NVIDIA's upcoming Rubin and Feynman GPU generations will primarily use copper SerDes for intra-rack connections, with Co-Packaged Optics (CPO) for inter-rack links (NVL1152) in Feynman, pushing full CPO ramp to late 2028-2029 due to manufacturing problems and yield issues. This delay is expected to benefit copper interconnect suppliers like Amphenol, which are projected to perform better than previously anticipated.
Co-packaged optics adoption is delayed by manufacturing challenges, extending the runway for copper interconnect suppliers in AI server clusters and impacting NVIDIA's next-gen GPU architecture.