Rack-Scale AI Servers Go All-Liquid as Rubin Density Rises and AI-Server Component Costs Climb
NVIDIA's Rubin-generation rack designs shown by ASUS in Seoul move eight GPUs into a single 1U liquid-cooled tray and push fans out of the chassis entirely, a density shift that raises demand for cold plates, manifolds and CDUs. Component-level pricing is splitting, with AI-server MLCCs rising while general-purpose passives fall.
At ASUS AI Tech 2026 in Seoul the central hardware change was mechanical rather than silicon. ASUS showed liquid-cooled NVIDIA HGX Rubin NVL8 designs that flip GPU orientation on the baseboard so cold plates sit directly over the packages, fitting eight GPUs into 1U where air-cooled HGX baseboards previously needed six rack units or more. Each node also carries a new generation of NVLink Switch onboard, keeping more communication inside the NVL8 domain instead of pushing it out to the scale-out fabric.
The rack-level trend is the same. ASUS displayed ORv3 liquid-cooled trays from its GB200, GB300 NVL72 and Vera Rubin NVL72 configurations under the ASUS AI Pod brand, and the generational shift is toward removing fans from the chassis and liquid-cooling the entire rack. Speakers at the event framed efficiency, not raw nameplate power, as the constraint on how much performance a given data centre can deliver, which is why hybrid and full-liquid variants are being offered side by side to fit sites with different facility capabilities.
Component pricing is moving with the density. DIGITIMES reports that multilayer ceramic capacitor prices for AI servers have continued to climb while prices for general-purpose components used in consumer appliances have dropped sharply, as major manufacturers shift capacity toward higher-value products and leave midrange categories to Chinese suppliers. That split means bill-of-materials pressure concentrates in the AI server segment even as the broader passive-components market softens.
For procurement the practical consequence is that power and thermal envelopes, not chip availability alone, now set deployment timelines. A rack drawing well beyond traditional air-cooling limits requires liquid distribution, higher-capacity power shelves and leakage monitoring before any GPUs are installed, and each new accelerator generation shortens the window in which a facility design remains current.