Optics, round 2: the 3 names I'm working through

Hi {{first_name|Investor}}

The optical stocks that led the market's rally earlier this year spent the summer giving most of it back. As of September 3, 13 of them sat about 42% below their 52-week highs on average. The S&P 500 was under 1% off its own high.

Across the same stretch, the businesses kept growing. All 11 that reported in August grew revenue more than 20% year over year, and Corning ($GLW) was the only one to guide the next quarter below consensus.

The technology story underneath is one I keep coming back to: the point where an electrical signal turns into light is walking from the front panel of a switch toward the chip. This week is about what the reset did to that map, and which names I'm working through as a result.

One line before any of it: my positions in this chain are Lumentum ($LITE) and Marvell ($MRVL). Every other name below is research in progress, with the condition that would move it and the signal that would end it.

The frame that has held up for me all year is that the AI rack has one budget, and the marginal dollar inside it is moving from memory density toward data movement.

NVIDIA is reviewing Rubin Ultra's memory downward, from 12 memory dies per HBM stack toward 8, per TrendForce and The Information. NVIDIA says the roadmap is intact and the final spec is unset. What could make less local memory workable is a wider scale-up domain (NVIDIA's planned NVL576 binds 576 GPUs across 8 racks into one NVLink fabric, with optical links between racks, a roadmap design rather than a deployed one) and memory tiers outside the GPU package. My inference from that: less memory per chip, more spent on the links between chips.

A Nature Electronics review published August 19 puts the physics behind it: compute throughput has roughly tripled every 2 years while interconnect bandwidth advanced about 1.4x. One of the corresponding authors works at SK hynix, a memory company, which tells you where the industry thinks the constraint has moved.

I still hold the memory trade. My read is that the next dollar of system spending lands in switching and optics, and the supply chain that would catch it is the one that just got repriced.

The selloff had a story attached, and the story was that optical dollars and margins were shrinking together. It was built from 3 headlines.

NVIDIA's co-packaged switch uses a quarter as many lasers. Fabrinet's data center line came in flat, and an analyst on its call noted its NVIDIA revenue ran down 20% or more for the year. Credo's GAAP gross margin slid to 64.5% from 68.2%.

Those come from 3 different layers: light sources, assembly, and connectivity chips. Read separately, each one is dollars moving between layers.

Fewer lasers, but the one that's left has to feed light to many channels at hundreds of milliwatts, so it's a different and pricier product, and whether price offsets count is the open question. Less assembly at one vendor while NVIDIA's chain names Foxconn and SPIL for the current design. A GAAP margin move at Credo while its non-GAAP margin held at 68.0%.

The dollars moved seats inside the same building.

Two forces did the repricing itself. The 30-year Treasury has held above 5% since July, and profits that land in 2027 and 2028 get discounted harder than profits landing now. And several names funded growth with equity or heavy capex, so the market started applying capital-intensive-cycle multiples across the whole group.

That's the setup where I want a way to separate what a company has actually confirmed from what the market has already paid for. I keep a card with 5 questions on one side and one number on the other.

Side one: what's confirmed.

Each question gets a yes, a partial, or a no. Then I weight the whole card by how much of the company's revenue is actually optics. A perfect card on 15% of revenue moves a stock a lot less than a decent card on 100%.

Side two: what's already paid.

One number: forward earnings multiple, put on a common 2027 basis so companies with January and June fiscal years compare cleanly.

The card does one job: where side one is long and side two is short, the name deserves the work. Where side two is long, the price already assumes the execution, and the question becomes whether it arrives.

A worked example, on a name I hold. Lumentum fills side one almost completely: it's in the first seat on external light sources, on optical circuit switches and on substrate security (AXT capacity reserved through 2031, with two $43.5 million deposits). NVIDIA named it as laser supplier in its August production update. Its September quarter is guided to $1.225 to $1.275 billion at a 39.5% to 40.5% non-GAAP operating margin. Card: yes, yes, partial (the deposits run from Lumentum to its supplier, and I haven't seen customer prepayments disclosed the other way), yes, yes. Side two is also long: the stock trades at a multiple in the low 30s on 2027 consensus. So Lumentum is a name where I already own the execution, and what I check each quarter is whether it lands.

Coherent ($COHR) is the reverse shape, and it's the first of the 3 names below. It holds every layer from the indium phosphide wafer to the switch, and in fiscal 2026 it generated $79.5 million of operating cash flow against $1.1 billion of capex, with the gap covered by about $2.0 billion of new shares. That gap is the whole question.

The other 2 names: a connectivity chip maker whose GAAP margin drop set off the selloff while its adjusted margin didn't move, and the only pure-optics company in the group trading under 5 times sales, sitting in the exact layer the technology is walking away from.

The rest of this issue is for Insiders.

Below the line: the full card on Coherent, and on the 2 names I haven't named, each with the one line I check every quarter and the signal that would end my interest. Plus where I stand on all 4, what would tell me I'm wrong, and what I'm tracking into the year.