NVIDIA builds next-gen CPO prototypes that put optics directly on the interposer

Optical engines move to the interposer

NVIDIA presented "Optical Integration for the Next Generation of AI" at SEMICON West 2026, outlining progress in CPO (co-packaged optics) and DWDM (dense wavelength-division multiplexing) technologies intended for future chips.

In current architectures, optical engines sit on the substrate beneath the chip. NVIDIA's next-gen approach places them directly on the interposer, the layer between the chip and the circuit board, bringing optical components closer to the processor.

Optical speeds up to 256 Gb/s

With DWDM, the company aims to push optical speeds up to 256 Gb/s. DWDM sends multiple signals at different wavelengths over a single fiber to increase bandwidth.

NVIDIA says first prototypes have already been built and will be used in future chips. The company frames optical integration as the next step in addressing scaled AI and compute demands.

Why it matters

As AI models grow, chip-to-chip data transfer in data centers is becoming a bottleneck. Moving optical links closer to the chip is seen as critical for both energy efficiency and bandwidth. NVIDIA has not yet said which products will use the technology or when.