News

Nvidia Feynman architecture expected to replace copper NVLink with co-packaged optics

Sunday, February 22, 2026 at 11:00 PM

Discussion regarding the potential transition from copper-based NVLink to co-packaged optics (CPO) in Nvidia's future Feynman architecture for data center scale-up.

Context

At GTC 2025, Nvidia unveiled its next-generation Feynman microarchitecture, slated for release in 2028. This roadmap milestone represents a major shift toward "Inference Sovereignty," focusing on token-level latency and power efficiency rather than raw throughput. A critical evolution for Feynman is the expected transition of NVLink from copper-based interconnects to co-packaged optics (CPO). While current NVL72 systems rely on over 5,000 copper cables per rack, physical limitations at speeds of 448 Gbps and beyond are driving the move to silicon photonics. To secure this transition, Nvidia has invested $4 billion in photonics leaders Coherent and Lumentum. By integrating optical transceivers directly into the GPU package, the company aims to achieve 3.5x lower power consumption and 10x better network resiliency. This shift is essential for scaling "AI factories" to millions of GPUs, enabling the real-time reasoning and agentic AI workloads that the industry expects to dominate by 2028.

Related Companies

TSMC
TSMC
2330
TW
Nvidia
Nvidia
NVDA
US