Nvidia recently unveiled a prototype of its upcoming Rubin Ultra GPUs alongside the NVL576 Kyber rack infrastructure during GTC 2025. These GPUs are scheduled for release in the latter half of 2027. As an AI-focused infrastructure leader, Nvidia is actively strategizing the transition from current capabilities to future advancements. This includes plans for GPU servers that may require up to 600kW of power per rack.
Current and Future Server Capacities
Currently, the Blackwell B200 server racks consume about 120kW of power per unit. The initial Vera Rubin solutions, expected in late 2026, will utilize the same structural framework as the Grace Blackwell series. However, the forthcoming Rubin Ultra solutions aim to significantly increase the GPU capacity per rack by fourfold. Jensen Huang, during a Q&A session, confirmed that some solutions could reach up to 600kW per rack, with full SuperPODS potentially requiring multi-megawatt power levels.
Performance Expectations
While detailed specifications for the Rubin Ultra are not yet available, Nvidia has outlined ambitious performance targets. The roadmap suggests that the upcoming Rubin NVL144 racks will achieve up to 3.6 EFLOPS of FP4 inference by 2026, while the Rubin Ultra NVL576 racks are projected to deliver an impressive 15 EFLOPS of FP4 by 2027, showcasing a significant leap in both computational and power density.
Architecture of Rubin Ultra
Each Rubin Ultra rack will comprise four ‘pods,’ each exceeding the computational power of an entire Rubin NVL144 rack. Each pod will contain 18 blades, with each blade capable of housing up to eight Rubin Ultra GPUs, along with potentially two Vera CPUs, although this detail remains unconfirmed. In total, each rack could support 576 GPUs, with significant computational output from each pod.
Advancements in NVLink Technology
Enhancements are also on the horizon for NVLink technology, with the new units set to feature three next-generation NVLink connections, an upgrade from the two connections available in current 1U rack-mount units. Early versions of the NVLink and Rubin Ultra blades have been showcased alongside the Kyber rack infrastructure.
Projected Power Requirements
While specific power specifications have yet to be disclosed, Jensen Huang indicated that future data centers might require megawatts of power for each server rack. Although the Kyber infrastructure targets around 600kW, future developments could see power demands surpassing 1MW per rack, especially given the current 1000-1400 watts per GPU consumption seen in the Blackwell series.