The optical engine has lived on the package substrate for years, a compromise between proximity and manufacturability. Nvidia’s researchers want it off the substrate entirely.
At SEMICON West in October the company showed an optics-on-interposer prototype, placing the engines on the same silicon interposer that carries the compute dies. Removing the organic substrate from the electrical path shortens the connection and tightens the pitch, which should raise bandwidth density and cut power per bit.
Eight colors on one strand
Alongside it, Nvidia demonstrated dense wavelength-division multiplexing running eight wavelengths at up to 32 Gb/s each, for 256 Gb/s on a single fiber.
Both remain research. No commercialization timeline was offered, and the company says it is evaluating them for future products. Its researcher Nikola Nedovic noted that efficient optical links for AI systems still come down to balancing device and circuit design against thermal management and packaging.
Commercial optics is already shipping
The prototype matters because of what is in production behind it. Spectrum-X Ethernet Photonics switches went to full production in May, pairing 200 Gb/s SerDes with co-packaged optics for the Vera Rubin platform, with CoreWeave, Lambda and Oracle Cloud among early users.
The supply chain is unusually deep for a research program. TSMC fabricates the photonics, SPIL handles chip-scale packaging and test, TFC packages laser dies, Foxconn integrates finished switches, and Coherent, Lumentum and Sumitomo Electric supply external lasers.
The forward ramp points the same way. Kyber NVL1152 joins eight racks into a single 1,152-GPU NVLink domain, and the Feynman platform that follows is expected around 2028.
The substrate was always the slowest thing in the chain. Deleting it is the point.
