Razor's Edge: Optics not Software Are The Real Shitcos..

I am increasingly convinced markets right now are dominated by a new generation card counting traders chasing kpi's in a manner without much regard for ownership.

Someone asked me yesterday why would I short AAOI after having been long it and knowing full well the tape is optics drunk, and that's because having "invested" in optics businesses I generally have more a of a sustained shitco economics view on the industry.

I did totally get they would be the last mania in dc capex names and that's fine and dandy, but I do think people should remember this space has a 90%+ probability of blowing itself up. And I am amazed how they are being compared to memory stocks right now considering capacity expansion is a whole different world.

Adding step-function capacity in memory is HARD:

• The manufacturing process runs on bleeding-edge semiconductor fabs.
• HBM requires advanced DRAM dies fabricated at sub-20nm nodes, then stacked using through-silicon vias (TSVs) and hybrid bonding — processes that only three companies in the world can do (Samsung, SK Hynix, Micron).
• Building a new DRAM fab costs $15-20B and takes 2-3 years from groundbreaking to volume production.
• The equipment is specialized (EUV lithography, advanced etch, deposition), lead times on tool orders from ASML and Applied Materials are 18+ months.
• Yield learning on a new node takes another year after tools are installed.
• The physics of shrinking DRAM cells gets harder every generation. And the TSV/bonding steps for HBM add another layer of yield complexity on top of the base DRAM process.
• The entire industry can maybe add 15-20% capacity per year, and even that requires tens of billions in coordinated capex across three players.

Adding Optics capacity is A LOT EASIER:

(opus 4.6 got this)

The assembly step is the volume bottleneck, and assembly scales linearly. A transceiver is mostly a packaging and integration exercise . You're putting a laser, a DSP, a PCB, passive components, and fiber coupling into a metal housing. Adding assembly lines is similar to adding manufacturing lines at a Fabrinet or Foxconn. Buy more pick-and-place machines, hire more technicians, lease more cleanroom space.

The key components are mature semiconductor processes. The DSP comes from TSMC on 5nm or 3nm — these are standard digital CMOS wafers on existing process nodes, not new node development. EML and VCSEL lasers are InP and GaAs compound semiconductors manufactured on 2-to-6 inch wafers using processes that have been refined for 20+ years. The lithography is I-line or DUV stepper — nothing exotic. The epitaxy tools (MBE, MOCVD) are commercially available and have lead times of 6-12 months, not 18-24 months like EUV scanners. Yes, laser yield is tricky, but the process technology itself is mature.

No equivalent of the "only three companies can do this" constraint. In memory, Samsung/SK Hynix/Micron is a hard oligopoly. In optics, there are 15-20 companies worldwide that can assemble transceivers, 5-8 that can fabricate lasers, and 3-4 foundries (TSMC, GlobalFoundries, Tower Semi) that can do the SiPh processing. The supplier base is much more fragmented, which means capacity can come from many directions simultaneously. When InnoLight, Eoptolink, Coherent, Lumentum, Jabil, Fabrinet, Source Photonics, Hisense, Cambridge, and Accelink all add capacity at once — which they are — the aggregate supply increase is massive.

Silicon photonics shifts the bottleneck to standard CMOS fabs. As the industry moves from EML-based transceivers to SiPh, the manufacturing shifts from exotic InP wafer fabs to standard silicon foundries. Tower Semi is spending $650M to scale SiPh on 8-inch and 12-inch wafers. GlobalFoundries has a SiPh platform. These are fabs that already exist and are being repurposed or expanded, not built from scratch. The CW laser that feeds SiPh is a DFB on InP — the simplest, most mature laser technology in the stack. This is precisely why SiPh share is going from 33% to 50%+ by 2026 — it's the path of least manufacturing resistance.

The capex requirements are orders of magnitude smaller. A new HBM-capable DRAM fab is $15-20B. AAOI's entire capacity expansion — which they claim will make them the largest U.S. 800G/1.6T producer — costs $400-600M. Coherent's Vietnam facility is $127M. Even Nvidia's $4B combined investment in Coherent and Lumentum is small relative to what Samsung spends in a single year on memory fabs. Lower capex means more players can fund expansion, and shorter payback periods mean the incentive to build is stronger.

Time to production is 12-18 months, not 3 years. From ordering equipment to producing transceivers at volume, the cycle is roughly 12-18 months for assembly capacity and maybe 18-24 months for a new laser fab line. For memory, it's 2-3 years minimum from fab construction start to meaningful volume output. The optical supply response is faster, which means any demand-driven pricing premium gets competed away more quickly.

Now back to me..

Why is AAOI my favorite short in this sector now?

Past criticism of AAOI was all fair and obvious, but shit risk reward analysis and not technically nuanced. A short-thesis in early 2025 focused largely on their failures and unlikely ability to fullfill 800g orders. Ok, that has largely beenknown but priced for near death stocks with massive sector tailwinds forming are just horrible setups.

Also, name dropping works very well in capex mania cycles as all anyone cares about first is whether you gonna get orders. Takes a little time to focus on how much owner earnings will come from that. Optics has a major problem on that front and this will eventually come into focus. I have yet to have one person say anything about AAOI other then the rev potential thrown out by the CEO if they 5x capacity.

I am amazed at the lack of common sense on all of this.

Remember this company is effectively telling you that for $200ml in capex they have built assembly capacity to take massive hyperscale share and that they will 5x this capacity in under 18 months for a yet undisclosed capex number.

So, without EML laser IP that relevant, and without leading edge fab capacity or in house sip ip, you gonna create 20bl in mkt cap for a couple hundred million?

You would think the first question for the CEO would be if you can 5X capacity for a few hundred million why do you think pricing won't implode by 2028? The entire bull case depends on the supply-demand imbalance persisting essentially indefinitely. In memory, structural constraints (oligopoly, extreme capex, physics constrains, super long build cycles..etc) mean supply gaps can persist for years. In optics, the barriers to adding capacity are lower on every dimension. Cost, time, technical complexity, and number of capable players point to obvious industry economics moat problems. AAOI is suppsoedly racing to build 500K units/month of capacity that will arrive into a market where the supply response is already structurally faster than its ever been in the history of this industry.

And you can assume that they are looking at 1bl+ capex so this company can only recoup this investment with any decent irr if you are short capacity till the 2030's otherwise ur looking at maybe 6 months post full buildout before margin pressure becomes a big deal. You could also argue there is some cost pressure here cause of this environment so their capex could run as high as 1.6bl which means you need clean sailing till 2031 with their margin structure before you have recouped this.

So, spare me the p/e multiple on 2028 cause if your at teens today on that estimate after 10x revenue you are seriously in fucking trouble as a "investor". But again that not what anyone is doing here, as it really boils down to the narratives.

Lite, the dominant player of the space, which is the old sdl from JDSU days has generated a little over a 1bl in fcf past decade. I think this whole space will see red by 2029 as it gets frack sanded fast by capacity explosion.

Which bring me to the risk/reward on AAOI being asymettric...

They now have an adjusted EV based on dilution/capex equal to what HPE paid for Juniper. Let's assume they get to 1bl by year end this year and the capacity comes online smoothly by summer 2027. That's 2bl or so revenue next year and then 4-5bl by 2028. Silly projections to be honest, but I will tolerate them because they make my point. If they can earn this much in 2028 with this little investment, then there is no better IRR then assembly capex for transcievers which means there is no way you do those numbers or that ur margins look way different as all the folks higher up in the supply chain take your remaining margin.

So, in a functioning brain bull case this already over 20x 2028 cause no way the margins are not 10% or less by then.

But what if a company with no history of delivering continues to fail to execute?

Well, then u have total collapse. That's a great bet.