Nvidia revives Rubin CPX chip program with 2027 launch planned, analyst says

Nvidia has revived its Rubin CPX product program, which was previously believed to have been canceled, according to supply chain research by TF International Securities analyst Ming-Chi Kuo. Production of the next-generation AI accelerator is now scheduled to commence in the first quarter of 2027. The revived architecture features substantial modifications to GPU specifications and rack architecture, tailored specifically to handle compute-intensive AI inference workloads. Under the updated specifications, each CPX GPU delivers computing performance comparable to the primary Rubin GPU while operating at a maximum power rating of 2,300 watts. Nvidia upgraded the memory configuration to 168 GB of HBM4, moving away from the earlier plan of 128 GB GDDR7. The system uses a dedicated MGX ETL rack architecture accommodating configurations of up to 256 GPUs and requires CPX units to pair 1:1 with Vera Rubin NVL72 racks. CPX will manage prefill tasks and generate KV cache—which represents over 50% of current inference workload—before routing data to Rubin GPUs for decode processing, boosting Nvidia's long-term competitive moat in data center AI infrastructure.