Overview: Nvidia's Dual Push for Next-Gen AI Infrastructure

At the recent OCP summit, Nvidia unveiled the Vera Rubin rack architecture featuring liquid-cooled busbars supporting 5000 amperes and up to 250kW per rack. Concurrently, Nvidia is collaborating with the industry to advance 800V high-voltage direct current (HVDC) power distribution. These dual initiatives address a critical bottleneck: as GPU power demands surge, traditional power supply equipment increasingly consumes valuable rack space.

Technical Details: The Physics of Power and the Copper Limit

The fundamental physics of power distribution dictates this shift. Delivering 1 megawatt (MW) at 54V requires approximately 18,519 amperes, whereas 800V reduces this to just 1,250 amperes—a 15-fold difference. High current causes severe resistive heating in copper busbars. Nvidia estimates that a 1MW rack at 54V might require up to 200 kilograms of copper, and traditional power shelves could occupy up to 64U of rack space. This spatial inefficiency directly conflicts with the dense deployment of GPUs and NVLink interconnects.

Architectural Shift: Moving Power Out of the Rack

The 800V paradigm fundamentally alters the power delivery path. Instead of converting AC to DC inside the compute rack, 800V consolidates AC-DC conversion into independent Power Racks. These distribute 800V DC via busbars to compute racks, which retain DC-DC step-down modules. Nvidia's roadmap targets H2 2026 for 800V Power Racks compatible with existing AC infrastructure, 2027 for row-level centralized power delivering up to 2MW per row, and long-term integration of solid-state transformers for direct medium-voltage AC to 800V DC conversion.

Market Context: Why the Transition is Happening Now

The urgency for this transition stems from escalating rack power densities. TrendForce estimated in June 2026 that while the GB300 operates at roughly 150kW and the Vera Rubin VR200 at 225kW, future Rubin Ultra designs could reach 660kW, with some racks approaching the megawatt scale. Furthermore, AI workloads involve rapid transitions between compute, communication, and idle states, demanding dynamic load management and robust energy buffering. While 54V systems can still support the 225kW Vera Rubin phase, the leap to 660kW and beyond necessitates a reevaluation of low-voltage, high-current economics.

Industry Impact: Redefining the Power Supply Chain

This architectural overhaul is reshaping the power supply chain. Functions previously handled by in-rack power modules are shifting to centralized Power Racks, elevating the importance of high-power DC-DC converters and DC fault protection. Global manufacturers like Delta, Vertiv, Eaton, Schneider, and Lite-On are actively advancing these technologies. In the Chinese market, Megmeet (002851) showcased 800V Power Racks at the OCP summit, reporting 1.841 billion CNY in power product revenues for H1 2026. Zhongheng Electric (002364) saw its data center power revenue surge 98.48% to 808 million CNY, though its 800V series remains in the industrialization phase. Oultons (300870) reported 1.136 billion CNY in data center power revenues, with high-power server power contributing 936 million CNY.

Future Outlook: The Economic Equation and Adoption Timeline

The economic viability of 800V hinges on multiple factors. Nvidia claims 800V DC can reduce copper material requirements by approximately 45% compared to 415V AC, and improve end-to-end efficiency by up to 5% over 54V systems. However, the capital expenditure for centralized power racks must be offset by these operational savings and the value of reclaimed rack space. While the technological trajectory is clear, widespread deployment of 800V Power Racks is expected to gradually scale post-2027 with Rubin Ultra, reaching broader adoption around 2028. Legacy data centers must carefully weigh the total cost of ownership against the tangible benefits of high-density AI computing.