I directed Claude 4.6 opus in writing a long-to-short report on AAOI as an exercise. I am very impressed with the work and do wonder if this is the future of activist investing long/short. I didn't refine this much just for sake of objectivity, and figured I'd share it.

1. Why I Went Long at $30

Applied Optoelectronics was, by most conventional measures, a dying company. I went long anyway.

The reasoning was not about fundamentals—it was about market psychology. When AI infrastructure investing began broadening & accelerating in late 2025, every company with the word “optical” in its description was getting a second look from thematic investors chasing the picks-and-shovels AI trade. AAOI, a company that had peaked at $100 in 2017 on an Amazon transceiver deal and spent the next seven years declining to single digits, suddenly had a narrative again. In a datacenter capex cycle, you buy LITE/COHR/GLW first, and AAOI last. But the beauty of last name is you can maybe size with way more conviction then you did for the compounders 1yrs ago, and they tend to be a rotational missed out/catch-up beta trade.

The thesis was simple and deliberately cynical: investors chasing AI exposure care about topline growth first and unit economics second. In the early innings of a thematic cycle, the market rewards revenue acceleration and TAM expansion stories. Margin structure, competitive positioning, and laser chip sourcing are second-order concerns that only get scrutinized when the growth story matures or the stock gets expensive enough to invite short interest.

What I Knew Was Wrong With the Company

Conversations with multiple former AOI engineers—spanning R&D, chip design, and manufacturing, with tenures ranging from four to eight years at the company—painted a consistent and damning picture of its competitive position. The problems they identified were structural, not cyclical:

Failed laser R&D for eight consecutive years. From 2017 to 2025, AOI’s chip group proposed new laser designs in internal meetings repeatedly, and none reached volume production. The company remained stuck on 25Gb DML lasers—a product from roughly 2015. The experts at the company could not advance beyond this, and when those engineers left for competitors, they succeeded elsewhere. The problem was seemingly organizational, not human capital driven.

Buying the core component from direct competitors. For its 800G transceivers, AOI was purchasing EML (electro-absorption modulated laser) chips from Coherent—the same company it then had to compete against on transceiver pricing. As one former engineer put it plainly: “You buy chips from your competitors and have to compete on price to Coherent. Can you imagine how this is impossible?” Multiple industry sources independently confirmed the same dynamic.

Entrenched leadership unwilling to make hard decisions. CEO Thompson Lin, who is also the founder and chairman, has controlled the company since inception. Former engineers noted that underperforming employees in the chip group were never let go, new technology investments were chronically underfunded (one engineer assigned to study silicon photonics), and the management culture prioritized retention over results. These accounts were consistent across multiple expert calls.

Products generations behind the market. While Coherent and InnoLight were shipping 1.6T transceivers, AOI was still selling 100G products with contractual annual price declines, and its “new” 800G product used purchased lasers rather than anything developed in-house.

A Company in Financial Peril

The structural problems above were not abstract competitive concerns—they had brought AAOI to the edge of financial extinction. By late 2024, the company carried an accumulated deficit approaching $500 million. Revenue had collapsed from the 2017–2018 peak. The stock traded in single digits was 90% off its peak in a market where its sector was flourishing. The 25Gb DML—their only production laser—was a declining product with contractual annual price deflation, and they had no chip capable of competing at 400G or above. The silicon photonics program, according to former engineers, consisted of a single employee assigned to study it. Cash was burning. The company had been unprofitable for years.

This was not a company navigating a difficult transition. This was a company facing insolvency. The AI infrastructure spending wave did not merely help AAOI—it saved it from extinction.

The absence of acquisition interest during this period is itself telling. Between 2020 and 2025, the optical components industry underwent significant consolidation: the Coherent/II-VI merger, Lumentum acquiring multiple assets, Broadcom building out its photonics division. U.S.-based InP laser fabrication with competitive IP is strategically scarce—exactly the kind of asset that attracts acquirers in a consolidating market. Yet nobody acquired AAOI, even when the stock was under $5 and the enterprise value was negligible.

The reason is that informed acquirers understood what the market did not: manufacturing infrastructure without competitive IP is just a building with clean rooms. AAOI has 199 patents and a functional InP fab in Texas. But the IP portfolio is anchored to a 25Gb DML product line in a market moving to 100Gb-per-lane speeds and beyond. Anyone can build a fab. What you cannot easily replicate is the process IP, yield curves, and design library that produce competitive chips—and AAOI did not have these at the speeds that mattered. Sophisticated buyers evaluated the technology and passed. The market spoke through the absence of a bid.

Why None of That Mattered at $30

The bet was never that AOI would become a great company. It was that hyperscale customers would want an alternative transceiver supplier who was desperate enough to work on their terms—and AAOI’s desperation was precisely the point.

In a supply-constrained environment where InnoLight had 50%+ of Nvidia’s wallet share and Coherent was capacity-limited, a mid-tier U.S.-based manufacturer willing to eat thin margins for volume was exactly what procurement teams at Amazon and Microsoft needed. Not for technology leadership, but for leverage and supply chain diversification. AAOI’s weakness was its selling point: a company with no bargaining power would accept terms that a healthy competitor would refuse.

The stock offered the AI-adjacent topline growth story that thematic investors wanted, combined with catalysts (800G qualification, Made-in-USA narrative, the Amazon warrant deal) that would drive multiple expansion before anyone asked hard questions about whose laser chips were inside the transceivers.

The Amazon Warrant Deal: Desperation Disguised as Validation

The bull narrative treats Amazon’s $4 billion purchase agreement with warrant component as a blue-chip endorsement of AAOI’s technology. The reality is the opposite. The warrant structure reveals which party had leverage and which did not.

Amazon gave AAOI a purchase commitment in exchange for warrants—meaning Amazon receives equity upside if AAOI’s stock rises on the back of Amazon’s own orders. This is a classic hyperscaler procurement structure designed to lock in a desperate supplier on favorable terms. Amazon has used similar warrant-based deals with other suppliers (most notably Rivian), and in each case the motivation was supply security and pricing leverage over a capital-starved counterparty—not an endorsement of the supplier’s technological superiority.

A company with proprietary IP that multiple customers want does not give away equity to secure a purchase agreement. A company facing potential insolvency does. The warrants vest based on purchase milestones, so Amazon only dilutes existing AAOI shareholders to the extent Amazon is actually buying product—heads Amazon wins through volume commitments at favorable pricing plus equity upside, tails Amazon walks away with no cost. Amazon obtained optionality for free from a counterparty that had no alternatives.

The warrant deal should be read not as Amazon validating AAOI’s technology, but as Amazon identifying a supplier desperate enough to accept terms that transfer economic value to the buyer. It is evidence of AAOI’s weakness, not its strength.

I bought at ~$30 and sold today at $80. The thesis played out. Now I believe the setup has inverted.

 

 

2. The Laser Economics Nobody Is Discussing

The central question for AAOI’s profitability—and the one that neither sell-side analysts nor the company’s earnings call Q&A has adequately addressed—is straightforward: for each product shipping today, does AOI use its own laser chips or purchased ones?

The answer determines whether AAOI has the margin structure of a vertically integrated manufacturer or a contract assembler. And the evidence strongly suggests that for the products driving the 2026 revenue ramp, it is the latter.

Mapping the Laser Architecture by Product

Product	Laser Type	Source	Evidence	Margin Impact
100G	25Gb DML	In-house (Texas)	Legacy product; former engineers confirmed stable production	Decent but declining ASPs erode margin annually
400G SR8 (multimode)	8x 50Gb VCSEL	Potentially in-house	Multimode short-reach; VCSEL-based; AOI has VCSEL capability; Credo partnership for DSP	Moderate; lower ASP but in-house light source
400G FR4/LR4 (single-mode)	4x 100G EML	Purchased externally (required)	Same 100G/lane constraint as 800G; Thompson guided 50–70K/mo single-mode 400G capacity by mid-2025	Structurally inferior: same purchased-EML dynamic as 800G
800G DR8	8x 100G EML	Purchased externally (required)	No in-house laser at 100G/lane; 25Gb DML is 4x too slow; 100G EML never reached production	Structurally inferior: buying from competitors you price against
1.6T (future)	CW + SiPh	In-house CW laser planned	Dec 2025 400mW CW announcement; volume production "later in 2026"	Highest margin potential IF successfully scaled

 

The critical insight is that both 400G single-mode and 800G—the products driving the data center revenue ramp—cannot use any laser chip AAOI currently manufactures. This is not an inference or a reading between the lines of management commentary. It is a hardware constraint. The purchased-EML dependency predates the 800G story entirely: in late 2024, Thompson Lin was already guiding 50,000–70,000 units per month of 400G single-mode capacity by mid-2025, a product requiring the same 100G-per-lane EML chips AAOI cannot fabricate. The 800G ramp simply doubles down on the same structural problem—more lanes per module (8 versus 4), more EMLs per unit, more dollars flowing to competitors.

An 800G DR8 transceiver operates at 8 lanes of 100G per lane. Each lane requires a laser capable of 100Gb/s PAM4 modulation—either a 100G EML or a CW laser paired with a silicon photonics modulator. AAOI’s highest-capability production laser is a 25Gb directly modulated laser (DML), a product that has been in production for approximately a decade. A 25Gb device cannot operate at 100Gb/s per lane. There is no configuration, multiplexing, or design workaround that bridges a 4x gap in per-lane data rate. AAOI announced a 100G EML prototype at OFC in 2018, but according to multiple former engineers who worked at the company through 2025, no new laser design beyond the 25Gb DML ever reached volume production. The 400mW CW laser announced in December 2025 remains in customer sample stage, with volume production guided for “later in 2026.”

Therefore, every 800G transceiver AAOI ships must contain externally purchased laser chips. This is a technical certainty, not a matter of interpretation.

Management’s language on the Q4 2025 earnings call is consistent with this conclusion without stating it directly. When asked about laser sourcing, Thompson said “most of our laser production will be in-house by the end of this year”—future tense. The 10-K states that “the majority of data center optical transceivers we sell utilize our own lasers.” This is true in unit terms because 100G products (25Gb DML) and multimode 400G products (VCSELs) still constitute the majority of units shipped. But the growth products—400G single-mode and 800G—both require 100G-per-lane EML chips that AAOI cannot manufacture. The hedged language tracks the legacy product mix. As single-mode products dominate the revenue ramp, the share of revenue dependent on purchased lasers will increase dramatically even if the unit count statement remains technically accurate.

The EML Supplier Dynamic

The likely suppliers of AAOI’s 100G EML chips include Coherent (the most probable primary source, identified independently by multiple former engineers and the #2 global EML supplier), Lumentum (less likely given Nvidia has locked most of their capacity), and Mitsubishi Electric or Sumitomo Electric (Japanese III-V suppliers with less constrained allocation). The 10-K language—“we depend on a limited number of suppliers” plural—suggests multiple sources, which is standard practice for supply chain risk management.

The competitive dynamic this creates is perverse. AAOI’s primary EML supplier is very likely also its direct competitor on finished transceiver products. Coherent sells EML chips to AAOI at a margin, then competes against AAOI’s assembled 800G transceivers using those same EMLs—made at internal transfer cost, not market price. Coherent has no incentive to give AAOI favorable pricing and every incentive to keep their highest-performing bins for internal use. AAOI is in the position of funding a competitor’s margin while trying to underprice them on the final product.

 

 

3. The Margin Structure: Assembler Economics at a Premium Multiple

The cost difference between in-house and purchased laser architectures is not marginal—it is the difference between a viable long-term business model and a structurally challenged one. To understand why, it is necessary to build the unit economics from the component level up.

800G BOM Comparison: Vertically Integrated vs. AAOI

Component	Coherent (Vertically Integrated)	AAOI (Purchased Lasers)	Delta
8x 100G EML laser chips	$40–120 (internal transfer cost)	$80–240+ (market pricing)	$40–120+ disadvantage
DSP, TIA, drivers	$150–200	$150–200	Parity
PCB, housing, fiber assembly	$120–200	$120–200	Parity
Total BOM	$310–520	$350–640	$40–120+ gap
ASP (hyperscaler volume)	~$900	~$900	—
Estimated Gross Margin	40–55%	25–28%	15–27 pt gap

 

The 15–27 percentage point margin gap is structural, not operational. It cannot be closed through better manufacturing efficiency, higher yields on assembly, or scale leverage on non-laser components. The gap exists because AAOI is buying the single most expensive component in the BOM at market price from a supplier who makes it at a fraction of that cost for internal use. Every 800G transceiver AAOI ships sends $40–120+ in economic value to a competitor.

The Competitive Pecking Order

Within the 800G transceiver market, there is a clear hierarchy of margin structures that maps directly to laser sourcing:

Supplier	EML Source	Competitive Advantage	Est. Gross Margin	Position
Coherent	In-house (II-VI merger)	Full vertical integration; controls scarce EML supply	40–55%	Strongest
Lumentum	In-house	Largest laser fab base; Nvidia relationship; 4-inch wafer migration	40–50%	Strong
InnoLight	Purchased (but volume leverage)	50%+ Nvidia wallet share; massive volume drives supplier pricing power	30–40%	Viable
AAOI	Purchased (minimal leverage)	Small volume; no supply leverage; "Made in USA" + desperation pricing	25–28%	Weakest

 

Even InnoLight, which also purchases EMLs externally, has a fundamentally better position than AAOI because its volume (50%+ of Nvidia’s optical allocation) gives it purchasing leverage that AAOI lacks. When InnoLight negotiates with EML suppliers, it does so from a position of being the largest single buyer. AAOI negotiates from a position of being a small, recently insolvent alternative supplier that hyperscalers maintain for leverage purposes.

The Price Compression Trap

Every optical transceiver generation follows a predictable pricing trajectory: 30–40% ASP decline within 18–24 months of volume production. This is the dynamic former engineers described on AAOI’s 100G products—contractual annual price declines embedded in customer agreements. The same dynamic will apply to 800G.

The critical question is how ASP compression interacts with different cost structures:

Scenario	800G ASP	Coherent Gross Margin	AAOI Gross Margin
Today (2026)	~$900	~50% ($450 gross profit)	~28% ($252 gross profit)
Late 2027 (−30% ASP)	~$630	~40% ($252 gross profit)	~17% ($107 gross profit)
2028 (−40% ASP)	~$540	~35% ($189 gross profit)	~10% ($54 gross profit)

 

The mechanism is straightforward. Coherent’s internal EML transfer cost stays relatively fixed as ASPs decline—their laser fab cost structure does not change because a competitor lowered finished module pricing. AAOI’s purchased EML cost is sticky: capacity-constrained EML suppliers have limited incentive to cut component prices proportionally when the constraint is on their side. The result is that as 800G ASPs compress, AAOI’s gross margin gets squeezed from both directions—falling revenue per unit and a floor on the largest cost component.

At a 17% gross margin, AAOI’s 800G transceiver business barely covers variable manufacturing costs. At 10%, it does not. This is the mathematical inevitability of competing on assembly when you do not control the most expensive input.

Management’s own guidance already confirms the dynamic implicitly. Q1 2026 gross margin was guided at 29–31% with a “slight headwind from data center mix.” If 800G carried attractive margins, ramping it into the revenue mix would improve the blend, not dilute it. The headwind exists because 800G margin is below the corporate average—and the reason is purchased laser chips sitting in COGS.

Where to Find Confirmation in Financial Filings

AAOI does not disclose EML suppliers by name in any public filing, nor is it required to unless a single supplier crosses materiality thresholds (typically 10%+ of COGS). The company will avoid naming them because doing so would shatter the “vertical integration” positioning underpinning the bull narrative. However, the financial statements contain indirect confirmation:

Inventory composition shifts. AAOI is required to break out raw materials, work-in-progress, and finished goods in its inventory footnote. When the company primarily sold 100G transceivers with in-house DMLs, raw materials as a percentage of total inventory should have been relatively low (the most valuable component was manufactured internally, appearing as WIP or finished goods, not raw materials). As 800G ramps with purchased EMLs, raw materials should spike disproportionately—EML chips at $10–50+ each, times 8 per unit, times tens of thousands of units, flowing directly into purchased component inventory. Management already signaled this: inventory rose to $183M in Q4 2025, attributed to “raw material purchases for increased production.” Track raw materials as a percentage of total inventory across quarters. A jump from ~30% to 50%+ as 800G scales would confirm purchased EMLs driving the cost structure.

Gross margin trajectory as 800G dominates mix. If 800G used in-house lasers, ramping it would be margin-accretive or neutral. If gross margin compresses as 800G becomes a larger share of revenue in Q2–Q3 2026, it confirms the margin structure of an assembler, not a vertically integrated manufacturer.

10-K supplier risk language. The filing already contains two revealing statements. First: “We incorporate our own components into our transceivers wherever possible. In instances where we do not produce components ourselves, we source them from external suppliers.” Second: “We depend on a limited number of suppliers, including in some cases our own internal supply, for certain raw materials and components.” The phrase “including in some cases our own internal supply” tells you that for some products internal supply is the source, and for others it is not. Paired with the 25Gb DML limitation, the “others” include 800G.

Import/export records. Services like ImportGenius or Panjiva track customs manifests. EML laser chips shipped from Japan (Mitsubishi/Sumitomo) or from Coherent facilities would appear as imports to AAOI’s Sugar Land or Taiwan addresses. The relevant HS code for semiconductor laser diodes is 8541.40.

 

 

4. The Capex Crisis: Building a $1B Business on a $216M Balance Sheet

Even if every unit of 800G revenue materializes on schedule, AAOI faces a fundamental arithmetic problem: the capital expenditure required to build the capacity for that revenue far exceeds the company’s ability to fund it from operations.

The Capacity Math

Metric	Value
End 2025 capacity (800G)	~90,000 units/month (~31% Texas, ~69% Asia)
End 2026 target capacity (800G + 1.6T)	500,000+ units/month (~25% Texas, ~75% Asia)
Incremental capacity required	~410,000 units/month (5.6x increase in 12 months)
2025 actual capex	$209M (vs. $120–150M original guidance)
Cash on hand (Q4 2025)	$216M
Current debt	$67M

 

A 5.6x capacity increase in 12 months is an extraordinary manufacturing scaling challenge. It requires not merely acquiring and installing equipment, but building new clean room space (the Sugar Land building lease), tripling InP laser fabrication capacity for the CW transition, qualifying 1.6T production lines, and doing much of this at U.S. labor and construction costs.

The 2025 capex of $209M—already 40–75% above original guidance—delivered only the first 90,000 units of monthly capacity. Scaling an additional 410,000 units requires conservatively $400–600M+ in 2026 capex, and potentially more if 1.6T line equipment and CW laser fab buildout are included.

The Funding Gap

Even under optimistic assumptions, AAOI cannot fund this buildout from operations:

Cash Flow Component (2026 Estimate)	Amount
Revenue (if $1B guide achieved)	$1,000M
Gross profit at 30% margin	$300M
Less: operating expenses	($200–228M)
Operating income	$72–100M
Required capex	($400–600M+)
Funding gap	($300–528M)

 

A $300–528M funding gap cannot be bridged with $216M in existing cash and $67M in debt capacity. The company needs external capital, which means one or more of the following:

Equity raises. The share count has already expanded from approximately 46 million to over 76 million shares in two years. At $80 per share, every 10 million incremental shares raises $700–800M but represents another 13% dilution to existing shareholders. The stock price is the ATM machine.

Debt financing. With only $67M in current debt, there is theoretical capacity to lever up. But lenders typically want demonstrated profitability before extending large credit facilities, and AAOI is not yet consistently profitable on a GAAP basis.

Customer-funded capex. The Amazon warrant structure may include milestone payments or advance purchase commitments that effectively fund the buildout. This would explain why AAOI accepted the warrant terms—but it also means Amazon controls even more of the economic arrangement.

The Reflexive Loop

This creates a reflexive dynamic between the stock price and the business trajectory. Thompson Lin is guiding to $1B revenue and $120M operating profit while needing $400–600M+ in capex that does not exist on the balance sheet. The stock price is the funding mechanism. If the stock drops significantly, the equity raise window narrows or closes, the capacity buildout stalls, revenue guidance gets missed, and the stock drops further.

This is the same dynamic that played out in 2017–2018: capex was spent, capacity was built on the back of an elevated stock price, customer concentration shifted, and the company was left with overcapacity and debt when the cycle turned. The stock went from $100 to $5. The only difference now is that the numbers are bigger.

Five executives, including the CEO and CFO, sold a combined 45,348 shares at $38.38 on January 22, 2026. The stock subsequently ran to $59.25. Insider selling patterns in the context of a promotional narrative and a stock-price-dependent business model warrant close attention.

 

 

5. The CW Laser Pivot: Right Strategy, Wrong Moat

AAOI’s strategic response to the 800G margin problem is to transition to CW (continuous wave) lasers paired with silicon photonics for 1.6T transceivers. In December 2025, they announced a 400mW narrow-linewidth CW pump laser built on their buried heterostructure InP platform. The strategic logic is sound. The competitive implications are not what the market assumes.

Why CW+SiPh Changes the Economics

Component	800G with Purchased EMLs	1.6T with In-House CW + SiPh
Laser cost per module	$80–$400+ (8 EMLs at $10–50+ each)	$6–20 (2–4 CW lasers at $3–5 internal cost)
Laser supply control	Dependent on Coherent, Lumentum, others	In-house manufacturing in Texas
Competitive dynamic	Buying from direct competitors	Self-sourced; SiPh PIC suppliers are foundries, not competitors
Estimated gross margin	25–32%	40%+ (per management targets)
Availability	Shipping now (with firmware delays)	Not shipping; volume targeted H2 2026 at earliest

 

I

Why CW Fabrication Is Feasible for AAOI

A CW laser is fundamentally a DFB (distributed feedback) laser operating in continuous wave mode—no integrated modulator. AAOI has made DFB-based lasers for over a decade; the 25Gb DML is built on a DFB structure. The 400mW CW pump laser uses the same buried heterostructure InP platform the company already runs in its Texas fab. The core physics and process technology are familiar. This is not AAOI attempting to learn a new material system or process architecture—it is adapting existing capabilities to a simpler device structure.

This is a genuine and important distinction from the failed 100G EML program. An EML integrates an electroabsorption modulator monolithically with the laser—requiring atomically precise bandgap engineering at the boundary between two functional regions on a single chip. The manufacturing tolerances are tighter, the yield sensitivities higher, and the process complexity substantially greater. CW lasers have none of these integration challenges. The device is simpler by design.

What Can Go Wrong: Execution Risks

Power and linewidth yield at volume. Getting a lab sample to 400mW with narrow linewidth is not the same as achieving 95%+ yield across thousands of wafers. High-power CW operation stresses the semiconductor material; thermal management at the die level is non-trivial. Burn-in failure rates or power degradation over qualification lifetime (2,000–5,000 hours at elevated temperature) can disqualify an entire production run. AAOI’s historical problem, as multiple former engineers independently emphasized, was not designing lasers—it was getting designs to volume production quality. The same organizational weakness could repeat.

Coupling efficiency to silicon photonics PICs. CW lasers must couple efficiently into silicon photonics chips manufactured by someone else (GlobalFoundries, Tower Semiconductor, TSMC). The interface requires sub-micron precision optical alignment. Each SiPh PIC vendor has different coupling specifications. If AAOI’s laser package does not align optimally with the PIC their hyperscale customer uses, the package must be redesigned—adding quarters of delay.

Reliability qualification. Telcordia GR-468 qualification typically requires 6–12 months of accelerated life testing. AAOI can ship samples now, but volume production requires passing customer-specific reliability gates. Failure modes during qualification are common with new high-power laser designs, and each failure resets the clock. “Volume production later in 2026” could easily become “early 2027.”

Wafer scaling disadvantage. AAOI’s InP fab runs smaller wafers (likely 2-inch). Lumentum is migrating to 4-inch InP wafers, which yield roughly 4x the die output per wafer run. If AAOI stays on 2-inch wafers while competitors move to 4-inch, their cost per laser is structurally higher and capacity scales more slowly per capex dollar. Equipment lead times compound this: MOCVD reactors and MBE systems have 6–12+ month lead times. “Tripling InP laser capacity by mid-2027” requires equipment that may or may not already be on order.

None of these risks are likely fatal individually. CW fabrication is genuinely simpler than EML, and the failure modes are yield and reliability, not fundamental physics. AAOI will almost certainly produce CW lasers a somet volume eventually. The question is when, at what cost, and at what scale relative to competitors.

The Real Problem: CW Competitive Landscape

This is where the thesis turns from “execution risk on the transition” to “structural problem with the destination.”

The very attribute that makes CW lasers feasible for AAOI—they are simpler to manufacture than EMLs—also means the competitive moat around CW production is thinner. The EML market is a genuine oligopoly: five suppliers globally, Nvidia locking most allocation, barriers to entry measured in decades of process development. That oligopoly structure is what gives EML suppliers pricing power and what makes AAOI’s purchased-EML cost so sticky.

The CW laser market for silicon photonics will be structurally more competitive:

Supplier	CW Laser Capability	Advantage over AAOI
Lumentum	Decades of InP experience; largest laser fab base globally; already shipping CW for SiPh; migrating to 4-inch wafers	Existing Nvidia/hyperscaler relationships; 4x wafer output advantage; capacity alone could dwarf AAOI
Coherent	Already in CW production; massive InP infrastructure from II-VI merger	Vertically integrated: can offer complete solution (laser + PIC + transceiver) that AAOI cannot match
Broadcom	In-house silicon photonics platform AND laser capability	Single-vendor chip-to-module solution for hyperscalers
Yuanjie Semi (688498.SH)	Chinese CW laser manufacturer already shipping volume to domestic customers	Low-cost competitor if tariff dynamics shift or hyperscalers tolerate Chinese components
Shijia Photonics (688313.SH)	Growing Chinese CW laser volume	Price competition on commodity CW devices
Multiple others	Various InP shops can produce CW at required power levels	Lower barriers = broader supplier base = less pricing power for any single supplier

 

The Central Irony

AAOI failed at the hard thing (EML) that would have given it a real moat, because only five companies worldwide can produce EMLs at volume. Now it is pivoting to the easy thing (CW) that a much broader set of competitors can do.

The market is pricing AAOI as though CW+SiPh vertical integration is a proprietary advantage. In reality, it is the one laser architecture where AAOI’s competitive position is least differentiated. AAOI’s differentiation in CW is “Made in USA” combined with vertical integration into transceiver assembly. This is real but narrow—it is a tariff and geopolitical play, not a technology moat. And it is not unique: Lumentum has a fab in San Jose. Coherent has fabs in multiple U.S. locations. “Made in USA” is not AAOI’s alone—AAOI is simply cheaper and more desperate.

 

 

6. What Hyperscalers Actually Care About

A common objection to the laser sourcing thesis is that hyperscale customers do not care about the internal architecture of a transceiver—they care about spec compliance, volume delivery, and price per port. This is correct, and it cuts both ways.

The Indifference Principle

Amazon, Microsoft, Google, and Meta do not care whether photons inside a transceiver are generated by an EML, a CW laser hitting a silicon photonics modulator, or any other architecture. They care about four things: does the transceiver meet the power budget, bit error rate, reach, and thermal envelope specifications? Can the supplier deliver volume on time? What is the price per port? And does this supplier provide diversification against concentration in other parts of the supply chain?

This validates the original long thesis. Amazon and Microsoft did not choose AAOI because of its laser technology. They chose AAOI because it was a desperate U.S.-based alternative supplier willing to accept unfavorable commercial terms. The architecture inside the box was irrelevant to the purchasing decision.

Why This Is Also the Short Thesis

Hyperscaler indifference to architecture means AAOI does not need to “win” any laser technology race to keep its customers. As long as it can deliver a working module at an acceptable cost, customer relationships persist even with inferior margin structure. But—and this is the critical point—hyperscalers are absolutely sensitive to price per port. And price per port is where margin structure becomes AAOI’s problem, not the customer’s problem.

Amazon does not care that AAOI is paying Coherent $15 per EML chip. Amazon cares that the finished transceiver costs $X. If InnoLight offers the same 800G module for $800 and AAOI needs $950 to preserve the same margin percentage, Amazon pressures AAOI to match $800. AAOI either matches the price and takes the margin hit, or loses the volume. The EML-versus-CW distinction matters not because the customer cares about the architecture, but because it determines who has cost leadership. And cost leadership in a commodity hardware market is the only durable advantage.

Where CW Transition Does Matter Strategically

There is one scenario where architecture becomes strategically relevant to hyperscalers: when EML supply is constrained and CW+SiPh is the only path to sufficient volume. This is approximately the current situation. Nvidia has locked most global EML capacity, so non-Nvidia hyperscalers (Amazon, Google, Microsoft, Meta) are actively pushing CW+SiPh adoption because it is the only architecture with available capacity growth.

In this environment, AAOI’s CW laser program has strategic value—not because CW is technically superior, but because it is the only laser architecture with available capacity to scale. If AAOI can ramp CW production in Texas ahead of competitors, it becomes a strategically important supplier for a different reason than its current role: a domestic source of a scarce input (CW lasers for silicon photonics) rather than a dependent buyer of someone else’s scarce input (EMLs).

The real question is whether AAOI can make this transition—from “useful but margin-disadvantaged assembler of EML-based transceivers” to “strategically important domestic CW laser supplier for silicon photonics”—before the 800G margin compression cycle squeezes it out. That is a race against time, and the stock at $100 is priced as though the race has already been dominated by them.

 

 

7. The Promotional Cycle: Seven Quarters of “Next Quarter”

A company that narrowly escaped financial extinction now projects itself as a $4.5 billion revenue run-rate capacity business within 18 months. The trajectory from near-insolvency to aspirational mega-cap projections has been enabled by a stock price that serves simultaneously as scorecard, financing vehicle, and promotional tool. A close reading of seven consecutive earnings calls reveals a consistent pattern: 800G volume shipments are always about to happen, capacity targets keep expanding while being missed, the language around laser sourcing is carefully constructed to avoid direct answers, and the forward projections grow more aggressive even as the timeline slips.

The 800G Timeline: A Year of Slippage

Q4 2024 (February 2025): Thompson guided that 800G “body manufacture” would begin in “Q2, first of Q3” 2025. Capacity target: 200,000–250,000 units per month by end of 2025. Capex guidance: $120–150M for the year. “We believe we are well positioned to capitalize on this opportunity.”

Q1 2025 (May 2025): Timeline pushed back: “We continue to believe that we will produce meaningful shipments of 800G products sometime in the second half of 2025, likely in late Q3 or Q4.” Factory audit completed by hyperscale customer. Still in qualification. Thompson confirmed: “The vast majority of our 400G business is for single-mode transceivers.” This establishes the purchased-EML dependency was already the core of the 400G growth story.

Q2 2025 (August 2025): Timeline unchanged: “We expect to achieve meaningful 800G product shipments in H2 2025.” 800G revenue remained immaterial—only qualification deliveries. But the forward narrative accelerated: “By Q2, Q3 next year, 800G will be even bigger than 400G, okay? That’s my point.” Three additional hyperscaler customers in qualification at various stages.

Q3 2025 (November 2025): Timeline now “meaningful shipments in the fourth quarter.” Specific expectation: $4–10M in 800G revenue in Q4. Data center revenue came in below expectations due to “shipping and receiving delays.” Thompson’s language turned promotional: “Right now, customers give us crazy numbers... They’re talking about more than 300,000 of 800G plus 1.6T single mode transceivers, just AOI share.” Capacity target for 2026: over 200,000 units per month by mid-2026, primarily in Texas.

Q4 2025 (February 2026): 800G revenue came in “a lot below” $4M—not even close to the $4–10M expectation from the prior quarter. Cause: “ongoing firmware optimizations.” The volume ramp is now “starting in Q2” 2026. But on the same call, Thompson projected $378M monthly transceiver revenue by mid-2027—a $4.5B annualized run-rate, roughly 10x the full-year 2025 revenue. “I’m very confident with $25 million [800G in Q2], but customer demand could be $35 million–$40 million.”

The 800G volume ramp has slipped from “Q2–Q3 2025” to “H2 2025” to “Q4 2025” to “Q2 2026”—a full year of delays. Each quarter, the near-term miss was accompanied by a larger forward projection. The aspiration grew in direct proportion to the distance between promise and delivery.

Capacity Targets: Always Growing, Always Behind

Quarter	Capacity Target	Actual at Period End	Capex
Q4 2024	200–250K/mo by end 2025	—	$120–150M guided
Q3 2025	200K+/mo by mid-2026	—	—
Q4 2025 (actual)	500K+/mo by end 2026	~90K/mo (missed by >50%)	$209M actual

 

AAOI ended 2025 at approximately 90,000 units per month of 800G capacity—less than half the 200,000–250,000 units originally guided for that date. The response was not to temper expectations but to set an even more aggressive target: 500,000+ units by end of 2026, a 5.6x increase in 12 months from further behind. Meanwhile, actual capex of $209M exceeded the original $120–150M guidance by 40–75%, establishing a pattern where the cost of building out is consistently underestimated.

The Laser Language: Careful Misdirection

Across seven earnings calls, management’s language on laser sourcing follows a consistent pattern of technically accurate statements that create a misleading impression of vertical integration:

“Many key components, like our laser chips, are already manufactured in the United States.” (Q4 2025) True for 25Gb DMLs and VCSELs produced in the Texas fab. Not true for the 100G EMLs that go into every 800G and 400G single-mode transceiver—the products driving the revenue ramp.

“In-house laser manufacturing as a strategic advantage... helped the company avoid shortages affecting others.” (Q4 2025, paraphrased from CFO) This frames the laser situation as a strength, implying AAOI makes its own lasers for the products that matter. The shortage they are supposedly avoiding is the EML shortage—but for 800G, they are purchasing EMLs from the same constrained supply pool, not avoiding it.

“Most of our laser production will be in-house by the end of this year.” (Q4 2025) Future tense. Refers to the CW laser program for 1.6T, not the 800G products shipping today. Conflates unit count (where legacy DMLs and VCSELs dominate) with revenue-relevant production. By the time CW lasers are in volume production—if on schedule—the 800G cycle may be well into margin compression.

“The gross margin for 1.6 terabit products is higher than others.” (Q4 2025) Thompson volunteers that 1.6T margins are better, which implicitly tells you 800G margins are worse. He was never asked to compare them—he offered this framing to redirect the conversation toward a product that does not yet ship. Nobody on the call asked the follow-up: if 1.6T margins are higher, what are 800G margins specifically?

The margin headwind confession is the most direct confirmation. Across both Q3 and Q4 2025 calls, management repeatedly stated that “data center revenue mix in the next few quarters will be a slight headwind” to gross margins. If 800G and 400G single-mode products used in-house lasers, ramping these high-ASP products into the revenue mix would be margin-accretive, not dilutive. The headwind exists because purchased EML costs sit in COGS for the growth products, dragging the blended margin as they become a larger share of revenue.

The Escalating Forward Narrative

Thompson’s promotional language escalated in direct proportion to the delay in 800G delivery:

Q4 2024: “We believe we are well positioned to capitalize on this opportunity.”

Q3 2025: “Customers give us crazy numbers.”

Q4 2025: $378 million monthly revenue by mid-2027. “The demand from the customer... even how I say, we got the order from them pretty soon. We’ll see in few weeks.”

The Q4 2025 earnings call was the apex: a $4.5 billion annualized run-rate projection, delivered on a call where actual 800G revenue was “a lot below” $4 million. The 2026 revenue target of $1 billion and operating profit target of $120 million were offered without binding orders in hand—only “loading forecasts” from customers and demand signals that Thompson characterized as capacity-constrained rather than demand-constrained. As of the call date, firm 800G orders would arrive “pretty soon, within a few weeks.”

The pattern is familiar. AOI’s stock went to $100 in 2017 when Amazon bought its 100G transceivers. Those transceivers had quality issues. The stock subsequently declined over 95% to under $5. Thompson Lin remained CEO throughout. The current cycle—stock appreciation driven by AI-adjacent narrative, funded by equity dilution, supported by aspirational forward projections that grow each quarter while near-term delivery slips—shares structural similarities with that earlier episode. The difference is that the numbers are larger, the capex requirements more demanding, and the laser chip dependency more acute.

 

 

8. The Short Thesis

The long thesis was: thematic investors will chase AI-adjacent topline growth before scrutinizing margin structure. That phase is complete. At $80 per share (pre-earnings close), the stock prices in near-perfect execution of the $1 billion 2026 revenue guide and beyond. The short thesis is that the margin structure underlying that revenue is structurally inferior, the capital required to build it doesn’t exist on the balance sheet, and the strategic pivot intended to fix it leads to a less differentiated competitive position.

Core Arguments

1. The 2026 revenue ramp runs on assembler economics—by hardware necessity. AAOI’s most advanced production laser is a 25Gb DML. An 800G DR8 transceiver requires 100G per lane. No amount of management commentary changes the fact that the company does not possess a laser chip capable of operating at the speed its growth product requires. Every 800G unit ships with purchased EML chips, carrying gross margins in the 25–28% range—below the company’s blended 31% and well below the 40% target that justifies the current valuation. AAOI permanently gives away $40–120 per unit in economic value to an EML supplier who is also a direct competitor on finished transceivers.

2. Price compression will squeeze purchased-laser economics hardest. Every optical generation experiences 30–40% ASP declines within 18–24 months. If 800G pricing compresses from $900 to $600 by late 2027, Coherent’s internal laser cost stays fixed and margin compresses from 50% to 40%—healthy. AAOI’s purchased EML cost is sticky; if BOM only drops to $500 while ASP drops to $600, margin goes from 28% to 17%—barely covering variable costs. By 2028, continued ASP decline could push AAOI’s 800G margin below 10%.

3. The capex required to execute the plan does not currently exist on the balance sheet. Scaling from 90K to 500K+ units per month in 12 months requires an estimated $400–600M+ in capital expenditure. Cash on hand is $216M. Even if $1B revenue at 30% gross margin generates $72–100M in operating income, the funding gap is $300–528M. The stock price is the funding mechanism, creating a reflexive loop: if the stock falls, the equity raise window closes, the buildout stalls, guidance gets missed, and the stock falls further. Reflexivity cuts both ways here.

4. The CW pivot solves the wrong problem. CW+SiPh vertical integration is the right strategic direction for margin improvement. But CW lasers are simpler to manufacture than EMLs, which means a broader competitive field: Lumentum, Coherent, Broadcom, and multiple Chinese vendors all produce CW lasers. AAOI failed at the hard thing (EML) that would have given it a genuine moat, and is pivoting to the easy thing (CW) where its competitive differentiation is narrowest. The market is pricing CW vertical integration as proprietary advantage when it is commodity manufacturing plus a “Made in USA” label—a label shared by Lumentum and Coherent.

5. Customer concentration is extreme and worsening. Three customers represented 91% of Q4 2025 revenue (39%, 31%, and 21% respectively). The $1 billion 2026 guide depends on two hyperscale data center customers being “roughly equivalent” in size plus one smaller contributor. One qualification failure, order delay, or sourcing shift by any of these customers would materially impair the revenue trajectory.

6. The promotional cycle has historical precedent at this company. Thompson Lin guided to a $378 million monthly transceiver run-rate by mid-2027 on a call where actual 800G revenue came in “a lot below” $4 million due to firmware issues. The gap between the aspiration and the current reality is enormous, the stock price is being used as the financing mechanism to bridge it, and insiders are selling.

Key Risks to the Short

Hyperscaler support as strategic imperative. Despite the unfavorable economics described above, Amazon and Microsoft may sustain AAOI through the transition purely for supply chain diversification. Hyperscalers have deep pockets and long time horizons. If maintaining a domestic alternative supplier is strategically important enough, they may continue placing orders even if AAOI’s unit economics are subpar—effectively subsidizing the company’s transition to CW+SiPh. The hyperscalers’ purchasing behavior may override normal competitive dynamics for longer than expected.

Successful and early CW+SiPh transition. If AAOI scales 400mW CW laser production ahead of schedule, achieves strong yield, and 1.6T qualification proceeds faster than anticipated, the margin structure inverts and the vertical integration story becomes real. The OFC investor session on March 17, 2026 is the next key catalyst for technical disclosure on CW production timelines and customer design wins.

Sustained AI capex supercycle. If hyperscaler infrastructure spending continues accelerating beyond current forecasts, demand may be so overwhelming that even suboptimal suppliers like AAOI can grow into premium valuations. In a genuine shortage, the market does not discriminate on margin quality—every working transceiver finds a buyer at a price that works.

Tariff escalation benefiting domestic manufacturing. Escalating U.S.-China trade tensions could increase the strategic premium on U.S.-manufactured optical components, benefiting AAOI’s Texas fab operations even if unit economics remain suboptimal relative to vertically integrated competitors. Policy shifts could provide artificial competitive advantages that the technology alone does not warrant.

 

 

9. Catalysts and Monitoring Framework

Date/Event	What to Watch	Bull Signal	Bear Signal
Mid-March 2026	800G firmware qualification completion	On-time completion; volume orders confirmed	Further delays; scope expansion
March 17, 2026 (OFC)	CW laser production timeline; customer design wins; competitive positioning vs. Lumentum	Specific volume timeline; named design wins; yield data	Vague timelines; no CW specifics; generic slides
Q1 2026 earnings (May)	800G revenue contribution; gross margin trajectory; inventory composition	Meaningful 800G revenue; margin holds >30%; raw materials stable	800G still immaterial; margin compresses; raw materials spike
Q2 2026 earnings (Aug)	800G revenue dominance; breakeven; capex guidance update	800G >$25M; non-GAAP profitability; capex funded	800G disappoints; profitability pushed; equity raise announced
H2 2026	1.6T first revenue; CW laser volume production start	1.6T shipping with in-house CW; visible margin uplift	1.6T delayed; continued dependence on purchased EMLs
Ongoing	Insider selling; equity raises; capex vs. cash flow; import records	Insider buying; FCF approaching breakeven; no dilution	Continued insider sales; ATM offerings; cash burn acceleration

 

 

Bottom Line

AAOI was a good long at $30 because the market rewards narrative before it rewards fundamentals. It is a compelling short at $100 because the CEO has now fully played all his cards and brought eyeballs to the name. The company is spending hundreds of millions of dollars it does not have to build capacity for a product where it has no laser chip of its own—not by strategic choice, but because its R&D failed to produce anything beyond a 25Gb DML in over a decade. The transition to CW+SiPh is the right strategic move, but it leads to a less differentiated competitive position, not a more differentiated one—the exact opposite of what the market assumes. And it is a 2027 margin event being priced as a 2026 reality, guided by a CEO with a history of promotional projections, funded by equity dilution enabled by the very stock appreciation the projections sustain.

The question the market has not asked—and that management has not been forced to answer—is the simplest one: your most advanced production laser runs at 25 gigabits per second, your growth product requires 100 gigabits per lane, and you are guiding to $1 billion in revenue. Whose chips are in the box?

 

Disclaimer: This report reflects the author’s personal investment analysis and is not investment advice. The author has exited a long position in AAOI and may initiate a short position. All information is based on publicly available sources and expert interviews over past 4 months.