I’ve been spending a lot of time going through optical earnings calls and there is one thread that keeps showing up in different places.
It shows up in Ciena’s language around distributed AI training.
It shows up in Nokia’s AI and cloud orders.
It shows up in Lumentum’s comments around pump lasers and narrow-linewidth components.
It shows up in Coherent’s indium phosphide expansion plans.
It shows up in Corning’s long-term hyperscaler fiber agreements.
And it shows up in Lumen’s dark fiber contracts.
At first, these might all look like separate stories.
One company is talking about optical systems. Another is talking about lasers. Another is talking about fiber. Another is talking about physical routes. Another is talking about coherent DSPs.
But they are pointing toward the same architectural shift.
That is the layer I want to walk through in this report.
This is one of my favorite ways to have exposure to the optical supercycle and it is often overlooked.
But the companies at the forefront are pounding the table, and it’s time to listen.
In this report I will breakdown what this layer is, all the signal we are hearing, why this is coming into play, the significance of this market, and 12 companies that have exposure.
Disclaimer
This report is for educational and informational purposes only. It reflects my personal research process and interpretation of publicly available information. It is not financial advice, investment advice, or a recommendation to buy, sell, or hold any security. I may own positions in some of the companies discussed. Do your own research and make decisions based on your own risk tolerance, time horizon, and financial situation.
The topic is scale-across.
I have written about this in a few different ways before, but I want to create a comprehensive report around it.
The idea is really straightforward.
AI clusters are getting too large to keep in one place.
At a certain point, the constraint becomes physical. Power. Land. Cooling. Permitting. Fiber access. Route availability. The ability to get enough infrastructure in the same location at the same time. So the architecture starts to spread out.
Instead of building one giant AI data center and forcing everything into one site, hyperscalers start connecting multiple buildings, campuses, or facilities together so they can operate like one larger training environment.
Scale-across exposes a bottleneck. Once AI compute stretches across distance, the limiting layer becomes the optical infrastructure between those sites like coherent optics, fiber, pump lasers, narrow-linewidth lasers, optical amplification, dark fiber routes, and the indium phosphide capacity behind many of the most important components.
This is an architecture shift.
The Different Layers
I think the easiest way to understand this is to separate AI networking into three layers.
Scale-up.
Scale-out.
Scale-across.
They are very different from an investment perspective.
Scale-up: inside the rack
This is where GPUs sit close together and need to communicate as if they are one larger processor.
Distances here are short. We are talking centimeters to a few meters.
The main constraint is moving data across the rack with extremely high bandwidth and low latency.
Scale-out: inside the data center
Scale-out is the data center hall.
This is the network connecting thousands of GPU racks inside one large building. Distances can run up to roughly two kilometers. This is where 800G and 1.6T transceivers are ramping now.
The main optical technology here is PAM4.
PAM4 stands for pulse-amplitude modulation with four levels. Instead of sending light as only on or off, it uses four intensity levels. That lets the system carry more data through the same channel.
PAM4 works well inside the building. It is cost-effective, efficient, and already deployed at scale.
This is where a lot of the current optical trade sits.
Scale-across: between buildings and sites
Scale-across starts when the cluster stretches beyond one building.
At that point, PAM4 starts to run into reach limits, especially at higher speeds. At 1.6T speeds, common FR-class PAM4 links sit around the two-kilometer range. Longer direct-detect variants can extend farther, but coherent and coherent-lite become more practical as links move into campus and DCI distances.
DCI means data center interconnect. It is the optical networking used to connect one data center to another.
That is where coherent optics enters, and that is where this report really starts.
Why Distance Changes the Optical Stack
Inside the data center, the job is to move a huge amount of data over relatively short distances.
Between data centers, the problem changes.
The data still has to move quickly, but now the signal has to survive distance. It has to move across fiber routes and handle dispersion, noise, loss, amplification, and reliability requirements that become much more important as the link gets longer.
This is where coherent optics comes in.
PAM4 is mostly about reading the intensity of light, while Coherent optics reads more information from the light wave.
A coherent receiver mixes the incoming signal with a separate internal laser. That allows the system to recover information about both the intensity and the phase (timing) of the light wave.
Once you use both intensity and phase, you can carry much more information through the same optical channel. You also get more advanced digital signal processing (DSP), which helps correct for the distortion that builds as the signal travels over distance.
That is why coherent optics becomes important when AI data centers stretch across campuses or metro areas.
At 400G, coherent optics can move data around 80 kilometers without amplification. With amplification, the reach can extend much farther. The tradeoff is complexity.
Coherent modules require narrow-linewidth lasers, stronger DSPs, and more advanced optical assemblies. They cost more and use more power than basic short-reach optics.
But if the job is connecting AI buildings across 10, 50, or 100 kilometers, coherent becomes the practical default.
And within this, there is a progression.
400ZR is the base. It carries 400 gigabits per second and has been shipping in volume for years.
800ZR and 800ZR+ are the current ramp. These move 800 gigabits per second and are becoming more important for cloud and DCI networks.
Ciena specifically said three hyperscalers are using its optical solutions for distributed AI training, and it pointed to its RLS platform plus 800ZR pluggables as the products addressing that demand.
Then there is 1600ZR, 1600ZR+, and 1600CL.
These target 1.6 terabits per second. OIF work on 1600ZR, 1600ZR+, and 1600CL advanced through 2025 and 2026, with products now in development and interoperability work still progressing.
One part I find really interesting is coherent-lite.
Coherent-lite is built for the middle ground. It takes coherent technology and strips away some of the long-haul features that are unnecessary for a campus link.
The target is roughly 2 to 20 kilometers at around 30 watts, without external amplification.
That is kind of the sweet spot distance that starts to matter when AI compute spreads across multiple buildings or campuses.
PAM4 gets stretched, full coherent ZR can bring more reach and overhead than the link requires, and coherent-lite sits in the middle.
Multi-Rail Makes the Demand Bigger
There is another layer that I think is still underappreciated in multi-rail networking.
A rail is a complete, independent network fabric attached to the GPU cluster.
In NVIDIA’s reference architecture for its HGX AI factory, each GPU connects into two independent network planes. Depending on topology, those planes can improve redundancy, aggregate bandwidth, or both.
Each rail is its own optical fabric. That means another set of switches, transceivers, and set of fiber runs. And once distance increases, another set of amplification requirements.
This is why multi-rail can multiply optical demand.
Lumentum said on the May 2026 earnings call that the multi-rail opportunity is “huge,” and that it could be even bigger than what people are modeling.
Michael Hurlston (CEO) added that Lumentum is probably more constrained in scale-across components than even EMLs, especially pump lasers and narrow-linewidth lasers.
That is a major comment!
EMLs are the laser chips everyone already talks about for 800G and 1.6T transceivers. But Lumentum is saying the scale-across component layer may be even tighter.
Nokia also launched a multi-rail in-line amplifier at OFC 2026 that it says delivers an 8x increase in density without expanding physical infrastructure.
That is the kind of product you launch when customers are trying to push more fiber capacity through the same physical footprint.
So multi-rail is part of the same thesis.
More distributed compute creates more optical links. More rails multiply the number of links Distance adds amplification.
That brings the bottleneck back to lasers, optical components, and fiber.
The Bottleneck Underneath the Bottleneck: InP
Scale-across creates more demand for optical systems.
Those systems require lasers, coherent components, pump lasers, narrow-linewidth assemblies, gain chips, and photodetectors.
A lot of that supply chain touches indium phosphide.
Indium phosphide, or InP, is a compound semiconductor material used in many high-speed optical lasers, gain chips, coherent optical components, and certain photodetectors.
Meaningful InP epi and device capacity sits outside general-purpose CMOS foundry capacity. The manufacturing base is smaller, more specialized, and historically built around telecom demand. That is why scaling it quickly is hard.
During this reporting season, several companies pointed to the same broad constraint.
Fabrinet said shipments and revenue were below demand levels because of component shortages in lasers, memory, and ASICs. That is useful because Fabrinet is a contract manufacturer. It builds for many companies across the optical supply chain, so its commentary gives a broad read-through.
Coherent said indium phosphide has been a key constraint for multiple quarters and described it as an industry constraint.
Lumentum said the laser chip supply-demand imbalance was greater than 30%, and that customers wanted more output than the company could supply. It also said pump lasers and narrow-linewidth components for scale-across were more constrained than EMLs.
Nokia said that the scale at which the industry is building indium phosphide is driving demand back into the supply chain and requiring more capacity.
Applied Optoelectronics said industry InP laser shortages reinforce the need to accelerate internal capability expansion.
Semtech, through its newly acquired HieFo business, said gain chip demand was outpacing capacity by about 3x.
Why InP Capacity Is Hard to Add
Silicon scaling has a larger industrial base. There are more equipment suppliers, more fabs, more process knowledge, and a much larger ecosystem.
InP is more specialized. The manufacturing process often uses MOCVD, which stands for Metal Organic Chemical Vapor Deposition. This is a process used to grow compound semiconductor layers on a wafer.
The knowledge base is narrower. The wafer sizes are smaller. The supply chain has been moving from 3-inch to 4-inch to 6-inch wafers, while advanced silicon manufacturing runs on much larger 300-millimeter wafers.
Coherent gave the economic reason this transition counts. Management said moving from 3-inch to 6-inch production creates more than 4x as many devices at less than half the cost.
That is a major capacity lever. But these ramps take time. New fabs need process transfer, qualification, yield learning, equipment, and customer acceptance.
That is why the supply response is underway, while the constraint can still persist. Coherent is targeting a doubling of internal InP output by the end of calendar 2026, and another more-than-doubling by the end of calendar 2027.
Lumentum acquired a Greensboro facility and is converting it from gallium arsenide to indium phosphide, with production roughly six quarters away.
Nokia is ramping San Jose InP capacity later in 2026, though management said that capacity is more meaningful longer term.
Semtech is trying to expand HieFo capacity by 3x to 4x by the end of calendar 2026, then another 3x to 4x by the end of calendar 2027.
AAOI is expanding InP laser fabrication by roughly 350% by 2027.
The industry is responding. The question is whether it can respond fast enough.
This article is really long.
Continue reading here:
