LightPath Technologies ($LPTH), an Orlando-based IR optics company with $50.6M in revenue over the first nine months of FY26 (Jul 2025 to Mar 2026) and ongoing net losses, is trading near a $1B market cap. FY26 Q3 standalone revenue was $19.1M (YoY +109%), with backlog at $110.6M (+196%). The investment thesis here is that supply chain regulation and material substitution are pushing in the same direction: in the U.S. defense IR optics market, where dependence on Chinese-dominated germanium must be unwound, a domestically integrated chalcogenide glass producer commands a procurement eligibility premium. This article covers a material property comparison based on SPIE paper data, the regulatory drivers from NDAA Sections 834/844, and the competitive landscape.

Table of Contents

1. Why I Started Looking at This Stock

2. The Germanium Problem: Why Now

3. The Alternative Material: Chalcogenide Glass and What the SPIE Paper Shows

4. LightPath Technologies: Company Anatomy (Products, NRL License, M&A, Lockheed Martin Program)

5. Financials: Fast Growth, Unproven Profitability

6. Competitive Landscape: Not a Monopoly

7. Bull Case, Gap, Optionality

8. Valuation: Three Futures That a $1B Market Cap Demands

9. Scenario Framework 10. Monitoring Checklist

1. Why I Started Looking at This Stock

A company with $50.6M in FY26 nine-month revenue carries a $1B market cap. It is still losing money.

One of the most dramatic stock moves in the photonics sector this year was $AAOI (Applied Optoelectronics). It ran from a 52-week low near $15 to an intraday high of $233. AAOI is not a large platform company that dominates optical materials, lasers, and modules the way Lumentum or Coherent does. What the market bought was “a rare supply chain position: U.S.-anchored vertical integration from laser chips and subassemblies in Sugar Land, Texas to high-speed optical modules, with expanding AI datacenter transceiver capacity.” The onshoring premium kicked in across the AI datacenter supply chain.

I saw a similar structure in $LPTH. The sector is different. AAOI is datacenter optical interconnect. LPTH is defense IR optics. Both are vertically integrated in the U.S., and both are receiving multiples based on supply chain positioning rather than platform dominance. That said, AAOI is already at a $600M+ total revenue run-rate with Q1 2026 datacenter revenue of $81.4M, while LPTH is still at $67M annualized with defense backlog conversion yet to be proven. The comparison has its limits in scale and visibility.

I like doing DD on small-cap tech stocks. Holding $POET and $LWLG, I have seen firsthand that photonics small caps can work when the technology fits the cycle. I have been looking for the next one, and $LPTH caught my eye while I was building a defense sensor investment map.

It is an infrared optics company. It makes lenses and camera modules that go into EOIR (Electro-Optical/Infrared) systems: night vision, thermal imaging cameras, missile seekers. My first reaction was honest: “Lenses and mirrors? Isn’t that a commodity?”

Then I thought about ASML. A TWINSCAN DUV scanner costs $100M+, and a significant portion of that price comes from the lens column built by Zeiss. Dozens of SiO₂/CaF₂ glass elements, polished and aligned to nanometer precision. A “glass-cutting company” is one of the most irreplaceable suppliers in the semiconductor industry. ASML’s UV lithography optics and LightPath’s IR thermal optics are completely different in precision, materials, customers, and ASP. But one structural parallel holds: when optical materials and lens assemblies define the performance ceiling of a system, the “glass-cutting company” is not a commodity supplier. It becomes a bottleneck supplier. Detection range, resolution, and focus stability across temperature extremes in a thermal camera all start with the physical properties of the lens material.

$LPTH caught my eye during the defense sensor mapping work precisely because there was an interesting story at the material level. And this is not limited to defense. In AI datacenters, optical interconnects are becoming the key to unlocking power and bandwidth constraints in GPU cluster scaling. In semiconductor manufacturing, ASML/Zeiss optics push the physical limits of advanced nodes. In defense, IR lens materials set the performance ceiling of EOIR systems. The sectors are different, but the common thread is clear. Optics is no longer a peripheral component. It has become a constraint on system architecture.

But the interesting question about this stock is not “Is it a good company?” Good materials, good customers, good regulatory tailwinds are already substantially reflected in the stock price. A $1B market cap on a company with $50M-range nine-month revenue is the evidence. The real question is more uncomfortable. For a company with $50M in FY26 nine-month revenue to sustain a $1B market cap, what has to go right from here?

Simply selling more BlackDiamond lenses is not enough. Backlog has to convert to revenue. Assemblies and modules have to keep climbing as a share of the mix, pushing gross margin toward 40%. Large defense options like the Lockheed Martin program have to turn into actual production revenue. This article is less about “Is LPTH a good company?” and more about “Has the market already gone too far, or is this the beginning of a re-rating into a $1B defense optics platform?”

[Figure 1: LPTH Stock Price, 52-Week Range, Key Event Markers]

2. The Germanium Problem: Why Now

The traditional workhorse material for infrared lenses is germanium (Ge). It is a single-crystal semiconductor with strong IR transmission. The problem is who makes it.

China is estimated to account for roughly 60 to 70% of global germanium refining (USGS does not publish exact shares but describes China as the “leading producer”; independent estimates from Nordic geological surveys put the figure around 68%) [1]. Including Russia, some estimates place the combined share above 80% (based on 2021 data) [2]. The U.S. has virtually no domestic germanium refining infrastructure.

In July 2023, China introduced an export license regime for gallium and germanium [3]. Price spikes and procurement uncertainty hit defense and industrial IR programs worldwide.

Then in December 2025, the FY26 NDAA was signed into law, turning this into a regulatory mandate.

Section 834: Directs the DoD to develop and implement a strategy to eliminate dependence on covered nations (China, Russia, Iran, North Korea, Belarus) for optical glass and optical systems by January 1, 2030 [4].

Section 844: Adds germanium and gallium to the covered material list. The effective date is two years after enactment (December 2025) [5].

These provisions do not impose an immediate ban. They require strategy development and phased implementation. Chinese-sourced Ge lenses are not an overnight “compliance violation.” But over the 2027 to 2030 timeframe, procurement eligibility will be affected, and contract flow-down conditions in defense programs are likely to shift [4].

One more thing to know. In November 2025, China suspended certain export bans on gallium, germanium, and antimony to the U.S. until November 2026 [21]. The investment thesis is therefore closer to “long-term procurement risk created by the license regime and NDAA” than “immediate supply cutoff.” The uncertainty itself, the fact that China can reimpose restrictions at any time, is what motivates material substitution for defense program managers.

The market at stake is significant. The global IR and thermal imaging systems market is estimated at approximately $8B in 2025, projected to reach $10.6 to $11.7B by 2030 at a CAGR of 5.9 to 6.2% [13][14]. Defense accounts for roughly 35% of this, while automotive ADAS is the fastest-growing segment (CAGR 7.7 to 7.9%) [15]. Uncooled LWIR represents 72% of total revenue, and the core optical materials for this segment are germanium and chalcogenide glass [15]. In other words, the germanium supply chain issue touches a multi-billion-dollar market.

Defense EOIR systems have a 5 to 10-year design-to-deployment cycle. With a 2030 deadline, material decisions effectively need to be made now.

Germanium is not a bad material. The problem is that production is concentrated in China and Russia, and the NDAA has legislated a phased elimination of that dependence. This is a long-term procurement risk rather than an immediate ban, but given 5 to 10-year design cycles, the direction is already set.

[Figure 2: Global Germanium Supply Chain + NDAA Regulatory Timeline]

3. The Alternative Material: Chalcogenide Glass and What the SPIE Paper Shows

So what is the alternative? During this DD I found an SPIE conference paper.

“Investigation of As₄₀Se₆₀ chalcogenide glass in precision glass molding for high-volume thermal imaging lenses” (Huddleston, Novak, Moreshead, Symmons, Foote; LightPath Technologies, SPIE Proc. Vol. 9451, 2015) [6]

This paper is the key to understanding LightPath’s technical position. It contains head-to-head data comparing the mainstream chalcogenide glass composition Ge₂₈Sb₁₂Se₆₀ (contains germanium) against As₄₀Se₆₀ (germanium-free), which is the composition LightPath is pushing.

Chalcogenide glass is an amorphous glass made from Group 16 elements such as sulfur (S), selenium (Se), and tellurium (Te). Ordinary glass (SiO₂) blocks infrared, but this material transmits IR across the 1 to 18 μm range. It covers both MWIR (3 to 5 μm) and LWIR (8 to 12 μm), the key operating bands for thermal imaging cameras.

Three key differences versus germanium.

3-1. Athermalization: dn/dT

The first number that stands out in Table 2 of the paper is dn/dT, the rate at which refractive index changes with temperature. Ge₂₈Sb₁₂Se₆₀ has a dn/dT of 70 × 10⁻⁶/°C. As₄₀Se₆₀ comes in at 32 × 10⁻⁶/°C. Less than half. For reference, single-crystal germanium has a dn/dT of approximately 396, making As₄₀Se₆₀ roughly 1/12th of that value.

What this means in practice: military thermal cameras must operate from -40°C to +60°C. Germanium lenses shift focus across that range, requiring separate mechanical compensation mechanisms. The low dn/dT of As₄₀Se₆₀ increases the design freedom for passive athermal lens systems and can reduce the need for mechanical focus correction. Whether correction hardware disappears entirely in a real system depends on the optical prescription, housing CTE, and detector package design. But the direction is clear. Design freedom in SWaP-C (Size, Weight, Power, Cost) goes up.

3-2. Precision Glass Molding (PGM): Volume Production

Germanium is a single-crystal material. It has to be shaped one piece at a time using diamond turning. Chalcogenide glass is amorphous. Heat it up and it softens, allowing it to be pressed into a mold. This process is called PGM (Precision Glass Molding).

The glass transition temperature (Tg) of As₄₀Se₆₀ is 185°C, versus 285°C for the Ge-containing composition. A 100-degree difference. Lower molding temperature means less process energy and longer mold life. Complex shapes like aspheric lenses can be molded in a single step. Compared to diamond turning, production cost is inherently lower at the process level.

3-3. Transmission Range and Refractive Index

As₄₀Se₆₀ transmits from 1 to 18 μm, broader than Ge₂₈Sb₁₂Se₆₀’s 1 to 16 μm range (which has an absorption peak near 12.5 μm). Refractive index at 10 μm is 2.7777 versus 2.6023. Higher refractive index means the same optical effect can be achieved with a thinner lens. This is favorable from a SWaP-C perspective.

Tradeoffs

As₄₀Se₆₀ is not perfect. The same table in the paper reveals weaknesses. Vickers Hardness is 142 versus 189 for Ge₂₈Sb₁₂Se₆₀. It is mechanically softer and scratches more easily. CTE is 20.9 × 10⁻⁶/°C versus 14.5 × 10⁻⁶/°C. In harsh environments, DLC (diamond-like carbon) coatings are used to supplement surface hardness.

There is also a risk that should not be overlooked. Arsenic toxicity. As₄₀Se₆₀ is, as the name suggests, composed of arsenic and selenium. Arsenic is classified as “fatal if inhaled (H330)” under UN GHS, and strict safety controls are required during manufacturing to prevent dust and fume inhalation [16]. In finished lens form, the arsenic is locked within the glass matrix and poses low risk during routine handling. But regulatory burden during manufacturing and disposal is real.

As₄₀Se₆₀ is an arsenic-based glass, so substance-specific REACH/RoHS/CLP review is required at the manufacturing, polishing, and disposal stages. ECHA’s SVHC Candidate List is substance-specific rather than a blanket ban on all arsenic compounds. Finished lenses have arsenic locked in the glass matrix, but regulatory compliance verification remains a separate risk for expansion into commercial markets such as automotive and industrial applications. The industry is also developing arsenic-free compositions using antimony (Sb), and it is likely that some of the nine additional NRL-licensed compositions held by LightPath include such alternatives.

As₄₀Se₆₀ (BD6) offers germanium-free supply chain risk avoidance, athermal design freedom, and PGM volume production cost advantages, but comes with tradeoffs in mechanical durability and arsenic-related environmental regulation.

[Figure 3: Ge₂₈Sb₁₂Se₆₀ vs As₄₀Se₆₀ Material Properties Comparison]

That is the structure of the problem.

Germanium supply is concentrated in China. The NDAA will reshape procurement eligibility over 2027 to 2030. The most mature alternative material is chalcogenide glass, and within that category, the germanium-free composition As₄₀Se₆₀ has material property advantages.

The problem is not the material. The problem is how much of this material transition translates into actual backlog, gross margin, and per-share value after dilution. The harder questions start here.

First, which company can produce this material at volume within the United States? Second, is that company a simple material supplier, or can it be re-rated as a defense IR systems company? Third, and most importantly, at a $1B market cap today, is that re-rating still ahead of us?

LPTH’s story is compelling. But a good story and a good entry price are two different things.

Subscribe now

Read more

Lightwave Logic ($LWLG) raised its AI and data center optical transceiver TAM forecast from $17B to $47B and its SAM from $1B~$2.5B to $2B~$4B on the Q1 2026 earnings call. Four Fortune 500/Global 500 customers are now at Stage 3 (prototyping), with one to two additional Tier 1 customers expected by year end. Cash stands at $100M (as of 5/11). The stock hit a 52-week high of $18.71 intraday on earnings day, then fell to the $11 range within three trading days. This article breaks down Q1 results and earnings call commentary, analyzes the real implications of the TAM expansion, the shift in EO polymer competitive dynamics following the Marvell-Polariton acquisition, and the structural reasons behind the stock’s decline.

$LWLG: From $0.93 to $18.71 in 52 Weeks

The 52-week price range for $LWLG: low $0.93, high $18.71 [1]. Roughly a 20x spread.

That might look like pure volatility, but add a time axis and the picture changes. Through H2 2025, $LWLG traded between $1 and $3. The market valued this pre-scale commercial materials company, with quarterly revenue still measured in tens of thousands of dollars, at somewhere in the $200M to $400M range. Then 2026 happened. In January, the company hired Dr. Aref Chowdhury as CTO (formerly optical CTO at Nokia). In February, a fourth Fortune 500 company entered Stage 3. In March, PDK integrations with Tower Semiconductor and GlobalFoundries were announced back to back. By late April, the stock was trading in the $12 to $14 range.

On earnings day, May 13, the stock hit $18.71 intraday [2]. CEO Yves LeMaitre raised the TAM by roughly 2.8x and announced an expanding Fortune 500/Global 500 customer pipeline. The next day, it dropped 15 to 20% [2]. A resale registration for 402,500 shares by existing holders was filed [3], and the absence of a concrete contract announcement on the call gave traders the excuse to take profits.

As of May 19, the stock was around $11.1, putting the market cap at approximately $1.7B (based on 154M shares outstanding) [1][4].

This volatility tells the story of where $LWLG sits right now. The market is betting on the TAM expansion and SiPh ecosystem integration, but it is simultaneously wrestling with the reality of a “$29K revenue company at a $1.7B valuation.” Six months ago, the market cap was $300M. The fact that it rerated to $1.7B on expectations alone, and that those expectations could get shaved by nearly 40% in three trading days, says everything about the current tension.

Key takeaway: $LWLG’s 52-week range spans $0.93 to $18.71. The market cap rerated by more than 5x in six months, but dropped nearly 40% in three trading days after the earnings-day high. The tug-of-war between expectations and valuation gap remains live.

The Material on Top of SiPh: Why EO Polymer Now

The technical backdrop for this article connects directly to three previous PhotonCap pieces.

First, our analysis of the LWLG and Tower Semiconductor PH18 PDK integration, where we laid out Perkinamine’s technical moat (sub-1V drive voltage, BEOL compatibility, rare-earth-free) and the Design Win Cycle Stage 3 progress. Second, a polymer photonics deep dive comparing the materials technology across LWLG, NLM Photonics, and VisEra. Third, our Marvell-Polariton acquisition analysis, where we noted that Marvell’s official announcement emphasized plasmonics-based modulation technology but did not specify the IP scope for organic EO chromophores.

Layering the Q1 earnings call updates onto that context sharpens the picture.

TAM Expansion: The Structure Behind SiPh Going to 70%

LeMaitre projected that SiPh will grow from approximately 23% of the optical transceiver market in 2021 to over 70% by 2030 [5]. This framing matters for LWLG because Perkinamine does not compete with SiPh. It sits on top of SiPh. As SiPh foundries gain market share, the addressable market for Perkinamine integrated into those foundry PDKs expands in lockstep.

Understanding that structure changes how the TAM numbers should be read. The 2028 AI and data center optical transceiver TAM that LeMaitre raised to $47B (up from $17B, roughly 2.8x) is not “LWLG’s market.” It is the total transceiver market that LWLG could supply materials into [5]. The expansion is driven by upward revisions to 1.6T and 3.2T transceiver deployment forecasts, growing CPO demand above 200Gbps per lane, and rapid growth in the scale-across market (DCI, inter-campus connections) [5].

The SAM (Serviceable Addressable Market) that LWLG claims as actually penetrable rose from $1B~$2.5B to $2B~$4B [5]. Still a large number, but it represents roughly 4 to 9% of the TAM.

Foundry Ecosystem Expansion

LWLG now has PDK integrations completed or in progress at four foundries: Tower Semiconductor (PH18), GlobalFoundries (via GDSFactory), SilTerra (via Luceda Photonics), and AMF [5][6][7]. A year ago, the company depended on AMF alone. On the call, LeMaitre acknowledged that “foundry capacity constraints remain an industry-wide bottleneck” but projected that “new investments and new entrants will balance supply and demand within 12 months” [5].

Reliability Data: Telcordia 85/85 Passed

A significant technical update: device-level Telcordia stress testing was passed. Testing at 85°C and 85% relative humidity validated thermal stability and environmental robustness [5]. The oldest technical concern around EO polymers has been the long-term reliability of organic materials, and this result is a direct data point against that concern. LeMaitre commented that “recent reliability results at the device level represent another important step towards commercial readiness” [5].

Earnings Call Headline Numbers

Revenue $29K (YoY +27%), net loss $6.3M (EPS -$0.04), R&D $3.5M, G&A $3.3M (YoY +77%), cash $75M (Q1 end) to $100M (as of 5/11) [4][5]. Four Fortune 500/Global 500 customers at Stage 3, with one to two Tier 1 additions expected by year end [5]. Material supply and licensing agreement in negotiation with a lead customer, targeting 2027 high-volume production [5].

Key takeaway: SiPh market share expansion (23% to 70%+) is a direct TAM tailwind for LWLG. The foundry footprint has grown from one to four. Device-level Telcordia 85/85 testing passed. Of the $47B TAM, LWLG’s SAM is $2B to $4B.

[Figure 1: LWLG Design Win Pipeline & Foundry Ecosystem Map]

That covers the “what happened” from public sources.

The real questions sit beneath the surface.

A $47B TAM is nice, but what assumptions underpin the $2B to $4B SAM? The key variables management used to derive it, including SiPh modulator adoption rates, ASPs, and integration levels, were not disclosed. Whether those assumptions are realistic or anchored to a bull case changes the credibility of the SAM entirely.

What specifically needs to happen for the four Stage 3 customers to cross into Stage 4 (manufacturing)? What is the structure of the “material supply and licensing agreement” LeMaitre referenced, and what revenue scale does it imply for LWLG? Will the foundry bottleneck actually clear within 12 months, or does surging SiPh demand keep pushing emerging processes like polymer further down the priority queue?

Is the near-doubling of G&A in one year a sign of “commercialization readiness” or expense creep in a company with no scaled commercial revenue? How exactly has the competitive landscape changed now that Marvell acquired Polariton? A year ago, LWLG was the only EO polymer player with SiPh foundry PDK integrations. With NLM Photonics now active at both GF and Tower, and Marvell internalizing SOH modulators, what does the new map look like? And was the 40% post-earnings drop just profit-taking, or did the market detect a structural crack?

Below, we break down the Q1 cost structure, the path from Stage 3 to Stage 4, the limits of the TAM/SAM assumptions, the updated Marvell-Polariton competitive landscape, the structural reasons for the price decline, the emerging dual revenue model from IP strategy and the Denver production line, and scenario-specific monitoring points.

Read more

On May 13, 2026, two events landed on the same day in the humanoid cycle. Figure AI livestreamed Helix-02, three robots autonomously sorting 30,000 packages in 24 hours with zero failures. The livestream extended into a multi-day run. As of May 17, the counter showed 81 hours and 101,391 packages. The same day, Schaeffler announced two parallel deals with UK humanoid startup Humanoid AI, deploying up to 1,000 to 2,000 robots at Schaeffler global plants by 2032 plus a separate 5-year actuator supply agreement covering seven-digit (1M plus) joint actuators through 2031. This is an Investment Map of the 20 listed companies sitting on the cascade behind 11 humanoid makers globally. The thesis is that the cycle alpha is in the component cascade, not in the robot maker race. PhotonCap GVM (Global Value Map) scores 20 names across 5 axes (Technology Barrier, Supply Concentration, Bottleneck Value, Verifiable Incumbency, Catalyst Proximity) and reveals 2 Hidden Cards positioned for re-rating. Among the five tiers, Tier 2 Precision Components is the bottleneck where automotive valuations and humanoid component exposure intersect.

Contents

1. 24-Hour Livestream and the Fremont Pivot: Investment Hook

2. The 5-Tier Component Cascade Behind One Humanoid

3. Beyond the Public Framework: 4 Groups and the Hidden Card Label

4. Identifying the 20 Companies and the Multiple Gap vs Auto Components

5. 4-Group Classification, Differentiation, and Company-by-Company Analysis

6. Scenarios, Monitoring, and Closing

7. References and Sources

1. 24-Hour Livestream and the Fremont Pivot: Investment Hook

On the evening of May 13, 2026, Figure AI CEO Brett Adcock posted a short sentence on X: “Watch a team of humanoid robots running a full 8-hr shift at human performance levels. Fully autonomous running Helix-02.” In the video, three Figure 03 humanoids stood beside a conveyor belt, flipping packages so their barcodes faced down, pressing plastic-wrapped boxes so the scanner could read the labels, and walking to charging stations autonomously when batteries depleted. Business Insider reported that the stream passed 24 hours and 30,000 packages with over 3 million views and zero failures [2]. Seoul Economic Daily later reported that the run had extended to 81 hours and 101,391 packages as of May 17 [35]. TechRadar reported more than 10 million views on the original X video [34]. Business Insider has separately reported that Amazon and GXO have tested humanoid robots in warehouse settings, while Microsoft is better framed as a Figure AI investment context rather than a confirmed warehouse deployment [4]. This was one of the clearest public demonstrations yet that humanoid robots can perform a full warehouse-style shift without teleoperation.

[Figure 1: Humanoid 5-Tier Supply Chain Cascade]

[Figure 2: Humanoid BOM Value Distribution per Robot]

The robot maker landscape shows 11 companies spread globally across the same cycle. In the US, Tesla (Optimus), Figure AI (Figure 03), Boston Dynamics (Atlas, acquired by Hyundai Motor Group in 2021, with planned deployment at Hyundai Georgia plant starting 2028) [33], Agility Robotics (Digit, Amazon fulfillment and GXO Logistics paid pilots), and Apptronik (Apollo, Mercedes-Benz partnership plus Google DeepMind collaboration). In China, Unitree (G1 and H1, the lowest-price line) and UBTech (Walker S, with deployments at BYD and Geely automotive plants). Norway has 1X Technologies (NEO, OpenAI investment), Canada has Sanctuary AI (Phoenix, Magna partnership, relatively small scale), the UK has Humanoid AI (the one-year-old startup that emerged with the May 13 Schaeffler thousand-unit deal), and South Korea has Rainbow Robotics (RB-Y1 dual-arm humanoid plus the Hubo series, Samsung Electronics 35 percent subsidiary, descended from the KAIST HUBO program, the clearest pure-play public humanoid stock in Asia) [30]. These 11 companies have different robot designs and different customers, but even if the winners differ, the demand pool of actuators, reducers, magnets, sensors, and edge compute largely repeats. Supplier mix and geography do vary by platform. Tesla weights toward Sanhua and Leader Drive in China, Apptronik and Boston Dynamics lean toward automotive incumbents like Schaeffler and Allegro, Rainbow Robotics leans toward SPG and the Korean cluster. Vendor share per maker shifts, but the 20-supplier demand pool fills from the sum of all robot makers.

Earlier in 2026, Tesla added a separate capex signal. On January 29, Reuters reported that Tesla planned roughly $20 billion in 2026 capex and was shifting Fremont and Gigafactory Texas factory space toward Optimus production [32]. The Optimus Gen 3 reveal was originally expected in Q1 2026 but has slipped toward mid to late 2026 [36]. Goldman-linked supply-chain reporting has framed Chinese suppliers’ pre-order capacity buildout as a key signal to watch [7].

And on May 13, the exact same day as the Figure livestream, another signal landed on the same cycle. German automotive component centenarian incumbent Schaeffler announced two parallel deals with UK humanoid startup Humanoid AI. First, plans to deploy up to 1,000 to 2,000 robots at Schaeffler global manufacturing sites by 2032 (with phased deployment timing). Second, a separate 5-year actuator supply agreement valid through 2031, under which Schaeffler will supply more than half, over 1 million units, of Humanoid’s wheeled platform joint actuator demand [6][31]. An automotive component company started taking humanoid robots as a customer.

That is the investment question. Not who wins among robot makers (Tesla vs Figure vs Unitree). The question is which 20 component suppliers under that race will be re-rated first. The cycle is beginning to cross from prototype demos toward first commercial production. And in past tech cycles, the supply chain almost always gets re-rated before the platform maker. The same pattern that played out when launch providers were not the first to re-price in the LEO satellite cycle (III-V epi wafer and OISL terminal optics names like IQE, Coherent, AXT moved first), and when a single cable-laying vessel became a bottleneck for one hyperscaler in the subsea cable cycle, plays out here too. In this Investment Map we will look at the 20 companies sitting under that cascade.

2. The 5-Tier Component Cascade Behind One Humanoid

Before naming who wins, look at the cascade. A humanoid robot is, structurally, a stack of 5 vertical tiers. From the surface down to the unit-economics floor:

Tier 1: Critical Materials. The starting point is rare earth permanent magnets. PhotonCap estimates roughly 2 to 3 kg of NdFeB (neodymium iron boron) magnet content per humanoid depending on actuator count and hand dexterity, 2x to 3x the magnetic content of a typical EV per MP Materials management framing. The 25 to 30 major joint actuators, extending to 30+ with hands and linear axes, each require high-torque motor content, and torque density determines robot weight, agility, and battery life. China currently controls 89.6 percent of the global processing share of rare earths, structurally upstream. MP Materials (US, Mountain Pass mine plus Texas magnet plant), Lynas Rare Earths (Australia plus Malaysia), JL Mag Rare-Earth (China sintered NdFeB leader), and USA Rare Earth (Oklahoma plant commissioning in 2026) are the four listed companies in the Tier 1 layer. All four have already partially re-rated on rare earth supply chain narratives, but the humanoid demand uplift is yet to be priced in [5][7][11][12][24].

Tier 2 Precision Components are 5 companies, the actual bottleneck position in humanoid supply chains. Harmonic Drive Systems (6324.T, Japan) is the global incumbent in strain wave reducers, market cap ~$2.7B, 1Y +16% YTD. Leader Drive = Green Harmonic (688017.SS, China) makes the same strain wave reducers as primary supplier to Tesla Mexico, market cap ~$4.7B, +89% since KOID inception. Nabtesco (6268.T, Japan) is the global #1 in RV cycloidal reducers, strong in industrial robots, market cap ~$3.4B, 1Y +20%. Schaeffler (SHA.DE, Germany) is the integrated incumbent in bearings plus planetary roller screws plus rotary actuators, market cap ~$9.4B, 1Y +90 to 100%. SPG (058610.KQ, South Korea) is the only domestic player producing all three precision reducer types (planetary, strain wave SH, and cycloidal SR) in volume [28]. Even by global standards, it is a rare integrated challenger that combines planetary, strain wave, and RV lines into one portfolio (Germany Neugart specializes in planetary, Japan Harmonic Drive in SH, Japan Nabtesco in RV). Per Korean media reporting, SPG supplies 22 reducers (20 for joints plus 2 for wheels) to Rainbow Robotics RB-Y1 dual-arm humanoid [28]. Market cap ~$2.2B (KRW 3.0T), 1Y +420% [26][27][28].

PhotonCap illustrative bottom-up BOM model (model-based estimate, not a primary source, range varies with actuator count and hand dexterity per platform): Magnet ($200 to $400) into Reducer plus screw ($800 to $1,500) into Smart actuator ($3,000 to $5,000) into Sensor ($500 to $1,000) into Edge AI compute ($300 to $800). Component cascade total approximately $4,800 to $8,700 per robot (about 24 to 44 percent of the $20K target unit cost). The rest is robot maker assembly plus battery plus structural chassis plus gross margin. Component cascade vs robot maker assembly margin ratio is approximately 60:40 in the illustrated model. That is, supplier-side capture is roughly 60 percent of illustrated BOM.

Tier 3 Actuators and Motors are 6 companies. Sanhua Intelligent (002050.SZ / 2050.HK, China) has crossed from HVAC and EV thermal into linear actuators for humanoid, market cap ~$15.7B. Tuopu Group (601689.SS, China) covers rotary actuators plus dexterous hands plus electronic skin, market cap ~$15.2B, +33% YTD. Inovance (300124.SZ, China) is China’s #1 servo motor maker, supplying Unitree and UBTech upstream, market cap ~$22.3B. Nidec (6594.T, Japan) is the global incumbent in BLDC plus coreless motors, market cap ~$30B. Moog (MOG.A, US) crosses aerospace actuator capability into humanoid joints, market cap ~$9.7B, +83% since KOID inception. ROBOTIS (108490.KQ, South Korea) holds the global academic robotics standard position with its Dynamixel servo motor (used by NASA Astrobee, Boston Dynamics research labs, MIT/Stanford/CMU labs) and is crossing into Tier 2 reducer territory with its self-developed Cycloid gear DYD. Capacity expansion is underway with the Uzbekistan plant construction. Market cap ~$3.5B (KRW 4.76T), 1Y return +809.71%, the strongest outlier among the 20 listed names [29].

Tier 4 Sensors and Perception are 3 companies. TDK (6762.T, Japan) covers magnetic sensors plus image sensors plus IMU (with the InvenSense acquisition), market cap ~$60B. Melexis (MELE.BR, Belgium) is the global #1 in magnetic position sensor ICs, market cap ~$2.6B. Allegro Microsystems (ALGM, US) has a near-monopoly in BLDC motor current sensor ICs and a strong position in magnetic position sensor ICs, market cap ~$8.35B. Humanoid robots need 30 to 50 magnetic sensors per unit, one per joint, more than 2x the sensor density of an EV. None of these three companies have yet had robotics revenue priced in. Their P/E multiples are still in the automotive cyclical range (12 to 20x).

Tier 5 AI Compute, 2 companies, the platform layer. NVIDIA (NVDA, US) is the mega-cap covering Jetson edge SoC plus the GR00T VLA foundation plus the Cosmos world model, market cap ~$5.5T. Qualcomm (QCOM, US) covers the RB6 edge robotics platform, mega-cap ~$216B.

That is why the component suppliers may capture the earlier public-market alpha, even if the 11 robot makers (Tesla / Figure / Boston Dynamics+Hyundai / Agility / Apptronik / 1X / Unitree / UBTech / Sanctuary AI / Humanoid / Rainbow Robotics) capture the final system margin. The robot maker’s gross margin pool sits on final assembly plus software plus system integration, and the substantial portion of physical BOM value (in the PhotonCap illustrative model, component cascade is about 60 percent and robot maker side about 40 percent) flows to these 20 component suppliers. The cycle alpha is in the cascade, not in the robot maker race.

3. Beyond the Public Framework: 4 Groups and the Hidden Card Label

So far the picture is public information. The next layer separates the public framework from the investable framework.

The real question shifts. Among the 20 listed companies sitting on this cascade, who is fully priced and who is still auto-priced? And specifically, which 1 to 2 names in Group B Auto-Priced Hidden Layer are sitting on the largest delta between current valuation and humanoid revenue mix exposure?

[Figure 3: GVM Positioning Scatter (names withheld)]

This article scores the 20 names along PhotonCap’s GVM framework (Technology Barrier, Supply Concentration, Bottleneck Value, Verifiable Incumbency, Catalyst Proximity) and groups them into 4 thesis categories: Group A Fully Priced, Group B Auto-Priced Hidden Layer, Group C Material Choke Pure-play, Group D Platform Beneficiary.

Three thesis frames open in this article.

• Why does Hidden Card 1 sit in the same Tier 2 Precision Components position as Schaeffler? Specifically, the GVM scoring breakdown of why that position is a Visibility Discount Zone.

• Why is the geographic split (US 6 / China 5 / Japan 4 / South Korea 2 / Europe 2 / Australia 1) almost identical to the LEO satellite supply chain pattern? What that implies for re-rating sequence.

• Why does the Tier 2 Precision Components bottleneck position trade at 1x to 5x P/S while equivalent positions in semiconductor (ASML, Tokyo Electron, Applied Materials) trade at 8x to 12x P/S? The robotics premium gap.

Among the 20 names, the company sitting in the largest valuation gap right now is hidden under just one ticker. Hidden Card 1 is the single most-mispriced name in the Tier 2 Precision Components position. Hidden Card 2 is the secondary hidden position with rare integrated 3-type reducer pure-play exposure.

• Tesla Optimus Gen 3 reveal (originally expected in Q1 2026, slipped toward mid to late 2026) / Figure 03 follow-on deployment (rumored 2026 H2) / Goldman Sachs supply chain re-survey (scheduled). Which names are most directly exposed to each catalyst.

If you would like the full identification, valuation gap analysis, GVM scoring breakdown, scenario mapping, and monitoring points, the paywall opens below.

Subscribe now

Share

Read more

Coherent reported Q3 FY26 revenue of $1.806B, +27% YoY on a pro forma basis, an acceleration from the prior quarter. This was not a routine beat-and-raise. It was the quarter the company re-declared its SAM at $50B+ by CY2030, layered onto a +$20B+ incremental category that was not part of its previous framing. Compared with Lumentum’s +90% YoY in the same quarter, the two companies look like fundamentally different bets on the same industry cycle. This article walks through segment mix shift, the 8-quarter +530bps gross margin trajectory, the “one quarter ahead of plan” InP capacity signal, the actual structure of the NVIDIA $2B deal, and the $20B+ new SAM that re-positions Coherent as a platform rather than a transceiver pure-play.

Table of Contents

1. The Quarter in One Read

2. Why InP / EML Has Been the Cap on the Industry

3. Segment Decomposition: Industrial Shrinks, Revenue Still Grows +21%

4. Margin Trajectory: What Happened Over 8 Quarters

5. “One Quarter Ahead of Plan”: The Supply Layer Signal

6. The Actual Structure of the NVIDIA $2B Deal: COHR vs LITE

7. What the $50B+ SAM Declaration Says About Platform Re-positioning

8. COHR vs LITE: Same NVIDIA $2B, Different Companies

9. InP Substrate Supply Chain: The Sumitomo, AXT-Tongmei, JX 6-inch Race

10. Broader Implication: The Optical Layer NVDA Locked In Within Two Months

11. Q4 FY26 Guidance, Scenarios, and Monitoring

12. Closing

[Figure 1: Hybrid Overview - Event Timeline + Industry Stack]

The Quarter in One Read

On May 6, after market close, Coherent (NYSE: COHR) reported Q3 FY26 (March quarter).

• Revenue $1.806B (GAAP +20.5% YoY, pro forma +27% YoY, +7.1% QoQ) [1]

• Non-GAAP gross margin 39.6% (+105bps YoY, +530bps cumulative over 8 quarters) [1]

• Non-GAAP diluted EPS $1.41 (vs $0.91, +55% YoY) [1]

• Datacenter & Communications segment $1,362M (+41% YoY) [2]

• Industrial segment $444M (-16% YoY) [2]

• Q4 FY26 guidance: Revenue $1.91B to $2.05B (midpoint +9.5% QoQ), Non-GAAP EPS $1.52 to $1.72 [1]

This is not a single event quarter. Three different layers stacked up within two months.

1. March 2: NVIDIA announced a $2B equity investment in Coherent, a multi-year multi-billion-dollar purchase commitment, and future capacity rights on advanced laser and optical networking products [3]

2. March 23: COHR joined the S&P 500 [4]

3. May 6: Q3 earnings, with revenue acceleration, 105bps GM expansion, and a 1.6T transceiver ramp guide

The three layers operate on different axes. The NVDA deal locks in forward demand visibility, S&P 500 inclusion is a valuation re-rating trigger, and the earnings print is revenue realization. Most earnings calls weaken when any one of these is missing. This quarter had all three.

Lumentum (LITE) reported the same March quarter one day earlier, with $808.4M in revenue, +90.1% YoY [5]. The two companies are roughly 2x apart in absolute revenue, but LITE’s growth rate is more than 3x. And both received a $2B NVIDIA equity investment.

So the question is why the two companies that both received NVDA $2B had such different quarters, and what Coherent’s platform re-positioning actually means. That is the spine of this article.

Why InP/EML Has Been the Cap on the Industry

To understand this quarter, the location of InP/EML inside the AI optical stack matters.

Optical connectivity in AI datacenters depends on the laser source inside the transceiver. As lane speed scales from 800G to 1.6T to 3.2T, EML (Electro-absorption Modulated Laser), an InP-based laser, becomes the gating component. Moving from 100G/lane to 200G/lane does not increase EML count inside the same 8-lane configuration. What increases is the device-level burden: bandwidth, extinction ratio, yield requirement, and test complexity. Unit demand scales with optical lane count and transceiver volume, while lane speed raises the difficulty and cost per EML.

Concretely, an 800G transceiver typically uses 8 lanes x 100G, while 1.6T uses 8 lanes x 200G or 4 lanes x 400G. In an 8-lane 1.6T transceiver, 8 EMLs are required. The EML count itself is proportional to the optical lane count per transceiver. Lane speed is not a multiplier on EML count. Going from 100G to 200G per lane does not add EMLs to the same 8-lane configuration. What increases is the modulator bandwidth, extinction ratio, yield requirement, and test burden on each EML device. On top of that, transceiver count per hyperscaler datacenter scales with GPU rack count. EML unit demand scales multiplicatively with (lanes per transceiver x transceiver count), while the spec and cost burden per EML rises as a separate function of lane speed.

The constraint is that EML/InP has been globally supply constrained. In compound semiconductors, InP has smaller wafer size than GaAs or Si (traditionally 3-inch) and higher crystal defect density, so yields are low. The way to unlock it is the migration to 6-inch InP wafers, which is the technical inflection point where yield and throughput improve together.

Why 6-inch is an inflection. Wafer area scales with the square of diameter, so moving from 3-inch to 6-inch gives roughly 4x more potential die per wafer. Coherent has explicitly stated in IR materials that 6-inch lines run at higher yields than 3-inch lines [2]. Actual good die uplift depends on device layout, edge exclusion, defect density, and line maturity. Since wafer processing time does not scale linearly with die count, per-die cost falls in a step function. In compound semiconductors, this kind of wafer size migration resets the capacity curve for the entire industry once it happens.

Coherent is ramping 6-inch InP device / platform production across three sites: Sherman (Texas), Järfälla (Sweden), and Zürich (Switzerland). Lumentum is constructing a new InP fab in Greensboro [5]. Note that bulk InP substrate supply itself sits in a separate upstream layer that requires supplier qualification and committed supply. Both companies have spent the past year repeating the same message on earnings calls: demand is there, capacity is not. PhotonCap’s earlier Lumentum FY26 Q3 piece (on InP supply chain beta transmission) focused on the cap itself. This COHR quarter is the moment that cap visibly starts to lift.

In this Q3 earnings call, the single most important line from Coherent is this.

“On track to double internal InP output by year-end and more than double again by 2027” [2]

It sits on slide 6 of the investor presentation. On the earnings call, management reportedly commented that this doubling is roughly one quarter ahead of the original plan (call commentary, transcript-based).

In an industry where InP supply was the cap, supply ramping faster than plan does not mean “we built too much, who will buy it.” It means orders are stacked beyond capacity. Demand visibility did not pull supply visibility forward. The opposite happened. Demand was already there, and revenue recognition is paced by how fast supply unlocks.

This connects to LITE’s +90% YoY. Both companies are riding the same supply tailwind in the same window. LITE has a smaller starting revenue base and a near pure-play datacenter portfolio, so the same supply unlock shows up as a larger YoY percentage. COHR has revenue more than 2x LITE’s, and it is also simultaneously divesting its Industrial-adjacent businesses, so the headline growth rate looks lower. On a pro forma basis at +27% YoY, COHR actually accelerated from the prior quarter. Both are different expressions of the same cap lifting.

The difference is in starting size and segment structure, and more importantly, in how each company frames its own SAM. Both received NVIDIA $2B equity, but the COHR deal has a line that does not appear in the other company’s announcement. That single line is what re-positions COHR from a transceiver player to a platform company this quarter.

One additional point worth raising. In the same quarter, NVDA went beyond COHR and LITE. On May 6, the same day as COHR’s earnings, NVDA announced a multi-year commercial and technology partnership with Corning, including 10x expansion of US optical connectivity capacity and three new plants in North Carolina and Texas. Within two months, NVDA placed simultaneous commitments across three layers of the optical stack: EML / laser (LITE), transceiver / networking (COHR), and fiber / connectivity (Corning). This is not a single-company bet. It is a layer-level supply lock-in pattern.

One layer deeper still is the InP wafer substrate layer. COHR runs 6-inch InP device / platform production across Sherman, Järfälla, and Zürich, while LITE is expanding Greensboro. Bulk InP substrate wafer itself is sourced from upstream suppliers under separate qualification and committed supply arrangements. The global InP wafer market is extremely concentrated, with the top three (Sumitomo Electric, AXT / Beijing Tongmei, JX Advanced Metals) holding roughly 91 to 95% of the market depending on the source. AXT itself has already been the subject of two PhotonCap deep dives. Part 1, “AXT Inc. (AXTI) Deep Dive: The Hidden Bottleneck in AI Optical Interconnects“ (March 2, 2026), mapped the InP supply chain and AXT’s vertical integration. Part 2, “A Company With $88M in Annual Revenue Just Got a $6B Market Cap: What AXTI’s Q1 Changed, and What It Didn’t“ (May 4, 2026), examined Q1 results (InP revenue $13.6M, +70% QoQ, GM 29.9% with a 36pp YoY swing, backlog $100M+) and the asymmetry between narrative and valuation. This COHR quarter shows how the same industry cycle connects from fab level back to substrate level.

The 6-inch race carries two implications for the broader cap. First, the period when EML / InP supply was the binding constraint is in the process of unwinding. Second, the industry is entering a phase where multiple layers of the stack (substrate, epi, processing, module) can ramp in parallel, not just one company at a time. For the ramp curve from 1.6T to 3.2T and 6.4T to remain unbroken, both fab-side ramp at COHR / LITE and substrate-side ramp at Sumitomo / AXT-Tongmei / JX have to function together. COHR’s “one quarter ahead” signal suggests substrate supply has not yet become an immediate blocker for its fab-side ramp, though it does not by itself prove that the industry-wide substrate ramp is fully on schedule.

Optical exposure can no longer be valued as a single transceiver multiple. Layers with capacity rights attached have to be separated from layers without. What differs across stocks is which layer a company sits in, whether it is inside a named NVDA deal scope, and whether the portfolio is pure-play or platform. The company-level breakdown sits past the paywall.

[Figure 2: COHR 5-quarter segment revenue mix shift]

So far: Q3 earnings headlines (+27% pro forma, +55% EPS), the shared supply-unlock signal across both companies, the framing that the same NVDA $2B landed differently in the two companies, the pattern that NVDA placed simultaneous commitments across three optical layers (COHR + LITE + Corning) within two months, and the substrate-level Top 3 supplier structure.

From here: COHR’s 5-quarter segment decomposition, the two drivers behind the 8-quarter +530bps gross margin trajectory, the actual scope difference inside the NVDA $2B “capacity rights” clause (LITE’s OCS backlog is not named inside the disclosed NVDA deal scope), the 6 new growth engines under the $50B+ SAM declaration with their specific timing, the Sumitomo + AXT-Tongmei + JX 6-inch race quantitative comparison, how the broader thesis propagates to adjacent optical companies, and a quantitative COHR vs LITE scorecard with Q4 guidance scenarios.

Subscribe now

Share

Read more

On March 3, 2026, while KOSPI crashed 7.24%, LIG Nex1 hit the daily limit up. Hanwha Systems surged +29%, Hanwha Aerospace +20%. But among the names that rallied that day, RFHIC and i3system barely registered on most investors’ radar.

This article breaks the defense supply chain into 4 tiers and applies PhotonCap’s GVM (Global Value Map) framework across 28 listed companies from materials to system integration. The core thesis is simple. On the first day of war, platforms (fighters, tanks, warships) rally. When war drags on, the sensors and components inside those platforms make the money. Why nLIGHT (LASR) climbed +533% from its 52-week low near $12 to over $80, why ASELSAN hit a $47B market cap, why RFHIC landed a KRW 50.6B contract from Raytheon: the answer is in this tier structure.

If you have been following the PhotonCap series, familiar names will appear. IQE and AXT from the glass substrate series, Coherent from the subsea cable series. III-V compound semiconductor as a material tier connecting optical communications to defense: that is why PhotonCap covers this topic.

The GVM Score is a framework I built for my own investment decisions. Of the 28 companies analyzed, I held existing positions in 3 before writing, and initiated positions in 3 additional companies immediately after constructing the GVM table. Which companies those are is disclosed inside the article.

Contents

1. March 3, 2026: The Day Sensors Beat Platforms

2. The Physics of Modern Warfare: Why Sensors Are the Bottleneck

3. 28 Companies, 4 Tiers, and Names Not Yet Revealed

4. Company Identification + Peer Comparison

5. 4 Clusters + Differentiators + Company Analysis

6. GVM Dashboard + Scenarios + Monitoring

7. References & Sources

1. March 3, 2026: The Day Sensors Beat Platforms

On March 3, 2026, the market bought sensors before platforms.

On February 28, the U.S. and Israel struck Iran. When the Korean market opened on Monday, March 3, KOSPI crashed 7.24% [1]. Samsung Electronics, SK hynix, Hyundai Motor all fell in unison. But defense stocks went the opposite direction. LIG Nex1 hit the daily limit (+29.86%), Hanwha Systems gained +29.14%, and Hanwha Aerospace rose +19.83% [1]. The entire market was in panic mode while defense was celebrating.

That much was all over the news.

What caught my attention was something else.

Scanning the list of stocks that rallied that day, names like RFHIC (intraday +17%), i3system, FiberPro, and KOTS Technology appear [2]. These companies do not build fighters, tanks, or missiles. What they make is GaN power amplifiers for radar (RFHIC), infrared imaging sensors for guided weapons and tanks (i3system), and fiber optic gyroscopes for inertial navigation systems (FiberPro).

Hanwha Aerospace builds the K9 self-propelled howitzer. LIG Nex1 builds the Cheongung-II air defense system. In systems like the Cheongung-II, the multi-function radar (MFR) and active radar seeker determine core performance. The RF power amplifiers inside this radar system are the bottleneck. RFHIC signed a KRW 50.6B high-power amplifier supply contract with Raytheon Canada [3], and is positioned to benefit from expanding K-defense guided weapon exports. i3system supplies infrared imaging sensors and FPA (Focal Plane Array) units for the K2 tank gunner sight and guided weapon seekers [4]. A core name in the K-defense IR sensor supply chain.

Platforms are the visible weapons. Sensors are those weapons’ eyes and ears.

On the first day of war, the headline is “Hanwha Aerospace builds the K9.” But when war stretches to 3 months, 6 months, a year, the story changes. The platforms are already built. What runs short is replenishing missiles, upgrading radars, replacing electronic warfare suites. Every time a missile is fired, a seeker is consumed. Every time threats evolve, radars and EW suites get replaced. Platform lifespan is 20 to 30 years, but the sensors and RF components inside them are upgraded or consumed on 5 to 7 year cycles. And the companies making these sensors and components have market caps 10x to 100x smaller than the platform companies.

The Cheongung-II intercepted Iranian missiles in actual combat in the UAE. According to operational data disclosed by Representative Yoo Yong-won, the UAE-deployed Cheongung-II reportedly neutralized approximately 29 out of 30 targets, reportedly achieving roughly 96% interception effectiveness [5]. This was the first time in K-defense history that an exported weapon system proved its performance in live combat. The UAE requested additional batteries, and Qatar and Kuwait began showing interest.

But who benefits most directly from this combat validation? A 96% intercept rate is not simply the performance of a single missile. It is the combined result of the multi-function radar, engagement control, seeker, launch vehicle, and field operations. What this article focuses on is the sensor and RF component layer that is repeatedly consumed and upgraded. The supply chain of these components is the subject of this article.

Key takeaway: When war breaks out, platforms rally. When war drags on, sensors make the money. The investment map of 28 companies across the supply chain starts here.

Share

2. The Physics of Modern Warfare: Why Sensors Are the Bottleneck

Understanding the defense sensor supply chain requires knowing four technology axes. RF GaN, EO/IR, Fiber Laser, Electronic Warfare. These four are the “eyes, ears, fists, and brain” of modern weapons systems.

2.1 RF GaN: The Heart of Radar

Radar performance is determined by the transmit/receive (T/R) module. The industry is transitioning from silicon (Si) and gallium arsenide (GaAs) to gallium nitride (GaN). GaN can deliver over 10x the output power per unit area compared to Si. That means seeing farther, tracking more targets simultaneously, and withstanding stronger jamming.

The problem is that very few places in the world can fabricate GaN-on-SiC wafers. Growing GaN epitaxy on silicon carbide (SiC) substrates requires high temperature, high vacuum, and extreme crystal quality control. If this tier gets bottlenecked, the entire AESA (Active Electronically Scanned Array) radar supply chain halts.

MACOM fabricates GaN/GaAs RF semiconductors in its own fab. Its recent GBP 45M investment in IQE (a III-V epitaxial wafer specialist) with a long-term supply agreement signals an intent to directly manage upstream bottlenecks [6][7]. RFHIC is the only company in Korea that produces GaN transistors in-house, and it received a KRW 50.6B power amplifier supply contract from Raytheon Canada [3]. Based on analyst coverage, RFHIC’s GaN power amplifiers are also estimated to supply core RF components for the Cheongung-II radar [8].

This structure follows the same grammar as previous PhotonCap series. IQE and AXT from the glass substrate series make InP/GaAs substrates. Those substrates are used not only for optical communication lasers but also connect to radar GaN-on-SiC, SWIR sensor InP/InGaAs, and some cooled IR detector materials within the III-V compound semiconductor ecosystem. III-V compound semiconductor is the material tier that runs through optical communications, satellites, and defense.

2.2 EO/IR: The Eyes of Weapons

EO/IR (Electro-Optical/Infrared) is the technology that detects infrared radiation to identify targets. It converts thermal radiation invisible to the human eye into a camera image. Missile seekers, tank sights, drone surveillance cameras, naval surveillance systems: all use EO/IR sensors.

The core component is the FPA (Focal Plane Array). It is the detector that converts IR energy into electrical signals, and comes in two types.

Cooled: Uses InSb (indium antimonide) or MCT (mercury cadmium telluride) materials. Requires cryogenic cooling (77K) but offers high sensitivity for long-range target identification. Used in missile seekers and fighter IRST (Infrared Search and Track) systems. Uncooled: Uses VOx (vanadium oxide) microbolometers. No cooling required, making it suitable for miniaturization. Used in personal weapon sights, drone cameras, vehicle night vision.

Teledyne (formerly FLIR Systems) is the pioneer in this market. i3system is the only company in Korea that develops both cooled and uncooled IR sensors in-house, supplying FPA units for the K2 tank gunner sight and guided weapon seekers. i3system sensors are also mounted on the K2 tanks exported to Poland. The company is currently building a new uncooled factory, scheduled for full operation from Q2 2026 [4].

LightPath Technologies makes IR optical lenses. The company produces BlackDiamond, a proprietary chalcogenide glass, under an exclusive license from the U.S. Naval Research Laboratory (NRL) [9]. As a germanium (Ge) replacement material, it supplies IR cameras for L3Harris’s naval SPEIR program [9]. The most recent quarter (FY2026 Q3) showed revenue of $19.1M, up +109% year over year, with a backlog of $110.6M [10].

But LightPath’s real catalyst is elsewhere. The company was selected as a subcontractor to deliver IR thermal camera systems for Lockheed Martin’s NGSRI (Next Generation Short Range Interceptor) program, the next-generation Stinger missile replacement. It passed a qualification milestone in 2024 and began delivering flight-capable hardware to Lockheed Martin test units [11]. Following an initial order of 10,000 systems, annual follow-on orders of up to 10,000 rounds are possible over a 10-year program [11]. Per company forward guidance, LightPath’s contribution per missile is estimated at $5,000 to $10,000. If Lockheed Martin wins final selection, LightPath’s annual revenue could grow by $50M to $100M, with a lifetime value of $500M to $1B (author estimate). The CEO called this the program that could “nearly double the size of the company overnight.”

This example illustrates the core thesis of this article. Lockheed Martin builds the missile, but the missile’s eyes (IR seeker) are still made by a sub-$1B small-cap. This asymmetry is the investment logic of the sensor tier.

2.3 Fiber Laser: The Economics of the Drone Age

Directed Energy Weapons (DEW) use lasers to shoot down drones. The reason this matters comes down to economics.

Shooting down a single drone with a missile costs $0.5M to $3M. A single Patriot missile costs $4M. But defending against swarms of $500 commercial drones with dozens of Patriot interceptors is a cost structure the defender can never win.

Laser weapons cost $1 to $10 per shot. All they need is electricity. A 30kW laser can burn a small drone. A 70kW laser can intercept mortar rounds and rockets. The core of these laser systems is the high-power fiber laser, and nLIGHT is the pure-play in this market.

nLIGHT (LASR) was still doing industrial laser cutting and welding through 2025. Then it deliberately exited the industrial business and went all-in on defense DEW. The company launched the HADES directed-energy product line, with its first product being a 70kW-class high-energy laser effector [12]. It delivered lasers to the U.S. Army’s DE M-SHORAD (Directed Energy Maneuver Short-Range Air Defense) program and is expanding an assembly plant in Turin, Italy, for European customers.

From a 52-week low near $12 to over $80 recently. 1-year return is +533%. Q1 2026 revenue came in at $80.2M, up +55% year over year, with A&D (Aerospace and Defense) revenue nearly doubling [13]. Market cap is approximately $4.6B.

Coherent from the subsea cable series comes to mind here. Coherent was core to subsea cable coherent optics, but another business pillar is GaN-on-SiC substrates and high-power laser materials. The same photonics technology sends data through subsea cables and shoots down drones in defense.

2.4 Electronic Warfare: The Invisible War

Electronic Warfare (EW) is the technology that detects, jams, and deceives enemy radar and communications. In modern warfare, if EO/IR sensors are the “eyes,” EW is the “brain.” Blinding the enemy’s eyes while protecting your own.

An EW suite integrates multiple receivers, transmitters, signal processors, and antennas into a single system. It also uses GaN RF components (high power is needed to transmit jamming signals), IR sensors (missile approach warning), and software-defined architecture.

Elbit Systems (Israel) has combat-proven EW technology. The Mini-MUSIC DIRCM (Directed Infrared Countermeasure) system blinds the IR seeker of an incoming missile with a laser. In 2026, it received a $275M EW suite contract from an Asia-Pacific country [14]. Hensoldt (Germany) has the Kalaetron EW product family, and as of Q1 2026, new orders surged year over year, pushing the order backlog to EUR 9.8B [15].

[Figure 1: Defense Sensor 4-Tier Supply Chain Map]

2.5 The 4-Tier Structure: From Materials to Systems

Breaking these four technology axes into a supply chain yields 4 tiers.

T1 Compound Semiconductor and Critical Materials: Companies making GaN-on-SiC wafers, InP/GaAs substrates, and IR window materials. If this tier stops, neither radars nor sensors function.

T2 Component and Optics: Companies using T1 materials to build laser sources, IR detectors, RF chips, and optical lenses. Performance is determined at this tier.

T3 Sensor and Subsystem: Companies assembling T2 components into radar modules, EO/IR turrets, and EW suites. Units mounted directly onto weapons systems.

T4 System Integration: Companies integrating T3 sensors into final weapons systems (missiles, fighters, warships, air defense systems). Hanwha Aerospace, L3Harris, and Elbit sit here.

This structure follows the same grammar as previous PhotonCap series. The LEO satellite series had T1 (materials) through T5 (services). And at the T1 material tier, the same companies appear repeatedly. IQE made InP laser epi for LEO satellites; here it makes GaN epi for radar. AXT made InP substrates for optical communication lasers; here it makes GaAs substrates for IR detectors.

That is why this article reads defense through III-V materials and sensor supply chains, not through platforms. Other defense analyses do not show this connection. You have to understand the physics inside the sensors to see it.

Key takeaway: The defense sensor supply chain has a 4-tier structure where T1 materials are the most scarce and irreplaceable. And the companies at T1 overlap with the III-V semiconductor names PhotonCap has already covered across other series.

[Figure 2: III-V Compound Semiconductor Cross-Series Map: T1 material overlap across Glass Substrate, LEO Satellite, Subsea Cable, and Defense Sensor series]

3. 28 Companies, 4 Tiers, and Names Not Yet Revealed

Everything up to this point is visible from public information alone. That GaN is core to radar, that EO/IR sensors are the eyes of weapons, that laser weapon economics are changing missile defense: this is first-order structure that appears in sell-side defense reports.

The differentiation for investors begins here.

Place 28 listed companies across 4 tiers and score them with GVM, and four clusters emerge.

• [A] Bottleneck Component / Material: Companies with monopoly or near-monopoly positions in specific materials or sensor components. If these companies stop supplying, radars go dark or missiles cannot fly. Small market caps, but everything upstream depends on them.

• [B] Sensor Origin: Companies with deep technology moats. IP that competitors cannot replicate within 5 years.

• [C] System Integrator: Companies where backlog equals strength. The longer war lasts, the more backlog accumulates, converting to revenue over the next 3 to 5 years.

• [D] Market Cap Dislocation: Companies with high GVM scores but where the market has not yet re-rated. This gap is the cycle alpha.

Which companies belong to [A], which to [D]: that unfolds after the paywall.

One preview. Among the 28 companies, one recorded a 1Y return of +7,885%, and another hit a $47B market cap while having near-zero awareness among Korean investors. The GVM scores and re-rating velocity of these two companies are the most important data points in this series.

[Figure 3: Positioning Scatter Plot (dots only, no labels). Dashed circle on the Hidden Card]

One more thing. This series features companies from 9 countries. The United States, Korea, the United Kingdom, Germany, France, Israel, Turkey, Japan, and Sweden. Only 3 names are familiar to Korean investors (Hanwha Aerospace, LIG Nex1, Hanwha Systems). Among the remaining 25 companies, some have never been covered in a single Korean brokerage report.

Their names, tier placement, GVM scores, and 4-cluster mapping are revealed for the first time in this article.

[Figure 4: 9-Country Defense Sensor Ecosystem Map, country-by-tier distribution]

Subscribe now

Read more

The glass substrate investment map was optics. CPO transceivers, SiPh, InP lasers. Right in the center of what I do. The LEO satellite map was a step further out. The optical component inside an OISL terminal is my territory, but orbital mechanics and launch vehicle economics are not. The subsea cable map went further still. The III-V pump laser inside an EDFA repeater is my territory, but the cable-laying vessel market and national security framing are not.

To be honest, there are limits to interpreting fields outside my expertise from a perspective that is also outside my expertise. I can find the number that says 60 cable-laying vessels exist worldwide, but someone in the shipbuilding industry understands the operating economics of a single vessel better than I do. I can back-calculate that NEC’s subsea revenue is in the single-digit percentage range, but an analyst covering Japanese IT conglomerates will do a better job valuing that business unit than I can.

Sometimes subscribers reach out. “Analyze XX purely from XX perspective.” When I get these requests, it weighs on my conscience. Is it right to publish analysis, behind a paywall, on a topic where I am not certain whether my interpretation is correct or not?

So I thought about it seriously.

Here is what I can and cannot do. Explaining why a material is a bottleneck at the optical component tier, why a process is difficult, why a qualification takes years. That is something I can do. Macro, geopolitics, valuation models, shipbuilding, power infrastructure. These are areas where I have no hands-on engineering experience. In these areas, I am a researcher reading and cross-checking public sources, not a practitioner with field experience.

Until now, I have mostly interpreted things through an optics lens. Going forward, I want to widen the aperture slightly, expanding from optics into a broader engineering perspective. Not just photonics, but compound semiconductor manufacturing, packaging, system integration, reliability testing. A wider engineering layer for looking at supply chains, while being clear about which layers I have touched firsthand and which I have not.

What changes is the scope. What does not change is the principle. Primary sources first. Limits of inference stated explicitly. When I do not know, I say I do not know. That stays the same.

I am continuing to think about the topics you have been requesting. Within the range I can cover, I will write as honestly as I can.

Best Regards,

PhotonCap

Over 95% of international internet data travels through subsea fiber optic cables. Yet only about 60 dedicated cable-laying vessels exist worldwide, with an average fleet age of 25 years. New subsea cable investment for 2025 to 2027 is projected at $13B, nearly double the prior three-year period [1]. With Google, Meta, Amazon, and Microsoft all placing orders simultaneously, the bottleneck sits at cable-laying ships and EDFA repeaters. This article breaks down the subsea cable supply chain into 5 tiers (fiber, cable, wet plant, optical components, installation) and scores 8 listed companies using the PhotonCap GVM 5-axis framework. The market has already repriced fiber and cable. Wet plant scarcity has not yet been recognized as a standalone asset.

This is the third article in the PhotonCap Investment Map series. The glass substrate map covered the supply chain above AI data centers. The LEO satellite map covered the orbital layer. This one covers what sits beneath the ocean between data centers. The EDFA pump laser inside a subsea repeater sits in the same III-V compound semiconductor manufacturing ecosystem as the laser in CPO transceivers and the optical component in satellite OISL terminals. Three cycles converge at the component tier.

Subscribe now

The third installment of the PhotonCap Investment Map series is out. This is also the second investment map collaboration with Damnang and Nutty, following the glass substrate map. This time, we go beneath the ocean between AI data centers: subsea cables.

Contents

1. Introduction

2. Technical background: How subsea cables work

3. GVM Framework + Where the real difference begins

4. Company identification + scale comparison

5. GVM Score: 5-Axis Heatmap

6. Classification + differentiation + company analysis

7. Scenarios + monitoring + conclusion

8. References & Sources

1. Introduction

Over 95% of international data traffic flows through subsea fiber optic cables [2]. Financial transactions, video calls, cloud computing, and now AI training data. Subsea cables are the physical spine of the internet.

The problem: only about 60 dedicated cable-laying vessels exist worldwide to build that spine [3]. Average fleet age is 25 years [4]. A new build costs $200M to $500M per vessel [4]. Construction lead time exceeds two years. Right now, utilization rates for these ships exceed 85%, and the project queue extends years out [5].

New subsea cable investment is projected to reach $13B over 2025 to 2027. That is nearly double the prior three-year period (2022 to 2024) [1].

The driver is clear. AI. Building data centers is not enough. You need the physical backbone connecting them. As Meta VP of Engineering Gaya Nagarajan put it: “Without the connectivity that connects those data centers, what you have are really expensive warehouses” [2].

So hyperscalers are putting money directly into the ocean floor. In a previous article, “The Real Beneficiaries of Azure +40%, AWS +28%, GCP +63%”, we tracked where $580B in cloud capex flows inside data centers: GPUs, HBM, and optical transceivers. This article tracks the $13B flowing into the physical connection between data centers: subsea cables.

Google operates multiple proprietary subsea cables including Curie, Dunant, Grace Hopper, and Equiano, and announced TalayLink (Australia to Thailand) in late 2025 [6]. Meta officially unveiled Project Waterworth in February 2025: 5 continents, 50,000km, 24 fiber pairs, maximum depth 7,000m. Reported investment exceeds $10B [7][8]. Amazon is constructing Fastnet, a US-to-Ireland cable targeting 2028 completion [6].

A decade ago, subsea cables were built by telecom consortium cost-sharing. Today hyperscalers commission and own them directly [1][9]. That structural shift is creating investment opportunities across the entire supply chain.

Scope: this article breaks down the subsea cable supply chain into 5 tiers, scores 8 listed companies across those tiers using the GVM 5-axis framework, and identifies where the bottleneck sits. This is the third installment following the glass substrate and LEO satellite maps.

[Figure 1: AI Datacenter Interconnect via Subsea Cable (datacenter to landing station to submarine cable with repeaters to landing station to datacenter)]

2. Technical background: How subsea cables work

2.1 The physics of subsea cables

At the core of every subsea cable is optical fiber. Glass preforms are melted in drawing towers hundreds of meters tall, pulled into strands thinner than human hair, bundled into dozens of fiber pairs, then wrapped in steel wire armor and polyethylene sheathing. That produces a submarine cable.

The problem is signal attenuation. Light weakens as it travels through glass. Roughly every 60 to 80km, the signal degrades to the point where re-amplification is required. That is the job of the repeater, specifically the EDFA (Erbium-Doped Fiber Amplifier) [10].

[Figure 2: Subsea Cable Cross-Section + EDFA Repeater Block Diagram (fiber pair to EDFA to WDM to fiber pair, with III-V pump laser callout)]

An EDFA works by pumping energy from a high-reliability III-V compound semiconductor pump laser (typically 980nm or 1480nm) into erbium-doped specialty fiber, which amplifies the weakened optical signal back to full strength. A 10,000km trans-Pacific cable contains hundreds of these repeaters [10]. Each one is a combination of III-V pump laser, erbium-doped fiber, and WDM coupler.

Here is where PhotonCap readers should pay attention. The III-V pump laser in an EDFA sits in the same high-reliability compound semiconductor manufacturing ecosystem as the CPO/SiPh transceiver laser source covered in the glass substrate map and the satellite OISL terminal component covered in the LEO satellite map. Not every subsea pump is an InP device (980nm pumps are typically GaInAs/GaAs, while 1480nm pumps are InP-based), but all three cycles converge at the III-V compound semiconductor manufacturing tier. This is not coincidental. The optical component supply chain is application-agnostic.

2.2 Supply chain: 5-tier structure

The subsea cable supply chain breaks into 5 tiers.

T1. Material (fiber + preform): Raw material for subsea cables. Corning is the largest player in this tier. European and Japanese fiber manufacturers also participate.

T2. Cable Manufacturing: Wrapping optical fiber in steel wire armor, insulation, and outer sheathing to produce cables that survive decades on the ocean floor. Key listed companies sit in Europe, and one of them is the only Western cable manufacturer that owns its own cable-laying vessels.

T3. System Integration (wet plant): The narrowest bottleneck. This tier designs and integrates the wet plant (repeaters, branching units, and submerged optical equipment) while also coordinating dry-plant interfaces such as SLTE and power-feeding equipment. Only 4 companies globally can do this: SubCom (US), ASN (France), HMN Technologies (China), and NEC (Japan) [12][13]. Excluding China-linked HMN Tech from Western hyperscaler procurement, the non-China shortlist narrows to SubCom, ASN, and NEC. Only NEC is publicly listed. It is a $41B IT conglomerate where subsea cable revenue sits in the single-digit percentage range of total sales, meaning the market does not separately value this business. Where this company lands in the GVM framework is covered beyond the paywall.

T4. Optical Components + Coherent Systems: Supplying the pump lasers and WDM couplers for EDFA, plus coherent DSP/transponders for landing stations. The same component tier covered in the CPO map applies here.

T5. Installation (cable-laying vessels + subsea construction): The ships and crew that physically lay cable on the ocean floor. Of the roughly 60 dedicated vessels worldwide, far fewer can handle modern large-scale projects [3][4].

[Figure 3: Supply Chain 5-Tier Cascade (no company names, tier colors only)]

2.3 Why the cycle is happening now

Three forces are converging to create the largest subsea cable cycle in a decade.

First, AI training traffic. Distributed training demands bandwidth connecting data centers across continents. Existing cables cannot keep up.

Second, hyperscaler self-ownership. A decade ago, telecom consortia split the cost. Today Google, Meta, and Amazon individually invest billions to build and own proprietary cables [1][9]. Meta’s Waterworth ($10B+) is Meta’s first single-owner global-scale cable [7].

Third, subsea cables are being reclassified as national security infrastructure. European governments are directly nationalizing subsea cable assets. France acquired a major system integrator’s submarine business from a global telecom equipment company for €350M in December 2024 [13][14]. Italy received EU approval in April 2026 to acquire a domestic telecom operator’s subsea cable subsidiary (operating 600,000+ km of global network) for €700M [30][31]. Japan has begun reviewing subsidy support for new cable-laying vessel construction [5]. NATO launched “Baltic Sentry” in January 2025 to protect Baltic Sea cables [2].

Subsea cables are being reclassified at the level of energy and transportation infrastructure. This creates demand for alternative routing beyond existing paths, driving total order volume higher.

One more factor. Telecom subsea cables are not the only users of cable-laying vessels. Offshore wind, cross-border interconnectors, and HVDC subsea power cables all compete for the same offshore installation capacity, vessel slots, burial tools, and skilled operator pools [33]. AI data centers require not just communication backbone but power grid as well, and offshore wind and interconnector projects eat into the same cable-laying vessel schedule. In other words, the vessel bottleneck is not just an “AI traffic” bottleneck. It overlaps with AI power infrastructure.

These three factors combined are projected to grow the subsea cable market from $14.9B in 2025 to $26.4B by 2031 (CAGR 10%) [12].

Key takeaway: the drivers are AI traffic, hyperscaler self-ownership, and national security reclassification, all converging simultaneously. The cable-laying vessel bottleneck overlaps with HVDC power infrastructure. Listed investor access concentrates in T1 (Material), T2 (Cable), and T4 (Optical Components). There is exactly one listed path into T3, but it sits buried inside a conglomerate, largely invisible to the market.

3. GVM Framework + Where the real difference begins

Everything above can be assembled from public sources. Surging subsea cable demand. A narrow bottleneck. The picture is clear.

This article scores 8 listed companies across the supply chain using the PhotonCap GVM (Global Value Map) 5-axis framework. This is the subsea cable edition of the same framework used as PB/CQ/CR/OP/MA in the glass substrate map and TB/PQ/BR/LM/MO in the LEO satellite map.

GVM 5-axis definition (subsea cable)

TB (Technology Bottleneck): How many companies can replace this role. 5 = fewer than 3 qualified suppliers globally. 1 = 10+ alternatives.

SC (Subsea Concentration): Subsea cable revenue as a percentage of total revenue. 5 = 40%+. 1 = below 5%.

BV (Backlog Visibility): Subsea-related backlog coverage in years. 5 = 3+ years. 1 = below 1 year.

VI (Vertical Integration): Number of supply chain tiers simultaneously covered. 5 = 3+ tiers. 1 = single narrow tier.

CP (Customer Pricing Power): Bargaining power versus hyperscalers/system integrators. 5 = sole source, high switching cost. 1 = commodity, multiple alternatives.

Phase weighting

Current phase: “hyperscaler simultaneous ordering + vessel/wet plant shortage”

Heavy (x1.5): TB, CP. In a bottleneck phase, irreplaceability and pricing power determine alpha. Neutral (x1.0): BV, VI. Backlog visibility and integration are phase-independent. Light (x0.5): SC. Higher subsea concentration means higher beta, but mega-caps generate meaningful absolute revenue even at low SC.

[Figure-Table 4: GVM 5-Axis Definition]

Where the real difference begins

These 8 companies have fundamentally different profiles. They sort into 4 groups.

One group makes fiber and raw material. Another has vertical integration from cable manufacturing to installation. A third holds an oligopoly position in wet plant system integration. The fourth supplies optical engines and systems, from EDFA pump lasers to coherent DSP and networking equipment.

Among these 4 groups, the company with the highest GVM weighted score has received the least market re-rating. Only one company out of eight scored TB 5. That company generates single-digit percentage revenue from subsea cables, but that single-digit percentage represents a capability held by only 4 companies globally (3 non-China). While Corning rose +316% in one year and a European T2 company gained +172%, this company’s re-rating fell well short of half those figures.

Which company, and why the market has not priced it yet, is covered below the paywall.

[Figure 5: Positioning Matrix, Hook Version (dots only, no names, Hidden Card dashed circle)]

Share

Read more

On April 29, 2026, Microsoft, Amazon, and Alphabet reported earnings on the same day. Azure grew +40%, AWS +28%, and GCP +63%. Their combined 2026 capex guidance reached approximately $580B. Microsoft and Alphabet raised their guidance, while Amazon maintained its $200B target set in February. This article breaks down where that capex flows across 4 hardware layers and identifies the 10 companies where that spending becomes revenue. All three beat earnings, yet MSFT and AMZN fell -3% AH while GOOGL rose +5.7% AH (+10% next-day close). Understanding why these reactions diverged, and which layer-level winners that divergence points to, is the purpose of this article.

Contents

1. Same-day earnings, different stock reactions

2. The 4 layers where capex flows

3. Paywall hook: which layer benefits most

4. 10 companies across 4 layers, with scale comparison

5. Layer-by-layer classification, differentiation, and company analysis

6. Scenarios, monitoring, and conclusion

7. References & Sources

1. Same-day earnings, different stock reactions

On the evening of April 29, 2026, three 8-K filings hit the SEC within roughly two hours. Microsoft, Amazon, Alphabet. The Big 3 of the U.S. cloud market.

All three beat earnings. Across every line item.

Microsoft: Q3 FY26 revenue $82.9B (+18%), Azure +40% (above 37-38% guidance), Microsoft Cloud $54.5B (+29%), AI run rate $37B (+123% YoY) [1]

Amazon: Q1 2026 revenue $181.5B (+17%), AWS $37.6B (+28%, highest growth in 15 quarters), AI run rate $15B+ [2]

Alphabet: Q1 2026 revenue $109.9B (+22%), Google Cloud $20.0B (+63%), backlog $460B+ (nearly doubled from the prior quarter) [3]

On paper, all three delivered. But after-hours reactions diverged.

MSFT: -3% AH. AMZN: -3% AH. GOOGL: +5.7% AH (+10% next-day close).

All three beat, yet the market priced them differently.

The biggest differentiator was how the market priced each company’s capex ROI. All three maintained or raised their 2026 capex to historically high levels, but the market assessed each company’s return on that capex differently.

Microsoft: FY26 capex $190B (vs. prior consensus of $154.6B, +23%). CFO Hood specified that “$25B reflects component price impacts” [1][4].

Amazon: 2026 capex $200B (maintaining February guidance). $44.2B deployed in Q1 alone [2].

Alphabet: 2026 capex $180-190B (raised from $175-185B). 2027 capex to “significantly increase” [3].

Combined: approximately $580B. Including Meta, the four-hyperscaler total exceeds an estimated $700B [5].

This was the market’s question: “Is this spending justified? Who earns a return on this capital?”

The real question is not about hyperscaler stock prices. It is about the supply chain that receives that capex. Microsoft disclosed this directly on its earnings call: roughly 2/3 of this quarter’s capex went to short-lived assets, meaning GPU and CPU chips, while the remaining 1/3 went to long-lived assets, meaning data centers, land, and power infrastructure [4].

Applying Microsoft’s disclosure as a heuristic across all three, an estimated $380B of the $580B flows to chips and server hardware (conservative estimate). The remaining $200B goes to real estate and power. And within that $380B, there are further layers.

All three are running the same capex race, but their strategies differ. That difference drove the stock reaction divergence.

Why GOOGL reacted strongly on its own: GCP revenue of $20.0B at +63% is more than double the growth rate of AWS (+28%). Pichai saying “we are compute constrained” on the earnings call was a strong signal that demand exceeds supply [3]. Backlog nearly doubling to $460B gave the market confidence that Alphabet’s capex spending is still justified. The preview of “significantly increasing” 2027 capex was read as a reward, not a penalty.

Why MSFT and AMZN fell -3%: Microsoft’s $190B capex guidance exceeded the prior consensus by over $35B. CFO Hood carved out “$25B from component prices” [4], but the market had not yet decided whether that $25B was an unavoidable pass-through of GPU pricing power or a sign of inefficiency. Amazon faced a similar dynamic. Deploying $44.2B in Q1 alone (22% of the $200B guide in a single quarter), the market questioned whether AWS revenue growth of +28% was sufficient to justify the pace of spending.

One insight emerges from this. The market is not penalizing capex itself. It is watching how quickly capex converts to revenue growth. GCP at +63% read as fast ROI. Azure at +40% and AWS at +28% read as relatively slower returns.

But flip this to a supply chain perspective and the conclusion reverses entirely. All three are sustaining or increasing capex, and the vast majority flows into overlapping supply chains. Whether GOOGL goes +10% or MSFT goes -3%, both are drawing from overlapping pools of GPUs, HBM, networking switches, and optical components. Hyperscaler stock reactions diverged, but the total capital flowing into the hardware supply chain grew in aggregate.

Section 1 takeaway: The cloud Big 3 all beat earnings on the same day, yet stock reactions diverged. The difference was the market’s assessment of capex ROI velocity. But from a supply chain perspective, all three are increasing capex, and roughly 2/3 of that capital flows into the same chips and hardware. When hyperscaler stocks diverge, the alpha often shows up in supply chain names instead.

2. The 4 layers where capex flows

The hardware portion of hyperscaler capex breaks down into layers. This is the analytical framework for the article.

Layer 1: Compute (GPU) is the main chip that trains and runs AI models. Products like the NVIDIA Blackwell GPU sit here. For hyperscalers, this is the single largest cost item. Hyperscalers account for approximately 50% of NVIDIA’s Q4 FY26 data center revenue of $62.3B (+75% YoY) [6]. NVIDIA does not own this layer alone, however. There is one challenger, and the difference in 1Y returns between the two shows how layer dynamics work.

Layer 2: Custom ASIC (XPU) refers to chips that hyperscalers design in-house, optimized for their own workloads. Google’s TPU, Amazon’s Trainium, and Meta’s MTIA are the most prominent examples. Design IP, IO, and packaging are handled by fabless design partners, and two companies effectively split this layer. Revenue scale and margin trajectories diverge quarter by quarter. This layer is growing fast because hyperscalers want to reduce NVIDIA dependency and build chips better suited to their specific workloads.

Layer 3: Memory (HBM) is the high-bandwidth memory that sits next to GPUs and ASICs. Memory bandwidth is the bottleneck for both AI training and inference. As models grow larger, both HBM capacity and speed become critical. Three companies effectively split this market, and in the 2026 cycle, one of them nearly tripled its revenue in a single year. Which company that is sits at the core of the thesis.

Layer 4: Networking + Optical is the fabric that connects GPUs and ASICs into clusters. AI training clusters require tens of thousands to hundreds of thousands of GPUs running a single model simultaneously, making chip-to-chip communication as important as the chips themselves. This sub-divides further.

Networking switch: Ethernet-based switches (transitioning from 800G to 1.6T) or NVIDIA’s own fabric like NVLink and InfiniBand.

Optical interconnect: optical transceivers connecting switches to GPUs/ASICs. Also transitioning from 800G to 1.6T.

What makes this layer interesting is that the entire layer outperformed together. Other layers have one company dominating internally. This layer saw every constituent benefit from the cycle. And within the layer, one stock posted a 1Y return larger than anything in NVIDIA’s league. A mid-cap name with returns on a different order.

[Figure 1: AI compute stack 4-layer diagram]

These 4 layers are interdependent. You cannot just buy GPUs and have them work. HBM must come along, and networking switches and optical transceivers are needed to connect them into clusters. So capex flows into all 4 layers nearly simultaneously. The mix across layers differs, though, and that mix shifts over time.

One interesting datapoint from NVIDIA’s Q4 FY26. Of the $62.3B in data center revenue, networking revenue alone was $11B, up +263% YoY [6]. Networking’s share of data center revenue rose from under 9% a year ago to 18%. Even within a single company, the revenue mix is shifting rapidly.

That is one company’s story, but scaling it to the layer level creates a single question.

Roughly 2/3 of the $580B in capex, or about $380B, scatters across 4 layers. Which layer is growing the fastest, and who is winning inside that layer?

[Figure 2: Hyperscaler capex flow into 4 hardware layers]

One more point before moving on: the three companies’ capex strategies are not identical.

Microsoft made the most transparent disclosure. CFO Hood explicitly split capex into 2/3 short-lived assets (chips and servers) and 1/3 long-lived assets (data centers, land, power) [4]. He also specified that $25B of the short-lived portion reflects component price impacts. This means GPU price increases are a structural driver pushing capex higher. As long as NVIDIA holds pricing power, a significant portion of capex flows to Layer 1 by default. NVIDIA is not staying in Layer 1 alone. It has made $2-3.2B equity investments in Coherent, Lumentum, Marvell, and Corning, directly locking in Layer 4 supply chains. (The background and implications of these investments were covered in our previous article, “NVIDIA’s Four 2026 Photonics Investments.”)

Amazon takes a different angle. AWS is the most aggressive hyperscaler on in-house chips (Trainium, Graviton). Amazon deployed $44.2B in Q1. Amazon is pushing Trainium/Graviton harder than any peer, suggesting its capex mix likely tilts more toward Layer 2 (Custom ASIC) than Layer 1 (NVIDIA GPU), though exact layer-level capex allocation has not been disclosed.

Alphabet is the most aggressive overall. 2026 capex of $180-190B, followed by a “significant increase” in 2027 [3]. Google Cloud backlog of $460B signals that secured demand justifies years of future capex. Google is pursuing in-house design across Layer 1, 2, and 4 with TPU v6 (Trillium) and proprietary optical interconnects.

How these strategic differences affect layer beneficiaries: Microsoft has the highest NVIDIA dependency (Layer 1 heavy). Amazon has the highest ASIC dependency (Layer 2 heavy). Alphabet pursues in-house design across the most layers (most diversified). But all three rely on external procurement for HBM (Layer 3) and optical interconnects (Layer 4). Some layers simply cannot be brought in-house. Identifying those layers is the focus of the next section.

Section 2 takeaway: The AI compute stack splits into 4 layers: Compute, Custom ASIC, Memory, and Networking + Optical. The three hyperscalers’ capex strategies tilt differently across layers, but HBM and optical interconnects are externally sourced by all three. Just as networking mix within NVIDIA shifted +263%, the mix shift across layers is what determines stock selection.

3. Which layer benefits most

Everything above can be assembled from public earnings filings and quarterly reports. Cloud acceleration, capex raises, distribution across 4 layers.

The real differentiation lies in identifying who wins inside each layer. And the shape of the winner differs by layer.

In one layer, a single $5T company dominates the entire layer, and the challenger’s market cap is roughly 1/7th of that. Yet the challenger’s 1Y return is more than 4x higher. Why? Because a smaller market cap base means greater return leverage on the same revenue increment. That is Layer 1.

In another layer, two companies split the market, but one company’s AI revenue exceeds the other’s total revenue. Margin trajectories diverge for reasons directly tied to where design service value is created. That is Layer 2.

In yet another layer, all three companies share the same cycle, but one of them broke through $800B in market cap this week. The best weekly return since December 2008. That is Layer 3.

And in the final layer, the entire layer outperformed, with one stock posting a +1,309% 1Y return. It joins the Nasdaq-100 on May 18. But another stock in the same layer beat earnings and still dropped -14%. Same layer, opposite reactions. Why? That is Layer 4.

One more thing. Following the hyperscaler earnings on April 29, supply chain companies reported earnings throughout the first week of May, and layer-by-layer reactions split sharply. AMD surged +11%. Arista plunged -14%. LITE fell -7% on sector profit-taking despite beating. Micron gained +30% in a single week. All sharing the same $580B capex cycle. Why such divergence?

[Figure 3: 4 layers positioning + post-earnings market reaction]

Summarizing the 4 layers in one line each.

Layer 1 (Compute): NVIDIA + challenger. NVIDIA dominates on market cap. The challenger outperforms on returns.

Layer 2 (Custom ASIC): 2 design partners split the market. The market’s question is margins.

Layer 3 (Memory): 3 companies move together. The most leveraged name in the cycle moves fastest.

Layer 4 (Networking + Optical): Entire layer outperformed. Within it, +1,309% vs. -14%.

The following sections unpack each.

• 10-company identification table across 4 layers

• Market cap and revenue scale comparison

• 1Y return classification by layer

• Layer differentiation + Bull/Gap/Optionality for each company

• Capex scenario (Base / Alternative / Downside) mapping by layer

If you hold NVIDIA, the most critical variable is how fast hyperscalers shift to in-house chips (Layer 2). If you hold Micron, it is whether HBM4 yield ramp can close the gap with SK Hynix. If you hold Arista, it is when supply chain constraints ease. The questions differ by layer, and the answers determine which stock’s return diverges by an order of magnitude within the same $580B capex cycle.

Subscribe now

Read more

LEO mega-constellations are moving from a “deployment phase” to a “service phase.” Amazon Leo (formerly Project Kuiper) began its enterprise preview in November 2025 and confirmed a mid-2026 commercial launch [1][7]. AST SpaceMobile received FCC commercial authorization for its 248-satellite direct-to-device constellation [2]. Starlink operates an active constellation of roughly 10,000 satellites by tracker count [3]. This piece evaluates 26 listed names spread across nine or more countries (North America, Europe, Korea, Japan) using a satellite-comm-cycle-specific 5-axis GVM Score, maps them onto a 4-tier cascade, and distributes them across a Catalyst Calendar through the end of 2027. Identification, mapping, and scenario analysis for the 26 names are covered in the paid section.

This is the second investment map in the PhotonCap series, following Investment Map / 15 Companies in the Glass Substrate Supply Chain. Where the first piece covered the supply chain underneath the AI datacenter core, this one looks at the layer where that infrastructure extends into space. I have been tracking the intersection of SpaceX and optics (OISL, SiPh) for a long time, and as a 2-year RKLB investor, the space cycle remains a personal interest. Recent threads have started to converge into a single thesis: Elon Musk’s mentions of space datacenters, SpaceX’s silicon photonics adoption, and the photonic semiconductor supply chain are now lining up. This piece is an extension of that work, written to push the analysis further and stay invested.

Contents

1. Intro: A Production Cycle in Orbit

2. Technology Background: OISL, Rad-Hard, and InP/SiPh

3. Why the Cycle Is Starting Now

4. Other Applications on the Same Supply Chain

5. GVM Score, Cascade Sequencing, Catalyst Calendar

6. 26 Carriers: Identification and Market Cap Comparison

7. GVM Score: Satellite-Comm-Specific 5-Axis Evaluation

8. Cascade Sequencing: Time-Lag Between Tiers

9. Catalyst Calendar: Trigger Events Through End of 2027

10. Company Cards: Top 5 Deep + 21 Compact

11. Asymmetric Outcome: Scenarios and Monitoring

12. Where the Framework Breaks: Hypothesis-Level Validation Signals

13. Watchlist: Private + Adjacent

14. Conclusion: The Mismatch in the Orbit Timeline

15. References & Sources

1. Intro: A Production Cycle in Orbit

On April 14, 2026, two acquisitions were announced on the same day. Rocket Lab completed its acquisition of OISL terminal company Mynaric for $155.3M [4]. Amazon announced its $11.6B acquisition of satellite-comm operator Globalstar [17]. One internalized a bottleneck. The other locked in spectrum and infrastructure. Same day, same industry, two different layers sealed in.

That same month, names sitting on the LEO satellite supply chain moved like this. +314% off the 52-week low. +280%. +185%. +112%. The names behind those moves are unpacked in this piece.

Since 2020, LEO mega-constellations have mostly been in the “we’re launching satellites” stage. Starlink put dozens of satellites up every month, the market got as far as “oh, satellite internet works,” and that was about it. For a supply chain investor, the layer that mattered was elsewhere. Not the launch vehicle. Not the satellite bus itself. The optical inter-satellite link (OISL) terminal that connects satellites to satellites, and the rad-hard semiconductors that survive the radiation environment. That is the bottleneck of this cycle.

In April 2026, four things landed in the same month. Amazon Leo reconfirmed its mid-2026 commercial launch [7]. Rocket Lab completed its $155.3M acquisition of OISL terminal specialist Mynaric [4]. Amazon announced its $11.6B Globalstar acquisition [17]. The Space Development Agency (SDA) accelerated Tranche 2/3 awards following its Tranche 1 first launch in September 2025 [5]. The transition from “R&D awards” to “repeat production awards” is happening right now.

Core thesis: As LEO mega-constellation transitions from “deployment phase” to “service / production phase” in 2026, the cycle alpha is unlocked not in launch vehicles but in the Tier 2 bottleneck made up of OISL terminals and rad-hard semiconductors. The 5 axes of the GVM Score are Technology Bottleneck, Program Qualification, Backlog-to-Revenue, Launch Manifest Depth, and Multi-Orbit Readiness.

The scope of this piece: 26 listed names exposed to the production cycle, a private watchlist, and supply chain positioning. The 26 names span the US, Korea, France, Germany, the UK, Japan, Italy, Luxembourg, and Canada (nine or more countries), with market caps ranging from roughly $220M to $136B. About 620x apart in size, all exposed to the same cycle.

[Figure 1: LEO satellite supply chain / 4-tier schematic, no company names]

2. Technology Background: OISL, Rad-Hard, and InP/SiPh

Why OISL, Not RF

There are two ways for satellites to exchange data with each other. Radio (RF inter-satellite link) and laser (optical inter-satellite link). Early constellations used RF. Either via ground station hops, or via Ka/Ku-band RF cross-links between satellites.

The bandwidth gap is the issue. RF ISL caps out at hundreds of Mbps, maybe 1 to 2 Gbps at best. OISL today runs at 2.5 to 100 Gbps, and the SDA spec is rising from 2.5 Gbps class to next-generation 10 Gbps [5]. Starlink V2 Mini onward carries OISL terminals. SDA has locked OISL as the standard from Tranche 1.

The difference is not just carrier frequency. It also comes from aperture efficiency, beam divergence, link budget, and pointing stability. RF ISL is constrained by spectrum allocation and antenna aperture, while OISL uses narrow beams and a higher optical carrier to deliver much higher point-to-point throughput. On top of that, no spectrum interference, hard to intercept (extremely narrow beam), and better SWaP (size, weight, power).

What Goes Inside an OISL Terminal

Open up an OISL terminal and there are five core parts.

1. InP laser diode: 1550 nm class light source. Same InP (indium phosphide) base used in terrestrial fiber-comm transceivers

2. Optical receiver / photonic front-end: converts 1550 nm optical signal to electrical. Coherent or direct-detect architectures both possible. Some implementations overlap with SiPh receiver stacks

3. Fiber collimator / FAU (Fiber Array Unit): optical alignment

4. Fine-pointing MEMS mirror or FSM: beam pointing between satellites. Tracks at micro-radian precision over hundreds to thousands of kilometers

5. Rad-hard FPGA/ASIC: the processor that controls all of the above. Operates without SEU (single-event upset) in the space radiation environment

Items 1 and 2 are the ones that matter. The technology stack of InP laser, 1550 nm photonic receiver, fiber coupling, and precision pointing all goes through space qualification and ends up inside the OISL terminal. Coherent, Lumentum, and II-VI (now Coherent) legacy lines build both terrestrial datacenter transceivers and space laser terminals. Different applications on the same supply chain. PhotonCap has been tracking this stack’s SpaceX exposure early. SpaceX’s adoption of laser ISLs was first identified in the community in SpaceX, the Secret Behind the $800 Billion Valuation: Space Connected by Lasers, and Finding Hidden Beneficiaries, and the company’s silicon photonics adoption was analyzed within hours of the relevant job descriptions appearing in SpaceX-xAI: The $1.25T Valuation Built on Silicon Photonics.

Rad-Hard Semiconductors: Why a Separate Layer

Terrestrial semiconductors can be used in some LEO missions with screening, shielding, redundancy, and fault-tolerant design, but mission-critical control electronics often require radiation-tolerant or radiation-hardened parts. High-energy particles (protons, heavy ions) hit transistors directly and cause bit flips (SEU), and accumulated radiation dose (TID) drifts threshold voltage until the chip dies. So rad-hard FPGA/ASIC forms its own market. Microchip Technology’s RT PolarFire and BAE Systems’ RH12 (built on GlobalFoundries 12 nm FinFET) are examples in this market [6].

Characteristics: low-volume production, high margins, long qualification cycles (5 to 7 years), strong customer lock-in. Once a chip gets in, it stays in for the life of the program.

[Figure 2: OISL terminal block diagram / InP laser + SiPh receiver + MEMS mirror + rad-hard FPGA]

[Figure 3: RF ISL vs OISL comparison / bandwidth, SWaP, security, interference]

3. Why the Cycle Is Starting Now

LEO satellite constellation has been a topic for over ten years. Iridium has been around since the 1990s, OneWeb went bankrupt in 2020 and came back, Starlink started launching in 2019. So why is 2026 a turning point from a supply chain investing perspective. Four triggers landed in the same month. Just April.

First, Amazon Leo’s commercial launch reconfirmation (April 9). After rebranding from Project Kuiper to Amazon Leo and starting an enterprise preview in November 2025, Amazon CEO Andy Jassy’s April 9, 2026 shareholder letter reconfirmed the mid-2026 commercial launch and revenue commitments from enterprise, government, and telco customers [7]. Amazon disclosed that its Kirkland satellite manufacturing facility was designed for peak capacity of 5 satellites per day production (annual capacity of roughly 1,000 satellites, author estimate) [7]. Launches use ULA Atlas V, Arianespace Ariane 64, Blue Origin New Glenn, and even SpaceX Falcon 9, and after the LE-02 launch in late April 2026, roughly 302 production satellites are in orbit [8]. Amazon has filed for a 2-year extension of its FCC first-deadline (1,618 satellites by July 30, 2026, half of the constellation) to 2028.

What this means for the supply chain: once Amazon goes into a monthly launch cadence, repeat demand emerges across satellite component classes such as OISL, rad-hard electronics, phased-array antennas, solar arrays, and propulsion. The shift from one-off R&D purchases to repeat production demand is the starting point of the supply chain investing thesis.

Second, Rocket Lab’s Mynaric acquisition completed (April 14). That same week, Rocket Lab completed its acquisition of Munich-based OISL terminal specialist Mynaric AG [4]. $155.3M in aggregate consideration (cash plus 2,277,002 RKLB common shares). Mynaric’s CONDOR Mk3 terminals were already going into Rocket Lab’s SDA Tranche 2 Transport Layer-Beta program as the OISL component, and this acquisition internalized that capability. Rocket Lab’s SDA position now consists of Tranche 2 Transport Layer-Beta 18 satellites ($515M, January 2024) plus Tranche 3 Tracking Layer 18 satellites ($816M, December 2025), totaling $1.3B+ in SDA prime contracts [4].

This acquisition reshaped the OISL market structure. There were four SDA-qualified OISL terminal suppliers: Mynaric, Tesat-Spacecom US, Skyloom, and CACI [9]. With Mynaric absorbed into Rocket Lab, the count of independent OISL suppliers dropped by one, and Rocket Lab became the only listed company that vertically integrates launch, spacecraft, and OISL.

Third, Amazon’s Globalstar acquisition announcement (April 14). That same day, Amazon announced its $11.6B acquisition of satellite-comm operator Globalstar at $90 per share (or 0.3210 Amazon shares) [17]. A 117% premium to the late October 2025 close. The deal absorbs Globalstar’s L-band and S-band spectrum licenses, 24 operational LEO satellites, and 24 global ground gateways in one move. The company that was the backbone of Apple iPhone Emergency SOS (with 85% of Globalstar capacity reserved for Apple) now becomes the direct-to-device infrastructure for Amazon Leo. Closing expected in 2027. Bloomberg reported that Amazon stated it “will launch a direct-to-device service” in 2028 [19].

The implication: Amazon is signaling entry into the D2D (direct-to-device) market alongside LEO broadband. With spectrum, satellites, and ground stations secured at once, additional demand emerges for D2D supply chain components (antennas, beam-forming ASICs, rad-hard semiconductors). The market read this deal as a commitment to both satellite broadband and D2D cellular.

Fourth, SDA Tranche 1 first launch + Tranche 2/3 acceleration. The Space Development Agency, under US Space Force, is building a LEO military mesh network (Proliferated Warfighter Space Architecture). Tranche 1 first 21 satellites launched on September 10, 2025 (built by York Space Systems, on SpaceX Falcon 9) [5]. Tranche 1 total (150+ satellites) runs as a 10-month launch campaign, and Tranche 2 Transport Layer Beta first launch is scheduled for September 2026 (ILC). Tranche 3 Tracking Layer was awarded December 19, 2025: $3.5B / 72 satellites across four primes [5]. However, Tranche 3 Transport Layer is paused under Pentagon review. SDA is the anchor buyer of OISL demand.

Additional: AST SpaceMobile FCC authorization plus launch incident (same week). On April 21, 2026, the FCC issued AST SpaceMobile authorization for SCS (Supplemental Coverage from Space) and direct-to-device commercial service across its 248-satellite constellation [2]. The deadlines: 124 satellites by August 2, 2030; full 248 by August 2, 2033. But two days before the authorization, on April 19, BlueBird 7 was placed in a lower-than-planned orbit on the third Blue Origin New Glenn flight after a second-stage burn anomaly, and was classified as expected to de-orbit [10]. AST announced an expected $30M insurance claim, and the FAA placed New Glenn under mishap classification (grounded). Both the bull case and the risk of the D2D cycle landed in the same week.

April was the month LEO supply chain shifted from a simple theme to a contract-driven cycle. Amazon Leo commercial launch reconfirmation, the Globalstar acquisition, the Rocket Lab-Mynaric acquisition, the imminent SDA Tranche 2 ILC, and the ASTS FCC authorization plus launch incident all landed simultaneously.

[Figure 4: April 2026 Cycle Trigger Map / five events, four triggers plus one mishap]

4. Other Applications on the Same Supply Chain

If you look at the satellite-comm cycle as just LEO broadband, the picture shrinks. There are at least four applications sitting on the same OISL, rad-hard, and launch supply chain.

1. LEO broadband (Starlink, Amazon Leo), now. The largest volume driver. Starlink active constellation roughly 10,000 satellites, Amazon Leo 3,236 satellites planned. These two constellations create the production demand for OISL, antennas, solar panels, and propulsion.

2. Direct-to-Device (AST SpaceMobile, Amazon/Globalstar, Starlink Direct-to-Cell), 2026 and later. Satellite broadband through existing smartphones. With Amazon’s $11.6B Globalstar acquisition, a hyperscaler entered the D2D market directly [17]. Globalstar’s L+S-band spectrum plus Amazon Leo’s constellation equals D2D infrastructure. AST SpaceMobile is going for broadband-class D2D independently. SpaceX secured its own D2D spectrum through a $17B EchoStar AWS-4 + H-block acquisition in September 2025 [15]. Demand for antennas and beam-forming ASICs.

3. Defense mesh networks (SDA, IRIS²), 2026 to 2030. The US SDA’s Proliferated Warfighter Space Architecture plus Europe’s IRIS² (Infrastructure for Resilience, Interconnectivity and Security by Satellite). SDA Tranche 2 ILC is September 2026. IRIS² signed a €10.6B 12-year concession contract with the SpaceRISE consortium (led by SES, Eutelsat, Hispasat) on December 16, 2024, covering a 290-satellite multi-orbit constellation, first launch in 2029, services starting 2030 [20]. For defense use, OISL is required (anti-intercept and low-latency). The largest customer for rad-hard semiconductors.

4. Cislunar / deep space, 2028 and beyond. NASA Artemis, Lunar Gateway, deep space relay. Volumes are small but unit pricing is extremely high, and flight heritage stacks up here.

The baton passes from “LEO broadband to D2D cellular to defense mesh to cislunar / deep space.” The first application creates volume; the next application expands TAM.

The key point: the same exact product is not repeated, but the technology stack of InP laser, coherent receiver, fine-pointing optics, and rad-hard control electronics gets reused across LEO broadband, SDA mesh, IRIS², and cislunar relay as the baton passes. Companies with high Technology Bottleneck scores have a high probability of carrying their qualification and flight heritage forward to the next program even when the first application changes.

5. GVM Score, Cascade Sequencing, Catalyst Calendar

The market saw the LEO satellite cycle as a “theme.” ASTS rallied, RKLB rallied, the “space-related stock” label was traded back and forth. But when the cycle moves from “deployment phase” to “production phase,” theme pricing breaks. Spread widens between companies with the same LEO exposure but different analytical quality.

The paid section of this piece applies three frameworks simultaneously to the 26 names.

Framework 1: GVM Score (Global Value Map, 5-axis). Each company is scored 1 to 5 across Technology Bottleneck, Program Qualification, Backlog-to-Revenue, Launch Manifest Depth, and Multi-Orbit Readiness. Total range 5 to 25. The exact difference between a 1 and a 5 on each axis, and which axes carry more weight by cycle phase (phase weighting), are unpacked in the paid section.

Framework 2: Cascade Sequencing (4-tier with time-lag). Tier 1 (Launch) to Tier 2 (OISL + Rad-Hard) to Tier 3 (Satellite OEM) to Tier 4 (Service / Terminal). One feature: this cascade is closer to parallel than sequential. Hyperscalers (Amazon) are placing orders into Tier 2 and Tier 3 simultaneously, and SDA is overlapping Tranche awards. Where this parallel structure creates alpha across tiers is the core analysis of the paid section.

Framework 3: Catalyst Calendar (through end of 2027). Catalyst events landing roughly 19 months from 2026 Q2 to 2027 Q4 are plotted on a timeline. Amazon Leo commercial service, Neutron maiden flight, SDA Tranche 2 ILC, IRIS² prime contractor selection, ASTS commercial service ramp, Amazon-Globalstar deal closing, SDA T2 Gamma launches. Which company’s catalyst lands when, assembled.

This is as far as public materials alone get you.

The real difference starts here. What score each of the 26 names received on each of the 5 axes, and how the ranking flips when phase weighting is applied. Where the valuation gap sits across the four tiers. And one structural inefficiency the market has not priced in yet.

If you split the 26 into four groups:

• “Vertically integrated launch”: launch + spacecraft + OISL inside a single company

• “OISL bottleneck holder”: a handful of companies supplying flight-proven OISL terminals globally

• “Satellite production prime”: large defense / space OEMs producing satellite buses for SDA / commercial constellations at scale

• “D2D service bet”: early revenue or pre-commercial-scale. Optionality on service ramp

The market gave a premium to “vertically integrated launch” and “D2D service bet.” RKLB sits at P/S of roughly 93x (TTM revenue $602M, market cap $56B). ASTS sits at hundreds of x P/S on FY25 revenue. But within the “OISL bottleneck holder” group, there is a name that is GVM top tier yet still priced at defense-IT multiples by the market. Its 1-year return diverges sharply from other names in the same GVM cluster. The market has not yet separated OISL scarcity as a distinct valuation factor.

And there is one more Hidden Card. Among the 26 names, there is one with a high TB (Technology Bottleneck) score but with space revenue under 5% of total, so the market reads it purely as a “terrestrial semiconductor cycle recovery” play. That name’s space revenue ramps when constellation production volume kicks in around 2027 to 2028. The time gap between the terrestrial semiconductor cycle and the space semiconductor cycle is creating its mispricing.

What these names are, why the TB x PQ combination is the cycle alpha, and how phase weighting flips the ranking, are unpacked in the paid section.

[Figure 5-A: Positioning Matrix / Hook Version (dots only, no names, CACI position with dashed amber Hidden Card ring)]

Subscribe now

Read more

For the past 30 years, high-performance CPU/GPU package substrate stacks have been built almost entirely on organic build-up substrates, specifically ABF (Ajinomoto Build-up Film) variants. 2026 is the first year glass substrates cross from R&D into pilot/qualification phase. Intel has committed to a “second half of this decade” production window [1], TSMC is reportedly preparing its CoPoS pilot line for mid-2026 completion [2], and Absolics has entered AMD certification at its Georgia Covington fab [3]. This article evaluates 15 publicly traded companies across six countries using a glass-substrate-specific GVM Score (5-axis framework), maps them to a 4-tier cascade, and distributes them across a 12-month Catalyst Calendar. Identification, mapping, and scenario analysis for all 15 companies are covered in the paid section.

This article was prepared in collaboration with and . The GVM Score is a framework I built for my own investment decisions. Of the 15 companies analyzed, I held existing positions in 3 before writing, and initiated positions in 2 additional companies immediately after constructing the GVM table. Which companies those are is disclosed inside the article.

Contents

1. Intro: The Substrate Is Changing for the First Time in 30 Years

2. Technical Background: Organic vs Glass, and TGV

3. Why the Cycle Starts Now

4. Multiple Applications on the Same Supply Chain

5. GVM Score, Cascade Sequencing, Catalyst Calendar

6. 15 Carriers: Company Identification + Market Cap Comparison

7. GVM Score: Glass-Substrate-Specific 5-Axis Evaluation

8. Cascade Sequencing: Time-Lag Between Tiers

9. Catalyst Calendar: Trigger Events Over the Next 12 Months

10. Company Cards: Top 5 Deep + 10 Compact

11. Asymmetric Outcome: Scenarios and Monitoring

12. Where the Framework Breaks: Validation Signals by Hypothesis

13. Watchlist: Unlisted + Adjacent

14. Conclusion: The Substrate Timeline Mismatch

15. References & Sources

1. Intro: The Substrate Is Changing for the First Time in 30 Years

One-year returns. +850%. +675%. +554%. +341%. +328%. These are stocks sitting on the glass substrate supply chain. The lowest of the 15 still returned +10%. Over the same period, SMH (semiconductor ETF) returned roughly +30%. Which companies these are unfolds in this article.

For 30 years, the substrate stack underneath high-performance CPUs, GPUs, and server accelerators has been built mostly on organic build-up substrates, particularly ABF. 2026 is the first year glass substrates cross from R&D into pilot/qualification phase. Volume ramp follows in 2027 to 2030.

The physics make sense. AI accelerator packages now exceed 100mm x 100mm, and organic substrates cannot handle the warpage. CTE (coefficient of thermal expansion) mismatch with silicon causes cracking under thermal cycling. Glass can be CTE-matched to silicon, has an order of magnitude lower dielectric loss, and can be processed in large panels [4].

But the problem is not just one material. When one layer of the substrate stack changes, every process above and below it changes too. Desmear equipment built for organic, laser drilling built for organic, plating built for organic, inspection built for organic. All of it has to be rebuilt for glass. Fifteen companies sit at the chokepoints of this cycle. They span South Korea, Japan, Germany, the United States, Taiwan, and Austria. Market caps range from $483M to $157B, a 325x spread, yet all of them are exposed to the same cycle.

Core thesis: The AI accelerator substrate transition from organic build-up to glass begins with a 2026 pilot/qualification phase and extends into a 2028 to 2030 volume ramp window. Cycle alpha is not captured by generic financial metrics (margin, valuation) but by five glass-substrate-specific axes: process bottleneck control, customer qualification depth, capex recovery visibility, order pipeline depth, and multi-application readiness.

This article covers 15 publicly listed companies exposed to the production cycle, an unlisted watchlist, and Chinese competitor developments.

[Figure 1: Glass substrate value chain / 5-layer schematic, no company names]

2. Technical Background: Organic vs Glass, and TGV

To understand why glass substrates replace organic in one view, look at three dimensions.

First, flatness and warpage. Organic resin warps in large packages. As AI accelerator die areas exceed the reticle limit (858mm²) into 2-reticle and 3-reticle territory, organic substrates face increasing warpage, interconnect reliability degradation, and joint crack risk under thermal cycling. Glass has an order of magnitude lower surface roughness and near-constant warpage behavior [4].

Second, dielectric properties and signal integrity. When AI accelerators run SerDes above 100GHz, substrate dielectric loss converts directly into wasted power. Glass has an order of magnitude lower dielectric loss than organic, delivering the same bandwidth at lower power [4].

Third, panel-level processing. Silicon interposers are wafer-based, limited to 300mm. Glass can be processed in large panels (e.g., 510mm x 515mm), offering several times more usable area than a 300mm wafer. This is why TSMC is pursuing CoPoS (Chip-on-Panel-on-Substrate) [2].

But glass has one big problem. You cannot drill vias through it. More precisely, you cannot mechanically drill glass like organic or plasma-etch it like silicon. Glass is brittle. A single micro-crack kills yield immediately.

This is where a new process category emerged. TGV (Through-Glass Via) processing. A laser selectively weakens a zone inside the glass, then a chemical etch removes only that zone to form a clean via. LIDE (Laser-Induced Deep Etching), LMCE, femtosecond laser processes, and other variants all fall into this category [5].

The cycle’s bottleneck shows up in TGV and metallization before it shows up in raw glass panels. The number of companies worldwide that can build TGV drilling equipment can be counted on one hand. The same goes for the RDL (redistribution layer) lithography and bonding equipment that follows the drill. This is why small-cap companies receive large beta in this cycle.

Supply chain layers:

1. Glass material layer: global glass panel makers (3 listed + Schott AG, unlisted)

2. TGV + process equipment layer: via drilling, RDL lithography, bonding, and inspection equipment makers

3. Substrate fabricator layer: companies producing finished glass substrates for customers

4. Adjacent substrate layer: incumbent organic substrate leaders exposed to the glass transition

5. Customer / Platform layer: Intel, TSMC, AMD, NVIDIA, Broadcom

Layer 2 is critical. It is not just TGV drilling. After via formation comes metallization (seed layer + plating), RDL patterning (lithography), glass-to-glass bonding, and inspection. Watching only the TGV drilling equipment maker means missing half of Tier 2. (A separate deep-dive on the leading TGV equipment company in this cycle is available in The Glass Beneath AI Chips.)

[Figure 2: Organic vs Glass substrate cross-section / TGV process flow (drilling, metallization, RDL, bonding, inspection)]

For a deeper look at the equipment and platform positions in the glass substrate value chain, see

For the full picture of why AI accelerator power density is reshaping packaging architecture, from grid to transistor across six orders of magnitude in voltage, see Nutty's article:

Nutty

The AI Power Crisis — Part 4

In Part 1, we followed electricity from power plants into the transmission grid. DC, which lost the current war 140 years ago, is coming back as power semiconductors mature…

Read more

a month ago · 46 likes · 1 comment · Nutty

3. Why the Cycle Starts Now

Glass substrate R&D has been running for over a decade. Intel has had its ATTD (Assembly and Test Technology Development) team working on glass substrates since 2014 [6]. So why is 2026 the inflection point? Three triggers landed in the same quarter.

First, Intel’s production commitment. After the September 2023 announcement, the market doubted Intel’s follow-through. A rumor circulated in mid-2024 that Intel had abandoned glass substrates [7]. Then Intel reaffirmed the commitment in September 2025, and at NEPCON Japan in January 2026, it demonstrated a thick-core glass substrate combined with EMIB [8]. The package measured 78mm x 77mm, roughly 1,716mm² (2 x 858mm²) of silicon area. That is a package exceeding a single reticle limit.

Second, TSMC’s CoPoS acceleration. TSMC had been constrained by organic substrates in its CoWoS architecture. Starting in late 2025, TSMC pulled forward the CoPoS (Chip-on-Panel-on-Substrate) roadmap. The CoPoS pilot line is reportedly targeted for mid-2026 completion. Volume production expectation is 2028 to 2029 [2]. As NVIDIA Rubin/Blackwell architectures are increasingly likely to sit on CoPoS, glass demand will absorb a portion of existing organic substrate maker backlogs at some point.

Third, Absolics AMD sampling. As of January 2026, Absolics had completed major equipment installation at its Georgia Covington fab and began supplying mass-production samples to AMD [3]. Korean reporting indicates that the exclusive TGV contract with F&S Electron expired in March 2026, and Absolics is transitioning to dual sourcing [9]. This signals a shift from R&D samples to customer qualification and early commercial sampling.

These three triggers landing simultaneously led the market to frame 2026 to 2027 as the “qualification phase” and 2027 to 2028 as the start of high-volume manufacturing [10].

Two caveats. First, this cascade may not be sequential. If hyperscalers place simultaneous orders at Tier 2 and Tier 3 without waiting for lag, the sequential cascade turns into a parallel ramp. This is a key validation point in the paid cascade analysis.

Second, the P in CoPoS does not necessarily mean glass. TSMC’s official language is “large-area panel.” The initial panel could be a continuation of silicon interposer or glass substrate. Whether glass wins on cost efficiency for panel substrates, or silicon returns in a large panel format, is the single biggest fork in the cycle thesis.

[Figure 3: Glass substrate cycle timeline / 2023 to 2030]

4. Multiple Applications on the Same Supply Chain

Looking at the glass substrate cycle through only one application, AI accelerator substrates, makes the picture too small. At least three more applications are queued on the same TGV processes, the same panel processing infrastructure, and the same glass panels.

HBM4 glass interposer. HBM currently sits on silicon interposers. For HBM4, stack height, warpage, interposer routing density, and thermal-mechanical reliability are all simultaneously critical. Glass has lower thermal conductivity than silicon, so its adoption logic is not heat spreading superiority but rather CTE tuning, flatness, low-loss routing, and panel scalability [10]. Market participants are discussing HBM4 glass interposer adoption beginning in 2028.

Co-Packaged Optics (CPO). CPO places optical transceivers within sub-millimeter distance of a switch ASIC. Glass opens the door to processing electrical RDL and optical waveguides on the same substrate, and is discussed as a platform with potential advantages in cost, panel scalability, and assembly yield relative to silicon photonics interposers [13]. CPO commercial deployment discussions are expected to accelerate from 2026 to 2027, but mainstream production of glass-based integrated electro-optical substrates may come later.

Photonic integration. Mounting both EIC (Electronic IC) and PIC (Photonic IC) on a single package is the next stage. A glass interposer carrying both EIC and PIC could bring optical fiber connectivity all the way into the substrate. Production is further out, but first commercial products are possible between 2028 and 2030 [13].

These three applications queued on the same glass supply chain have one implication. Even if the AI accelerator substrate cycle cools, the glass supply chain passes the baton to the next application. This is a multi-application supply chain cycle, not a single-application one.

This is the biggest difference from the optical semiconductor deposition cycle. That cycle’s beta was driven by a single application, optical communications, and when the application cycle cooled, the beta faded with it. Glass substrates have applications queued, which extends supply chain duration.

Key takeaway: The glass substrate supply chain is a multi-application supply-chain beta, not a single-application cycle. AI accelerator substrates are the first application. HBM4 interposer, CPO, and photonic integration are queued next.

[Figure 4: Multi-application supply chain stack on glass substrate]

5. GVM Score, Cascade Sequencing, Catalyst Calendar

The market has been evaluating the glass substrate cycle using generic financial metrics: margin, valuation, P/E. But when a cycle is crossing from R&D into commercialization, generic metrics misjudge it. Alpha in the glass substrate cycle resolves across five axes: process bottleneck control, customer qualification depth, capex recovery visibility, order pipeline depth, and multi-application readiness. The evaluation dimensions are not revenue share or ASP times volume, but how far yield has progressed, how deep customer certification runs, and whether capex can be recovered.

The paid section applies three frameworks simultaneously to all 15 companies.

Framework 1: GVM Score (Glass Value-chain Maturity, 5-axis). Each company is scored 1 to 5 on five dimensions. Raw total ranges from 5 to 25. The five axes are Process Bottleneck, Customer Qualification, Capex Recovery, Order Pipeline, and Multi-App Readiness. What exactly each axis measures, how 1-point differs from 5-point, and how phase weighting shifts the ranking are covered in the paid section.

Framework 2: Cascade Sequencing (4-tier with time-lag). The glass cycle is not a single tier. Tier 1 (Material) feeds Tier 2 (Equipment) feeds Tier 3 (Fabricator) feeds Tier 4 (Adjacent/Customer), with 6 to 12 month lags at each step. Where we are now determines alpha.

Framework 3: Catalyst Calendar (catalyst event timeline). Catalyst events expected within 12 months are plotted on a timeline, identifying which company’s catalyst lands when.

Everything up to this point is visible from public information alone.

The real differentiation starts here. How each of the 15 companies scores on five axes, and how rankings shift when phase weighting is applied. Where the valuation gap sits across the four tiers. There is a GVM top-tier company that the market still prices as an organic stock rather than a glass substrate stock. And there is a mid-tier company on raw score whose ranking jumps when phase weighting is applied.

One market inefficiency stands out. A GVM top-tier company is being priced by the market purely on organic stock multiples, producing the widest price spread within its GVM cluster. Which company that is unfolds in the paid section.

[Figure 5: Glass substrate cycle positioning matrix / Y-axis stake size, X-axis substrate exposure, no company names]

Subscribe now

Read more

Contents

1. 100 Seconds per PIC: Why CPO Test Is the Bottleneck

2. Background: The Physics of Why CPO Testing Is Hard

3. The 4 Insertion Framework and the Key Questions (paywall hook)

4. Insertion 1: PIC Wafer-Level Test

5. Insertion 2: EIC-PIC Combined Wafer-Level Test

6. Insertion 3: Optical Engine-Level Test, ficonTEC vs Chroma

7. Insertion 4: Module/System-Level Test

8. Burn-In: The Mandatory Step Outside the 4 Insertions

9. Two Competing Axes: Bull, Gap, Optionality

10. Scenario Analysis

11. Monitoring Points

12. Summary Table + Closing

13. References & Sources

1. 100 Seconds per PIC: Why CPO Test Is the Bottleneck

Per a recent TrendForce report, a full 100% optical inspection of a single PIC (photonic integrated circuit) going into a CPO module takes over 100 seconds on average.[1]

100 seconds. That is an order of magnitude longer than conventional semiconductor test. Over the past year, as this bottleneck became visible to the market, the equipment companies behind it moved at the opposite speed. One-year stock returns:

$AEHR +1,033%, Chroma (2360.TW) +757%, Advantest (6857.T) +389%, $FORM +468%, $TER +389%, $KEYS +150%. (as of 2026-05-06 close, price return basis, Yahoo Finance/Google Finance/Investing.com)

Market cap scales vary wildly. Advantest $128B, Keysight $62B, Teradyne $56B. These three are large-cap semiconductor equipment names. FormFactor sits at $11.6B, Aehr at $2.9B. ficonTEC is private (Germany-based, subsidiary of China’s Robo Technik). Chroma is listed in Taiwan at roughly TWD 979B (approximately $30B). Within the same CPO test ecosystem, market caps span orders of magnitude.

TSMC’s COUPE platform is entering CPO chip production in 2026.[1] NVIDIA invested $2B in Lumentum[2] and $2B in Coherent[3] to secure laser and optical networking capacity. A separate $2B strategic investment in Marvell covers NVLink Fusion, custom XPU, and silicon photonics/optical interconnect.[4] Across these three deals, $6B in AI connectivity and optics supply chain investment. Every one of these optical components needs to pass testing before it hits a production line.

Testing is becoming the single largest bottleneck in CPO mass production, and a competitive landscape is forming rapidly among equipment companies racing to solve it.

For context, consider the front-end equipment large caps. AMAT $327B, LAM $372B, KLA $236B (as of 2026-05-06). Against this group, CPO test companies sit at a distinctly different scale. Keysight ($62B), Teradyne ($56B), and Chroma (~$30B) are large test and measurement companies, but still 5 to 12 times smaller than front-end equipment leaders. FormFactor ($11.6B) and Aehr ($2.9B) are one to two orders of magnitude smaller. If CPO test grows into a multi-billion-dollar annual market comparable to front-end equipment, the market cap gap for pure-play smaller names like FormFactor and Aehr could be due for reassessment. The opposite is also possible. If CPO production gets delayed, test equipment capex gets deferred with it.

FormFactor reported Q1 2026 revenue of $226.1M (+32% YoY), beating guidance, and raised its 2026 CPO revenue guide to the high end of the $10-20M range on its earnings call.[5][6] Teradyne reported Q1 2026 revenue of $1.282B (+87% YoY), with AI-related revenue accounting for roughly 70% of the total.[7] Aehr disclosed a record $41M order from a hyperscale AI customer in April 2026 (for custom AI processor ASIC package-level burn-in).[8] None of these data points existed a quarter ago.

Bottom line: 1Y returns for the 7 CPO test equipment companies range from +150% to +1,033%. Market caps range from $2.9B to $128B. Compared to front-end equipment large caps (AMAT/LAM/KLA at $236B to $372B), pure-play names like FormFactor ($11.6B) and Aehr ($2.9B) are one to two orders of magnitude smaller. The speed at which NVIDIA’s $6B AI connectivity investment converts to actual production volume will define the next 12 months for these companies.

2. Background: The Physics of Why CPO Testing Is Hard

Start with why it takes 100 seconds.

A single-mode fiber core has a cross-sectional area of roughly 78.5 square micrometers. A silicon photonic strip waveguide is roughly 0.099 square micrometers. That is an 800x difference. Bridging this gap requires nanometer-precision alignment to couple light in and out of the device under test. Much of this process is still partially manual. There is no unified industry standard.

Conventional semiconductor test only deals with electrical signals. Touch the probe card to the pads and you are done. SiPh/CPO test requires simultaneous electrical and optical probing. You drive the modulator with electrical signals while simultaneously coupling light through an optical coupler, then receive the modulated optical signal on the output side. This requires six or more axes of precision positioning, with alignment tolerances in the tens of nanometers.

Add 3D hybrid bonding (the TSMC COUPE architecture) and it gets worse. When PIC and EIC are bonded face-to-face, the optical coupler ends up on the bottom side of the wafer. Electrical contacts are on top. You need to probe from both sides simultaneously. Existing semiconductor test equipment cannot do this.

From the perspective of established equipment companies, this is a completely different technology stack from the deposition, etch, and metrology tools that AMAT, LAM Research, and KLA ($200B to $370B market caps) dominate. In conventional wafer test, probe card pin placement accuracy at the micrometer level is sufficient. SiPh optical coupler alignment tolerance is in the tens of nanometers, roughly 100 times more stringent. Factor in thermal expansion from temperature changes (silicon CTE is roughly 2.6 ppm/K) and you cannot simply bolt an optical module onto an existing probe station. You need equipment designed from scratch for optical alignment.

[Figure 1: Electrical vs Optical Probing Comparison, Mode Size Mismatch]

There is another issue. Conventional electrical test has high repeatability. Align the probe card once and you can automatically test the entire wafer. In optical test, coupling conditions change from die to die. Waveguide position, edge coupler angle, and grating coupler pitch all carry process variation at the die level, so optical alignment must be independently optimized for each die. This accounts for a significant portion of the 100 seconds. FormFactor’s claim of sub-5-second per-die test time on the CM300xi-SiPh is a data point showing progress in alignment automation.[9] That said, TrendForce’s “100 seconds” and FormFactor’s “sub-5 seconds” likely refer to different test content scopes. The 100 seconds appears to cover full optical inspection, while the 5 seconds likely refers to alignment plus basic parameter measurement for a specific recipe. Direct comparison requires caution.

This is why the CPO test market is being driven not by the existing equipment large caps, but by these seven companies with specialized capabilities in photonic probing, optical alignment, and burn-in.

Previous PhotonCap and articles covered this technical background in detail:

• Why SiPh wafer test is hard, and how AEHR, FormFactor, and Keysight each cover a different layer: “The Three Pillars of SiPh Wafer Test”[9]

• Three signals from FormFactor’s Q1 2026 earnings (SK hynix 29.5%, NVIDIA 10.2% first 10% disclosure, CPO revenue high-end guide), plus TRITON production entry, Keystone acquisition, Teradyne Photon 100 competition: “FormFactor Q1 2026”[5]

• F2C assembly companies in CPO optical packaging: PC101 Lecture 4 Part 2[10]

• Seven bonding equipment companies accelerating across HBM4 and CPO cycles: “Seven Bonding Equipment Companies”[11]

Damnang’s Substack

Why You Should Be Watching Optical Test Right Now

Since publishing the Optical Investment Map, I’ve heard from a lot of readers. The common thread: the optical theme has already run up significantly, and it was hard to know which names to look at or how to approach them. The investment map helped by laying out the full value chain and the competitive dynamics at each layer. What I found most rewarding was hearing that people used it not just to pick individual stocks, but to build their own investment strategy and portfolio…

Read more

23 days ago · 57 likes · 6 comments · Damnang

This article builds on that technical foundation. The scope here is mapping the 4-stage CPO test insertion structure and identifying which companies occupy which roles at each stage.

One more piece of context. CPO test market growth is not simply proportional to CPO chip shipments. Longer test times mean more equipment is needed. Testing 1 million PICs at 100 seconds each requires 100 million seconds, roughly 27,778 equipment-hours. At 24-hour operation and 85% uptime, a single test cell processes roughly 734 PICs per day, or about 22,000 per month. 100,000 per month requires roughly 5 single-site test cells. 1 million per month requires roughly 45. Add retesting, multiple insertions, and multi-engine architectures, and actual equipment demand exceeds these figures. The equipment competition is a test time compression competition. The company that cuts 100 seconds to 50, then to 25, can serve the same customer volume with less floor space and capex. This is why FormFactor emphasizes per-die test time reduction and ficonTEC pushes multi-site parallel testing. Throughput competition is the equipment market competition.

Bottom line: The fundamental reasons CPO test is hard are simultaneous electrical plus optical probing, nanometer alignment, and double-sided access. 100 seconds of test time per PIC directly determines how many production test cells are needed. These technical requirements open the market to specialized companies rather than the existing equipment large caps.

3. The 4 Insertion Framework and the Key Questions

CPO testing is divided into four test insertions, from wafer to system.[12]

[Figure 2: CPO Test 4 Insertion Flow Diagram]

Outside this four-stage framework, there is one more step. Burn-in. This is the process step that screens for infant mortality in InP lasers, performed before Insertion 1.

Everything up to this point is visible from public data and the TrendForce report. Four insertions, plus burn-in as a separate step. You can figure this out from industry conference slides and equipment company press releases.

The real differentiation starts here.

These seven companies form two distinct competing axes within the four insertion stages, with the remaining companies classified as cross-axis layer players that are not tied to either axis. The technical strengths and weaknesses of each axis are precisely inverted.

As a subscriber asked, are ficonTEC and Chroma competitors or coexisting at Insertion 3? The answer becomes clear once you understand the two-axis structure.

One more thing. One of these seven companies has a Chinese ownership structure. With CPO test becoming a choke point in production, how this structural risk could influence customer vendor selection is covered in the paid section below.

The paid section below breaks down the test station structure for each insertion and maps Bull/Gap/Optionality for each axis. How each axis’s technical strength can flip into a weakness under specific conditions, and why cross-axis layer players remain relatively stable regardless of scenario, with quantitative data.

Do the order-of-magnitude differences in market cap among these seven companies reflect their actual positioning in the CPO test market, or is the market still mispricing them? The answer requires understanding the insertion-by-insertion landscape first.

One preview: the most interesting point of competition between the two axes is not technology. It is corporate structure. How one axis’s key partner’s ownership structure could distort customer vendor selection may be the most underpriced variable in the investment equation. Especially in the current environment where US-China technology competition is extending into the semiconductor equipment supply chain.

Subscribe now

Share

Read more

Contents

1. Intro: NVIDIA’s fourth photonics bet

2. Technical context: the path of light in AI infrastructure

3. What looks the same when you group the four, and what splits them apart

4. NVIDIA’s four photonics bets: identification + order-of-magnitude comparison

5. Four-group classification + differentiation + Corning deep dive

6. Scenarios + monitoring + closing

7. References & Sources

1. Intro: NVIDIA’s fourth photonics bet

NVIDIA’s fourth photonics bet was not a laser. It was a road.

Coherent, Lumentum, and Marvell are companies that generate, process, and interface light with electrical signals. Corning is the company that locks in the physical road that light actually travels on. If the next bottleneck in AI clusters is not just transceivers but also fiber, connectors, and FAUs, then GLW is not just a glass company. It becomes the connectivity toll road of AI infrastructure.

NVIDIA made four direct investments into photonics companies in 2026. In chronological order:

• 2026-03-02: Coherent ($COHR) $2B equity, Lumentum ($LITE) $2B Series A Convertible Preferred. PhotonCap covered both [Internal Reference 1].

• 2026-03-31: Marvell ($MRVL) $2B Series A Convertible Preferred + NVLink Fusion integration. Also covered by PhotonCap [Internal Reference 2].

• 2026-05-06: Corning ($GLW) $500M warrant deal + multi-year commercial partnership.

1-year stock total returns (as of 2026-05-06, market data):

• $LITE: roughly +1,326%

• $COHR: roughly +395%

• $GLW: roughly +305%

• $MRVL: roughly +206%

Here is what stands out. The first three are in the photonics layer that either generates light (Lumentum’s EML and lasers), processes and transmits it (Coherent’s advanced optics and optical networking products), or bridges electrical signals and the optical interconnect (Marvell’s DSP / custom silicon + silicon photonics). The fourth one, Corning, makes the physical layer that the light actually flows through. Optical fiber, cable, and connectors.

The thesis in one line: the connectivity layer is as critical as the transceiver layer in AI infrastructure. NVIDIA’s fourth bet landing on that layer is not a coincidence.

This article looks at how NVIDIA’s four photonics investments occupy different layers, and why Corning sits in the same cycle as the transceiver pure-plays.

2. Technical context: Corning is the vascular system of AI data centers

How data moves inside a GPU cluster

If you trace how data moves inside a GPU cluster, you can see exactly where each of the photonics companies that NVIDIA invested in sits.

Electrical signals from the GPU pass through SerDes, DSP, and electrical-to-optical interconnect (Marvell’s territory: custom silicon + silicon photonics platform). Then a laser or EML converts them to light (Lumentum and Coherent territory). Light is modulated through a silicon photonics modulator, loaded onto the transceiver, and travels along optical fiber (Corning territory). On the other GPU, the path runs in reverse, back to electrical signals.

[Figure 1: AI data center optical signal path diagram, with the 4 NVIDIA-invested companies positioned on each stage]

The four companies NVIDIA directly invested in each occupy one of these stages. The transceiver/silicon layer (Coherent, Lumentum, Marvell) generates, controls, and interfaces with light. The connectivity layer (Corning) carries it.

The metaphor: Corning is the vascular system

Map an AI data center to the human body and the layers line up like this:

• GPU: brain (computation)

• Transceiver / Laser / DSP: heart (pumps electrical signals onto the optical medium)

• Optical fiber + cable + connector: blood vessels (the conduit that carries light)

A great heart is meaningless if the vessels are blocked or insufficient. A 1.6T transceiver is meaningless if the fiber connecting it to another GPU node is missing. The cluster simply stops. And the number of fiber connections an AI data center needs is roughly 10x what a typical cloud data center needs. Corning’s own materials state that GenAI data centers require at least 10x more fiber connections than legacy networks.

Why connectivity matters, in numbers

A 16,384 GPU cluster (the standard hyperscale base size now) needs 16,384 cables just between compute nodes and leaf switches. Add leaf-to-spine and spine-to-core, and the fiber and connector counts grow by an order of magnitude. Hyperscale inter-campus links typically run cables with 3,000+ fiber strands (Corning official data).

Going back to the body analogy: as the GPU brain gets smarter, it consumes more oxygen. So fiber count and connector density have to go up, while cable outer diameter per fiber and attenuation budget have to come down. You need more fibers in the same conduit space, with lower loss. Upgrading transceivers alone does not solve this.

Why Meta committed up to $6B to Corning

On January 27, 2026, Meta signed a multi-year, up-to-$6B (through 2030) agreement with Corning. Meta became the anchor customer of the Hickory, North Carolina fiber cable plant, which Corning CEO Wendell Weeks said will become the largest fiber cable plant in the world once complete.

The signal here: hyperscalers are now worrying about locking in fiber supply before they worry about which transceiver vendor to use. Meta’s Prometheus campus (Ohio, 1GW) and Hyperion campus (Louisiana, 5GW) are both filed to use Corning fiber.

In the Q1 2026 earnings call Corning announced that “two additional hyperscale customers have signed agreements of similar size and duration to Meta’s.” So this is not one hyperscaler, it is multiple hyperscalers behaving the same way.

What Corning actually makes (the vessels)

Corning invented low-loss optical fiber in 1970. For more than 50 years it has held a long-standing supplier position in fiber and cable, and today it sells the entire vessel system, not just the fiber. Mapping the product portfolio to the vascular metaphor:

Aorta (subsea backbone)

• Vascade EX2000 / EX2500: ultra-low-loss, large-effective-area subsea fibers, 200μm options available. The thickest vessels carrying intercontinental data flows.

Artery (high-fiber-count trunk)

• RocketRibbon: stacked 12-fiber ribbons packed into a single cable, 432 to 864 fibers, FastAccess jacket for fast splicing.

• RocketRibbon Extreme-Density: up to 6,912 fibers in a single cable. The standard for hyperscale inter-campus DCI.

• RocketRibbon 200: SMF-28 Ultra 200 (200μm) fiber base, 15% smaller diameter at the same fiber count. 1,728 / 3,456 fiber options.

Veins and capillaries (inside data center)

• SMF-28 ULL: G.652-compliant ultra-low-loss fiber, terrestrial long-haul standard.

• SMF-28 Contour fiber: 190μm outer diameter, 40% smaller than standard, doubling fibers per conduit. Capillary-density for AI data centers.

• Multicore Fiber (MCF): multiple cores in a single fiber strand, 4x capacity per fiber inside the same 125μm cladding.

• Contour Flow micro cable: high-density mini-cable for intra-city and intra-campus interconnect.

Perfusion interface (fiber-to-chip)

• MMC connector with PRIZM TMT ferrule: expanded-beam connector. 36x fiber connection density vs legacy LC, or 3x vs MTP, in the same housing.

• Fiber Array Unit (FAU): ultra-compact fiber-to-chip coupling. The interface for switches and silicon photonics chips. A core component of CPO architectures.

• EDGE / EDGE8 Distribution System: pre-engineered structured cabling, cuts server row cabling installation time by 70%.

Neural pathway (latency-critical)

• Hollow Core Fiber (HCF): a structure designed for lower latency than conventional solid-core SMF. Corning is in a manufacturing collaboration with Microsoft to scale HCF production for Azure deployment, using the North Carolina production base.

This is bundled as the GlassWorks AI portfolio. The point is that Corning sells the full stack of fiber + cable + connectivity + CPO interface from one vendor. From a hyperscaler’s perspective, locking in one vendor means matching fiber count, cable diameter, connector density, and fiber-to-chip interface to a single spec.

[Figure 2: Corning’s fiber/cable product portfolio mapped to a vascular system diagram]

Summary: the fiber and connector layer matters as much as the transceiver layer in AI infrastructure. Corning has held a long-standing supplier position in the vascular layer for over 50 years, and today the portfolio runs full-stack from aorta (Vascade) to artery (RocketRibbon) to capillary (Contour fiber) to perfusion interface (MMC, FAU) to neural pathway (HCF). NVIDIA’s fourth bet landing on this layer is a continuation of the same trend.

3. What looks the same when you group the four, and what splits them apart

What public sources let you see is that NVIDIA invested in photonics 4 times, and the last one was in connectivity rather than the transceiver layer.

From here on, this is no longer just a “list of NVIDIA investments.” It is the point where the beta and risk of each layer split apart.

NVIDIA’s four bets all get grouped as “photonics investments,” but the instruments are not identical:

• Coherent: $2B equity

• Lumentum: $2B Series A Convertible Preferred ($695.31/share)

• Marvell: $2B Series A Convertible Preferred (bundled with NVLink Fusion)

• Corning: $500M warrant deal (Traditional Warrant + Pre-Funded Warrant)

There is a reason a $500M warrant carries different meaning than a $2B equity stake, and a reason this structure was applied to Corning.

If you just lump the four together as “companies NVIDIA invested in,” you also miss the way market cap, revenue mix, and AI exposure differ by an order of magnitude across them. Transceiver pure-plays and Corning are in the same cycle, but their betas are completely different.

And among the four, which one is the “most conservative way to get exposure to the AI photonics cycle” already has an answer that hyperscalers are showing through their LTAs. There is a layer where the answer exists, and a layer where it does not.

The next sections cover:

• Identification table for the 4 companies (instrument / market cap / 1Y / positioning)

• Order-of-magnitude comparison: transceiver layer vs connectivity layer

• Differentiation table (each company’s core weapon + key catalyst + one-line character)

• Corning deep dive (Optical Communications dynamics + the meaning of the Meta-NVIDIA double lock-in)

• Scenarios (Base / Alternative / Downside) + monitoring points

[Figure 3: 4-company positioning matrix, X-axis transceiver↔connectivity, Y-axis market cap]

Subscribe now

Share

Read more

Contents

1. Intro: What $808M in a Quarter Means

2. Five Key Signals from Lumentum’s Earnings

3. What OFC 2026 Showed Six Weeks Earlier

4. The Structure of the InP Cycle: Why All Five Signals Move the Entire Supply Chain

5. Five Groups, and One Unexpected Company

6. Company Identification + Large-Cap Comparison

7. Five-Group Classification + Differentiation + Company Analysis

8. Scenarios + Monitoring + Closing

9. References

1. Intro: What $808M in a Quarter Means

Quarterly revenue $808.4M, +90.1% YoY. After market close today, Lumentum (NASDAQ: LITE) reported FY26 Q3 results [1].

• Revenue: $808.4M (+90.1% YoY), near the top of $780M to $830M guidance

• Non-GAAP EPS: $2.37, exceeding the $2.15 to $2.35 guidance range

• Non-GAAP OPM: 32.2% (+700bps QoQ)

• Q4 Guidance: Revenue $960M to $1,010M (upper bound crosses $1B for the first time), OPM 35% to 36%, EPS $2.85 to $3.05

As of writing, LITE closed at $1,002.65 with a market cap of $71.55B (May 5, 2026, Robinhood) [2]. The trailing one-year stock return is approximately +1,500% [3]. However, in after-hours trading following the earnings release, LITE dropped to around $940 (roughly -6%) before recovering to the $975 range (author’s direct observation, May 5, 2026 PT). EPS and Q4 guidance came in strong and revenue landed near the top of guidance, but the stock reacted first to sell-the-news dynamics and dilution concerns. The significance of this pattern will be revisited in the company analysis section.

What makes this number interesting is that it is not just a Lumentum story. Lumentum doubling its 200G EML quarterly revenue means the InP wafer area carrying those EMLs came under strong pressure. EML integrates a DFB laser section and an EAM section on a single chip, so die-area burden and epi/process/test complexity trend higher than for a simple CW DFB laser. When revenue doubles on top of that structural complexity, wafer-start and epi capacity both face simultaneous pressure. That said, ASP, yield, and die-size mix all interact, so revenue growth and wafer utilization growth do not map 1:1. The same single earnings headline therefore sends a message simultaneously to the companies that make InP wafers, the companies that deposit epi on those wafers, the companies that sell MOCVD and MBE tools, and the companies that OEM laser chips on top of it all.

But that message does not arrive at the same magnitude everywhere. One layer responded at +297%. Another at +50%. Another surged to +5,000%. This asymmetry is the starting point for this article.

[Figure 1: Lumentum Q3 FY26 earnings at-a-glance. Segment ($533M / $275M) + 5 drivers + capital flow ($1.16B to $3.17B with Greensboro 5th InP fab)]

Subscribe now

Share

2. Five Key Signals from Lumentum’s Earnings

According to the press release and supplemental presentation [1][5], five drivers powered this quarter.

(1) 200G EML quarterly revenue doubled in a single quarter. Both 100G and 200G EML hit record quarterly shipments, with 200G EML revenue more than doubling sequentially. This matters beyond the headline number because of EML’s structural complexity. EML integrates a DFB laser section and an EAM section on a single chip, so die-area burden and epi/process/test complexity trend higher than for a simple CW DFB laser. Doubled EML revenue therefore signals strong pressure on InP wafer-start and epi capacity, though ASP, yield, and die-size mix prevent a 1:1 conversion from revenue growth to wafer utilization growth. In the earnings call, Hurlston added another data point: EML units are on a trajectory to increase 50% or more in the December 2026 quarter compared to December 2025 [25]. The sequential doubling is a QoQ number; the +50% is a YoY base. Both point to strong pressure on InP wafer area. Q3 laser chip shipments themselves doubled year over year. Japanese wafer fab capacity is fully allocated and trading at premium pricing. Critically, the company had disclosed a 30%+ EML supply gap in the prior quarter call, and this quarter acknowledged the gap has widened [25]. Wafer fabs running at double the rate still cannot keep up with demand.

(2) Narrow linewidth laser posted nine consecutive quarters of sequential growth, +120% YoY. And Lumentum introduced a new category: “scale-across.” Narrow linewidth lasers go into coherent transponders and DCI (Data Center Interconnect) applications, all built on InP DFB lasers. In the call, Lumentum bundled narrow linewidth and pump lasers into a scale-across category. This is the third axis after scale-out (more lanes per transceiver) and scale-up (intra-rack link intensification): when hyperscalers hit the power and space limits of a single building and move to multi-building distributed architecture, scale-across provides the high-bandwidth synchronization infrastructure. Narrow linewidth lasers deliver the precision needed for 1.6T speeds and higher-order modulation, pump lasers handle multi-rail optical amplification, and WSS (Wavelength Selectable Switch) manages traffic routing. The company stated this category is the largest contributor to gross and operating margin expansion [25]. The +120% YoY therefore connects directly to the +700bps QoQ OPM jump to 32.2%.

(3) Pump laser revenue grew +80% YoY, and this is where the strongest supply signal emerged. Pump lasers are III-V laser components used in EDFAs (Erbium-Doped Fiber Amplifiers) and inline optical amplifiers, with material mix varying by wavelength. 980nm pumps typically use GaAs/AlGaAs/InGaAs Fabry-Perot structures, while 1480nm pumps use InGaAsP/InP. Rather than mapping +80% entirely to InP wafer beta, the accurate read is that scale-across plus subsea optical amplification cycles are pressuring the broader III-V laser supply chain. Hurlston acknowledged in the call that pump laser supply constraints are more severe than EML constraints, and that this gap emerged as unanticipated [25]. The company has reached the stage of turning away customer orders. Multi-rail architecture acts as a multiplier for pump laser content. Lumentum plans to ramp pump capacity at the Rose Orchard facility in phased steps over the next four quarters [25], with output step-up velocity faster than EML. This is less a datacenter cycle and more a telecom backbone plus AI scale-across cycle accelerating simultaneously.

(4) CPO (Co-Packaged Optics) ultra-high-power CW laser at the 1W class entering production. Laser chip ramp is already in sequential growth. CEO Hurlston stated in the press release that meaningful revenue begins “exiting CY 2026” [1]. The earnings call went further: CPO ultra-high-power laser chip manufacturing ramp already entered sequential growth in fiscal Q3, and the company has begun fulfilling multi-hundred-million-dollar purchase orders for 1H 2027 delivery [25]. This is not “revenue will come” but “revenue is already ramping sequentially.” Simultaneously, the ELS (External Laser Source) module format is being developed as a turn-key solution for multiple CPO customers. Hurlston specified that for non-primary customer engagements, the ELS module is the key entry point [25]. Output doubled in one year: 400mW (2025) to 800mW SHP / 1W class (2026) [11]. CPO ultra-high-power laser chip manufacturing ramp is already in sequential growth, with December quarter meaningful revenue and 1H 2027 multi-hundred-million-dollar PO fulfillment proceeding on schedule [25]. Above that, the second layer of vertical integration (laser to ELS module) is in progress.

(5) 1.6T transceiver Q4 ramp plus internal CW laser integration already started in fiscal Q3. This is the most important single line. The supplemental (p.3) notes that 1.6T transceiver Q4 FY26 ramp will proceed alongside early-stage internal CW laser integration [5]. The call went further: Lumentum already began supplying its own CW lasers to some of its cloud transceivers in fiscal Q3, with approximately 20% of Q4 guidance transceiver mix using internal CW lasers [25]. The trigger for this internal capture is revealing. The original plan was to introduce internal lasers in calendar Q2 2026, but EML line tightness forced a delay. This time, external CW laser supply also tightened, prompting re-acceleration of internal conversion [25]. Internal capture is not purely a margin optimization strategy; it is also a response to external supply pressure. This inflection directly affects supply chain distribution.

Capital structure also shifted significantly. The balance sheet shows a large footprint [5].

• Cash and short-term investments: Q2 $1,155M to Q3 $3,172M (+$2,017M)

• Shareholder’s Equity: Q2 $847M to Q3 $2,973M (+$2,126M)

• Q4 guidance diluted shares: 95.2M to 102.0M (+6.8M)

Cash increased approximately $2B in a single quarter. CFO Wajid Ali identified the primary driver in the call: NVIDIA’s direct investment [25]. This is not market inference but an official company disclosure. Simultaneously, approximately 7M additional shares are projected in Q4 guidance. Where this capital goes was also clarified: a fifth InP fab in Greensboro, North Carolina was acquired in mid-March, converting an existing GaAs facility to InP [25]. Some existing tools can be reused. Ramp timing is early 2028, and this capacity is incremental to current guidance numbers. Hurlston specified additional revenue potential of $5B or more in incremental scale-up CPO revenue [25]. On top of the quarterly $2B target, a separate capacity layer is being prepared.

The call also reaffirmed the quarterly $2 billion revenue goal, articulated at OFC and still on track [25]. That is approximately 2.5x the current $808M quarter. With the Greensboro fab’s $5B+ incremental scale-up CPO revenue potential stacked on top, the company’s multi-year trajectory is a three-stage ramp structure: current quarter to $2B quarterly to additional CPO scale-up layer.

The strongest emphasis throughout was on supply control. Lumentum disclosed that long-term agreements for substrate have already been secured [25]. The 30%+ supply gap for EML and pump lasers sits within the company’s control envelope (capacity ramp), while substrate (external wafer purchases) is locked in via LTA. For supply chain companies, this is a significant signal: LTAs at this stage involve prepayment, take-or-pay, and volume plus price floor structures [25], meaning external substrate and epi companies are moving toward multi-year revenue and margin visibility. Counterparties were not disclosed, but the pattern is structural to the entire InP supply cycle, not unique to a single company.

One-line summary of Q3 FY26: The external supply cycle for InP-based optical components is accelerating, while Lumentum simultaneously begins pulling that InP demand internally. And the ramp trajectory toward quarterly $2B is already being locked in at the contractual layer. These three dynamics happening concurrently are the starting point for supply chain analysis.

3. What OFC 2026 Showed Six Weeks Earlier

The five drivers from this quarter were actually previewed six weeks earlier at OFC 2026 (Los Angeles, March 16 to 19) in portfolio form. OFC General Chair Tetsuya Hayashi (Sumitomo Electric) summarized in the wrap-up: “1.6T moving closer to real-world deployment across the AI networking ecosystem” [10].

Lumentum demonstrated a 1.6T DR4 OSFP pluggable transceiver prototype at OFC 2026, using four 400G differential EML lasers [11]. At the same booth (#1439), an 800mW SHP (Super-High-Power) laser was announced: 1310nm device delivering over 1W optical output at 25C, 800mW+ at 50C, less than 100kHz linewidth, greater than 40dB SMSR. A 16-channel DWDM UHP laser source was also demonstrated in an ELSFP (External Laser Source Form Factor Pluggable) module, with 200 GHz grid centered at 1310nm, approximately 24 dBm per channel, CW-WDM MSA compatible [11]. Lumentum explicitly used the phrase “advanced components, world-class packaging, and high-volume manufacturing expertise,” signaling production-stage readiness rather than R&D demonstration. Output progressed from 400mW UHP laser (2025) to 800mW SHP / 1W class (2026): a 2x jump in one year reaching production stage.

Coherent demonstrated its 1.6T transceiver portfolio at the same OFC across four platforms simultaneously: SiPh PIC + high-power InP CW laser + 200G InP EML + 200G GaAs VCSEL [12]. A 6.4T slot CPO was demonstrated using an in-house high-power InP CW laser external light source. A new 12.8T XPO MSA form factor was announced (12.8Tbps liquid-cooled module, rack-level 204.8Tbps front-panel density). One additional development stands out: a 400G InP modulator running on silicon [12]. This signals hybrid integration of InP modulators onto SiPh platforms. SiPh is not unidirectionally replacing InP; InP is extending into the SiPh modulator domain in a bidirectional coupling. This paradox will be revisited in the scenario analysis.

OFC 2026’s portfolio started transmitting into earnings revenue this quarter, Q3 FY26. This is the commercial-stage entry signal for the InP cycle.

4. The Structure of the InP Cycle: Why All Five Signals Move the Entire Supply Chain

Let’s step back and address the basics. Why does AI optical networking require InP, and why do all five drivers above converge on a single InP wafer?

InP is the active material that directly generates light in optical communications. Silicon has an indirect bandgap and cannot efficiently produce light. Datacenter transceivers operating at 1.3 and 1.55 micrometer wavelengths therefore build their EMLs and CW DFB lasers almost entirely on InP wafers [4]. Short-reach VCSELs at 850nm and 940nm use GaAs, and SiPh transceivers process their modulators on silicon, but the light source still depends on an external InP CW laser. Every layer in an optical transceiver that produces light sits on III-V compound semiconductors.

The key to this cycle is the transition from 4-inch to 6-inch InP wafers. Simple arithmetic gives approximately 2.25x the area, but when edge exclusion, automation, usable area, and yield improvements are included, a single wafer yields four times or more devices under the same process. This was stated in Coherent’s March 2024 announcement [7] and Oxford Instruments’ November 2025 release [8]. Die cost drops 60% or more. The 6-inch transition therefore creates a multi-year greenfield capex cycle for deposition tool companies: 4-inch era tool sets cannot handle 6-inch production.

Mapping Lumentum’s five drivers onto this structure:

(1) 200G EML quarterly revenue doubled signals strong pressure on InP wafer-start and epi capacity. EML integrates a laser and modulator on one chip, so die area, process complexity, and yield sensitivity can all increase relative to CW DFB. Direct beneficiaries include InP substrate companies (such as AXT), InP epi companies, and epi tool companies (such as Aixtron MOCVD).

(2) Narrow linewidth laser +120% YoY feeds the same InP DFB layer.

(3) Pump laser +80% YoY spans III-V more broadly (GaAs/AlGaAs/InGaAs or InGaAsP/InP depending on wavelength). This pressures the broader III-V supply chain rather than mapping directly to InP wafer beta alone.

(4) CPO ultra-high-power CW laser, CY26 revenue start is InP DFB CW laser. If CPO adoption accelerates, EML-centric InP demand could be reallocated toward CW laser and ELS architectures. However, if absolute transceiver volumes grow faster, total InP wafer demand continues to increase. The separation of absolute demand from per-unit demand begins here.

(5) Lumentum internal CW laser integration maintains InP wafer demand but shifts it from external InP suppliers to Lumentum’s internal operations. Coherent is moving in the same direction. CEO Anderson stated in the Q2 FY26 call that “internal indium phosphide sourcing [will] increase as a proportion of total supply” [9]. Coherent has ramped 6-inch InP lines at Sherman, Texas and Yarfalla, Sweden to 80% of target capacity, maintaining external partnerships for supply resilience and customer requirements while increasing internal share [9].

Lumentum Driver Direct Beneficiary Indirect Beneficiary Potential Displacement 200G EML doubled InP substrate, InP epi, MOCVD tool Etch/depo tool (none) Narrow linewidth laser InP substrate, InP epi MOCVD tool (none) Pump laser III-V substrate (GaAs + InP), III-V epi, MBE/MOCVD tool (none) (none) CPO CW laser CY26 InP substrate, InP epi, MOCVD tool ELS / packaging EML-heavy demand mix (long-term) Internal CW integration Lumentum/Coherent in-house InP (none) External InP epi service

[Figure 2: InP supply chain 4-layer diagram (substrate, epi tool/service, laser chip OEM, transceiver/module). Asymmetric arrows showing different beta thickness per layer.]

The same earnings headline arrives as “absolute volume surge” in one layer, “internal capture signal” in another, and “CPO transition signal” in yet another. This is the starting point for why beta splits by orders of magnitude.

5. Five Groups, and One Unexpected Company

Everything up to this point is visible by following Lumentum’s earnings release and publicly available 6-inch InP transition data. The consensus already recognizes that the AI optical cycle is an InP cycle. The real question is: where in the supply chain is the strongest beta right now?

The real differentiation starts here. Within the same InP supply chain, one-year stock returns range from +50% to +5,000%, splitting by orders of magnitude. Beta per layer has been completely different since the AI optical cycle began in 2024.

This asymmetry is not simply a matter of “which company is better.” It is a matter of the same cycle transmitting through layer structure with different multiple effects. Five distinct groups form:

• Pure-play InP volume beta (substrate): nearly 100% exposure to a single layer. Every wafer directly becomes revenue.

• Pure-play epi service: an external III-V epi foundry exposed to a single layer. Short-term upside and mix-down risk coexist.

• Multi-segment beneficiary: InP is a portion, but other cycles (LED legacy, power, quantum, advanced packaging) are simultaneously strong. Combined capex tool cycle beta.

• Lower-beta adjacent: deposition/etch tools where InP/photonic mix is small, diluting beta.

• Internal capture (winner side): the layer pulling external supply internally. This becomes mix-down risk for some external suppliers.

Among these five groups, one contains a company paradoxically exposed to the most direct mix-down risk. The Lumentum plus Coherent internal integration signal is simultaneously good news and bad news for that company. This is the entry point for this article.

This is the boundary of what earnings data and public disclosures can show. How the same cycle splits into orders-of-magnitude-different beta, which companies belong to which group, who the real internal capture card is, who faces mix-down exposure, and where each group’s valuation currently stands requires sorting the data layer by layer.

Read more

The real headline was elsewhere. On the same day, Veeco separately announced $250M+ in multi-product orders for InP laser manufacturing. In the company’s own wording, “a substantial portion of the orders are for the Spector® IBD” system, used for laser facet coating. The customers are leading manufacturers of 800G and 1.6T optical transceivers for hyperscale data centers.

This piece updates PhotonCap’s six-company piece, where Veeco’s SiPh exposure was framed mainly through MOCVD epitaxy. Q1 made the next layer clearer: the bigger surprise in the order book and the 2027 ramp was not MOCVD, but Spector IBD. The same ion beam technology is already the process tool of record across leading EUV mask-blank manufacturers, and that same asset is now extending into InP laser facet coating.

Ticker referenced: $VECO

Contents

1. Quarterly Headline

2. Quarterly data summary (results, segment mix, guidance, B/S, commentary)

3. The next question

4. PhotonCap reading: SiPh exposure 3-step + Spector IBD core lever + thesis update

5. Scenarios + monitoring + close

6. References & Sources

1. Quarterly Headline

The most direct variable behind Q1’s gross-margin miss was one LSA system blocked from shipment to a China customer ($8M, BIS export license). The real headline was elsewhere: Veeco received $250M+ in multi-product orders for InP laser manufacturing, with a substantial portion in Spector IBD for laser facet coating.

The PhotonCap six-company piece framed Veeco’s SiPh exposure mainly through MOCVD (epitaxy). Q1 data makes the next layer clearer: the bigger surprise in the order book and the 2027 ramp was Spector IBD, one process step downstream from epi. The same ion-beam tooling platform that is already the process tool of record across leading EUV mask-blank manufacturers is now extending into InP laser facet coating.

Scope: $VECO Q1 2026 earnings + the separate $250M+ orders PR + call transcript Q&A + six-company piece thesis update.

2. Quarterly data summary

2.1 Q1 results [1][4]

Non-GAAP EPS $0.14 lands at the low end of guidance. Non-GAAP gross margin 36.2% sits below the guidance range.

Segment mix (Q1 2026) [4]:

Compound semi at 12% maps directly onto the six-company piece‘s “photonics ~11%” framing. That bucket includes SiPh InP laser, GaN power, microLED, AR/VR, and LEO satellite solar cells.

Gross-margin miss explained [4]: Per CFO John Kiernan, BIS notified Veeco that one LSA system shipment to a specific China customer fab required an export license. That event cost about $8M in top line and was the most direct variable behind the gross-margin miss. Q2 gross margin is guided to 38 to 40%, with recovery expected.

2.2 Guidance [1][4]

Q2 2026: Revenue $170 to $190M, gross margin 38 to 40%, Non-GAAP EPS $0.20 to $0.32. FY 2026 (unchanged): Revenue $740 to $800M, Non-GAAP EPS $1.50 to $1.85.

The CFO and CEO directly named 2026 segment growth on the call: semiconductor mid-teens YoY, compound semi roughly 50% YoY, data storage 2x YoY [4].

Q1 actual plus the Q2 guidance midpoint gives 1H ≈ $338M. Subtract from the FY midpoint of $770M and 2H ≈ $432M, +28% versus 1H. Revenue recognition is weighted heavily into H2.

2.3 Balance sheet shifts (Dec 31, 2025 to Mar 31, 2026) [1][4]

Accounts receivable rose from $111M to $151M (+$40M), and customer deposits within contract liabilities grew $19M to $69M. Inventories ticked from $275M to $282M (+$7M), while cash plus short-term investments was roughly flat at $390M moving to $383M.

AR growth shows the timing of collection on revenue already shipped or recognized. Customer deposits show prepayments not yet recognized. Combined with the $250M+ orders and the H2 weighting embedded in full-year guidance, the picture leans toward revenue recognition concentrated in H2 and into 2027.

2.4 CEO commentary + Q&A core points

SiPh entered the highlight paragraph of the PR, with “particularly strong momentum in silicon photonics” [1].

On the same day, Veeco issued a separate PR announcing $250M+ in multi-product orders for InP laser manufacturing [2]. The product mix spans Spector IBD, Lumina MOCVD, and WaferEtch, and per the company’s own wording, “a substantial portion of the orders are for the Spector® IBD”. Customers are leading manufacturers of 800G and 1.6T optical transceivers for hyperscale data centers. Deliveries begin in 2026 and significantly accelerate in 2027 [2]. Per the Q&A, shipments start in Q3 2026 and ramp in Q1 2027 [4].

In the Q&A, Bill Miller explicitly committed to a 10x Spector IBD capacity expansion by early 2027, with potential for another 2x beyond that [4].

Veeco’s InP laser exposure splits into three process steps. At epitaxy (MOCVD), Aixtron is the incumbent and Veeco runs as a second source. At wet processing (etch / clean), Veeco holds a strong position with multiple leader customers. At laser facet coating, Spector IBD is the incumbent, and that is where the core lever sits [4].

[Figure 1: Veeco’s InP laser 3-step exposure (epi / wet / facet coating)]

Section 2 in one line: the most direct variable behind the Q1 miss is a single event. The actual headline is $250M+ in InP laser orders, with Spector IBD as the substantial portion.

3. The next question

That is what the PR plus call transcript can tell you about the quarter.

The real difference starts here.

The six-company piece framed Veeco’s SiPh exposure through MOCVD epitaxy. Q1 data suggests the larger lever was one process step downstream: InP laser facet coating via Spector IBD, on the same ion-beam tooling platform that is already the process tool of record across leading EUV mask-blank manufacturers.

Behind the paywall, four threads. The semi view covers the Q1 gross-margin miss as a single event and the three streams’ H2 ramp. The photonics view breaks Veeco’s SiPh exposure into 3 process steps, explains why Spector IBD is the core lever, and traces how the previously dormant DWDM-era IBD line reactivates into the AI cycle. The six-company thesis update splits the MOCVD epitaxy line and the IBD facet-coating line, then adds a new thesis line on a single ion-beam tooling platform spanning two AI-cycle markets. The piece closes with Spector IBD ramp scenarios A/B/C and monitoring points.

Read more

Contents

1. Intro: What YTD +255% Is Telling Us

2. Technical Context: Why Glass Is Needed (EIC, CPO, EPIC)

3. Three Layers of the Value Chain, One Complement, and Two Hidden Cards

4. Glass Substrate Majors

5. Watchlist + Photonics-SOI Complement

6. Monitoring Points + Closing

7. References & Sources

1. Intro: What YTD +255% Is Telling Us

LPKF Laser & Electronics (LPKFF / LPK.DE). A 50-year-old industrial laser equipment company headquartered in Garbsen, Germany. Annual revenue of €115M (FY2025), market cap roughly €481M. Its stock is up 255% year-to-date in 2026. [1][2]

Over the same period, SMH (VanEck Semiconductor ETF) is up +42%, Applied Materials +46%, and ASML +19%. This has been a strong year for semiconductor equipment stocks. And yet a €115M-revenue German mid-cap is delivering multiples of the sector’s best names. Why?

The answer is not LPKF’s brand. It is one process LPKF controls: LIDE (Laser-Induced Deep Etching). This is the most advanced production-grade process for drilling Through Glass Vias (TGVs) in glass substrates. Other approaches exist, including CO₂ laser and laser modification plus etch. But LIDE is the most advanced commercial equipment platform capable of forming crack-free, debris-free vias at production speed, and LPKF is expanding it across the NEXAR LIDE 5000, Ablate 5000, and Bond 5000 product lines. As AI chip packaging begins shifting from organic substrates to glass, this process is becoming one of the key bottlenecks.

The thesis of this article is straightforward. Glass substrates are already needed in EIC (electronic packaging), and as the transition moves to CPO (co-packaged optics) and EPIC (electronic-photonic IC), glass’s structural advantage only grows. This article maps the equipment, material, and substrate companies sitting at the narrowest chokepoints of that transition.

So where is the most advantaged position in this transition?

To answer that, we first need to understand why glass is becoming necessary.

Subscribe now

Share

2. Technical Context: Why Glass Is Needed (EIC, CPO, EPIC)

2.1 Starting With EIC: Why Organic and Silicon Interposers Hit Their Limits

AI accelerator die sizes are growing. NVIDIA’s Blackwell B200 exceeded the reticle limit (~858mm²) by placing two dies in a single package, and the next generation Rubin is likely to carry even more chiplets. More dies require a larger interposer or intermediate substrate, and that is where existing materials begin to break down.

The organic substrate problem: CTE (coefficient of thermal expansion) is about 17 ppm/°C. In plain terms, for every 1°C temperature increase, the material expands by 17 millionths of its length. That sounds small, but across a 100mm package, these differences add up to microns. Silicon dies are about 2.6 ppm/°C. Bond the two together, and they expand and contract at different rates with every temperature change. At small sizes, this is manageable. At 110x110mm package sizes, warpage can reach hundreds of microns. Stack HBM 8 to 12 layers high next to an ASIC, and the substrate bows enough to crack bump joints.

The silicon interposer problem: CTE matches the silicon die, which is great. But it is expensive. Cutting interposers from 300mm wafers means high cost per area, and scaling to larger sizes is limited. The TSV (Through Silicon Via) process itself is not cheap either.

Glass solves both problems at once.

• CTE: tunable from 3 to 5 ppm/°C. Composition can be adjusted to sit between silicon (2.6) and organic (17). [3]

• Dielectric properties: Dk ~5, Df <0.005. Lower high-frequency signal loss than organic substrates.

• Flatness: Glass surface roughness is in the single-nm range, more than an order of magnitude better than organic.

• Panel-level processing: Glass can be processed on panels of 510x515mm and larger, covering 3 to 4 times the area of a 300mm wafer. That means more packages can be processed in each panel-level run.

This is why Intel officially announced glass substrates in 2023. Intel’s target is 10x interconnect density over organic, with production planned “before 2030” on its official roadmap. [4]

[Figure 1: EIC packaging substrate evolution: organic → silicon interposer → glass]

Key point: Glass is already recognized as a better substrate at the EIC stage because it addresses organic’s warpage and silicon interposer’s cost simultaneously. But glass’s real differentiation does not stop here.

2.2 When Light Enters the Picture, the Problem Changes: The Fiber-to-Chip Bottleneck in CPO

In EIC, only electrical signals need to pass through the substrate. In CPO (Co-Packaged Optics), optical fibers must be connected directly to photonic chips on the package substrate. That introduces an entirely different class of problem.

As discussed in PhotonCap PC101 Lecture 4, the hardest part of CPO packaging is the “last millimeter,” the fiber-to-chip coupling. [5] The core of this problem is not thermal expansion. It is mode field diameter (mode size) mismatch.

A single-mode optical fiber (SMF-28) has a mode field diameter of about 8 to 10 µm at 1310nm. A silicon photonic chip waveguide, by contrast, has a cross-section roughly 0.5 µm wide and 0.2 µm tall. In mode area terms, that is about a 200x difference. Light coming out of the fiber is spread wide. The SiPh waveguide confines light in a tiny space. If the two are butt-coupled directly, coupling loss can reach tens of dB. The signal effectively vanishes.

This is why fiber-to-chip coupling always requires a mode size converter (spot size converter). The standard approach is to build edge couplers or grating couplers on the SiPh chip to expand the mode. But this alone has limits. Edge couplers can stretch the mode to ~3 to 4 µm, which still leaves a large gap against the fiber’s ~9 µm, and alignment tolerance stays within ±1 µm.

This is where glass enters. Ion-exchange (IOX) waveguides formed inside the glass substrate can be designed with a mode field diameter nearly matching the fiber, around 8 to 10 µm. That is why Corning’s IOX waveguide achieved <0.65 dB coupling loss against SMF-28 fiber. [6][20] The glass waveguide acts as an intermediate mode converter between fiber and PIC. The light path runs: fiber (~9 µm mode) into glass waveguide (~9 µm mode), then adiabatic taper down, then into PIC edge coupler (~3 to 4 µm mode). The mode shrinks in stages. Coupling is horizontal, through edge coupling.

Think of it this way: merging directly from an 8-lane highway (fiber) into a 1-lane alley (SiPh waveguide) creates a massive bottleneck. But routing through a 4-lane on-ramp (glass waveguide) first makes the merge smooth. In other words, glass provides an optical transition layer rather than forcing the fiber mode to collapse directly into the SiPh waveguide.

CTE (coefficient of thermal expansion) matching deserves a closer look here. Although both optical fiber and glass substrates are “glass,” they are not the same glass. Optical fiber is silica-rich glass with an extremely low CTE of ~0.5 ppm/°C, while packaging glass substrates are engineered borosilicate or aluminosilicate-type glasses with CTE typically in the 3 to 5 ppm/°C range. On CTE alone, silicon interposer (~2.6 ppm/°C) is actually closer to fiber than glass substrate is.

So where does glass’s CPO advantage come from, if not CTE? The combination, and mode size. Silicon interposers can also carry photonic waveguides. That is the logic behind silicon photonics and SiN photonic interposers. But the native silicon waveguide mode is far smaller than the fiber mode, so fiber coupling still requires additional spot-size converters, edge couplers, or other mode-transition structures. Glass IOX waveguides are different because they can be designed with mode sizes of ~8 to 10 µm, directly matching fiber. Add panel-level processing (510x515mm) and TGV electrical routing on the same substrate, and the picture changes. In an NVIDIA Spectrum-X class CPO package with 36 or more optical engines per package [19], the combination of fiber-matched waveguide mode, panel scalability, and integrated E/O routing is where glass has the clearest architectural advantage over silicon interposer.

[Figure 2: Mode field diameter mismatch and glass waveguide as intermediate coupler]

Key point: The fundamental bottleneck in fiber-to-chip coupling is mode field diameter mismatch (fiber ~9 µm vs. SiPh ~0.5 µm). Silicon interposers can carry photonic waveguides, but native silicon waveguides still require mode-transition structures for efficient fiber coupling. Glass IOX waveguides provide a fiber-matched mode size while enabling panel scalability and integrated E/O routing on a single substrate.

2.3 At EPIC, Glass’s Structural Advantage Grows

At the EIC stage, glass was “a better foundation.” At the CPO stage, it was “a foundation that can safely attach fibers.” At EPIC (Electronic-Photonic Integrated Circuit), glass becomes not just a support structure but a routing platform that carries both electrical vias and optical waveguides.

Glass has a distinctive capability: ion-exchange can form optical waveguides directly inside the substrate body.

In alkali glass (containing Na), replacing sodium ions with silver (Ag) or potassium (K) ions raises the refractive index of the exchanged region. That refractive index contrast creates a waveguide that confines light. Corning’s IOX process has demonstrated propagation loss <0.1 dB/cm at 1310nm and chip coupling loss <0.65 dB. [6][20]

Why does this matter? Because a single glass substrate can do three things at once:

• TGV (Through Glass Via) for vertical electrical signal routing

• IOX waveguide for horizontal optical signal routing

• Surface cavities for edge-coupling PICs (Photonic ICs) optically and flip-chip mounting ASICs electrically (bump bonding)

Corning’s OFC 2026 paper by Brusberg et al. demonstrated that the IOX-based GlassBridge platform can extend the glass substrate from a passive carrier to an active optical routing platform. [20] Based on PhotonCap’s analysis of the OFC 2026 presentation, the demonstrated structure placed 16 PICs and 1 ASIC on a 515x515mm glass panel, kept all electrical paths within 13mm, and achieved sub-5-minute placement time per PIC. [6]

Alternative integration routes do exist. Polymer optical waveguides on organic substrates and SiN photonic waveguides on silicon interposers are both being explored. What differentiates glass is that the substrate itself becomes the optical routing layer while simultaneously supporting TGV-based electrical routing and large-area panel processing, all within a single platform. [21]

This is not a matter of material preference. It is a packaging architecture question. For a platform that routes electrical and optical signals at the substrate level while enabling panel-scale processing, glass is the most advanced candidate among publicly disclosed approaches.

[Figure 3: Glass substrate plan view showing TGV + IOX waveguide + PIC edge coupling]

Key point: From EIC to CPO to EPIC, the structural advantage of glass compounds. At EIC, glass competes with organic and silicon. At EPIC, glass’s unique ability to integrate substrate-level optical routing becomes a clear differentiator against alternatives.

2.4 Why Now: The Timing Case

Glass substrate technology itself is not new. TGV processes appeared in academic papers by the mid-2010s. So why did it suddenly become an investment theme in 2025 and 2026?

Three things opened up on the demand side at the same time. First, NVIDIA Spectrum-X Photonics and Broadcom Bailly confirmed CPO on production roadmaps. Second, AI accelerator package sizes reached a point where organic substrate warpage becomes a real constraint. Third, glass core substrate pilot lines started spinning up simultaneously across Korea, the US, and Japan. [9]

Mass production timelines mostly target 2027, with some at 2028 to 2029. The current phase is pre-production equipment ordering and qualification. That is why the equipment companies are moving first.

Key point: Glass substrates are technically ready. Demand is pulling. Supply is following. Mass production targets are 2027 to 2029. What is moving right now is equipment and materials.

3. Three Layers of the Value Chain, One Complement, and Two Hidden Cards

Everything up to this point is visible from public sources alone.

The real differentiation starts here.

The glass substrate value chain breaks into three layers:

• Layer A: The company with the most advanced TGV equipment position (equipment)

• Layer B: The company embedding waveguides and building an integrated platform (integrated platform)

• Layer C: Companies making the glass panels themselves (material / substrate fabrication)

Where is the most asymmetric position in this structure?

From public sources, glass substrates look like a long-term roadmap item for Intel, Samsung, and TSMC. But the real investment leverage is not in the material companies. It is in the equipment and routing platform players that control the pre-qualification bottleneck. Two hidden cards stand out.

Hidden Card 1: The Layer A (equipment) company’s patent coverage extends beyond TGV. Development of 3D waveguide formation inside glass is underway, meaning the equipment player could penetrate into integrated platform territory.

Hidden Card 2: The Layer B (integrated platform) company’s 16-PIC structure demonstrated at OFC 2026, if it reaches production, would put 4 to 8 times more PICs per panel than current CPO architectures. That changes assembly density entirely.

Layer A and Layer B offer very different risk/reward profiles within the same glass substrate theme.

[Figure 4: Glass substrate value chain, three layers, no company names mapped]

One More: A Company That Does Not Make Glass, but Grows With It

When glass packaging unlocks CPO yields, demand for PICs (Photonic ICs) increases. The starting wafer for those PICs is Photonics-SOI, and volume supply of that wafer is currently controlled by a single company. Glass substrates and Photonics-SOI are not horizontal competitors. They are vertical complements. (Detailed analysis: PhotonCap Soitec article [24])

The paid section connects this complementary position to the glass substrate investment theme.

Read more

Contents

1. +361% in one year, 2.7x the semiconductor index. The market’s thesis confirmed in earnings

2. Q1 2026 earnings key numbers and 3-axis definitions

3. Why wafer test sits at the chokepoint of the cycle

4. Where the real difference begins

5. Axis 1: SK hynix 29.5%, Korea 35.6%, DRAM +70%

6. Axis 2: CPO revenue high-end guide, Triton ramp, Keystone integration, COUPE matching

7. Axis 3: NVIDIA 10.2% first disclosure and Foundry & Logic +30%

8. 3-axis alignment and the next 12 months

9. References & Sources

1. +361% in one year, 2.7x the semiconductor index. The market’s thesis confirmed in earnings

FormFactor (FORM) returned +361.57% over one year.[1] Over the same period, the semiconductor index (SMH) returned +134% and the S&P 500 returned +14.9%.[1] That is roughly 2.7x the semiconductor index. The thesis the market had been pricing in was confirmed for the first time in numbers when Q1 2026 earnings were released on April 29, 2026.[2]

SK hynix at 29.5% of revenue, NVIDIA at 10.2%, Triton CPO production test guide raised to the high end, and revenue at $226.1M record (+32.0% YoY). HBM, CPO, and AI networking. Three cycles bent together in a single quarter.[2][3][4]

This article takes FormFactor’s Q1 2026 earnings and breaks them into three axes, HBM, CPO, and AI networking, to analyze (a) primary-source verification of each axis, (b) alignment across the three, (c) the dual nature of SK hynix concentration risk and Triton ramp upside, and (d) monitoring points for the next 12 months. This is a single-stock earnings analysis, but at the cycle level, signals from silicon photonics, HBM, and AI networking all land simultaneously.

2. Q1 2026 earnings key numbers and 3-axis definitions

Q1 2026 (fiscal quarter ended 2026-03-28) key numbers are as follows.[2]

• Revenue $226.1M, +5.1% QoQ ($215.2M), +32.0% YoY ($171.4M). All-time record.

• Non-GAAP gross margin 49.0%, +510 bps QoQ. +250 bps above the high end of guidance.

• Non-GAAP EPS $0.56, more than double Q1 2025’s $0.23.

• Cash + marketable securities $303.2M ($123.5M + $179.7M).

• Q2 2026 guidance: revenue $240M +/- $5M, GM 49.5% +/- 1.5%, EPS $0.61 +/- $0.04.

• 2026 CPO revenue guidance: $10-20M range high end (approximately $20M).[3]

The three axes are defined as follows.

Axis 1: HBM (memory cycle)

• DRAM probe card revenue $82.9M, +70% YoY.[4]

• Korea revenue $80.6M (35.6% of total), nearly doubled from Q1 2025’s $43.2M.[4]

• SK hynix as a single customer at 29.5% of revenue (up from 23.3% in Q1 2025, +620 bps).[4]

• In an HBM4 ramp + DDR5 supply-constrained environment, probe card demand entering priority allocation.[3]

Axis 2: CPO (silicon photonics cycle)

• Triton production test platform Insertion 1 entering production-ready stage.[3]

• 2026 CPO revenue guidance raised to high end ($20M).[3]

• Co-developed with Advantest + Tokyo Electron.[3]

• Keystone Photonics (optical probing specialist, Germany), acquired in Q4 2025, integration complete.[5][6]

Axis 3: AI networking (compute cycle)

• NVIDIA disclosed as a 10%+ customer for the first time, at 10.2%.[4]

• Foundry & Logic probe card revenue $111.2M, +30% YoY ($85.3M to $111.2M).[4]

• Networking probe card growth cited as the driver of NVIDIA’s 10% entry.[3]

[Figure 1: Q1 2026 key numbers card + 3-axis diagram]

3. Why wafer test sits at the chokepoint of the cycle

Semiconductor wafer test is the step where logic, DRAM, and Flash die are electrically validated at the wafer level before moving to back-end packaging. It sits immediately upstream of bonding and packaging. When bonding capacity scales, probe test capacity must scale with it. This structure was covered in a previous PhotonCap article comparing three wafer test equipment companies (AEHR, FORM, KEYS).[7]

In silicon photonics (SiPh), wafer test is more complex. Standard logic die only need electrical probes touching the wafer, but photonic integrated circuits (PICs) require optical signal measurement alongside electrical. Optical coupling must be precisely controlled at the wafer level, both edge coupling and grating coupling must be supported, and measurement throughput requires automation compatibility (OHT, Overhead Hoist Transport).[11]

CPO (co-packaged optics) adds another layer of difficulty. EIC (electrical IC) and PIC are integrated in the same package, but EIC test equipment (Advantest, Teradyne) and PIC probing equipment (FormFactor, ficonTEC) cover different domains. On a production line, a single platform must handle both.[8] FormFactor’s Triton platform handles PIC probing, with Advantest providing electrical test integration and Tokyo Electron providing wafer handling and production automation in a co-development structure.[3]

The bonding cycle accelerating simultaneously in HBM4 and CPO was covered in a previous PhotonCap article analyzing seven bonding equipment companies including BESI.[9] As bonding capacity scales, probe test capacity must follow. Within that layer, FormFactor as a comprehensive wafer test player is exposed to HBM probe, CPO probe, and Foundry & Logic probe simultaneously.

[Figure 2: Wafer test, bonding, packaging cycle layer diagram. Where the 3 axes enter]

4. Where the real difference begins

Everything up to this point is visible from press releases and IR pages. Record quarter, Q2 guidance at $240M, 2026 CPO guidance at $20M, SK hynix at 29.5%, NVIDIA at 10.2% for the first time. That is the surface of publicly available data.

What matters is that the three numbers are not moving independently. They point to the same supply chain bottleneck. Whether HBM concentration is a direct signal of SK hynix HBM4 ramp rather than just a generic memory cycle, whether the 2026 CPO revenue high-end guide is not just first revenue but positions Triton as a potential production test platform candidate for TSMC’s COUPE production line, and whether NVIDIA’s 10.2% is a one-time networking ASIC order or the starting point of a 12-month share expansion. Those are the questions visible beyond a single quarter’s earnings.

This analysis has two hidden cards.

First, SK hynix 29.5% + Korea 35.6%, a dual-edged structure. Direct beta to HBM4 ramp, but single-customer + single-country risk is simultaneously exposed.

Second, TSMC COUPE-Triton material matching. The material stacks align, but this is not a customer relationship confirmation.

These two are the core variables for the next 12 months of this thesis.

From here, the analysis moves into axis-by-axis data, Bull/Gap/Optionality breakdowns, scenario branches, and monitoring points.

[Figure 3: 3-axis diagram. HBM, CPO, AI networking signals captured simultaneously in one quarter]

So how will these three axes diverge over the next 12 months? The answer is in the axis-level data.

Read more

Contents

1. Intro: An $88M Revenue Company With a $6B Market Cap

2. Q1 Results: The Margin Inflection and $100M Backlog

3. Export Permits + China Domestic: The Structure Is Shifting

4. Demand Is Confirmed. But What About the Price?

5. Valuation Reverse-Engineering: The Weight of Three P/S Ratios

6. Insider Selling + Dilution + Competitor Expansion

7. Scenarios + Monitoring + Closing

8. References & Sources

1. Intro: An $88M Revenue Company With a $6B Market Cap

An InP substrate company with $88M in annual revenue and a $21M net loss is carrying a post-offering pro-forma market cap of over $6B (based on Q2 estimated diluted share count of approximately 63.5M at roughly $98 per share).[1][2]

AXT Inc. (AXTI) went from a 52-week low of $1.23 to $97.81.[2] That is a trailing 12-month return above 6,000%. Five months ago, the company raised capital at $12.25 per share. Last month, it raised again at $64.25.[3]

In the deep dive published on March 2, I analyzed the InP supply chain and AXTI’s vertically integrated model.[4] The stock was at $46.32 then, trading at a trailing P/S of roughly 28x. Two months later, the FY2025 trailing P/S has expanded to roughly 70x.

This time, though, earnings started catching up to the narrative. The Q1 2026 results confirmed an InP revenue rebound, a margin inflection point, and backlog crossing $100M.[5][6] Management also guided for the company’s first ever profitable quarter in Q2.[6]

The narrative-to-earnings conversion has started. That is a fact. The question is whether $98 already prices in the final outcome of that conversion.

[Figure 1: AXTI Stock Price with Key Event Markers]

2. Q1 Results: The Margin Inflection and $100M Backlog

2.1 Revenue: InP Crossed 50% for the First Time

Q1 total revenue came in at $26.9M (per Q1 2026 SEC 8-K).[5] That is +17% QoQ and +39% YoY.

By product, InP hit $13.6M, rebounding 70% from $8.0M in Q4 and crossing 50.6% of total revenue for the first time.[5] GaAs was $5.4M (down 23% QoQ), Ge $0.2M (flat), and raw material JVs $7.6M (flat). The export permit bottleneck appears to have eased somewhat relative to Q4.

By geography, Asia Pacific accounted for 78%, Europe 21%, and North America 1%.[5] The 1% North America figure is consistent with U.S.-bound permit constraints not yet being fully resolved.

2.2 Margins: A 36 Percentage-Point Swing in One Year

The real story of this quarter is not the revenue number. It is margins.

Non-GAAP gross margin hit 29.9%.[5] In Q1 2025, that number was negative 6.4%. A 36 percentage-point swing. CFO Gary Fischer attributed this to fixed-cost absorption from higher InP volumes and the structurally higher margins that InP carries relative to GaAs.[6] With InP now above 50% of revenue, the overall margin profile has shifted structurally upward.

When a Wedbush analyst asked whether Q2 margins could reach 40%, Fischer drew the line: “too aggressive.”[6] Mid-term target is 35%, with 40%+ as a longer-term internal goal.

2.3 EPS and the Path to Profitability

GAAP EPS was negative $0.03 (versus consensus of negative $0.05, a roughly 40% beat). Non-GAAP EPS was negative $0.01 (versus consensus of negative $0.04 to $0.05, a 75 to 80% beat).[7][8] Compare that to negative $0.20 in Q1 2025 and negative $0.08 in Q4 2025. The trajectory is clear: negative $0.20, then negative $0.08, then negative $0.03. Q2 guidance calls for GAAP EPS of positive $0.05 to $0.07.[6]

2.4 Backlog Tops $100M, Q2 Guidance Points to First Profit

InP backlog crossed $100M, up 67% from $60M+ in the prior quarter.[6] Management indicated that long-term supply agreements (LTSAs) with tier-one laser manufacturers and hyperscalers could be signed “in the very near future.”[6]

Q2 guidance calls for revenue of at least $34M, InP revenue of $17M+ (a company record), non-GAAP gross margin above 30%, non-GAAP EPS of $0.06 to $0.08, and GAAP EPS of $0.05 to $0.07.[5][6] This would mark the company’s first ever profitable quarter on both a GAAP and non-GAAP basis.

Bottom line on Q1: record InP revenue, confirmed margin inflection, backlog above $100M, first profitable quarter guided for Q2. The demand-side bull case received its first validation from actual numbers.

3. Export Permits + China Domestic: The Structure Is Shifting

The export permit variable remains. But the structure around it is changing.

Q1 permit issuance came in slightly above guidance, and Q2 is off to a good start.[6] The key development is on U.S.-bound permits. Until now, direct shipments to U.S. customers have been effectively blocked (North America is 1% of revenue). MOFCOM has now requested additional documentation on pending U.S. applications.[6] VP Tim Bettles described this as “an encouraging sign.” It signals active review rather than outright rejection. Permits for non-U.S. international customers (including overseas facilities of American companies) continue to flow smoothly.[6]

At the same time, China domestic demand is accelerating rapidly. The prior quarter’s guide of “60%+ QoQ growth in China data center revenue” was exceeded. Actual Q1 growth was over 100% QoQ, and management expects another doubling in Q2.[6] Domestic China sales require no export permits. The “natural hedge” structure described in the original deep dive has started to function in practice.

Export permits are improving, and China domestic revenue is growing faster than expected, partially offsetting export control risk.

4. Demand Is Confirmed. But What About the Price?

That is the picture you can see from public information alone. The Q1 margin inflection and $100M backlog are facts. Whether a $6B market cap is a reasonable price for those facts is an entirely separate question.

The rest of this article addresses three things: what trailing P/S of 70x and run-rate P/S of 46x actually imply, the pattern of consecutive insider sales with zero purchases and 45% dilution, and what Q2 needs to deliver for $98 to hold.

So where, at this price, is the most favorable position in the value chain?

Read more

This piece walks through that structure in four steps: (1) how the device works, (2) what exactly Marvell did not acquire, (3) how Marvell’s supplier relationship reorganization aligns with LWLG’s operational stance, and (4) whether $LWLG’s value comes from Polariton exposure or from the standalone foundry track (Tower / GlobalFoundries / AMF). The full landscape of EO polymer modulator industry and the comparison with NLM Photonics was covered in detail in a previous piece [5], so this piece narrows down to the Marvell-Polariton-LWLG vector.

Contents

1. Intro: The day Marvell bought Polariton, and the landscape one week later

2. What a POH modulator is, and why µm scale is decisive

3. Why this mapping is the key variable for the next 12 months

4. The one axis Marvell did not acquire: the chromophore layer

5. Six modulator platforms and LWLG’s standalone track

6. $LWLG (Perkinamine) vs NLM Photonics (Selerion): real second-source risk and CPO suitability

7. Supplier relationship reorganization and LWLG’s disclosure discipline

8. 12-24 month scenarios and monitoring

9. References & Sources

1. Intro: The Day Marvell Bought Polariton, and the Landscape One Week Later

April 22, 2026. Marvell announced its acquisition of Polariton Technologies, an ETH Zurich spinoff. The deal price was undisclosed. The press release was titled “Advancing Optical Performance Scaling to 3.2T and Beyond,” with quotes from Sandeep Bharathi, President of Marvell Data Center Group.[1] The Marvell IR press release explicitly names the application target as “scale-across, DCI, ZR/ZR+ coherent, 3.2T and beyond.” That is, Polariton POH is, from Marvell’s perspective, a device-level modulation IP aimed at scale-across, DCI, ZR/ZR+ coherent optical interconnect. On the same day, the market interpreted this transaction as a strengthening of Marvell’s optical roadmap.

April 22, around 5 PM ET. A POET Technologies ($POET) CFO interview aired on Stocktwits TV. He explicitly named Marvell as the “industry-leading customer” for the Sivers ELS PIC and disclosed Foxconn / Luxshare partnership negotiations.[3]

April 23-27. Marvell sent POET an NDA violation notice and canceled all Celestial AI purchase orders.[2] According to POET, Marvell stated that disclosures regarding POs and shipping information violated confidentiality obligations. External media coverage pointed to the CFO interview context.[3] This event is not the main subject of this piece, but it works as a reference point for how Marvell operates supplier-relationship disclosure rules after acquisitions, which is revisited in §7.

April 21. LWLG filed an 8-K disclosing an amendment to its Roth Capital Partners ATM Sales Agreement, increasing the program capacity to $51.4045M. The same filing disclosed that 8,079,319 shares had already been sold under the prior agreement for gross proceeds of approximately $35M.[18]

April 29. LWLG announced the engagement of Michael Best & Friedrich LLP as outside IP counsel. The stated purpose: building a “licensing-friendly ecosystem” for foundry and design partners. Aref Chowdhury, who joined LWLG as CTO & Head of Strategy on January 1, 2026 (formerly Nokia Network Infrastructure VP & CTO), was the named voice on the announcement.[4]

The same industry, the same week, multiple events. Different headlines, but the same underlying pattern. PhotonCap’s reading: Marvell is internalizing the layers where differentiated IP condenses (optical modulator, system interconnect), and reclassifying specific external assembly/interposer supply relationships into options it can control more strictly. Having secured DSP[15] and switch layers in 2021 with Inphi and Innovium, Marvell moved through the December 2025 Celestial AI announcement / February 2026 closing (upfront approximately $3.25B + earnout up to $2.25B)[16] and arrived at the April 2026 Polariton acquisition. While the first three were DSP, switch, and system interconnect, Polariton stands apart in that Marvell directly secured plasmonics-based device-level modulation IP.

[Figure 1: Marvell Optical IP Reorganization Timeline 2021-2026]

This piece walks through that structure in four steps: (1) how the device works, (2) what exactly Marvell did not acquire, (3) how Marvell’s supplier relationship reorganization aligns with LWLG’s operational stance, and (4) whether $LWLG’s value comes from Polariton exposure or from the standalone foundry track (Tower / GlobalFoundries / AMF). The full landscape of EO polymer modulator industry and the comparison with NLM Photonics was covered in detail in a previous piece [5], so this piece narrows down to the Marvell-Polariton-LWLG vector.

§1 Summary: One week of events (Marvell-Polariton acquisition, POET PO cancellation, LWLG ATM + Michael Best engagement) read as a single current. Marvell pulls device-level modulation IP inside while reorganizing external assembly/supplier relationships into more strictly controlled options. PhotonCap’s reading.

Subscribe now

Share

2. What a POH Modulator Is, and Why µm Scale Is Decisive

POH (Plasmonic-Organic Hybrid) is a modulator that places organic EO chromophore inside a 100 nm metal slot,[9] then uses surface plasmon polariton (SPP) field compression to dramatically amplify EO efficiency. The mechanism itself is not new. Linear EO modulation based on Pockels effect (χ⁽²⁾) is the same approach as TFLN (Thin-Film Lithium Niobate), but POH’s distinct point is µm-scale active length.

The Optica 2025 record device (Polariton + ETH joint result) reported measurements up to 1.14 THz, 997 GHz 3-dB EO bandwidth, with active length of just 10-15 µm.[6] The slot width is 100 nm. The 100 nm slot is filled with organic EO chromophore, and the SPP field is concentrated only inside that slot, dramatically amplifying the field intensity. Compared to silicon Mach-Zehnder modulators that require lengths of several mm, this is roughly two orders of magnitude shorter.

The implication is two-fold. First, footprint shrinks dramatically, allowing high-speed modulators to be densely integrated into transceiver/packaging environments. Second, driving voltage drops sharply. The Optica 2025 paper reports 117 V·µm phase modulator equivalent VπL (= ~0.013 V·cm in MZM equivalent).[6] Compared to TFLN’s typical Vπ·L of ~1 V·cm, this is also two orders of magnitude smaller.

[Figure 2: 6 Modulator Platforms, Vπ·L vs Active Length scatter]

Why does this matter for AI optical interconnect? GPU, HBM, and switch ASIC are the visible capex of AI infrastructure. The next bottleneck is moving toward the electrical-to-optical conversion layer that connects them. The optical modulator that converts electrical signals into light, a µm-scale device, is the next bottleneck.

The key constraint is power per bit. As lane rates climb from 800G to 1.6T to 3.2T+, two metrics tighten simultaneously: per-lane data rate and per-bit power consumption. Conventional silicon modulators (carrier-depletion or carrier-injection, plasma dispersion based) hit the ~50 GHz bandwidth limit. TFLN is faster and has linear Pockels response, but its mm-cm active length keeps the modulator footprint and driving voltage above what hyperscaler optical interconnect roadmaps target. POH is the platform that crosses both bottlenecks: ~1 THz bandwidth and µm-scale length.

§2 Summary: POH = SPP field compression in 100 nm slot + organic EO chromophore. Active length 10-15 µm, ~1 THz bandwidth, sub-fJ/bit class drive. From the AI optical interconnect bottleneck perspective, POH is the platform that crosses both bandwidth and footprint/power bottlenecks. The chromophore inside the slot is the source of all this performance.

3. Why This Mapping Is the Key Variable for the Next 12 Months

Up to this point, the public-record facts are straightforward: Marvell acquired Polariton, the deal price is undisclosed, the application target is scale-across / DCI / ZR/ZR+ coherent / 3.2T and beyond. POH device performance is approximately 1 THz bandwidth, µm scale.

The reading from §4 onwards starts diverging here. Did Marvell really acquire 100 percent of the technology that delivered the Optica 2025 record device? The press release says “Plasmonic-Organic Hybrid technology and the engineering team,” but the chromophore inside the slot is a chemically synthesized organic EO material. Polariton’s slot structure and the chromophore inside that slot are not the same IP. Acquiring the slot and acquiring the chromophore are separate questions.

[Figure 3: POH Device Cross-Section, What Marvell Bought vs What Stayed Outside]

This is where the asymmetry that creates the LWLG opportunity opens up. Specifically, three things: (1) the structural reason why the chromophore must remain external, (2) the gap between standalone track value and Polariton-exposure option value in the LWLG valuation function, (3) whether NLM Photonics’ Selerion can close the second-source risk vector.

This is where the investment-relevant boundary appears. So far, public materials have established the device performance and the acquisition target. The real difference starts in §4. Marvell acquired the slot, but did not acquire the chromophore. Within the next 12 months, the answer to the question “what chromophore goes into Marvell’s POH production line” is the asymmetric position variable. Two hidden cards live in this cycle. One is the single line in the Optica 2025 acknowledgment. The other is the multi-vendor POH ecosystem confirmed in the Nokia ECOC 2023 paper as a primary source. These two are the variables for the next 12 months of asymmetric position.

Paid Section Guide

• §4. The one axis Marvell did not acquire (chromophore layer): Optica 2025 acknowledgment primary source + Nokia ECOC 2023 paper primary source. The exact citation showing external OEMs evaluated the Polariton + LWLG combination from approximately 2.5 years before Marvell’s acquisition.

• §5. Six modulator platforms and LWLG’s standalone track: Si carrier / InP MZM / TFLN / BTO / EO Polymer / POH order-of-magnitude comparison. The structure where LWLG builds value through Tower / GF / AMF separately from Polariton exposure, plus LWLG company position table (market cap, cash, monthly burn, 4 Stage 3 customers, all primary source).

• §6. Perkinamine vs Selerion (real second-source risk and CPO suitability): Material-by-material r₃₃, Tg, GR-468 pass status, and 100°C+ CPO environment suitability. Where NLM Selerion has the upper hand by metric, head-on.

• §7. Supplier relationship reorganization and LWLG’s disclosure discipline: What the POET event revealed about Marvell’s post-acquisition supplier operating rules. How LWLG’s anonymous customer naming + Chowdhury CTO appointment + Michael Best IP advisory line align toward a licensing revenue model.

• §8. 12-24 month scenarios and monitoring: Scenarios A/B/C classification with LWLG / NLM impact mapping. Seven monitoring points for 2026 H2 (OFC, ECOC, quarterly earnings, Tower/GF tape-out, etc.).

Read more

Contents

1. Intro: Two paths in the same fab

2. Monolithic vs Bonding (technology background)

3. Five domains where bonding lives

4. Where the real divergence begins

5. Identifying the 7 bonding equipment names + comparison

6. 4-group classification + differentiation table + company-by-company analysis

7. Mapping the demand side that is locked into bonding

8. Scenarios + monitoring points + closing

9. References

1. Intro: Two paths in the same fab

Computing cycles have always split at assembly. The 1990s DRAM cycle showed up first at the wire bonder. The 2010s mobile SoC era surfaced first at flip chip. Now the AI cycle is moving to hybrid bonding and TCB. With monolithic scaling effectively halted, bonding has become the new chokepoint, and seven bonding equipment names sit directly in the path of HBM4 and CPO packaging capacity bottlenecks. This is the investment universe for this article. The article covers how monolithic and bonding diverge inside the same fab, what the five bonding domains (HBM, chiplet, Silicon Photonics (SiPh), CPO, optical interposer) are, how these seven companies are differentiated, and which one is most leveraged in which scenario. Which seven, how they are split, and who fits into which group are covered in the paid section.

[Figure 1: Evolution of bonding by computing cycle, 1990s DRAM (wire bond) → 2010s mobile SoC (flip chip) → 2020s AI (hybrid bonding/TCB)]

The seven companies face very different cycle exposure. Two are bonding pure-plays at roughly $20B market cap each, with the highest beta. One is a $320B WFE giant whose bonding exposure is single-digit percentage of revenue but locks in a strategic ecosystem position. Two are diversified backend players where bonding sits as one segment among several. One is in transition from a wire bonder cash cow to fluxless TCB and HBM. One is private but is a standard-grade W2W bonder supplier and a peer valuation comp.

One-line takeaway: AI’s real chokepoint is not wafer fab capacity but advanced packaging bonding capacity. This article picks seven bonding equipment names directly exposed to HBM4, CPO, and photonic packaging cycles. Which seven, how they are split into four groups, and which one is most leveraged in which scenario are covered in the paid section.

2. Monolithic vs Bonding: two paths in the same fab

2.1 The limits of monolithic

Trying to fit everything onto one GPU die has already hit the reticle limit (about 858 mm²). Cramming logic plus memory plus I/O onto the same wafer node breaks down on both yield and cost. Whether it is NVIDIA Rubin, AMD MI400, or Apple M3 Ultra UltraFusion class SoC, leading-edge compute is moving away from monolithic single-die toward die-to-die integration and advanced packaging. HBM is even more dramatic. Going from 12-Hi to 16-Hi requires wafer thinning down to 30 μm [3], and packing 16 layers under the JEDEC HBM4 package height limit (775 μm) puts real pressure on the stack. Conventional micro-bumps and TCB are running out of margin, and hybrid bonding (Cu-Cu direct contact) is becoming an increasingly important option.

2.2 The bonding answer

Bonding fundamentally means “stacking different dies tightly together”. There are several techniques:

• Wire Bonding: connecting chip pads to a lead frame with thin gold or copper wires. Used in legacy memory, automotive, and SMT for decades.

• Flip Chip: flipping a chip upside down and joining it to a substrate via solder bumps. The standard for mobile SoCs and GPU substrates from the 2000s.

• TCB (Thermo-Compression Bonding): melting micro-bumps with heat and pressure. Hits voids issues at fine pitch, evolving into fluxless TCB. The primary technique for HBM3E 12-Hi, partially extending to HBM4 12-Hi [4]. Tool price about $1-2M per system.

• Hybrid Bonding: directly joining Cu pads to Cu pads at atomic level without micro-bumps, with the dielectric chemically bonded. Interconnect pitch comes down to sub-10 μm, and electrical and thermal characteristics are inherently better. Tool price about $3M, requires CMP step, demands higher cleanroom class [4]. Already used in AMD X3D, TSMC SoIC, and partially in HBM4 16-Hi.

[Figure 2: Monolithic vs Bonding comparison flowchart]

2.3 Two paths in the same fab

Same fab, two paths. Monolithic uses EUV / DUV lithography, etch, and deposition to print everything onto a single wafer. Bonding finishes individual dies on different wafers, then physically and chemically joins them. The two are not entirely substitutes (advanced logic chips still need monolithic for the base die), but bonding capacity is what determines the ramp speed of HBM stacks, AI chiplet integration, and photonic / CPO modules.

Worth noting: a front-end WFE giant has reached down into back-end bonding by acquiring a stake in one of the bonding pure-play companies, locking in one ecosystem axis [2]. Read in reverse, this means advanced packaging bonding is no longer a simple “back-end” step but an area as critical as wafer fab itself. Important enough that a front-end giant has to either enter directly or acquire. Which company acquired which, and what the ecosystem implication is, are covered in the paid section.

Key: same fab, two paths. Monolithic scaling is hitting reticle and HBM height limits, opening room for the bonding side. A front-end giant reaching down into the bonding pure-play ecosystem signals that bonding is now strategic territory.

3. Five domains where bonding lives

Bonding is not one market. It is five separate markets, where bonding requirements differ by precision, throughput, and process domain.

3.1 HBM memory stack

The current generation (HBM3E) splits in two. SK Hynix uses MR-MUF (Mass Reflow Molded Underfill), which reflows and underfills all layers in a single step [5], while Samsung and Micron use TC-NCF (Thermal Compression with Non-Conductive Film), pressing one layer at a time. SK Hynix took the lead on both yield and thermal performance, which translated into roughly 50% HBM market share advantage.

The next generation (HBM4) is going to 16-Hi, which requires wafer thinning to 30 μm to fit 16 layers under the JEDEC 775 μm height limit [3]. That is roughly half of the 50 μm used at the 12-Hi era. Three different paths emerge:

• SK Hynix: keeping its MR-MUF strength as the primary path for HBM4 / HBM4E 12 to 16-Hi, while adopting hybrid bonding mass production in parallel for next-gen high-stack HBM (a two-track strategy). As of April 2026, SK Hynix has completed 12-high HBM hybrid bonding validation [35], and reportedly placed its first mass-production order for an inline hybrid bonding system co-developed by two global bonding equipment companies, at roughly KRW 20 billion [36]. This is moving past backup status into actual mass-production deployment.

• Samsung: aggressive bet on hybrid bonding. As of April 2026, supplying 16-Hi HBM4 hybrid bonded samples to NVIDIA but with reported yields around 10% [3]. Hybrid bonding goes mainstream from HBM4E (2028).

• Micron: continuing TC-NCF, with the largest hybrid bonding patent portfolio (180+ filings, Adeia IP license) [7].

There is one variable here. The JEDEC HBM4 standard was officially released as JESD270-4 on April 16, 2025, allowing 775 μm package height for both 12-high and 16-high (already loosened from HBM3E’s 720 μm) [41]. And on April 1, 2026, TrendForce reported (citing Newsis / Chosun Ilbo) that JEDEC is considering further relaxation to roughly 900 μm starting from HBM4E [6]. If this happens, the hybrid bonding timeline could slip by a generation. This is a variable that directly affects bonder ordering schedules.

On the equipment side, two companies hold TCB market share, one is targeting the next generation with hybrid bonding, and one has entered with fluxless TCB (first HBM system shipped to a large memory customer in December 2025) [8]. The exact identification table is in the paid section.

3.2 AI logic chiplet

This is one of the biggest current revenue drivers. AI accelerators increasingly depend on advanced packaging: Hopper-class GPUs already rely on large logic dies packaged with HBM, while Blackwell / Rubin-class platforms move deeper into multi-die integration. AMD MI300 / MI400 and X3D-class products make the chiplet and hybrid-bonding angle even more explicit. Across the board, leading AI compute chips are now built by bonding logic, memory, and I/O onto a single substrate or interposer.

TSMC effectively dominates this market with SoIC (System on Integrated Chips) and CoWoS (Chip-on-Wafer-on-Substrate), and CoWoS capacity is expanding toward 100,000 to 120,000 wpm by 2026 [4]. Intel runs its own line with Foveros Direct + EMIB, and Samsung Foundry is following with X-Cube.

AMD X3D (V-Cache memory die hybrid bonded onto a processor) is already a mass-production case using hybrid bonding [9]. An integrated D2W (Die-to-Wafer) inline line jointly developed by a front-end giant and a bonding pure-play is targeting high-volume manufacturing in this domain, and the D2W hybrid bonding market is estimated to grow from about $275M in 2025 to $2.4B+ in 2030 (CAGR 57%) [10]. Which two companies are behind this integrated line is covered in the paid section.

3.3 Silicon Photonics heterogeneous integration

This is the area where photonic chips (SiPh) and III-V light sources (GaAs, InP) get put on top of silicon. Monolithic absolutely cannot do this, so bonding is mandatory.

The classic heterogeneous integration example is bonding an InP laser onto a silicon photonics waveguide. Two global standard-grade suppliers (one listed, one private) hold positions in W2W (Wafer-to-Wafer) bonding for this domain [20], and one company supports silicon photonics CPO bonding via fluxless TCB [11]. Which company sits where is identified in the paid section.

3.4 CPO (Co-Packaged Optics)

This is the area where optical engines are co-packaged with switch ASICs / GPUs in the same package, replacing traditional pluggable optical transceivers. Reduces signal loss, lowers power, and increases bandwidth.

TSMC’s COUPE (Compact Universal Photonic Engine) platform was targeting 2026 production [12], and Samsung Foundry announced TC bonding-based OE (Optical Engine) for 2027 and CPO turnkey for 2029 [12]. NVIDIA Spectrum-X Photonics CPO switches, and Marvell / Broadcom CPO optical engines all fall in this category. NVIDIA has officially announced its own CPO switches, but the exact production split with external foundry platforms such as COUPE is reported by industry sources rather than formally specified, so it remains a monitoring item.

3.5 Optical Interposer (POET-class)

The newest category. Optical waveguides are drawn on a silicon or other substrate, with various optical and electronic chips bonded on top.

POET Technologies’ POET Optical Interposer is the representative case. A platform that integrates electronic and photonic devices (lasers, modulators, photodiodes, DSPs, etc.) into a single multi-chip module on a silicon interposer, with sub-micron precision die bonding as the core [13]. The business model itself sits on top of die bonding. If bonder capacity is short, POET’s commercialization speed is directly affected.

Lightmatter Passage falls in the same category. A 4,000 mm² active photonic interposer (GF Fotonix process) with logic chiplets stacked on top, creating in-package optical I/O [42]. Bigger than POET’s interposer and the logic chiplets are bonded directly on top, but the “bonding capacity = commercialization speed” structure is the same. Ayar Labs TeraPHY is structurally different. CMOS and photonic are monolithically integrated into a single die chiplet (also GF Fotonix-based), then co-packaged with the host SoC in the same multi-chip module package. This is chiplet-level co-packaging rather than stacking on an interposer, depending more on advanced packaging chiplet integration capacity than on bonding capacity [43].

Key: bonding is five separate markets, each requiring a different type of bonding and a different precision metric. No single company is #1 across all five.

[Figure 3: The five domains where bonding lives. HBM memory stack, AI logic chiplet, SiPh heterogeneous integration, CPO, optical interposer in one visual]

Subscribe now

4. Where the real divergence begins

The five domains above can be followed using public company disclosures. From here on, the real differentiation begins.

The bonding equipment market is converging on two specific events that will shift the whole picture over the next 12 months.

One is a takeover competition for one of the bonding pure-plays. According to a Reuters single-source report (London March 12, 2026 / NY March 13), one bonding pure-play company is evaluating takeover interest from two global WFE giants via Morgan Stanley, and that company’s stock hit an all-time intraday high of +14% the day of the report [14]. This is the biggest variable in bonding equipment ecosystem reshuffling. Rankings of the seven companies could shift by an order of magnitude depending on who acquires, or whether the deal falls through. Which company is the target, and which giant already holds a 9% stake, are covered in the paid section’s company-by-company analysis.

The other is JEDEC HBM4E height relaxation. According to TrendForce on April 1, 2026, JEDEC is considering further relaxation of the package height to around 900 μm starting from HBM4E (7th generation) [6]. If this happens, MR-MUF / TC-NCF could survive one more generation even at 16-Hi, and full hybrid bonding adoption gets pushed past HBM4E (2028). Samsung’s aggressive bet on hybrid bonding could turn out to be wasted, or could become a Samsung-only differentiation opportunity.

How these two cards affect each bonding equipment company, which company falls into which of the four groups, and which demand-side names locked into bonding (HBM 3 majors, AI compute, photonic, foundry) are most leveraged are the core variables for asymmetric positioning over the next 12 months.

[Figure 4: Why precision matters, micron to sub-100nm progression and the applications it unlocks]

Subscribe now

Share

Read more

POET dropped 47.35 percent on April 27. The surface reason is an NDA breach, but the real question this piece raises is different. Why did Marvell, the day after announcing the Polariton acquisition, claw back every purchase order POET had received from Celestial AI in a single stroke? And why did the cancellation reach all the way back to the initial production order first disclosed on April 25, 2023? Pure retaliation would only have required cutting the recent batch. The 24-hour gap and the three-year-deep PO recall together look closer to a layer-cleanup signature.

This piece breaks the event apart through the lens of optical-stack realignment. The free section unpacks the following:

• §1 7 days in April: the timeline from the Polariton acquisition announcement (4/22) to the written notice to POET (4/23) to the -47.35% move (4/27). Combined with the Celestial AI acquisition close (2/2), the sequence shows modulator IP being pulled inside Marvell’s stack.

• §2 Why μm is the key: the POH (Plasmonic-Organic Hybrid) modulator has an active section of 4 × 25 μm × 3 μm. A device capacitance of about 3 fF is what produces the sub-fJ/bit and 1 THz bandwidth values. This is a fundamentally different length scale from the sub-mm to mm-scale InP DFB light source.

• §3 Why μm devices cannot be chiplet-bonded: bond-pad capacitance, RF interconnect parasitics, electrode transition, and organic-chromophore reliability are four reasons monolithic or BEOL integration is the natural fit. POET’s pick-and-place + passive-alignment paradigm has little obvious room here.

• §4 Same polymer family, different integration schemes: the chromophore linkage between LWLG and Polariton, and the structural reason POET, being modulator-material agnostic, never locked in to any one camp.

Core thesis: this event is not a simple customer dispute. It is the first signal of a cycle where Marvell aligns which layers it pulls inside and which layers it leaves outside.

That is the surface of the event. The real divide starts in the paid section. The mechanism-level reason Ge EAM and POH do not collide inside the same stack (Franz-Keldysh vs Pockels), the multi-vendor POH ecosystem confirmed in the Nokia ECOC 2023 paper as a primary source (the Polariton-LWLG-Nokia link plus the Aref Chowdhury Nokia → LWLG CTO move plus the Karin Raj Nokia CTO Europe ↔ Sivers board concurrent seats), why SIVE escapes the same pressure, POET’s path forward and the cumulative disclosure-pattern risk, a 7-ticker + Polariton-internal positioning map, and a 3-scenario benefit/loss matrix are all unpacked in the paid section.

Table of Contents

1. Timeline: 7 days in April

2. Plasmonic-Organic Hybrid modulators: why μm is the key

3. Why μm devices cannot be chiplet-bonded

4. Same polymer family, different integration schemes

5. Why SIVE is not in the same layer

6. Celestial Ge EAM and Polariton POH: two modulators, two layers in the same stack

7. Nokia ECOC 2023 and the multi-vendor POH ecosystem

8. POET’s light-source integration was a strength. Why did it fail to land at Marvell?

9. POET’s path forward

10. Closing

POET dropped 47.35 percent on April 27, closing at $7.95.[1][2] The surface reason is an NDA breach. POET CFO Thomas Mika went on Stocktwits TV and publicly disclosed both the Marvell purchase order and the state of Foxconn/Luxshare negotiations, and Marvell labeled this a violation of confidentiality obligations regarding the purchase orders and shipping information, then issued written notice on April 23 canceling every purchase order POET had received from Celestial AI.[1][3][4]

That is the official explanation.

But the question investors should be asking is different.

Why was every PO canceled, all the way back to the initial production order first disclosed on April 25, 2023, instead of just the recent batch? Pure retaliation would only have required cutting recent orders. Marvell reached back three years to wipe the entire history. Stranger still is the timing. Marvell announced the Polariton Technologies acquisition on April 22. Marvell issued written notice to POET on April 23. One day apart.[1][11]

The core hypothesis of this piece starts here. POET cancellation is hard to read as just an NDA breach. It looks more naturally as an external optical-engine slot being cleared right after Marvell brought modulator-and-system stack in-house through Celestial AI (closed 2026.2) and Polariton (announced 2026.4.22).

The official cause is the NDA breach. The reading below is PhotonCap’s inference based on the timing, scope, and technology layer in public materials. We have not seen Marvell’s internal decision memo.

So this piece narrows to two questions.

One, Celestial AI already had Ge EAM, so why did Marvell pay for another modulator? Aren’t they overlapping? Two, is this POH platform really Marvell-only, or was it already in the Nokia ECOC 2023 paper?

Once those are answered, the LWLG and SIVE optionality expands one notch beyond the prior PhotonCap take. And the real alpha here is not a single PO cancellation. It is which layers of the optical stack Marvell pulls in-house and which it leaves outside. The 12-month winner-and-loser line will fall along that boundary.

Summary: The official reason is an NDA breach. The timing (one day after Polariton) and the scope (entire history including the 2023 PO) are hard to read as simple retaliation. The framing of this piece is which layers Marvell is erasing from its optical stack.

1. Timeline

Marvell’s optical asset acquisition timeline, laid out in order:

• 2021: Inphi $10B (coherent DSP plus SiPh full stack) [5]

• 2021: Innovium $1.1B (Ethernet switch silicon) [6]

• 2024.10: Celestial AI absorbs Rockley Photonics IP, including over 200 patents on EAMs and optical switching [7]

• 2025.08: Celestial AI publicly discloses Photonic Fabric Module Gen1 with OMIB packaging at Hot Chips 2025, naming GeSi EAM as its chosen modulator [8][9]

• 2025.12 announce / 2026.2 close: Celestial AI $3.25B plus $2.25B milestone (Photonic Fabric, scale-up optical interconnect) [10]

• 2026.4.22: Polariton Technologies (undisclosed terms, Swiss POH modulator, ETH Leuthold spin-off) [11][12]

• 2026.4.23: Marvell sends written notice canceling every purchase order POET had received from Celestial AI [1]

• 2026.4.27: POET 6-K disclosure, stock down 47.35 percent, closes at $7.95 [1][2]

The interval between the Polariton acquisition announcement on April 22 and the PO cancellation written notice on April 23 is exactly one day. Two modulator IP acquisitions in eight months, on two different layers, paid for separately, followed immediately by clearing an external optical engine vendor. The sequencing is too clean to read as random.

Summary: Across eight months, Celestial (Ge EAM, intra-package) and Polariton (POH, transceiver/coherent) were absorbed as two separate modulator-IP layers, and POET’s entire PO history was canceled in the 24-hour window between them. The surface is NDA breach. The sequencing reads more like layer cleanup.

2. Plasmonic-Organic Hybrid Modulators: Why μm Is the Whole Point

This is the technical anchor. The plasmonic-organic hybrid (POH) modulator developed by the Polariton / ETH Leuthold lineage has an active region measured in micrometers. The numbers:

Heni et al., Nature Communications 10:1694 (2019)[13]

• Active section 4 × 25 μm × 3 μm

• Sub-1V driving electronics

• 0.07 fJ/bit at 50 Gbit/s

• 0.3 fJ/bit at 200 Gbit/s

• 2 fJ/bit at 400 Gbit/s

• Device capacitance about 3 fF

Melikyan et al., Nature Photonics (2014)[14]

• 29 μm long plasmonic phase modulator

• 65 GHz bandwidth

• Foundational POH paper

Haffner et al., Nature Photonics (2015)[15]

• 10 μm long all-plasmonic Mach-Zehnder

• Microscale optical communication demonstration

Burla et al. / Ummethala et al., 500 GHz and THz plasmonic series[16]

• Plasmonic modulation pushed into the sub-THz / THz domain

Horst et al., Optica (2025), Polariton + ETH joint[17]

• 3-dB EO bandwidth 997 GHz

• 6-dB EO bandwidth above 1 THz

• Frequency response measured up to 1.1 THz

Wiley Advanced Materials Technologies (2025), silicon MRM comparison[18]

• Conventional silicon micro-ring modulators bottleneck around 42.5 to 54 GHz EO bandwidth. Order-of-magnitude gap to POH.

[Figure 1: POH μm active region vs InP DFB sub-mm/mm-scale light source, RF parasitic contrast]

The point is not the headline numbers. The point is that the active region is in the micrometer scale. Conventional silicon Mach-Zehnder modulators need millimeter-scale length. Thin-film LiNbO3 also needs millimeter scale. POH puts an EO chromophore inside a metal slot. The electromagnetic field is squeezed into a slot tens of nanometers wide, and the EO efficiency rises sharply. The same input produces a much larger output.

This architecture is the result of fifteen years of continuous work in the Leuthold group: plasmonic absorption modulator at KIT in 2011 (Optics Express), the 2014 Nature Photonics phase modulator at ETH, the 2015 Nature Photonics all-plasmonic MZM, the 2018 Nature low-loss plasmon-assisted modulator, the 2019 Nature Communications IQ modulator, and the 2025 Optica 1 THz result.[13][14][15][17][19]

Summary: The core POH value is sub-fJ/bit and 100 GHz plus bandwidth from a 4 × 25 μm active region. The device length scale is fundamentally different from mm-scale silicon and TFLN, and that fact dictates the integration scheme.

3. Why μm Devices Cannot Be Chiplet-Bonded

POET’s core IP is the combination of Optical Interposer, elastic averaging passive alignment, and Laser-Assisted Bonding (LAB). Mohamed et al. in JLT 2023 reported, on a 100-sample basis, post-bond X 0.09 μm, Y 0.39 μm, θ 0.0002° at 3σ using a triangle slide-stop geometry.[20] At ECOC 2025, POET demonstrated a 1.6T 2xFR4 PIC with S_sd21 of 53.7 GHz, 8 dBm per channel at 100 mA, 4 dB ER, and 2 dB TDECQ at 106 Gbaud PAM4 SSPRQ.[21] The equipment stack is ASMPT AMICRA NANO (±0.2 μm), Vanguard FaML (Mycronic), and SMBI wafer-level burn-in.[22]

This paradigm fits mm-scale devices very well. Sivers DFB lasers and Mitsubishi dual-facet EMLs are millimeter-scale, so the interface loss from bonding is small relative to device loss itself. The full equipment-and-paper analysis was covered in the prior PhotonCap POET deep-dive piece.[22]

The same paradigm becomes much less natural for μm-scale POH. There are four reasons.

1. Bond-pad capacitance can easily overwhelm device capacitance. The Heni 2019 POH device capacitance is around 3 fF. Once the device is attached as a discrete chiplet, the bond pad, solder bump, wire, redistribution metal, and package trace can add tens to hundreds of fF of parasitic capacitance. The driver is no longer charging a 3 fF modulator. It is charging the package interface as well. That erodes the sub-fJ/bit value proposition at the interface.

2. RF interconnect parasitics can become larger than the device parasitics. A μm-scale modulator only preserves its advantage if the RF path stays extremely short and well controlled. Chiplet bonding introduces pads, bumps, short traces, vias, and package transitions. Their impedance discontinuities, reflections, and losses can dominate the high-speed signal. Even a device with THz-class EO bandwidth loses much of its system-level value if the package transition becomes the bandwidth bottleneck.

3. RF launch and electrode transition enter the same length scale as the active device. When the POH active section is roughly 25 μm long, the bond pad, electrode taper, and nearby signal-line transition are no longer remote packaging details. At that length scale, the modulator response is no longer determined by the active slot alone. The RF launch and electrode geometry must be co-designed with the device. That makes monolithic or BEOL co-design much more natural than a discrete chiplet path.

4. Organic chromophore reliability requires package-level co-design. EO chromophores can be sensitive to humidity, UV exposure, thermal cycling, and long-term poling stability. BEOL-level integration lets passivation, encapsulation, and thermal path be designed together with the device process. Chiplet bonding adds underfill, edge sealing, and CTE-mismatch failure modes. This is one reason the Polariton/ETH lineage repeatedly emphasizes CMOS-compatible BEOL integration.[23]

The implication: when Marvell acquired Polariton, the modulator IP at minimum will be processed inside monolithic SiPh, not through an external packaging house. POET’s pick-and-place plus passive-alignment paradigm therefore has little obvious entry point at the device level for this specific asset.

📊 Summary: A 25 μm device routed through a millimeter-scale bonding infrastructure loses its value proposition to bonding parasitics. Monolithic or BEOL-level SiPh integration is the structurally natural fit, far more than an external packaging-house path.

4. Same Polymer Family, Different Integration Schemes

This is where the LWLG bet and the Polariton bet diverge.

When the modulator is millimeter scale, the bonding interface loss is tolerable. LWLG and TFLN families have die-level heterogeneous integration as a viable production path. POET should have aimed at the thin-film EO polymer slot, but its InP DBR plus interposer model is modulator material agnostic, so it never locked into any single material camp.

EO polymer chemistry exists at multiple suppliers, including Lightwave Logic, NLM, and others. POH does not differentiate at the polymer molecule. It differentiates at how that polymer is placed into the slot. Polariton’s integration recipe with Perkinamine series 3 is ten years of accumulated work from the Leuthold group at ETH Zurich. The March 2025 expansion of the LWLG-Polariton partnership was specifically about taking that recipe into a 200 mm foundry PDK.[24]

The Polariton acquisition created Marvell’s indirect exposure to LWLG. That topic was covered in the prior Polariton acquisition piece.[23] If the chromophore inside POH is the Perkinamine family, LWLG holds an indirect supplier position into the Marvell stack as the modulator material vendor. (This hypothesis is then confirmed at primary-source level in §7 via the Nokia paper.)

📊 Summary: Device length determines whether chiplet bonding is viable. POH naturally favors monolithic or BEOL integration. EO polymer thin films and TFLN allow chiplet. InP DFB at mm scale also allows chiplet. The integration architecture is dictated by device length, not material chemistry alone.

This is where the surface of the event ends.

[Figure 2: 7-company positioning matrix, photonic packaging layer × device material layer]

What we have unpacked so far:

• Marvell canceled every POET PO the day after announcing Polariton (timing + scope)

• POH’s μm device + sub-fF capacitance value is preserved by monolithic/BEOL integration (chiplet is possible but RF parasitics erode the value)

• POET, being modulator-material agnostic, did not lock in to any one camp

The real divide starts here. The seven tickers (POET, SIVE, LITE, COHR, MRVL, NOK, LWLG), plus Polariton as a Marvell-internal asset, all sit on the same event, but they split into four distinct groups: the company already pushed out of the packaging layer, the company on a different layer that escapes the pressure, the company carrying multi-OEM material optionality, and the company that sits on both ends of the two acquisitions. Their risk/reward profiles diverge sharply.

Two Hidden Cards live in this cycle. One is the mechanism-level reason Ge EAM and POH do not collide inside the same stack. The other is the multi-vendor POH ecosystem confirmed in the Nokia ECOC 2023 paper as a primary source. These two are the asymmetry-position variables for the next 12 months.

The whole question reduces to one line.

Which layers is Marvell pulling inside, and which layers is it leaving outside?

— PAID SECTION begins —

The paid section unpacks that question into a layer-realignment map across the seven companies.

What the paid section covers:

• §5 Why SIVE is not in the same layer: what it means that Marvell absorbed modulator IP while leaving light-source IP outside

• §6 Celestial Ge EAM and Polariton POH mechanism split: the mechanism-level reason the two modulators do not collide inside the same stack (Franz-Keldysh vs Pockels)

• §7 Nokia ECOC 2023 + multi-vendor POH ecosystem: primary-source confirmation of the LWLG-Nokia-Polariton link, plus the Aref Chowdhury personnel move

• §8-§9 The nature of POET’s layer pressure and its path forward: why POET’s strength in light-source integration did not land at Marvell, and the MSA pluggable + module-vendor pivot

• Additional analysis: 7-ticker identification table plus Polariton-internal mapping, 5-group classification, Ge EAM vs POH wavelength mapping, 3-scenario benefit/loss matrix (Polariton in-house success / TSMC COUPE fallback / integration delay), Hidden Card events (256 GBd PAM trade-off, Polariton price estimate + LWLG royalty scenarios, Nokia next-gen coherent POH adoption odds), quarterly monitoring points

• §10 Closing: thesis wrap-up

Read more