Nvidia moves optics onto the interposer: light beats copper

AI clusters no longer grow because transistors got faster. They grow because data still manages to travel between chips. That runway is running out — copper is hitting physics. Nvidia just showed the detour: light. ### What went on the stand At SEMICON West 2026 Nvidia unveiled a prototype of OOI — Optics-on-Interposer. The optical engine no longer sits on a conventional organic package substrate. It moved onto the silicon interposer, right where the compute chips live. What that buys: higher bandwidth density, lower power draw per bit moved. For racks where every watt turns into heat instead of work, that matters more than any clock bump. ### Why this is not just another part The problem with AI systems stopped being how much you can compute. It is how fast you can hand results to the next chip without burning half your power budget on the trip. Copper at high speeds runs hot, noisy and hungry. Optics fixes that — but until now it was kept at the edge of the package, away from the hot silicon. OOI drags it into the middle. Put simply, light now goes where the real work happens. The chip is not what changes. The way chips talk to each other is. ### What is solid, what is not Solid: a prototype exists, it was shown, the design logic is spelled out. It targets systems after 2028 — meaning Nvidia is building a foundation, not a product for next quarter's shelf. Not solid: calling it a release. DigiTimes reports a prototype and a roadmap, not volume shipments. Nvidia has not confirmed it officially. Between a prototype and a product sit yields, laser cost, test, and years of engineering grind. And here is the part that scares rivals: whoever gets co-packaged optics to volume first does not win a feature war. They win the whole AI infrastructure market. ### Who feels it The supply chain shifts. Who builds interposers. Who builds photonic components. Who can test the whole sandwich without drowning in scrap. Asia sits deep in that chain, and any OOI move lands on orders far beyond Taiwan. While competitors argue over benchmark numbers, Nvidia changes the subject: who can wire thousands of chips together without starving them of power. Boring? Only at first glance. This is what sets the price of training the next model. Prediction: by 2028 co-packaged optics becomes the baseline for top-end accelerators, with copper left to the mid-range. Disclaimer: supply chain leak, not officially confirmed <a href="https://www.digitimes.com/news/a20261023PD240/nvidia-cpo-interposer-technology-2026.html" rel="noopener" target="_blank">DigiTimes</a>

🔗 Read on www.digitimes.com