Photonics: the full basket, name by name

Hey {{first_name|Investor}} -

On Thursday I laid out the map: the AI buildout is hitting a wall made of copper, light is how it gets around that wall, and the supply chain splits into four jobs. Today I put real names on that map. If you missed Thursday, start there.

Read the supply chain as four jobs rather than a list of parts. Something has to make the light, which means the lasers and the materials they are grown from. Something has to load data onto that light and read it back off, which means the modulators, detectors, and the silicon that drives them. The pieces converge into a finished transceiver (the module that converts electrical signals into light and back) or a switch that goes into the rack. And every seam along the way has to be tested before anyone trusts it in a live network.

I weight that last job most heavily, and here is why. Light fails differently than electricity. An electrical signal mostly passes or fails a check. Light degrades as loss and noise across a range, so an optical part carries far more points that have to be measured, and more of them appear as speeds climb. A transceiver that fails inside a running cluster can idle thousands of GPUs. The companies that build the test instruments get paid more per wafer and per deployment as speeds rise, whichever optical architecture ends up winning. The makers compete over the design. The testers collect either way. Bet the bottlenecks, and hold the architecture itself as optionality.

Two layers carry the thesis. The component layer is where the visible demand sits, the lasers and modules the fastest AI networks are buying right now as they move from 800 gigabits to 1.6 terabits. The testing layer sits quietly above it, and it's the one I keep coming back to, because it gets paid whether the winning optical design is one flavor or another. The name I anchor it with is Viavi Solutions ($VIAV), which I've held and disclosed before.

You can already see where the money is moving. NVIDIA ($NVDA) put $4 billion across two optical suppliers, Lumentum ($LITE) and Coherent ($COHR), each deal carrying a multi-year purchase commitment and rights to future capacity. When the company building the GPUs prepays to reserve the parts that feed them, that tells you where the shortage is.

One risk matters more than which optical design wins, and it sits a layer below the whole stack, in the raw material. Almost every optical laser in an AI network is grown on a wafer called indium phosphide, or InP. Both the metal it starts from, indium, and the finished wafer are concentrated in China, which refines roughly 70% of the world's indium and, since early 2025, has required export licenses for it, releasing supply in measured batches rather than freely. Since those controls came in, the price of a standard InP wafer has climbed about 250%, to around $5,000.

This binds tighter than the architecture race, because both sides of that race run on InP. Pluggable modules use InP lasers, and co-packaged designs use InP lasers sitting just outside the package, so the material is unavoidable either way. Silicon, the cheaper substitute, cannot do this job at the short distances where most AI-cluster traffic runs. Whoever controls the substrate controls the buildout, and the pressure runs one direction: the cost lands on the companies that buy the material, and the benefit accrues to whoever owns it.

That reframes a phrase you will hear constantly: sold out. When a laser maker is sold out through 2028, it can mean two different things. The good version is that demand is strong enough to grow both price and volume. The harder version is that the company cannot get enough substrate to build more, which caps volume and leaves only price to move, with that extra price flowing upstream to whoever holds the material. Watch whether growth keeps coming from units or only from price, and which way gross margin moves. Margin compressing while a company is sold out is the tell that the rent is migrating upstream.

Two things keep this in proportion. The constraint is geopolitical, not geological. There is plenty of indium in the ground; the lever sits in Beijing, and it moves with the state of U.S.-China trade relations, so it can tighten as easily as it loosens. A fresh batch was released as recently as late May, but the same dynamic that eased it can just as quickly reverse. And the strongest names are insulating themselves. Coherent makes its own InP and is doubling its U.S. wafer capacity, and Lumentum is building its own U.S. capacity, which is a large part of why NVIDIA's capital and its rights to future capacity matter. Owning the substrate, or locking up access to it, is the real moat. Treat this as a brake on the theme and a risk to weigh on every name in it, not a wall that ends the growth.

Below is the full basket, name by name: the two I'd anchor on, the component and networking-silicon names around them, the foundries that make the optical chips, the copper-side names crossing into optics, and the speculative tier I'd keep on a watchlist rather than buy. Each comes with the reasoning, the role it plays, and the specific risk I'd watch.