Edition #2. Published every Tuesday. Last week: we talked about power. This week: the photonics layer just got tighter, and the shortage is rotating upstream.

If you watched some of the photonics stocks after earnings last week, you'd have probably concluded: “well, it was a good run.”

AAOI dropped 11.7% Thursday on an EPS miss of two cents. LITE dipped on its print. The pure-play optical trade looked like it was rolling over.

Then, on Monday open, LITE hit a new all-time high. COHR, new all-time high. AAOI up roughly 30% from Friday's close.

Every time someone tries to call a top on photonics, the market reminds them the sector is a bottleneck in AI right now.

This week, that bottleneck just got tighter, with a critical shortage forming in its own supply chain.

If you listened to the earnings calls from those three names last week, you'd have heard all of them say roughly the same thing. Demand is great, order books stretch out years and that growth is accelerating.

But there's a critical bottleneck forming in their supply chain and it's the same one across all three and a few others in the photonics sector:

Indium phosphide (InP) lasers.

InP is the semiconductor platform behind many of the lasers and photonic devices that make high-speed datacenter optics work. Silicon photonics can guide, modulate, and integrate light, but silicon is still a poor light emitter. The light often has to come from somewhere else.

In many 800G and 1.6T optical modules, that means InP-based lasers.

The 200,000-GPU clusters hyperscalers are designing for 2026–2027 are, in fiber terms, lit by InP.

In the earnings calls from last week, all three companies flagged their InP laser chip fab capacity, the device fabs where substrate wafers get processed into lasers.

InP sits at the core of nearly every AI data center product these companies ship. That's where the immediate bottleneck sits.

All three are racing to close the gap with aggressive capacity expansion:

Three reasons the gap may not close on the timeline the consensus is pricing.

1. CPO demand grows faster than InP fab capacity can: Goldman Sachs now forecasts the optical networking TAM expanding from ~$15B in 2026 to ~$154B by 2028, ~10x jump in two years, with CPO contributing roughly $91B of the $154B.

Silicon photonics penetration climbs from 6% in Q1 2024 to 46% by Q4 2028. Every CPO module still needs an external InP laser source. A single Rubin Ultra NVL576 super-node uses 162 of them.

2. The bottleneck rotates upstream: Even if LITE, COHR, AAOI and others hit every capex number, they all draw their substrate wafers from the same three companies:

Substrate capacity is expanding, but materially slower than device-fab demand. AAOI's CFO said on last week's call they hold about a year of substrate inventory and are "making calls with all the suppliers."

AXT made the gap explicit on its own call. Customers are asking CEO Morris Young why he can't triple or quadruple capacity. His answer: "Our need is 10x. It's not easy." They just raised $632.5M to double InP capacity by end of 2026 and double again by end of 2027. Even that pace lags what customers are asking for.

The HBM playbook is showing up one layer upstream too. Hyperscalers, the customer's customer, are now reaching past LITE/COHR/AAOI to engage AXT directly on long-term supply agreements. Today's device-fab constraint becomes 2027's substrate constraint. The shortage moves up the stack.

3. InP fab buildout is an 18-to-24-month minimum: Even with aggressive capex, you can't accelerate the critical infrastructure buildout for a new fab like reactor installation, cleanroom qualification, or yield ramp.

Lumentum's Greensboro acquisition is the shortcut, i.e. repurpose an existing GaAs fab rather than go greenfield and even that requires multi-quarter process conversion. Coherent's 6-inch ramp, the actual structural unlock, started in early 2024 and took years of engineering before yields reportedly exceeded the 3-inch baseline. The fab-side response to today's order book lands in 2027 or later, not 2026.

Pays rent: NVIDIA, hyperscalers, neoclouds buying at scarcity pricing. NVIDIA's $4B March stake in LITE and COHR was the buyer financing the bottleneck, similar to the the HBM playbook.

Captures rent: The InP-native names, at the layer where they sit.

Watch three things: COHR's 6-inch yield ramp, ELSFP volumes through 2026, and substrate capacity adds at Sumitomo, JX, and AXT. Until the 6-inch lines scale and substrate catches up, the gap doesn't close.

Here’s how the constraints moved in the other layers:

Korean DRAM and NAND export prices kept ripping higher as Samsung and SK Hynix continued diverting wafer capacity to HBM. Q1 contract prices were already up 60–70% QoQ, and partial normalization isn't expected before 2027.

You can read the capacity shift directly in the customs data. The HBM-to-DRAM premium has collapsed from 18x in June 2025 to 2.6x by end of April 2026. This is mainly let by commodity DRAM catching up to it rather than HBM cooling off.

Then layer a union strike threat on top of that. Samsung's largest chip union has voted for an 18-day walkout starting May 21 if pay talks stay deadlocked. JPMorgan puts the hit at up to 12% of quarterly profit. DRAM spot may already be pricing it in. Nine days out. Every session without a deal tightens spot further.

A union vote at a single fab is now a supply chain event for the entire AI stack. That's how thin the memory layer has gotten.

A few important earnings this week. Read our pre-call briefs here and stay prepared.

We track Indium Phosphide Substrate, and every other cell in the constraint map, daily. When the next big move happens, we'll know first.

Want the live data behind The Chokepoint? Tessara is the AI infrastructure operating system for investors.

We track what's binding in the supply chain and what it means for what you own. 270+ companies across compute, memory, foundry, networking, and power. We tracked InP from substrate to module before three earnings calls confirmed it this week.

Apply for early access →

See you next week,

Teng

This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.