Every AI model you've ever used runs on a data center where tens of thousands of GPUs talk to each other constantly, at speeds copper wire physically cannot sustain. The moment data travels more than a couple of meters, electrons give way to photons - light through glass. The devices that convert electrical signals into light and back are called optical transceivers, and they're becoming one of the bottlenecks in the AI supply chain.
Applied Optoelectronics ($AAOI) makes them. More importantly, it makes the lasers inside them, the single hardest component to source, in its own fabs.
1. How a transceiver actually works
To understand why AAOI matters, understand what it sells. A data-center optical transceiver does one job: it sits at the edge of a switch or server and translates between the language of chips (electrical signals) and the language of long-distance data movement (light).
The chain: Electrical in → a switch ASIC or GPU sends a high-speed electrical signal into the module. Laser fires → an InP laser converts those electrons into precisely modulated pulses of light. Light travels → photons race down a glass fiber across the rack, hall, or building. Photodetector → at the far end, the light becomes an electrical signal again.
The critical component is the laser. Specifically an indium phosphide (InP) laser. InP is the material that emits light at the wavelengths and speeds data centers need. It's difficult to manufacture, capacity is scarce, and it's the part most prone to shortage. LightCounting has repeatedly named InP laser and EML chip capacity, not demand, as the factor limiting the entire market's growth.
Most transceiver vendors buy their lasers from third parties and assemble modules. AAOI makes its own lasers, chips, and assemblies in-house. When the scarce component in the whole industry is the laser, the company that manufactures its own controls its destiny in a way assembly-only rivals cannot.
Why "speed" is really "revenue per unit": the industry is climbing a ladder 400G → 800G → 1.6T → 3.2T. Each rung roughly doubles the data a module carries, and roughly doubles its price. AAOI's entire growth story is moving up that ladder into demand that already exists.
2. Why the growth is happening now
The demand isn't speculative. The AI optical transceiver market roughly doubled in two years, from ~$5B in 2024 to ~$10B in 2026, with ~60% forecast growth in 2026.
Where the firms disagree is on the number, not the direction:
The spread from ~$24B to a $100B bull case by 2030 is itself the signal: no serious analyst disputes the market is growing fast; they only disagree on how insanely fast.
The demand mechanism that makes this cycle different is the GPUs themselves. Every new GPU generation roughly doubles the optical bandwidth each chip needs. The scale-out network bandwidth per GPU has gone from 400G on Hopper (H100) → 800G on Blackwell (GB200/GB300) → 1.6T per GPU on the coming Vera Rubin NVL72 (per SemiAnalysis). More bandwidth per GPU means more and faster, higher-priced transceivers per GPU.
This isn't abstract. SemiAnalysis models an 18,000-GPU GB300 cluster (two-layer InfiniBand) as requiring 18,432 800G transceivers plus 27,648 1.6T transceivers. Roughly 2.5 transceivers per GPU just for the scale-out network, and that ratio climbs to 1:3 in three-layer configurations. And it compounds two ways: as clusters grow, you add both more GPUs and more (faster) links per GPU. The newer the GPU, the more optical content it drags along and the roadmap only points up (NVIDIA's Rubin Ultra "Kyber" rack scales to 144 GPU packages, 4× denser than today's NVL72).
There's a second, larger frontier behind it. Connectivity inside the rack ("scale-up") currently runs on copper but copper is hitting a hard physical wall (~9 meters is roughly the ceiling at 800G). As scale-up domains grow from tens to hundreds of GPUs, SemiAnalysis calls the shift to optics "a physical inevitability" opening an entirely new, and even larger, optical market on top of scale-out.
The 1.6T bars are small today and large tomorrow. NVIDIA alone is estimated to need 5M+ 1.6T units in 2026; Google ~4M. AAOI is building capacity for it.
3. Where AAOI fits in the CPO shift
The next architectural step in optics is co-packaged optics (CPO). Moving the optical engine off the front panel and onto the switch chip package itself. It's the industry's answer to the power and density limits of pluggable modules at the highest speeds, and NVIDIA, Broadcom, and other hyperscalers are all building toward it. The occasional bear take is that CPO makes transceiver vendors obsolete. For AAOI, the opposite is true, it's positioned to sell more content into a CPO switch, not less.
CPO doesn't eliminate the optics, It relocates and disaggregates them, and AAOI builds the two pieces that matter most. At OFC 2026 the company demonstrated both halves of the optical side of a CPO system: its 6.4T On-Board Optics (OBO) optical engine, which sits right next to the switch ASIC, and its 25dBm external laser source (ELSFP, 400mW CW), which keeps the temperature-sensitive laser safely outside the hot ASIC. In AAOI's own words, these solutions "serve as the foundation for 102.4T+ switching architectures." The only thing AAOI doesn't make is the switch chip itself, essentially every optical component around it, it does.
If pluggable optics stays dominant, AAOI sells transceivers with its lasers inside. If CPO ramps, AAOI sells the optical engine and the external laser source those CPO systems require. And denser CPO-enabled racks scale into bigger clusters that need more rack-to-rack optical links anyway. Across every branch, the one component that must be manufactured is the InP laser - the exact part in shortage, and the one AAOI makes in-house.
Timing helps too. CPO is still early: LightCounting expects meaningful deployment to begin in 2026–2027 and reach high volume only around 2028, and even then projects CPO at roughly 30% of 800G/1.6T ports by 2027 - with pluggables dominating for years beyond. NVIDIA has reaffirmed pluggable optics through 2027. So the pluggable business AAOI is ramping today has a long runway, and its OBO and ELSFP products position it for the co-packaged future as it arrives, rather than being displaced by it.
4. How AAOI captures share
Three pillars:
Pillar 1 - Vertical integration (the laser moat). InP laser capacity is the binding constraint on the whole industry. AAOI makes its own, and is expanding laser fabrication capacity ~350% by 2027. In a shortage, the vertically integrated maker supplies itself while assembly-only competitors wait in line. The most durable element of the thesis.
Pillar 2 - US manufacturing (the tariff and trust moat). AAOI is building "the largest AI-focused data-center transceiver production capacity in the U.S." in Sugar Land, Texas. By end-2026, ~30% of 800G/1.6T output is targeted from the US; by end-2027, over 50%. CFO Stefan Murry's point: US-made product is the one category "not going to be tariffed." Against the Chinese leaders who dominate today (InnoLight, Eoptolink), a US laser-to-module supply chain is a genuine differentiator for hyperscalers wary of geopolitical risk.
Pillar 3 - Orders already confirmed. This isn't a hope-and-dream capacity story. AAOI has disclosed over $324M in confirmed 800G and 1.6T orders from multiple hyperscalers:
- First 800G volume order - hyperscaler, Dec 2025
- 800G order growth - $124M cumulative for one customer (Mar–Apr 2026)
- 1.6T order - $200M+ from a long-term hyperscaler; ships Q3–Q4 2026
- First 800G volume shipment completed to a large hyperscaler, Q1 2026 - shipped
- Second hyperscaler — first 10,000 units of 800G — shipped
- Third hyperscaler — signaled intent for 800G — pipeline
5. The projections - revenue and the ramp
Management has put specific numbers on the table: >$1.1B FY2026 revenue guidance (raised from $1.0B), $471M forecast monthly revenue by mid-2027, and 60–80% guided sequential growth in Q3 and Q4 2026.
CFO Stefan Murry gave the exact composition on the Q1 call (a revenue forecast, not capacity):
AAOI is not betting the farm on 1.6T. The engine is 800G ($217M/month, ~46%) - proven, shipping, in a demand cycle that already exists. 1.6T is the upside layered on top. If you worried AAOI's story depends on a still-nascent 1.6T ramp executing perfectly, the mix says otherwise: 800G alone carries the plan.
And the capacity behind that revenue - with demand guided to exceed supply the entire way:
6. The margin story - where profit comes from
The bull case needs margins to expand as the mix shifts to higher-value 800G/1.6T. Management has guided exactly that.
Two forces drive it. Mix: 800G/1.6T carry higher margins, so as they grow from a sliver to the majority, the blend rises. (800G was just $4.6M, 5.6% of data-center revenue, in Q1 2026. The ramp has barely begun.) Scale and automation: AAOI's automated lines and vertical integration mean unit costs fall as volume rises. Management is explicit: margins improve as capacity expands.
7. The Sold-Out Market: Why Small Is an Advantage
AAOI is a small player in a market where every major supplier is completely sold out.
Start with the size gap. AAOI's last reported quarter was $151M in revenue. Lumentum (LITE) is guiding its current quarter to a record ~$985M. Coherent (COHR) just posted $1.81B . Next to those two, AAOI looks tiny. But in a supply-constrained market, being small isn't a weakness - it's the position with the most room to run.
Because the entire industry is sold out. This is what changes the calculus. AI optics isn't a share war over a fixed pie, it's a market where demand exceeds what every leader combined can physically produce:
- Lumentum is, in its CEO's own words, "completely sold out through end of 2027," and even at full capacity is under-shipping demand by 25–30%. It just bought a fifth indium phosphide fab (Greensboro, NC) solely to catch up, took a $2B investment from NVIDIA, and is targeting $8B in revenue by FY2028.
- Coherent reported a data-center book-to-bill above 4x , four dollars of orders for every dollar it can ship , with a record backlog stretching into calendar 2028 and long-term agreements running to the end of the decade. It's racing to quadruple its InP capacity by end-2027 and also took a $2B NVIDIA investment.
- AAOI guides that demand exceeds its own capacity through mid-2027 -the same story, which is exactly why it's expanding so hard.
When the two dominant players are each turning away a quarter to a third of the demand placed on them and booking orders three-plus years out, all of that unmet demand has to go somewhere. A sold-out market is the single best environment for a smaller, credible supplier to take share - because hyperscalers who can't get enough from LITE and COHR need another qualified source, and they are actively multi-sourcing to de-risk precisely this bottleneck. In a supply-constrained market you don't win by being cheapest; you win by being able to ship when the incumbents can't.
And AAOI's wedge into that gap is the same thing that constrains everyone else: the laser. The binding limit across the whole industry is InP laser capacity. LITE and COHR are both pouring billions into their own laser fabs and still can't meet demand. AAOI makes its own lasers and is expanding fabrication ~350% by 2027. The supplier that controls its own scarce input can grow shipments when supply-limited rivals cannot. Layer on US manufacturing (Sugar Land, tariff-insulated) at a moment when hyperscalers want domestic, non-China supply, and AAOI has a credible path to capture the incremental orders the giants physically can't fill.
The small base is the accelerant. Because AAOI starts at ~$151M a quarter instead of $1.8B, it takes only a sliver of this overflowing demand to transform its revenue. This is a company whose entire addressable opportunity is the demand its far-larger rivals are already turning away, in a market growing 60%+ a year. The smallest credible supplier, holding its own laser supply, has the most room to run of anyone in the group.
8. The Buildout: The Physical Proof Behind the Ramp
Can AAOI actually build the capacity?
AAOI has expanded its Houston-area manufacturing footprint to roughly 900,000 square feet across four buildings:
135,000 sq ft operating at the Sugar Land HQ today; a new 210,000 sq ft Sugar Land facility online summer 2026; a leased 154,000 sq ft Blue Ridge building; and 388,000 sq ft across two adjacent Pearland buildings announced in April 2026. In May, AAOI added three more Houston buildings under long-term leases with purchase options, all on top of parallel capacity additions in Taiwan.
Transceiver and laser production requires ISO-certified cleanroom - expensive, slow to build, and the true constraint on how much a facility can actually produce. In 2026 AAOI signed two design-build contracts with LCC3 Solution for exactly that:
The first, FAB2 (February 2026): ~92,674 sq ft of ISO-6 cleanroom, a $30.9M contract, targeting completion September 2026. The second, FAB4 / OMD 3 (filed June 25, 2026): ~195,591 sq ft of ISO-6 cleanroom plus office and testing space, a $94.1M contract, targeting substantial completion January 10, 2027. That's more than triple the size of the first contract, committed just four months later. A company doesn't sign $94M of cleanroom construction for demand it doesn't see.
Together this buildout is designed to enable monthly output of up to 700,000 units of 800G/1.6T transceivers and a ~350% expansion in laser fabrication by end-2027, plus ~400,000 ELSFP units/month for the CPO future.
Management is putting real money behind it: FY2025 CapEx hit $209M, well above the original $120–150M guidance. It's externally validated: a $20.85M Texas Semiconductor Innovation Fund grant (500+ jobs) helps underwrite the Sugar Land expansion. And the US location is strategic - onshoring directly serves the tariff-insulation and supply-trust argument with hyperscalers.
Not Financial Advice
