NVIDIA Fast-Forwarded Co-Packaged Optics Five Years Ahead of Schedule, Arriving First With Its Feynman GPUs

NVIDIA plans to bring Co-Packaged Optics to market first through its Feynman GPUs in 2028. Co-Packaged Optics, or CPO, integrates silicon photonics with GPUs to replace copper interconnects with light-based signaling. The approach targets lower latency and higher bandwidth links between CPUs and GPUs in large AI systems. Original commercialization targets called for 2033. NVIDIA advanced the timeline by five years according to a Nikkei Xtech report. The report notes that AI infrastructure growth can produce inter-platform distances exceeding 10 kilometers and data rates of several hundred gigabits per second or higher. The Optical Compute Interconnect Multi-Source Agreement formed in March. Members comprise NVIDIA, Broadcom, AMD, Meta, OpenAI and Microsoft. NVIDIA intends to field its initial CPO implementation alongside Feynman GPUs. NVIDIA confirmed at GTC 2026 that Feynman GPUs will adopt 3D Die Stacking. Intel will serve as a foundry partner using technologies such as EMIB. The GPUs will employ custom HBM rather than standard next-generation memory. A new CPU architecture called Rosa will support the GPUs. Additional components include BlueField-5, NVLink 8 CPO, Spectrum 7 204T CPO and CX10. AMD is pursuing its own CPO development with Global Foundries for MI500 GPUs around the same period.