To The Moon

hey everyone, Wonton Investor here. let's talk about the part of the AI trade that everyone's excited about right now, which is optics and power.

over the past few months i've been buying into the "plumbing" of AI data centres - optical modules (AAOI, LITE), connectivity chips (MXL), timing chips (SITM) and storage (STX, WDC). the reason is pretty simple. once a data centre has enough GPUs, the limit stops being how smart each chip is and starts being how fast the chips can talk to each other. copper wire runs out of distance and bandwidth, so you have to go optical, and the companies making that "nervous system" become the interesting ones.

and this earnings season, every single one of my six names beat and raised.

so you'd think i'd be super bullish. and i am on the demand. but i've actually changed my mind on something big, and i think a lot of people are missing it.

i think optics and power are going to commoditise

not this quarter. AAOI is still capacity-constrained into mid 2027, so pricing should be fine for a while. but look at what these companies are doing with all that money. AAOI is taking capacity from 200K to 650K units a month by the end of the year. one of the big optics players moved to bigger 6-inch indium phosphide wafers to ease its supply. everyone is building at the same time, and when everyone builds at once, the thing they're building stops being scarce.

we've literally watched this happen twice before.

the first time was solar. demand exploded for fifteen years, panels went everywhere, and the panel makers went bankrupt anyway because capacity flooded in and prices collapsed. the second time was optics itself in 2000. the internet really did need all that fibre, and it got laid and used, but the optical component makers still lost almost all their value because everyone built capacity at once.

there's also a new pressure this time called co-packaged optics. right now a transceiver is a separate box you plug into a switch. co-packaged optics moves the optics onto the chip package, so the transceiver stops being a product and becomes a line item for whoever owns the package. and energy has the same shape. power is short right now, but solar is the cheapest electricity ever and batteries are almost 90% cheaper than in 2010, so "we can deliver power fastest" is a race lots of people can join.

the thing that convinced me the market is already sensing this is that it's stopped paying for good quarters. Credo printed record revenue and record net income and fell about 18%. WDC beat and raised and fell 15%. MaxLinear grew revenue 55% and is still around half its June high. when growing revenue can't hold a stock up, the market is already pricing in peak margins.

i went through a whole list of evidence for and against this (memory contracts, geopolitics, power queues and a few others) in the full version on my substack, thewontoninvestor.substack.com - it's a lot longer than what fits here.

but it doesn't hit everyone at once

here's where i landed after going through all of it. commoditisation runs from the outside in. the assembly layer goes first: the transceiver module, the power box, anything you can build with money and a factory. the IP layer goes last, or never: the stuff that took decades to learn how to make.

you can already see it in the margins. Lumentum's gross margin crossed 50%, up over 12 points in a year, and i think that's because the scarce thing in optics is the laser, which is really hard to make, and not the module, which is much easier to assemble. SiTime runs a 67% gross margin and its operating margin went from 10% to 34% in a year. nobody is racing to build MEMS timing fabs (the factories for these tiny timing chips), so it's basically a toll booth.

so the useful question isn't "is photonics a commodity?". it's "which part of photonics, and when?"

so what am i doing about it?

i still own some of the optics names, because the demand is real and the ramp isn't done. but i now treat the component layer as the part that eventually gets competed away. the play i'm more interested in is finding the gaps these technologies can fill outside the AI data centre, and owning whoever holds the IP that stays scarce after the product doesn't.

the cleanest example i've found is Corning, a 175-year-old glass company (yes really lol). what they own is basically a recipe book for glass, and that recipe book shows up in fibre for co-packaged optics, in glass substrates for the next generation of chip packaging, and in non-AI demand like a multi-billion dollar fibre deal with Verizon. i don't own it yet, and i've written up a separate full post on it with exactly what would make me buy and what would make me walk away.

should YOU go buy optics stocks? i'm not a financial advisor, so i can't tell you that. but if you're looking at this space, i'd at least ask which layer you're buying, and whether that layer will still be scarce once everyone's new factories are running.

the full version on my substack also has the part that actually worries me most - how this whole AI buildout is being financed (roughly $4T of the ~$5.5T of planned capex is debt) and the AI fund that got margin-called while being right about everything. that's at thewontoninvestor.substack.com if you want the long read.

peace out,

The Wonton Investor

if you enjoyed this, follow me here for more, and on X at @thewontoninvest for more real-time thoughts.

disclosure: i hold AAOI and LITE. not financial advice - just my own reasoning about my own money.