OCP 2026 Turns Rack Power Into the Constraint as NVIDIA Opens Vera Rubin and 800 VDC Scales
NVIDIA opened its Vera Rubin rack and compute-tray designs and pushed an 800 VDC architecture at OCP 2026, while Delta showed a 660 kW in-row power rack and a 2.5 MW liquid-to-liquid cooling unit aimed at racks heading toward 250 kW.
The Open Compute Project Global Summit in San Jose on 13 October turned into a power and cooling show. NVIDIA opened the server rack and compute tray designs for its Vera Rubin platform to outside partners, covering power, liquid cooling and networking interfaces, and published an MGX 1.2 specification supporting rack deployments of up to 250 kW with a 5,000 A liquid-cooled busbar, rack-level blind-mate liquid interfaces and fanless liquid-cooled compute trays. It also published a Vera CPU rack design holding up to 256 Vera CPUs for agentic workloads, alongside the 72-GPU and 36-CPU Vera Rubin NVL72 configuration.
The electrical change is the substantive one. NVIDIA is proposing 800 VDC rack distribution instead of the 48-54 VDC that data centres use today, because at roughly a megawatt per rack the current at low voltage becomes the physical limit on scaling. Delta Electronics used the summit to show the hardware that makes that practical: a solid-state transformer converting medium-voltage AC straight to 800 VDC, a high-voltage distribution board stepping down to 54 VDC and 12 VDC at up to 98% efficiency and 1,000 W per cubic inch, and an 800 VDC in-row power rack built from six 7RU modules pairing 110 kW power shelves with 80 kW battery backup units for 660 kW of output and 480 kW of backup.
Cooling got the matching treatment. Delta showed a liquid-to-liquid cooling distribution unit rated up to 2.5 MW using three hot-swappable pumps in a 3+0 arrangement delivering up to 3,850 litres per minute at 60 psi, a microchannel cold plate for Vera Rubin NVL72, and a 5,000 A liquid-cooled busbar that cools the power delivery itself rather than only the processors. A 90 kW DC/DC shelf converts 800 VDC to 50 V at up to 98.5% efficiency.
AMD's side of the hall shows the same constraint being met differently. Its Helios reference design, developed with Schneider Electric, supports up to 246 kW on a hybrid air and liquid architecture, while the new OCP-standard Helion rack offers CPU-only and mixed configurations with up to 16 Instinct MI350P GPUs and 256-core sixth-generation EPYC parts at 400 W and 600 W. A Helios switch tray on the HPE floor carried two Broadcom Tomahawk 6 102.6T switches on 224G PAM4 SerDes, liquid-cooled blocks and retimers to reach rear fabric connectors through copper, an illustration of why co-packaged optics is the next step rather than a luxury.