What ECOC Showed Us About Where Density Comes From Now
I spent ECOC week in Málaga with partners, customers, analysts and investors, and almost every conversation started from the same premise. For inference, radix is the metric. Mixture-of-experts and agentic workloads need every node to reach every other node, and the number of endpoints a node can reach directly sets how many switching layers sit in between. Fewer layers means lower latency, lower power and a lower cost per inference. Few people in Málaga were still arguing that. The argument now is about where the density comes from.
The faceplate has run out of room. A switch in this generation brings thousands of fibers out of one box, and pluggables at the front panel cannot carry that density. So density is moving toward the package, and the contest has moved with it. At ECOC the battleground was near-package optics inside the box. Silicon photonics, wide-and-parallel VCSEL designs and newer light sources are all competing there, and each has a real case. It is the pluggable-era contest between VCSEL, EML and silicon photonics, moved a few inches inward.
From left to right: Rebecca Schaevitz, CPO | Vivek Raghuraman, CEO | Arrezo Azim, Head of People Operations
What decides it this time? At the faceplate, the question was how easily a module could be serviced. Inside the box, the question is whether it keeps working. Reliability is where near-package approaches will win or lose.
There is also a gate every one of these approaches has to clear, and it is electrical. Each still has to work with the SerDes the hyperscaler already runs. An optical technology that wants to win at hyperscale has to win with the SerDes, not in spite of it. I believe that is the next barrier to entry, and I will unpack it in a future piece.
Consider where the industry is in that move. Co-packaged optics is arriving in steps, through near-package optics, CPX and similar socketed approaches, and the step being decided now is optics inside the box, next to the package. Each step needs more fiber out of the box than the last, at a density the faceplate was never asked to carry, and each one brings the fiber closer to the silicon, where the reliability of that connection matters more.
Whichever engine wins the box, the fiber still has to get out of it.
The idea that drew the strongest response in Málaga, particularly from IC architects, was treating the rack as the endpoint. A chip has millions of electrical routes, and its interposer is where separate dies are integrated into one system. Design the rack the same way, and each server tray becomes the chiplet. What that rack needs is a backplane carrying optical routes at a comparable scale. Much of the industry is working on a photonic interposer for the package. For the rack, the backplane is the photonic interposer.
This is where Mixx fits, and why our first product to market is a connector. The SxC™ connector carries the density comparable to silicon scale interposer for high bandwidth. At today’s lane rates, that puts multi-petabit connectivity within reach of a single box, while most of the industry still plans in terabits. Few systems need that yet. The headroom is the point.
The bottom line: more fibers at the rack means higher radix, higher radix removes switching layers, and every layer removed takes its switches, transceivers and cables with it. What is left is lower latency, lower power and a lower cost for every inference the cluster serves. The architectural decisions for in-the-box optics are being made this year. The measure of every approach we saw in Málaga will be the same: whether the people who have to deploy it can trust it at scale.
Density is solved. What’s next? Fiber termination, the industry’s next bottleneck. Scaling optics isn’t just a silicon problem. Once photonics reaches semiconductor scale, fiber integration must reach OSAT scale.
