Why Optical Scale-Up Must Be Multiplexed

Previously in this series, we showed why copper has run out of road and why optics is no longer optional. The question now is not whether to go optical, but how to scale it. Our CEO Matt Crowley made the case in PIC Magazine: why optical scale-up has to be multiplexed, and why the answer is more wavelengths, not faster ones.

Once you accept that scale-up networks are going optical, the next question is how to scale the bandwidth. There are only two ways to move more data over a fiber. Go faster, by pushing symbol rates higher and packing more bits into each symbol. Or go wider, by running more wavelengths of light down the same fiber at once.

The industry has aligned on wider, and the reason is energy. Every step up in symbol rate and modulation complexity costs power, adds latency, and generates heat. Moving from simple NRZ encoding to denser PAM-4 packs more bits per symbol, but it pays for that in optical power, signal processing, and error correction. Across a coherent GPU cluster, those costs compound on every link. A single wavelength per fiber does not scale in scale-up.

Wavelength multiplexing takes the other road. Run each channel slow enough, around 50 gigabits per second, that simple NRZ encoding holds the error rate low, forward error correction stays shallow, and latency stays low and predictable. Then scale bandwidth by multiplication instead of escalation. Eight wavelengths at 50 gigabits per second deliver 400 gigabits per fiber in each direction. Sixteen deliver 800. The per-channel electronics never change. Only the number of wavelengths does.

This is also why adding fibers is a trap. More fibers means more attachments, more alignment, and more failure points, and the fiber count climbs fastest exactly where space is tightest. Multiplexing at the laser source raises bandwidth per fiber without that explosion. You scale capacity, not connectors.

Picture it as a staircase. The first-generation specification starts at four wavelengths, enough to prove the architecture in production silicon and align the supply chain. That is the floor, not the ceiling. Eight wavelengths double it. Sixteen double it again. The same fiber plant that carried four carries sixteen, with no change to the underlying link. Each step multiplies capacity on infrastructure already in place.

The payoff is not only throughput. Larger, flatter, low-latency scale-up domains enlarge a cluster's working memory, extend its context window, and leave room for deeper reasoning. The wavelength count an architect commits to today sets the ceiling on which models that cluster can run in 2028.

There is a deeper pattern here, and the semiconductor industry has lived it twice. Dennard scaling gave chips a free ride on frequency for thirty years. Moore's Law gave them a free ride on transistor cost. Each was a compounding curve, and each defined an era. Photonics has never had one. Every wavelength was its own discrete part with its own assembly cost, so the price of bandwidth never bent. Heterogeneous integration changes that. Bond the laser directly onto the silicon photonics wafer, and adding a wavelength becomes a design choice rather than another part to build and align by hand. The next wavelength rides the wafer's cost curve. For the first time, photonics gets a compounding curve of its own: more bandwidth per fiber, generation over generation, on the same infrastructure.

But a curve only exists if the manufacturing actually delivers it at volume, and not every approach does.

Read Matt Crowley's full piece in PIC Magazine. It walks through the two earlier times when the semiconductor industry watched a scaling curve break and return as something new, and lays out the full case for why wavelength scaling is photonics' turn: "Why Optical Scale-Up Must Be Multiplexed."

---

THE WAVELENGTH SCALING SERIES 

Part 1: AI Performance Now Depends on Optics, and CPO Is the Front Line - Why the network, not the processor, now decides how far an AI system can scale.

Part 2: 'Optics When You Must' Arrives for Data Centers - The copper physics that makes the move to optics no longer optional.

Part 3: Why Optical Scale-Up Must Be Multiplexed - Why bandwidth scales through more wavelengths, not faster channels.

Part 4: Optical Scale Up Fabrics Are Limited By Manufacturing, Not Architecture - With the architecture settled, why manufacturing now decides who keeps scaling and who stalls.