Optical Scale Up Fabrics Are Limited By Manufacturing, Not Architecture

Over the past three weeks, we traced one argument, from the network, to the physics, to the wavelength count. Our CEO Matt Crowley brings it home in The Next Platform: the architecture for optical scale-up is now settled, and what happens next is a manufacturing question.

Earlier this spring, AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI formed the Optical Compute Interconnect Multi-Source Agreement to set a common optical specification for scale-up networks. The architecture they aligned on is NRZ modulation with wavelength multiplexing, slow and wide, starting at four wavelengths per fiber. When six of the largest names in AI infrastructure agree on the same design, the architectural debate is over.

But settling the architecture does not settle how the industry climbs from four wavelengths to sixteen and beyond on the same fiber plant. The roadmap demands successive wavelength generations. It does not say which manufacturing approach delivers them. That question is not architectural. It is industrial.

It also hands system architects a new line on the evaluation sheet, alongside energy per bit, latency, and reach: wavelength-scaling headroom on the same manufacturing flow. Can a supplier's light source extend to eight, sixteen, and beyond without re-engineering? If the answer is no, the redesign is already on the calendar, and the two-year delta is baked in.

The hard part is volume. Discrete-laser supply chains were never built for hyperscale. Combining many lasers through a splitter network erodes the optical budget with every additional channel. Give each wavelength its own laser, and the assembly explodes: a single sixteen-wavelength source across eight fibers means roughly 128 lasers and 128 alignments, each holding micrometer tolerance across temperature and years of field life. Hyperscale CPO will need millions of these units a month, not thousands. That, not the architecture, is what has held DWDM back.

The way through is to stop treating the next wavelength as a part to assemble and start treating it as a circuit element on the wafer. Photonic integration has moved through three eras. First, discrete assembly, every wavelength aligned by hand. Then silicon photonics, which moved modulators and detectors onto the wafer but left the laser outside the foundry flow. Now heterogeneous integration, which bonds III-V gain material onto the silicon photonics wafer so the laser joins the same single foundry-aligned process. The next wavelength inherits the wafer's cost curve instead of carrying its own.

This is the manufacturing pattern CMOS established for electronics: one industrial process that absorbed new device types and compounded over generations. The free ride was never the transistor. It was the pattern. Heterogeneous integration is the CMOS of photonics. The claim is pattern equivalence, not scale equivalence.

That pattern is already running. SHIP™, Scintil Heterogeneous Integrated Photonics, sits on Tower Semiconductor's 200 millimeter silicon photonics lines today, the same lines that ship tens of millions of pluggable transceivers, with the 300 millimeter path next. It already reaches past the laser source. At OFC 2026, three different system vendors independently requested SHIP™ extensions across four device categories, none of which solved the same problem. The common factor was the pattern, not the device. LEAF Light™, our single-chip DWDM-native light engine, is the production proof, demonstrated in eight- and sixteen-wavelength configurations, with NVIDIA among the investors in our Series B.

The architecture is settled. The path to ship it runs through heterogeneous integration. The teams that put the manufacturing question on the evaluation sheet now will hold the position when this becomes the default. The teams that defer it will spend two generations rebuilding around the suppliers who did not.

Read Matt Crowley's full argument in The Next Platform, the capstone of this series. It lays out the full manufacturing teardown, why discrete-laser approaches hit a wall, and what the OFC customer pattern reveals about where photonic integration goes next: "Optical Scale Up Fabrics Are Limited By Manufacturing, Not Architecture"

---

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.