Blurring the Line Between Light and Silicon: Nvidia's OOI Redefines High-Speed Interconnect

At SEMICON West 2026, Nvidia pulled back the curtain on a potential game-changer for data center connectivity: its next-generation co-packaged optics (CPO) technology, dubbed OOI (Optics-on-Interposer). This isn't a minor tweak to existing designs; it's a fundamental rethinking of the package architecture itself.

The OOI Difference

Standard CPO approaches typically place optical modules side-by-side with the compute die on a shared organic substrate. OOI takes a different path by eliminating the organic substrate altogether. Instead, it integrates the optical engine—the component that converts electrical signals to light and back—directly onto the silicon interposer that holds the GPU chip.

The silicon interposer acts as a dense internal wiring layer within the package. Positioning the optical engine here brings the point of optical-electrical conversion dramatically closer to the computational heart of the chip.

The Engineering Payoff

  • Shorter Electrical Pathways: Drastically reducing the distance signals travel in the electrical domain slashes power consumption, cuts latency, and improves signal integrity.
  • Tighter Integration: Silicon interposers allow for much finer wiring than organic substrates, paving the way for packing more optical channels into a single package in the future.
  • Built for AI Clusters: As AI models grow, communication within thousand-chip clusters becomes a critical bottleneck. OOI's tighter photonic integration targets this bottleneck, aiming to enable the next leap in massive-scale AI training and inference.

A principal scientist from Nvidia Research shared that prototype samples featuring larger-area silicon interposers have already been produced, moving the technology from concept into a tangible engineering phase.

This demonstration signals a swift shift in high-performance computing from the era of "electrical interconnect" to "optical interconnect." With light engines sitting right next to GPU cores, the dense, power-hungry forests of copper cables inside today's data centers may soon give way to highly integrated, ultra-high-bandwidth optical links.