The AI Optics Market Has 2 Clocks

Hey {{first_name|Investor}} -

Almost every conversation about optics right now runs through one question: when does co-packaged optics take over. Analysts model it, and management gets asked about it on every call.

On the model I've been reading, the answer moves the 2030 market by about $4 billion out of $70 billion.

There are 2 optical transitions running, on 2 separate clocks, and the industry keeps discussing them as though they're the same event. One is largely settled and shifts the size of the market very little. The other adds roughly 60% to it, and it hasn't started yet.

Sorting those apart changed how I read the last 2 weeks of earnings.

Scale-out is the network between racks. Rack to rack across the building, then building to building. Most of what gets called "data center networking" is this, and it's where the pluggable transceivers live.

Scale-up is the fabric inside the machine. It binds a group of accelerators into something that behaves like one very large chip, so a model too big for any single GPU can sit across all of them at once.

Think about a restaurant. Scale-out is opening more locations and coordinating them: they place orders and share a supplier.

Scale-up is putting 72 cooks around one counter, all working the same dish, handing each other ingredients in the same second. That arrangement needs a completely different amount of communication per person.

That's the number that matters here. On the estimates I've seen, bandwidth required per accelerator inside a scale-up domain runs at least 9 times higher than scale-out, with port count scaling alongside it. So the same building generates far more optical content the moment the inside of the machine goes optical.

Today it's mostly copper. Short reach, no laser, no photodiode, no signal-processing chip. Copper works inside a rack and struggles the moment you ask it to go further at speed, and every doubling of bits per lane shrinks how far it can reach.

Within scale-out, you can dial the co-packaged share anywhere you like and the 2030 market lands in roughly the same place. That range holds between $70 billion and $74 billion regardless of the assumption. Those are one analyst's figures on one set of assumptions, so hold the specific numbers loosely and keep the shape.

3 things drive that.

Pluggable survives the transition. Even in a switch built with co-packaged optics, the ports on the card next to the accelerator stay pluggable, so the 2 methods coexist by design.

And when a module does get replaced, what fills the space is an optical engine, a fiber attach, an external laser module and a shuffle box. The optics vendor sells different parts into the same port.

Port volume grows underneath all of it. Optics tracks data center buildout, and buildout is going one direction. The model I'm working from has total optical port volume growing from 44 million in 2026 to 150.4 million in 2030.

Everything else expands while the architecture mix shifts.

A delay pays. If co-packaged integration proves harder than expected and the transition slips, bandwidth demand is unchanged, so the volume sits in pluggable longer.

Pluggable happens to be the structure with the highest optics vendor revenue per port. A slip means a longer run for a product line that's already shipping every unit it can build.

Put those together and timing the scale-out transition turns out to be a low-stakes call. You can be 2 years wrong in either direction and land in the same place.

One number from NVIDIA drives most of this debate. When it showed its co-packaged switches, it said the design uses a quarter as many lasers. Investors read that as laser suppliers selling a quarter as many units.

That comparison is measured against EML, which builds the light-emitting part and the data-modulating part on one indium phosphide chip. 1 chip per signal lane. An 800G port configured as 4 lanes of 200G needs 4 of them.

Current co-packaged designs use a different laser. The indium phosphide chip only makes light, and a silicon chip sitting next to it does the modulating, which lets one laser feed several lanes.

On that basis the count halves, going from 2 dies per port down to 1.

The dollars move the other way. Moving the light source outside the package means it has to push far more power to get the same signal where it's going. On the model I've been reading, that takes the laser content per port from around $8 to around $30, and the 8-wavelength specification used in scale-up brings a second die back, taking it toward $60.

You're swapping a row of reading lamps for a single stage light. Fewer fixtures, and the one that's left is a much harder thing to build.

So the same design change that cuts laser units by 75% on the old comparison roughly doubles laser dollars per port. That gap is most of why I keep paying attention to this layer.