NVIDIA Reportedly Cancels Four-Chip Rubin Ultra GPU

NVIDIA is adjusting its roadmap as reports indicate the cancellation of the four-compute-chip Rubin Ultra design ahead of its 2026 launch.

NVIDIA plans to deliver its first NVL72 systems featuring the Vera-Rubin architecture in the second half of 2026. This architecture includes the Vera CPU and a new Rubin GPU, which consists of two compute chips and eight HBM4 memory stacks. While NVIDIA confirmed a roadmap featuring Rubin and a subsequent Rubin Ultra, SemiAnalysis reports that the Rubin Ultra design with four compute chips connected via NVLink-2C2 has been canceled. This version was intended to double the HBM4 stacks compared to the standard Rubin. Although NVIDIA has not commented, TSMC is reportedly facing CoWoS-L package warpage challenges with 2+2 configurations. Such warpage can prevent chips from making full contact with connections, leading the industry to research glass substrates. NVIDIA continues to compete in the AI space against AMD's Instinct, Google's TPU, and Amazon's Trainium chips.