Vera Rubin Rack Headaches

NVIDIA’s Vera Rubin systems are pushing data centers past the point where new chips fit into old server rooms. The company’s own design calls for liquid cooling and new power architecture at rack scale. (nvidia.com) A rack is the metal cabinet that holds servers, power gear, and networking, and NVIDIA now treats that full cabinet as one computer instead of a stack of separate boxes. The flagship Vera Rubin NVL72 combines 72 Rubin graphics processors, 36 Vera central processors, NVLink 6 switching, ConnectX-9 networking, and BlueField-4 data processing units in one rack-scale system. (developer.nvidia.com) NVIDIA said on January 5, 2026 that Vera Rubin was designed around “power delivery and cooling” limits, not just raw chip speed. On March 16, 2026, it expanded that pitch to a full pod of five rack types and 40 racks that it says work together as one AI supercomputer. (developer.nvidia.com 1) (developer.nvidia.com 2) Cooling is the immediate constraint because these systems no longer rely on room air and fans the way older servers did. NVIDIA’s Vera Rubin rack design is “100% liquid-cooled,” and the company’s GTC 2026 session materials said Vera-Rubin deployments “extend the performance envelope of direct-to-chip liquid cooling.” (blogs.nvidia.com) (nvidia.com) Power is the second constraint because NVIDIA is pairing Rubin with 800-volt direct current infrastructure for future “gigawatt AI factories.” In October 2025, NVIDIA said more than 20 companies were joining its ecosystem for 800-volt direct current data centers and more than 50 MGX partners were preparing Vera Rubin NVL72 systems. (blogs.nvidia.com) That shift changes the retrofit math for operators that built around traditional 415- or 480-volt alternating current rooms and lower rack densities. NVIDIA said the move to 800-volt direct current offers higher capacity and efficiency, while Vertiv said in October 2025 that its 800-volt direct current reference architecture was already being used as a basis of design for large AI factory projects. (blogs.nvidia.com) (vertiv.com) NVIDIA’s public line is that Rubin’s heat can be managed with warmer water than older liquid-cooled systems needed. Jensen Huang said in January that Vera Rubin can use 45 degrees Celsius water and does not require chillers, a point NVIDIA repeated in materials describing “45C liquid cooling” and rack-level energy storage. (fierce-network.com) (blogs.nvidia.com) That does not mean existing facilities are automatically ready. Facilities Dive reported in January that chillers could still remain in halls built around Vera Rubin, even if the primary chip loop runs hotter, and Vertiv’s March 2026 release described operators facing pressure from rising density, complexity, and power demand as they compress deployment schedules. (facilitiesdive.com) (prnewswire.com) The argument around “rack headaches” is really about who pays to bridge that gap. NVIDIA says Vera Rubin-based products will start arriving from partners in the second half of 2026, so buyers now have to line up power distribution, liquid loops, and floor-ready designs before the hardware shows up. (nvidianews.nvidia.com)