be interchanged among different suppliers within a single facility.

This technological shift does not require abandoning existing data centers, as the power rack compatible with Nvidia MGX’s 800-volt direct-current architecture, scheduled to become available in the second half of 2026, serves as a hybrid solution that integrates ultra-dense compute capabilities into operational facilities without any modifications to the building’s electrical systems. Vladimir Trovoy, vice-president of infrastructure data-center architecture at Nvidia, said the architecture opens new possibilities for compute performance and the power density needed for AI at scale, noting that working with companies participating in the Open Compute Project provides a practical, realistic path for AI factories rather than a mere future vision. This hybrid approach also protects operators’ prior investments in land, power-supply rights and building infrastructure.

The roadmap includes three graduated expansion stages: it begins with updating power racks in the near term, moves by 2027 to a row-level power center that supports up to 2 megawatts per row via a top-level DC bus, and culminates with continuous-power campuses for fully-equipped facilities.For specialized facilities, the row-level power center provides a central station for distributing electricity across full racks, while facility-wide continuous-power campuses act as an integrated unit that converts grid electricity directly to 800 volt direct current in a single step, representing the architecture planned for the next generation of infrastructure over the coming decades. These developments coincide with Wood McKinsey’s forecast that global AI and data-infrastructure investment will reach $9 trillion by 2040, making facilities that secure their power architecture early the most prepared to absorb these massive investment flows.