NVIDIA Unveils Optics-on-Interposer Prototype for AI-System Interconnects

NVIDIA used SEMICON West 2026 to present Optics-on-Interposer, a co-packaged optics approach that moves optical components onto a silicon interposer rather than an organic substrate or a multi-chip module.

In a comparison for 200T ASIC packaging, the company indicated an optical-component placement span of 100 mm in a conventional organic-substrate design, versus 8 mm in the OOI silicon-interposer structure. The architectural objective is to bring key electrical and optical components closer together within the package.

NVIDIA also demonstrated a research DWDM scheme using eight optical wavelengths at 32 Gb/s per wavelength, for up to 256 Gb/s over a single fibre path. Both OOI and the DWDM work remain at the research stage. NVIDIA has said it is evaluating the barriers to commercializing these technologies, rather than presenting them as currently available products.

For AI data-center operators, the relevance is in the networking and packaging constraints emerging as GPU clusters grow. Interconnects between accelerators, memory and switching equipment are becoming a scaling bottleneck. Should NVIDIA translate OOI and its DWDM research into volume products, denser in-package connectivity and higher-capacity optical links could lower the energy burden of network infrastructure and enable higher cluster density.

The company plans to introduce its Rosa Feynman solution in 2028, with further CPO integration described as an expected direction for that platform. This remains a forward plan, not evidence of a commercially available or deployed product. More broadly, the work suggests that AI-infrastructure competition could continue to shift from standalone accelerators toward high-speed, energy-efficient interconnect architectures.