Nvidia Restarts Rubin CPX Project, Mass Production in 2027 to Boost AI Inference Prefill

According to analyst Ming-Chi Kuo's latest industry survey, Nvidia has restarted the Rubin CPX project, an AI inference prefill accelerator GPU, but with significant design changes, targeting production in Q1 2027. The new Rubin CPX offers computing power close to the standard Rubin GPU, improved prefill performance, same 2300W power draw, and 168GB HBM4 memory, up from 128GB GDDR7 in the old version.

In system architecture, 8 new Rubin CPX chips form one compute tray, and 8 trays make up a rack module, using Spectrum-6 all-copper Ethernet for scale-out. The new Rubin CPX uses a separate MGX ETL rack, with each rack containing 1 to 4 rack modules, connected via Spectrum-6 + OSFP optical solutions. Nvidia suggests pairing Rubin CPX with standard Rubin at a 1:1 ratio, with ETL racks connected to Vera Rubin NVL72 racks via Ethernet RDMA.

Kuo notes that over 50% of AI inference workloads involve processing input and building KV cache. The new Rubin CPX can handle prefill loads with more flexible deployment and lower cost. Its compute tray has 1.34TB HBM memory, enough for most long-context prefill needs.