NVIDIA plans to change how high-performance Artificial Intelligence servers are built by shipping fully finished L10 compute trays as part of its ‘Vera Rubin’ VR200 product stack. The company will move beyond supplying CPUs and accelerators to original equipment manufacturers and instead deliver trays pre-populated with ‘Vera’ CPUs, ‘Rubin’ accelerators, memory, 800G network interface cards, 110kW power delivery and liquid-cooling infrastructure. NVIDIA intends for these trays to be validated before they leave the factory, with the VR200 line entering volume production in late 2026.
Reports indicate that the ‘Rubin’ accelerators could have a thermal design power of up to 2.3 kW for the highest-end models, producing rack-level power requirements that exceed 250 kW. That level of heat and power density is cited as a factor making bespoke cooling designs economically unviable. NVIDIA is said to be standardizing cooling and power across two performance tiers and increasing network input and output density, decisions that reduce the feasibility and cost-effectiveness of third-party replication. By centralizing production with select partners, the company aims to shorten development cycles and improve manufacturing yields.
The shift will change the role of system vendors and OEMs. Rather than designing core server electronics, many will focus on rack-level assembly, power configuration, installation of rack cooling sidecars and final certification. The centralized approach limits some of the hardware differentiation options that OEMs and hyperscalers have traditionally used. Overall, the move redistributes responsibilities across the supply chain while emphasizing validated, turnkey compute trays for Artificial Intelligence deployments.
