citriniresearch.com api v1 post pdf postId=190712071

Let There Be Light Revisiting the Optics & AI Connectivity Theme CITRINI MAR 12, 2026 ∙ PAID For readers who have recently joined us, our job as analysts is to wade through the noise and identify where the puck is going. We’ve been doing this since we first began publishing research in May 2023 with “AI: Global Equity Beneficiaries”, opting to focus on the second and third-order effects of the massive data center buildout necessary for AI. Today, the AI buildout remains relentless, with an estimated $650B–$700B in capital expenditures to be outlaid by hyperscalers in 2026.

Since then, we have woven a thread through the AI supply chain backed by our thesis that infrastructure bottlenecks would dominate investor attention. As each layer of the supply chain is reinforced and expanded, the bottlenecks simultaneously respond to technological adoption and drive new demand in the infrastructure layer. In order to seek out these opportunities, we’ve continually deepened and expanded our existing coverage. Today, that leads us to update what was perhaps the single-most emblematic stance of our “secular growth at cyclical prices” approach: optics. CitriniResearch was one of the most timely on the street to outline the opportunity for optics.

For two months, with most of the stocks at or near 52 week lows, we pounded the table on this in two separate pieces. In “Interconnects 101” published in July 2024, we argued that AI’s next beneficiaries would become “more skewed towards who can provide the most marginal benefit,” specifically “optical interconnects and co-packaged optics (CPO)”. A few weeks later, Introduction in “Can You Hear Me Now? The Coming Optics Shortage & Telco Supplier’s New Dawn”, we made an even bolder and more bullish call that presented these cyclically blown out names as becoming a prominent AI bottleneck.

We even compared Ciena’s (CIEN US) revision cycle in 2024 to Nvidia’s going into 2023. In other words, our first call was that optics would matter a lot and the second was that the shortage would emerge before the market even noticed. This was not a popular take. At the time, optics sat inside a telecom supply chain most investors wanted nothing to do with. Consensus saw cyclical depression and inventory indigestion that was years from normalizing. We saw AI-driven connectivity demand colliding with years of underinvestment, setting up “a resulting optics shortage” precisely where the market was least interested in looking.

If you can believe it, given the way our optics names have traded since, many laughed at us for it. We were branded contrarians at best and tourists at worst, wading into a meat-grinder of a cycle that would end our streak of good calls. Well…they’re not laughing anymore. These calls were early, uncomfortable, and very right. Our connectivity basket has more than tripled since we published it – recently pulling ahead of our Interconnects basket, which has lagged year-to-date as the market has come to appreciate optical interconnects at the expense of names like Astera Labs (ALAB US) and Credo (CRDO US).

And it’s not done yet – the basket is up 35% YTD, led by now favorite Lumentum (LITE US). This piece is the next chapter. The optics trade has broadened and matured, certainly even crowded in some names. The risk-reward on some of the crowd favorites has become significantly more symmetrical. Since it is now well established that optics are core to the AI trade, we must ask where the next asymmetry sits. Despite more than enough reason to claim the win and walk off…we believe there are still opportunities further down the stack. As the bandwidth constraints of the “Memory Wall” pile up, the opportunities are in the companies enabling the next generation – photonic components, substrates, capital equipment, and the businesses positioned to benefit as CPO becomes a necessity.

We believe there is ample runway and enough yet-to-be crowded names in this sub- theme to provide continued outperformance for AI investors. Additionally, we think it’s time to be cautious and consider taking profits on some of the names we originally initiated on. Not because optics slows down, just because they’ve become priced for perfection. Below the paywall, we provide a brief overview of what “photonics” are and outline our favorites of the picks and shovels the market has overlooked that we think remain undervalued in the shadow of LITE and peer’s massive run. We’ve also teamed up again with financial investigative journalists Hunterbrook to dive into what we believe is a truly asymmetric play on the photonics supply chain.

You’ll remember them from our last collaboration covering the dual robotics and memory testing tailwind of Teradyne (TER US) that provided our readers with an extremely successful play. We try to make all our pieces lindy, reference texts that withstand the test of time. Even today, if you revisit our original AI primer or Interconnects sub-thematic, there’s information that’s still useful. So, while we’ll avoid a science lesson, we’re going to refresh you on what we covered in our Interconnects piece, summarize briefly what you need to know about photonics, and detail why you should care about this topic as an investor.

Data centers have historically used electrons to transmit information between GPUs, racks, and servers via traditional copper wires. So far, this technology has been adequate for contemporary data center architecture. The industry has accommodated expanding compute demands by shrinking and densifying copper interconnects. These demands have only grown given the emergence of artificial intelligence and high performance computing capabilities. However, while raw computing performance (FLOPs) have improved 60,000x over the past two decades, electrical interconnect bandwidth has only increased 30x. The contrast in bandwidth can be directly attributed to the physical constraints inherent in electrical interconnect technology.

Electrons traveling through copper wire generate heat, which constrains rack density and can create distortions in signal transmission. What’s more, electrons physically dissipate over long distances as they collide with other atoms, reducing cluster performance and reliability. 1) Photonics 101: Controlling Light Photons, by contrast, don’t have this problem. Photons are light packets that travel much faster, are more energy efficient, and can carry orders of magnitude more data. Because of this, our belief is that the entire semiconductor industry - from fabs to capital equipment manufacturers - will eventually transition to photonic based technologies.

Key to this transition are PICs. A Photonic Integrated Circuit (PIC) combines all of the components needed to generate, encode, route, and detect light onto a single device. Instead of electrons transmitting through copper wires, PICs use a laser source to beam photons through optical components. The end game for the semiconductor industry is to close the performance gap between computing and connectivity, ultimately surmounting the optical constraints of the Memory Wall. Similar to some of the other AI-adjacent technologies we’ve written about, photonic applications extend beyond just data centers. Sensing and LiDAR applications employ photonic components for autonomous vehicles, robotics, and defense technology.

Telecommunications specifically have relied on photonic technology for decades in long-haul fiber networks. Defense and electronic warfare applications like nLIGHT (LASR US) which we’ve covered in depth, employ III-V photonic devices across tactical communications, missile guidance, and signal intelligence. This is what excites us about the opportunity in photonics enablers: The core components, substrate materials, and equipment used to generate photonics can serve several markets simultaneously, creating a powerful confluence of tailwinds for companies servicing the photonic supply chain. While this has been appreciated in the more obvious winners, it’s less apparent in numbers for the recipients of capex spent on getting to the next generation.

Within data centers, however, the main application for photonics today is at the interconnect layer. When a GPU finishes a computation, the resulting data needs to travel to another GPU, to memory, to a switch, across a rack – you get the gist… That journey is where photonics shines (no pun intended). Today, the industry uses two primary approaches to convert electrical signals from chips into optical signals: pluggable transceivers and co-packaged optics (CPO). We’ve covered both of these before in Interconnects 101, so we won’t bore you too much. Just in case, though, (or if you want to avoid revisiting our admittedly heavy-handed kitchen analogy) here’s a refresher.

Pluggable transceivers are standalone modules that plug right into the front panel of a network switch or server. Think of them like a thumb drive for converting electrical signals into light and back again. They’re straightforward and easy to swap, which is why they have been the standard for over a decade. The problem with pluggables is physics. They simply sit too far away from the chip itself, meaning the signal has to travel a relatively long distance over copper traces on the circuit board before it ever reaches the optics. This introduces all sorts of problems… “The challenge with scaling out GPUs to many hundreds of thousands is the connection of the scale-out.

When the data centers are now the size of a stadium, we need something new — and that’s where silicon photonics comes ” “Each GPU requires six individual transceivers, consuming 180 watts and costing $6,000 per GPU. How do we scale up now to millions of GPUs? Because if we have a million GPUs, we need 6 million transceivers consuming 180 ” Jensen Huang, CEO, Nvidia, GTC - March 2025 This is where co-packaged optics (CPO) come into the picture. They move the optical engine directly onto the switch or processor itself. Instead of sending the electrical signal across the board and into a pluggable module at the edge, CPO converts the signal into light right next to the chip.

The copper path shrinks, cutting power consumption per port from 30 watts down to as low as 9 watts. Yet, even as the industry is increasingly bracing for CPOs, pluggable demand is not going away. In fact, it’s exploding. This is why our Interconnects and Connectivity baskets have done so astonishingly well since we first wrote about them. Every hyperscaler building GPU clusters today is buying pluggable transceivers by the hundreds of thousands from the companies we detailed. Innolight (300308 CH), the largest transceiver supplier to NVIDIA, can barely keep up with demand for its 800G and 1.6T modules.

Coherent (COHR US), Lumentum (LITE US), and Applied Optoelectronics (AAOI US) are all expanding production capacity as fast as they can build it. The reason is simple: pluggables are standardized and work everywhere. As a result, pluggable transceivers dominate scale across and scale out functions in datacenters and racks. Meanwhile, CPO is implemented at the bandwidth chokepoints highlighted by Huang, where traffic from dozens or hundreds of GPU racks converge into a single box. However, we will not be able to solely rely on pluggables for long. The big problem with pluggables is their extremely high power consumption.

At 6.4 Tb/s and beyond, power consumption rises substantially. This means two things. First, some of the interconnect names we originally covered may be too crowded, as they find themselves unable to scale up CPO as quickly as forecast. One that comes to mind is Applied Optoelectronics (AAOI US). We were very early to the turnaround for AAOI – first covering it when shares were around $3. Now, the stock sits above $120 and has become somewhat of a retail favorite – seemingly going up 10% every day this year. AAOI has a ramp that is concentrated and fragile; ten customers were 96.6% of 2025 revenue, with Microsoft making up 28.8%.

Meanwhile, the company has been scaling 800G capacity to more than 100,000 units per month, and its explicit CPO- aimed 400mW laser is only at the sample stage with volume production expected later in 2026. The bridge right now looks heroic, and it certainly was not appreciated until recently. It is fully appreciated now. If the market decides the merchant pluggable layer has gotten crowded before CPO revenue is material and scarcity rents migrate into CPO optical engines, external lasers, PICs, and advanced packaging… well, then there’s no reason the stock couldn’t trade 50% lower.

After all, their mid- 2027 targets require 5x current capacity on a technology they’ve never shipped. If Nvidia’s timeline on CPO is not delayed, we believe AAOI is likely closer to the top than recent price action might suggest. However, the bear case on AAOI (that they’re assembling external components rather than manufacturing proprietary lasers) only reinforces the fact that someone still has to grow those epiwafers and supply those tools. That brings us to our next point…. Second, the companies that benefit from CPO becoming a necessity instead of a “nice to have” will really benefit.

Source: ASE Technology (ASX US) There are a number of companies in the CPO and SiPh supply chain that have yet to fully appreciate how big this market will be. They remain mispriced, especially when compared to their better performing peers. So, in order to understand who’s getting paid that diverges from the obvious winners, let’s go a bit deeper into how exactly this technology works… CPO functions like a relay race routing light through the chip. Data starts as an electrical signal, gets encoded onto a beam of light, travels through a tiny waveguide etched into the chip, gets redirected upward by a grating coupler (think periscope), and is finally collected by a Fiber Array Unit (FAU) that funnels it into fiber optic cables.

Both CPO and pluggables consume the same upstream materials and equipment. Most of these names we have already covered, but some are unique… How Co-Packaged Optics (CPO) Actually Works And, like literally everything in the AI space, TSMC is critical. TSMC’s Compact Universal Photonics Engine, or COUPE, is a high performance silicon photonics packaging technology designed to bear the enormous computing demands of artificial intelligence. The compact part of the technology refers to the 3D “stacking” of the chip design. As we covered in the Advanced Packaging section of 26 Trades for 2026, this is integral for power efficiency and capability.

TSMC COUPE Platform The universal component of the COUPE technology refers to TSMC’s intent to standardize 3D packaging. TSMC is pushing the rest of the industry to adopt 3D packaging, with the end result being a 10x increase in connection density that is capable of supporting the 1.6T bandwidth required from leading-edge architectures like NVIDIA’s Blackwell and Vera Rubin. Currently, the COUPE platform is in the production validation and initial sample delivery stage. Full scale production is likely at least a year away. However, the writing is on the wall. We’d prefer to try to get ahead of the actual production ramp and expect that investors will do the same.

We’ve explained why tailwinds and plain physics are pushing photonics. But why aren’t we just content to rest on our laurels having called the bottom in COHR and LITE? Why are we focusing now on the rest of the supply chain?. All in all, we believe we have crossed the tipping point where demand for photonic manufacturing outpaces the supply chain, at a time when bifurcation and copper cost headwinds are tightening. The optical transceiver market for AI clusters is on track to exceed $10 billion in 2026, roughly doubling from 2024. COHR, LITE, and AAOI are all expanding production as fast as they can build it.

Tower Semi (TSEM US), in a similar vein, is investing $650 million to triple SiPh shipments by mid-2026. This is already happening now. NVIDIA starts shipping their first commercial CPO deployments, Quantum-X Photonics InfiniBand, in the first half of 2026. The Spectrum-X Photonics Ethernet switches follow in the second half of 2026, targeting 102.4 Tb/s spine switching with up to 512 ports of 800G. Broadcom is taking a similar path with Tomahawk 6 Why Now? Every single one of those pluggable transceivers consumes silicon photonics PICs. The upstream materials and equipment are identical whether the end product is a pluggable or a CPO module.

If CPO arrives on NVIDIA’s published timeline, that’s incremental demand on top of the pluggable wave. If it’s delayed, the pluggable demand alone more than justifies the substrate and equipment buildout. If this sounds familiar; that’s because it is. Yes, “picks and shovels” is cliché. It has also made anyone investing in AI an absolute pile of money. The bear case on CPO has a spokesperson: Broadcom’s CEO Hock Tan. On his most recent earnings call, Tan said: “You don’t need to go run into some bright shiny objects called CPO, even as we are the lead in CPOs.

CPOs will come in its time, not this year, maybe not next year, but in its ” Broadcom Q1 2026 Earnings Call - March 2026 While Tan may be right on CPO timing, it doesn’t matter for our thesis. We’re not making a pure bet on CPO. Instead, we are betting on the supply chain and the explosive demand for both pluggable transceivers and CPO capabilities. Ironically, Hock Tan is applying this logic to his own company. While he tells investors to be patient on CPO, Broadcom has already shipped three generations of CPO technology to Meta.

The bear case and the bull case are the same statement read at different time horizons. Our opinions on AAOI’s share price notwithstanding, their earnings call provides solid confirmation that photonics materials and the equipment beneath the PIC are now the bottleneck for scaling. “Demand for our 800G and 1.6T transceivers is well above what we can supply. We expect to generate over $1 billion in revenue in 2026, but that level is constrained by production capacity, manpower, and the supply chain, not end- market demand. At the same time, there is a major laser shortage, with some suppliers quoting lead times of at least one

” Applied Optoelectronics Q4 2025 Earnings Call - February 2026 Lumentum expanded Electro-absorption Modulated Laser (EML) production 40% in 2025 and is looking to ramp production another 40% in 2026. Tower Semi, as highlighted earlier, is investing $650 million to triple its Silicon Photonics shipments by mid-2026. Multiply this across every transceiver maker and foundry and you have ten companies pulling on the same two or three substrate and equipment suppliers at once. Whether AAOI executes or Coherent and Broadcom take the share, the demand for MOCVD reactors, InP crystals, and SOI wafers is identical.

Simultaneously, in March, NVIDIA announced a $2B investment into Lumentum to advance US-based manufacturing and drive R&D collaboration in data center optics. On the same day, NVIDIA announced another $2B investment into Coherent with the same underlying purpose. In February, Marvell finalized their acquisition of Celestial AI, expanding Marvell’s optical connectivity capabilities. Finally, in September last year, Ciena acquired Nubis to help usher Ciena into Scale Out and Scale Up computing architectures. The flurry of dealmaking, in our view, implies that the demand pull for PIC is ongoing, and is likely to accelerate. As a result, the entire supply chain is racing to secure photonic capability – even electronic IC companies.

This is reflected in Semtech’s (SMTC US) recent acquisition of HieFo, in which Hong Hou, the CEO of SemTech, called out customer demand for CPO. “The evolution to 1.6T and 3.2T data center architectures represents a fundamental shift in optical interconnect complexity. By combining HieFo’s proven InP technology, including lasers and gain chips, with Semtech’s industry- leading TIAs and laser drivers, we can offer customers a comprehensive solution for next-generation optical platforms, including co-packaged optics or near-packaged optics, strengthening Semtech’s position as a leader in high- bandwidth, low-power and low-latency networking ” Meanwhile, electronic interconnects are getting more expensive, not less.

The inputs of electronic interconnects (specifically, the cost of copper) have accelerated the industry’s push into photonic optics. Finally, there’s the geopolitical angle (there’s one of those in everything nowadays, it seems). A salient theme within the semiconductor “super cycle” has been the push to build parallel Western supply chains. Every reshored supply line needs its own stuff…this bifurcation doubles the equipment orders from upstream suppliers because you’re building two of everything. But we are not a semiconductor science education firm. We’re here to provide trades. We’re starting with our highest conviction, most asymmetric play we see in the space that we worked with our friends at Hunterbrook on uncovering.

Embedded within the COUPE roadmap and the massive amount of capital thrown towards the photonics buildout lives Himax (HIMX US), a Taiwanese-based company focused on display driver integrated circuits (DDICs). These chips are ubiquitous in the consumer electronics space, think televisions, PC monitors, mobile devices, and even smart glasses. A sleepy company dealing with LCD screens… sounds like fertile hunting ground, right? Not unlike other companies that capture our interest, Himax’s business is at the mercy of a rough cycle in both consumer electronics and automotive demand. The automotive segment specifically accounts for approximately 50% of Himax’s sales, with management providing muted guidance given a lapse in Chinese EV subsidies.

Consumer electronics, in a similar vein, remains uncertain – with customers running lean inventories amidst macroeconomic volatility. This much is clear by looking at how the shares have performed over the past five years. 2) Single Name Highlight: Himax (HIMX US) So if Himax is selling into these end markets, why are we interested in the stock? We’ve seen this play out countless times, with optics and the telecom cycle, memory testing and the automotive overhang, even early Nvidia and the crypto-miner GPU glut. Himax is priced as a cyclical loser due to its core business while playing an integral role in the proliferation of co-packaged optics, specifically via micro lens arrays (MLAs) and prisms.

Given the inherent cyclicality in their legacy business, the company is pivoting its focus towards higher-margin, “non-driver” products in co-packaged and wafer-level optics. In June 2024, Himax, in partnership with FOCI Fiber Optics (3363 TT), revealed their newest leading edge silicon photonics packaging technology. As part of this partnership, Himax took a 5% equity interest in the company via a $16 million cash infusion. Himax’s wafer-level optics (WLO) technology will bond with FOCI’s Reflowable Lensed Fiber Array (ReLFACon) to create a physical connection between fiber optics and multi-chip modules. We think that Himax’s micro lens array can become a key supplier for Nvidia’s Rubin, Feynman, and future generation platforms.

Management stated that 2025 served as a year of validation and trial production. 2026 is expected to be the first year of mass production, potentially contributing over $100 million to the top line. Management has pointed to accelerating sales growth in 2027-28 as they establish a foothold in CPO. “I’ve actually commented on the revenue potential in my last 2 earnings calls. So I think I will just quickly go through that again. So again, next year, 2026 is likely to be the first year of mass production. But it’s still too early to give an early indication for the year because, as I just said, exactly when MP will commence is yet to be determined by our customers.

I said in the Q&A of last earnings call that our annualized CPO revenue could reach over $100 million in the so- called early stage of mass ” Himax Q2 2025 Earnings Call As of this writing, HIMX trades at a mid-teens multiple of consensus 2026 EV/EBITDA, but the stock is lightly followed, with only five analysts providing sell side coverage. Further, the company sold off heavily during the end of 2025 on unfounded rumors pertaining to their partnership with FOCI. Management quelled those rumors earlier this year in a company press release. What is Himax supplying, and what does TSMC need?

TSMC needs a supplier that can make the ultra-precise optical parts that allow CPO to work in the real world. In this setup, the work is split between Himax and FOCI. Himax makes two of the most important hardware pieces inside the FAU. The first is the optics block which is the glass part that includes the microlenses and the 45- degree prism. The second is the V-groove baseplate, which holds the fibers in place. Himax makes both parts using its Nano-Imprint Lithography process. As both pieces are made in a tightly controlled way, the spacing of the optical channels can be matched very accurately from the start.

In other words, much of the alignment is built into the parts themselves rather than being left to difficult manual assembly later. FOCI then takes those parts and turns them into a finished FAU. It places the fibers into the V-groove plate, attaches the optics block, aligns the full structure, and delivers the completed unit back to TSMC for their optical packaging flow. The relationship is closer than a normal supplier partnership. Himax’s 5% stake in FOCI helps align their interests and supports closer cooperation. What makes this especially important is that Himax has been doing this kind of manufacturing for years.

The company has real production experience with Nano- Imprint Lithography going back to the iPhone X era, when it was making tiny dot- projector optics for Face ID. That is why moving from Face ID optics to CPO optics is not really a jump to a completely new manufacturing platform. The main change is the shape of the pattern being stamped, not the underlying production method. The tools, materials, and quality-control systems are largely the same. That is also why Himax’s management is confident that their existing capacity can support CPO demand. A new competitor would not just need similar equipment, they would also need years of process learning to reach the same level of yield, consistency, and production reliability.

Himax and FOCI are developing at least two FAU generations for different stages of the CPO roadmap. The current-generation product is the nearer-term design, with validation and mass-production-readiness work underway and only small-quantity shipments expected in 2026. Himax has indicated that early CPO adoption should begin primarily in AI switches, which suggests this first wave is likely switch-side (scale out) networking. The second-generation product is the bigger long-term opportunity: Himax says it is being finalized for production readiness, targets greater than 6.4T bandwidth, and is aimed at the GPU market (scale up), implying a more demanding, higher-channel- density use case.

Again, management has indicated that Gen 2 has the potential to become a nine-figure annual revenue opportunity even in the early stages of mass production. The exact mass production timing remains customer-dependent and is currently viewed as more likely in late 2027 or early 2028. The physical evidence goes beyond analyst conjecture. Three patents — two filed by Himax and one by TSMC — describe what appears to be the same FAU design and The Scale Patent Trail manufacturing process for CPO application. The documents were first unearthed by Latent Value. A Himax patent published in January 2026 details the nanoimprint lithography process for manufacturing the fiber array.

An optics block with lenses stamped on one face of a glass wafer and 45-degree prisms on the other, as well as a V-groove plate for holding the fibers in place. A second Himax patent, filed in Taiwan in November 2024, describes an automated wafer-level testing system that validates optics blocks — presumably before the optics blocks are diced and shipped to FOCI for assembly. Then there is TSMC’s January 2025 patent for a FAU inspection tool built specifically for CPO. The patent describes the schematic of the part being tested — without naming any supplier.

It’s the exact geometric structure that Himax describes: a prism/mirror on one face, microlenses on the other, sitting above a baseplate with fibers. More speculatively, it describes a step to inspect the “defect between mirror & lens” — language that implies TSMC is looking through a solid medium between two surfaces inside the same component. If the mirror and lens were separate pieces bonded together, you’d be inspecting an adhesive joint — and you’d call it just that. Inspecting for defects “between” two optical surfaces suggests a continuous material connecting them, which is what a monolithic glass block stamped by nanoimprint lithography would be.

A figure in TSMC's FAU inspection tool for CPOs that shows the schematic for quality checks on the emergent light path. By inspecting the mirror surface, the micro-lens surface, and "defect between mirror & ” The bottom section validates all 22 channels (ch1–ch22) for beam core pitch and quality. The 22-channel count matches the Himax manufacturing patent exactly. The former Lumentum engineer interviewed by Hunterbrook reviewed the patents and came away with the conclusion that “there is a high probability that Himax and FOCI are part of the FAU supply ” He cited three main benefits of this technology.

First: “It is a batch process. There are multiple units on a glass wafer that can be processed ” Second, this process has a lower cost than traditional lithography. Third, he added, the optical block using prisms also has higher precision than a mechanical process. He suspects Himax “is also making some of those micro lenses being used in the FAU. That’s why they wrote the wafer level testing ” The patents also reveal intriguing — if speculative — clues to the Himax/FOCI/TSMC supply chain centering on a custom FAU with an unusual specification: a 22-lens array.

In Himax’s January manufacturing patent, a figure (labeled as “Fig. 2C”) illustrates the setup showing the optics block with exactly 22 lenses. You can count the circles representing the lenses. A diagram in the January 2026 patent filed by Himax shows the complete fiber array assembly: the monolithic optics block (element 22) sits atop the V-groove baseplate (element 21). The illustration shows exactly 22 circles representing the micro-lenses stamped on the front face of the optics block (221) and a contiguous 45-degree prism (222) on the back. Light from the chip enters vertically, hits the prism, bends 90 degrees, and exits through the micro-lenses into fibers seated in the V-grooves (211) below.

Both pieces are manufactured by Himax using nanoimprint lithography. (Red lens count, yellow light path and fiber annotations added for clarity.) Annotations by Latent Value. Himax’s Taiwan testing patent says the prism under test “may also be equipped with multiple lenses, such as 16, 22, or 50 lenses, but the number is not limited ” Twenty-two is an unusual number in fiber optics, suggesting it might be custom design. Fiber counts are standardized around specific multiples like 12 or 16; both are numbers you’d see in product specs. Higher-density arrays generally scale in increments of those base units — 24, 32, 48, 72.

Then there’s TSMC’s patent — the closest thing to a smoking gun, though speculative. TSMC’s inspection schematic explicitly depicts 22 channels, labeled “ch1” through “ch22” (see above). Twenty-two channels appearing in both patents may be a coincidence. However, it’s interesting that they both describe the same part being inspected at two points in the supply chain: One by Himax before shipping, and another by TSMC after assembly, before it reaches a GPU. So how big is the opportunity? Morgan Stanley projects 5,000 next-generation NVIDIA Rubin Ultra racks shipping in 2027 and 28,000 in 2028. Even using conservative pricing assumptions, the FAU market scales from hundreds of millions in 2027 to billions of dollars in 2028, according to our analysis.

What’s more, NVIDIA’s Feynman architecture — the generation planned for after Rubin — is expected to multiply the unit count further. Image: Himax Patent TW I896409, FIG. 5. Himax's wafer-level testing system inspects dozens of prism blocks on a single glass wafer before they are diced and shipped to FOCI for assembly. Each block (element 22) will become the optics core of one FAU. The matching channel counts also suggest a coordinated design process rather than coincidence. The most likely interpretation is that TSMC defined the optical architecture and quality requirements, while Himax built its manufacturing and wafer-inspection flow around those specifications.

That would fit well with management’s description of the program as a joint development effort involving Himax, FOCI, and the end customer. There is also an important roadmap implication here. Himax’s wafer-level test patent is designed not only for 22-prism configurations, but also for 50-prism ones. That suggests the company was already preparing tooling for a higher-channel-count generation beyond the current design, potentially consistent with a later 12.8T CPO. Overall, the patents make the supply chain story much more concrete. Instead of relying only on commentary, this IP shows a plausible technical handoff: Himax fabricates and screens the monolithic optics blocks, FOCI assembles the finished FAU, and TSMC performs the final co-packaged-optics inspection before package

integration. That sequence lines up closely with how Jordan Wu (CEO of Himax) described the relationship on the earnings call, when he said the project had been developed jointly with Himax’s direct customer, its direct partner FOCI, and a shared anchor customer/partner. In summary, Himax stamps the optics and ships to FOCI (3363 TT), FOCI assembles the finished fiber array unit and engineers its connection to TSMC’s COUPE, TSMC bonds it onto the chip and it powers NVDA’s CPO platform. Beyond NVDA, there’s another underappreciated angle here that has flown under the radar. On the last earnings call, Wu said that “a leading brand’s smart glasses are poised to enter mass production later this

” He did not name the brand, but the implication was hard to miss. That comment fits with a DigiTimes report from early February 2026, which said Apple’s expected late-2026 smart-glasses launch is already starting to reshape Himaxʼs Apple Opportunity: Smart Glasses and Beyond Taiwan’s AR optics supply chain, with several suppliers increasing capex in preparation. There are also signs that Himax could be one of those suppliers. The company is one of the few players that appears to have all three of the key building blocks needed for smart glasses: WiseEye, its ultra-low-power AI sensing processor that runs on only single-digit milliwatts; a front-lit LCoS microdisplay that weighs less than a gram; and WLO nanoimprint waveguides, which use the same core nanoimprint manufacturing approach Himax applies in CPO optics, but adapted here to create the diffractive structures that steer images into the user’s eye.

On LinkedIn, one current Himax employee is listed as a senior product manager for “Advanced Optics & NPI Integration | Apple Project ” That same person previously worked on the Meta Quest 3 lens program, where they helped scale nanoimprint production from prototype stage to more than 7 million units. Taken together, that does not confirm Himax’s role, but it does suggest the company may be involved in a major smart-glasses program and has relevant experience in taking nanoimprint optics into high-volume production. Her recent posts talk about shipping against tight customer deadlines and fixing late- stage manufacturing issues, which is the kind of language you would expect from a program moving close to production ramp.

Industry sources estimate that Apple’s first smart glasses could ship around 3 to 5 million units in the launch year. In the near term, the main revenue opportunity for Himax would likely be WiseEye sensors. That could still be meaningful, but it would probably be modest, with an ASP in roughly the $10 to $15 per unit range. The much bigger opportunity would come later. Apple is widely expected to move toward display-equipped AR glasses around 2028, which would require technologies like LCoS microdisplays and nanoimprint waveguides — two areas where Himax is already active and has been showcasing capabilities with partners such as Vuzix.

If those products reach volume production, and if combined optics ASPs land in the $70 to $90 per unit range, the AR optics revenue opportunity could become large enough to rival CPO over time. Between Morgan Stanley’s Rubin Ultra forecasts and the highly-anticipated release of Apple’s smart glasses, we see the non-driver segment of Himax, which represented just under 20% of sales in 2025, emerging as the key growth vector for the company. We suspect this segment – which management has characterized as higher margin and less cyclical compared to its other operating segments – will ultimately comprise roughly 50% of the company’s sales by 2028.

HIMX consensus estimates point to just over $1 billion in sales for 2027, yielding ~$90 million in operating income. We think this estimate fails to incorporate the multi-hundred million dollar market opportunity in supporting the Vera Rubin ramp, in addition to the opportunity in switch sales and WiseEye adoption from Apple Smart Glasses. Assuming management executes properly on this opportunity, HIMX could end up trading at roughly 3-4x EV/EBITDA on 2028 numbers. What does this mean financially? As is with most trades, we want to ask: what are the picks and shovels? Again, we don’t intend to take a directional bet on any particular CPO architecture or end- market.

Rather, we’ve found success in identifying the “PICs and shovels” (forgive us) that equip and support the buildout of these technologies. The first part of the trade, getting long the “in-your-face” winners like Lumentum and Coherent, has paid handsomely. While these companies should see their shares continue to perform well, they’re also approaching a level of “priced in” that isn’t readily apparent in the rest of the supply chain. In this instance, PIC substrates (SOI, InP crystals, III-V epiwafers) all require specialized equipment to produce. At the equipment layer, there is even more market concentration relative to the substrates themselves.

The market is dominated by a handful of German companies, almost entirely invisible to S. generalists. Similar to other specialized sections of the semiconductor supply chain, these markets are characterized by monopolistic or duopolistic dynamics – each of which are structurally tied to the photonics demand wave. 3) PICs and Shovels Meanwhile, the market today is wasting time and energy on the wrong debate – “Which architecture ” At the equipment layer, the answer doesn’t matter. We are agnostic to who “wins” the transceiver or architecture race because both paths converge on the same suppliers.

Given this backdrop, we’re focused on owning the materials and equipment that both architectures depend upon. AIXTRON (AIXA GR) is a German company and the dominant supplier of MOCVD equipment, holding 70-90% market share. MOCVD, or metal-organic chemical vapor deposition, is the process of growing III-V layers onto substrates. This technology is particularly salient as Lumentum’s expansion, Coherent’s buildout, and Nokia’s InP capacity ramp all require new MOCVD tools supplied by AIXTRON. AIXTRON (AIXA GR): The MOCVD Monopoly The reason we’re interested in the company is because it’s in the middle of a transformation which we believe the market hasn’t fully caught on to yet.

AIXTRON’s revenue has historically been tied to the EV market, but optoelectronics, where photonics for AI data centers live, is rapidly becoming the lead growth driver. In their Q4 earnings release, management guided to an +100% increase in their photonics business: “And so we see that the cycle really has kicked off towards the end of ‘25. So you have seen that in the fourth Q4 ‘25, our photonics orders have significantly increased. Q1 to Q3, they were still on a relatively low level. In Q4 ‘25, our photonics orders have increased. We still, already now in Q1, we see continued order momentum from our customers… This is baked in our guidance.

So our guidance reflects that already. And we expect that the orders are coming in essentially throughout Q1 and continue to come in Q2 and covering then the revenues that we have forecasted for the year. And as you see, it’s a quite significant increase. It’s more than double year-over- year for the photonics ” AIXTRON SE Q4 2025 Earnings Call - February 2026 Photonics traction marks a clear inflection point for the company. If optoelectronics was roughly 23% of equipment revenue in 2025 (around €120 million), a double in that segment tracks to +€250 million, potentially making it the single largest segment by 2026, overtaking SiC and GaN combined.

Beyond photonics, we see a second catalyst the market is underweighting. NVIDIA’s proposed 800V HVDC data center architecture for Rubin Ultra requires GaN-based power semiconductors. Only GaN has the switching frequency and power density to make 800V power delivery work. If we take NVIDIA’s own roadmap at face value, Rubin Ultra shipping in 2027 with 800V power delivery, then GaN tool orders must begin in H2 2026, which is exactly what AIXTRON management indicated on their Q4 call: “AIXTRON maintains a clear market leadership position with more than 85% market share across GaN device classes, and we remain deeply engaged with customers expanding their GaN road map into the coming years.

Importantly, GaN is emerging as a central technology for AI-driven power architectures, particularly as hyperscale data centers plan the transition to high- efficiency 800-volt platforms. We expect additional volume from GaN from AI applications at some time in the 2027 and ‘28 time ” AIXTRON SE Q4 2025 Earnings Call - February 2026 AIXTRON is sitting at the convergence of photonics now and GaN power within six months. We think the consensus revenue trajectory from here is too low, particularly as the market remains uneasy about the demand from automotive and consumer end markets. The question boils down to whether we believe management’s own projections for optoelectronics and whether 2026 is the inflection year before 800V GaN ramps.

We think the activity across the supply chain corroborates their guidance. AAOI’s $209 million in capex for 2025 came in well above the $120-150 million anticipated by the street. Similarly, Coherent has tripled its InP capacity, with Lumentum, Nokia, and SMART Photonics all expanding InP production lines. Every one of those lines requires AIXTRON’s tools. We think the photonics equipment cycle is being pulled forward faster than the market expects. The transceiver industry is simultaneously scaling for 1.6T and beginning development on 3.2T. At AIXTRON’s current valuation, you’re paying roughly fair value for a photonics equipment monopoly and getting the GaN power ramp for free.

The stock has started to re-rate, but we think its dominant position makes it a must-own in the picks and shovels layer of the photonics supply chain. Veeco (VECO US) is the only credible MOCVD competitor to AIXTRON, and the two collectively control this market. We believe that Veeco also stands to benefit from rising photonics demand, albeit with a weaker setup. The company is more diversified across laser annealing, ion beam, wet processing, and lithography, and Veeco (VECO US) they’ve struggled to grow their semis revenue. They’re also mid-merger with the EV- exposed Axcelis Technologies (ACLS US), which creates somewhat of a distraction but also adds potential synergies.

Both companies are at cyclical troughs in their respective end markets, but post-merger you’re looking at one of the largest wafer fab equipment companies in the US. Together, they’d be sitting at a billion dollars or so in cash and in a very strong position to ride the GaN/800V power ramp. Overall though, we think Veeco looks attractive here, but Aixtron is still the purer expression of this theme. Given the pending merger, we will exclude from our basket at this time. SUSS MicroTec (SMHN GR) is the same setup we’ve seen across the photonics supply chain: a German equipment monopolist sitting at a cyclical trough, with a structural growth driver the market hasn’t priced because it’s sitting inside an HBM equipment company going through a down cycle.

SUSS holds a monopoly or near-monopoly position in equipment used within advanced packaging – specifically, wafer bonding and high end photomask cleaning. We won’t go too deep into the technicals here, but wafer bonding is what it sounds like: merging wafers together either horizontally or vertically. Photomasks are SUSS MicroTec (SMHN GR): Light Meets Silicon the stencils used to print onto those wafers. These need to be cleaned on a particle level before every use. If any particles are caught on the mask, they become defects. The reason we’re interested in SUSS is the market position.

The company supplies roughly 50% of the temporary bonding equipment used by Samsung and Micron for HBM production, while also holding approximately 90% share in photomask cleaning at TSMC. Around 70% of total revenue ties to advanced packaging, making SUSS arguably the purest play on this theme globally. That alone makes it interesting. But here’s where things get more compelling: The coming CPO wave requires a process called heterogeneous integration, physically bonding lasers onto wafers to create photonic integrated circuits. That bonding step is performed on SUSS’s XBS300 platform. Every silicon photonics foundry line being built or expanded by GFS, Tower Semi, and UMC needs SUSS to do this.

Today, photonics related orders are still a small fraction of total revenue, and the market hasn’t caught on to the coming platform shift. However, management has already flagged this as a potentially massive growth vector: “The market we play in, even with 14% CAGR, and the heterogeneous integration market with 32%. So that’s the fastest growing segment. And that’s probably also the fastest growing market for us to move forward. Depending on how good of a job we do there, we will benefit greatly there with our ” SUSS MicroTec Capital Markets Day - November 2025 Another growth lever we think is underestimated is hybrid bonding.

Hybrid bonding, as discussed in Muscle Memory, enables direct connections between stacked wafers at a much finer pitch. This massively increases interconnect density and bandwidth while lowering power consumption. SUSS’s XBC300 Gen2 system is purpose built for this transition, and competition is thin. “One element for this is our IP for hybrid bonding, where especially our hybrid bonding PIN SUSS proprietary IP is helping us to achieve this excellent performances. And last but not least, also here, you see the platform approach where we are able to offer die-to-wafer integrated hybrid bonding, wafer-to-wafer bonding tool that combine die-to-wafer and wafer-to-wafer hybrid bonding

” SUSS MicroTec Capital Markets Day - November 2025 In other words, SUSS owns proprietary IP that gives them a performance edge on bonding accuracy, and they’re the only company offering a single platform that covers every hybrid bonding configuration, meaning customers can standardize on one vendor rather than buying separate tools. We believe SUSS stands to be one of the big winners as the industry moves from pluggable transceivers to CPO, and from conventional thermo-compression to hybrid bonding. Both trends run directly through their product line. Right now, the setup is compelling because SUSS trades at roughly 16-18x trough EBIT on what management has flagged as a “transition year” caused by a HBM ordering pause ahead of 2027.

For 2030, management targets €750-900 million in revenue with 20-22% EBIT margins. At the midpoint, that’s roughly €173 million in EBIT, meaning you’re paying about 6x forward on management’s numbers. For context, BE Semiconductor (BESI NA), the closest European advanced packaging equipment comp, trades ~40x forward EBIT with lower market share concentrations in its niches. If SUSS executes on the transition, which we think they are more than well positioned for, the stock should re-rate toward BESI-like multiples from a much lower starting point. If we move further upstream in the photonics supply chain, we find Soitec (SOI FP), a French business that commands a near monopoly on the silicon on insulator market, the substrate that underlies all PICs.

Soitec has gone through an absolute rinse because of the downturn in smartphones, EVs, and other European cyclicals. Consequently, the market is pricing Soitec as a Soitec (SOI FP): SOI Substrate mobile/RF-SOI company going through a down cycle. Our belief is that the company is set up to be one of the biggest public beneficiaries of the photonics cycle. Every foundry making PICs today runs on Silicon-on-Insulator (SOI) wafers. For example, NVIDIA’s CPO platform runs on TSMC’s COUPE process, which fabricates the PIC on an SOI node. Soitec has a proprietary technology they call Smart Cut, which is patent protected and took decades to develop.

See Soitec’s video for a better visualization. While photonics-SOI is a relatively small part of Soitec’s revenue today, management’s 2026 outlook tells you where things are headed: “Photonics is currently in an inventory depletion mode, but by the end of this calendar year we expect customer inventories to have returned to pre-COVID levels. From there, we believe the business will reach an inflection point and begin growing again. More broadly, photonics is in a very dynamic momentum ” Soitec Q3 FY2026 Earnings Call - February 2026 Soitec isn’t standing still on their product roadmap. They’re broadening into compound semiconductor substrates (SiC, GaN, InP-on-silicon) through their SmartSiC and POI (Piezoelectric-on-Insulator) platforms.

This positions Soitec to capture more of the photonics supply chain over time, not just the passive waveguide substrate but potentially the active III-V integration layer. “We have more and more customers asking for photonics, and in larger volumes, across 200 millimeters and 300 millimeters. We are enhancing the capabilities of our photonics portfolio through LNOI technology, extending bandwidth and increasing capacity to reduce latency. Today, there is no specific limitation to support a market that is growing by around 25% to 30% per ” Soitec Q3 FY2026 Earnings Call - February 2026 The LNOI addition is particularly interesting because it extends Soitec’s content per PIC beyond the passive waveguide into the modulator layer.

This effectively doubles their addressable substrate opportunity, particularly as the industry pushes toward 200G and beyond. At current prices, you’re paying roughly 10x EV/EBITDA on trough earnings. While the stock has doubled from the beginning of this year, we feel that the current valuation is still too cheap for an 80% market share substrate monopoly that every silicon photonics fab is leaning on. Shin-Etsu Chemical (4063 JP) To understand how durable Soitec’s position really is, consider who the competition would need to be. Shin-Etsu Chemical (4063 JP) is the world’s largest silicon wafer manufacturer, with over three decades of SOI production experience and the largest thick bonded SOI capacity worldwide.

If anyone had the resources and expertise to replicate Smart Cut and challenge Soitec on advanced thin SOI, it would be Shin- Etsu. Instead, they license Smart Cut from Soitec. That should tell you everything about Soitec’s moat. For what it’s worth, we have been bullish on Shin-Etsu in its own right since last December. It sits at virtually every chokepoint in semi materials: the world’s largest silicon wafer producer, a leading supplier of EUV photoresists, dominant in photomask blanks, and a major provider of advanced packaging materials. The company generates enormous free cash flow, carries no net debt, and is executing a massive share buyback.

It’s not purely a photonics name, but it belongs in the broader “machines that build the machines” universe we’ve been building. We’ll likely have more to say on this name in the future. Nokia (NOK US): The InP Fab Hiding in Plain Sight Old and mostly forgotten, Nokia (NOK US) finds itself in a very interesting setup after years of underperformance. There are perhaps fewer companies positioned to reap the rewards from as many tailwinds as Nokia is. The market has pigeonholed Nokia as a 5G turnaround play or 6G beneficiary. Almost nobody, however, is looking at the company as a vertically integrated semiconductor company competing in AI data center optics.

Put simply, the Infinera acquisition fundamentally changed what the Nokia business is. Nokia now owns the only vertically integrated InP semiconductor fab in the West. The ex-Infinera Sunnyvale facility grows InP epitaxial layers, fabricates PIC, and packages them – all under one roof. This is the same class of compound semiconductor that NVIDIA is locking up capacity for at Lumentum and Coherent. We are now starting to see the rewards of the Infinera buyout come to fruition. What’s more exciting is Nokia’s effort to capture a share of the intra-data center optics market through the planned and ongoing construction of two new facilities in San Jose, California and Bethlehem, Pennsylvania.

This buildout is subsidized by the CHIPS and Science Act and lies downstream from Nokia’s $4 billion commitment to US manufacturing and R&D. The new 40,000 sq ft cleanroom and InP PIC fab in San Jose is set to complete in 2026/27. The product that will be scaled in San Jose is ICE-D, Nokia’s intra-data center optical connectivity platform. ICE-D competes directly against Lumentum and Coherent’s pluggable and CPO products, but takes a fundamentally different architectural approach: monolithic InP integration instead of silicon photonics with discrete lasers bonded on. Then there’s the NVIDIA angle: Nokia is the only company simultaneously receiving a $1 billion strategic investment from NVIDIA for 6G WHILE ALSO benefitting from NVIDIA’s $4 billion expenditure thrown across Lumentum and Coherent.

While the stock has started to move on recent press releases, we believe this old dog stands to perform well over the next 12-18 months as this transformation story continues to play out. Furukawa Electric (5801 JP) is no longer just a traditional fiber company, it is becoming a broader optics infrastructure supplier for the data center and AI buildout. Today, its portfolio includes optical cable, ultra-high-count rollable-ribbon cable, preconnectorized cable systems, MT ferrules, MPO/MMC-related connectivity, DFB laser chips, in addition to newer CPO building blocks such as ELS, TOSA, and compact CPO connectors. On the cable side, management is focused on rollable-ribbon cable, especially ultra- high-count cable for hyperscale data centers and dark-fiber providers.

Furukawa is also pushing preterminated and preconnectorized cable systems, where it combines cable with ferrules and connectors to offer a higher-value solution rather than just selling raw fiber. Management has said this part of the business should be a major earnings driver, with North American demand expected to see robust growth through FY2027 and beyond. Optical Cables & CPO Ecosystem Picks Furukawa Electric (5801 JP) Another key area is MT ferrules and connectivity. These are small, albeit essential parts of high-density optical links. After acquiring Hakusan in 2024, Furukawa said they became #2 globally in low-loss ferrules.

The company wants to use these ferrules not only as standalone components, but also inside bundled cable-plus- connector systems. More recently, Hakusan’s work on MMC/TMT very-small-form- factor connectivity shows that Furukawa is preparing for the next generation of denser optical interconnects, not just relying on today’s MPO/MT products. On the active-optics side, the most important product today is the DFB laser chip. Management has been very clear that demand is strong and that Furukawa already has orders from several transceiver makers. The company sees DFB chips as a major growth engine for 400G/800G and 800G/1.6T optics.

In late 2025, Furukawa announced a major expansion plan that will lift DFB-chip manufacturing capacity by more than 500% in 2028 versus 2025. This matters because the DFB chip is also the base for Furukawa’s future ELS and TOSA products. Nvidia has locked up long term contracts for DFB laser chips, which presents a huge opportunity for Furukawa. That is where the CPO angle comes in. Furukawa’s TOSA is the laser sub-assembly, while the ELS is the external laser module that sits away from the hot switch chip and sends light into the CPO system.

Furukawa’s strength is that it can go from DFB laser chip all the way to ELS module. The company has already announced a 16- channel blind-mate ELS for CPO, built from two 8-channel TOSAs. This is not yet the main profit driver, but it demonstrates Furukawa’s efforts in trying to move from being only a component supplier toward supplying higher-value optical engine building blocks for CPO. Furukawa is also developing a compact CPO connector. This connector is much smaller than MPO, can survive 260°C reflow, and is designed to attach densely to a CPO substrate or future glass waveguide substrate.

In simple terms, Furukawa is trying to capture not only the light source in CPO, but also the small optical port where fibers plug into future CPO hardware. We see a clear growth story at Furukawa in both the short and intermediate term. In the near term, the main earnings drivers are still cable, ferrules, and DFB chips. Management has spoken about these products most clearly and has called out expanding capacity. Looking out to 2027-28, Furukawa has three main growth vectors: bigger and denser cable systems, more connectivity content through ferrules and next-generation formats like MMC/TMT, and more active photonics content through DFB chips, ELS, TOSA, and eventually broader CPO-related products.

The bottom line being that Furukawa is turning into a picks-and-shovels supplier for AI optical infrastructure. It sells the cables, preconnected systems, ferrules, connectors, laser chips, and future light-source pieces needed for the next wave of optical scaling. Today’s profits are still driven by the more established products, but the upside in 2027-28 comes from Furukawa moving further into denser connectivity and CPO- adjacent photonics while still benefiting from the basic optical-cable buildout underneath it all. Sumitomo Electric, like Furukawa, is undergoing an evolution that extends their business beyond basic fiber-and-cable. It is increasingly becoming a multi-layer supplier to the AI data center optics stack, with exposure across optical cable, dense connectors, ferrules, chip-to-fiber coupling parts, optical devices, tunable light sources, and compound-semiconductor substrates.

The latest earnings already show this shift. In FY2025 3Q year-to-date, info- communications sales rose to ¥220.6 billion from ¥159.8 billion a year earlier (+38%), while operating profit increased to ¥46.1 billion from ¥10.4 billion (+360%). Management has attributed this growth to stronger sales of optical connectors and optical devices for data centers as generative AI buildouts expand. The company also raised its full-year FY2025 outlook for the segment to ¥310.0 billion of sales and ¥65.0 billion of operating profit. In other words, AI optics is already a major earnings driver even before CPO becomes meaningful.

Sumitomo Electric (5802 JP) Today, Sumitomo’s strongest position lies in the connectivity layer. The company sells MT/MPO/MMC ferrules, high-density multi-fiber connector systems, and related fiber-interconnect products such as 24-MPO, MMC, SN-MT, and AirMT. Management has been especially bullish on this segment. In its 2025 data center strategy deck, the company showed optical connector capacity rising 7x from 2023 to 2028. Planned capacity for 2026 was also raised by 5x the previously announced target, given stronger than expected demand. In the Q&A, management stated that they believe they have 30% share of the FY2025 MT ferrule market and that MMC already accounts for 70–80% of its own shipments.

This suggests Sumitomo is not just benefiting from market growth, but is also gaining share in the premium, high-density part of the connector market. The next layer is CPO and silicon-photonics attach. Sumitomo’s roadmap explicitly includes optical IC ferrules, optical IC connection components, connectors for external light sources (ELS), and low-loss, high-density, high-fiber-count connectors. Management says these products are being developed for next-generation GPU/server and rack optical interconnects, and the company has also been named one of NVIDIA’s silicon-photonics ecosystem partners. Sumitomo already sells chip-coupling products under its FlexBeamGuidE family, including fiber arrays for coupling photonic integrated circuits to optical fiber, couplers suited for co-packaged optics, OCS and fiber-array products aimed at 800G, 1.6T, coherent modules, and future CPO (3.2T+).

In CPO, Sumitomo looks less like a full optical-engine supplier and more like a company trying to own the fiber-to-chip interface, ELS connector layer, and dense attach hardware around the photonics engine. In optical devices, management has been clear that Sumitomo’s main intra-data center products are EML and CW-LD (Continuous Wave-Laser Diode). EML is still important today, and management noted that even at 1.6T, some transceiver vendors may still choose EML depending on design choices. But, the medium-term direction is shifting toward CW-LD feeding silicon photonics. In Sumitomo’s own estimates, the market mix moves from 76% EML / 24% CW-LD in 2024 to 31% EML / 69%

CW-LD by 2028, with CW-LD described as the mainstream in CPO. The company is targeting 200G/wavelength EML and 350mW+ class CW-LD, while expanding intra- DC optical device capacity by 12x from 2023 to 2028. That means management still sees a place for EML, but it clearly expects the bigger 2027–2028 growth wave to come from CW laser plus silicon-photonics architectures. One underappreciated piece of the story is InP substrates. Sumitomo mass-produces InP substrates in 2-, 3-, 4-, and 6-inch diameters, and management said capacity is being expanded for both internal use and external sales, with 6-inch InP already ready for mass production.

The company’s 2025 strategy deck shows InP substrate capacity up 2.4x by 2028. This positions Sumitomo one layer upstream from lasers and optical modules. On CW lasers, management said it already has a strong position in the CW laser market, and that CPO is a positive because CPO light sources need more than four times current output, which raises value-added. Sumitomo’s technical work also shows it has developed a 1.3 μm high-power laser for CPO using an integrated LD-DFB + SOA structure, reaching over 400 mW output. Sumitomo also has exposure to the DC interconnect and coherent side of the market.

Management says AI data centers are becoming more geographically distributed because of power and land constraints, which increases the need for wavelength- multiplexed coherent links between buildings and campuses. Sumitomo’s deck highlights its own tunable light source with about 40 nm tuning range and 19 dBm (90 mW) output for 800G. This means the company is not only exposed to short- reach optics inside the data center, but also to the longer-reach optical links that connect AI campuses (scale across) together. In summary, Sumitomo is moving up the optics value chain. Right now, the biggest drivers are still connectors, ferrules, cables, and optical devices.

But by 2027–2028, the more interesting growth opportunity should come from CW-LD for silicon photonics, optical IC ferrules and attach parts for CPO, ELS connectors, higher- power lasers, and InP substrates. Management also says CPO adoption should become more widespread around 2028–2029, and current connector expansion plans do not yet include CPO demand. This tells us that CPO is the upside rather than the base case. STMicroelectronics is a Swiss integrated device manufacturer (IDM) with a large foothold in automotive OEMs. Stop us if you’ve heard this before – we think the company is being unfairly punished for their exposure to the automobile downcycle, STMicroelectronics NV (STM US)

despite the company having an attractive opportunity in photonic fabrication. Earlier this month, the company announced that they’ve entered high-volume production of its industry-leading silicon photonics platform – the PIC100. Within this platform lives ST’s through-silicon via (TSV) product, a crucial component of the 3D packaging architecture. This technology is employed by leading hyperscaler businesses for optical interconnect and high-bandwidth computing demands. “The data center pluggable optics market continues to expand strongly, reaching $15.5 billion in 2025. We expect the market to grow at a compound annual growth rate (CAGR) of 17% from 2025 through 2030, surpassing $34 billion by the end of the forecast period.

In addition, co-packaged optics (CPO) will emerge as a rapidly growing segment, contributing more than $9 billion in revenue by 2030. Over the same period, the share of transceivers incorporating silicon photonics modulators is projected to increase from 43% in 2025 to 76% by 2030. ST’s leading silicon photonics platform coupled with its aggressive capacity expansion plan illustrates its capabilities to provide hyperscalers with secure, long-term supply, predictable quality, and manufacturing ” Dr. Vladimir Kozlov, CEO and Chief Analyst at LightCounting ST has the strategic backing of Amazon Web Services, which expanded their investment and partnership with ST last month.

This relationship is characterized by a multi-year, multi-billion dollar commercial contract focused on equipping AWS’s data centers with silicon photonics wafers and laser drivers. With this contractual anchor, STM is refining its PIC100 platform to help customers reduce total cost of ownership and ship products to market faster. During the Morgan Stanley technology conference, management highlighted the compounding effect of both the AWS contract and the red hot market for optical cable: “So compare what we said end of January, so that this year, we can deliver USD 500 million. So thanks to the first effect of the AWS contract, but amplified by the

acceleration of the demand on optical cable, this year, I can say we will be nicely above $1 billion -- $500 million nicely. And next year, well above $1 billion. So we have really boosting effect from the start of the optical cable and from the benefits of our overall AWS contract that I repeat is a multi[billion]-dollar contract for the next 5 years that will start this year and moving forward will ” STMicroelectronics Morgan Stanley Technology Conference - March 2026 The company expects to quadruple capacity by 2027, accelerating through 2028. Notably, STM’s PIC100 technology is expected to support several generations of pluggable, co-packaged, and near-packaged optics, providing ample runway and visibility into customer demand in years beyond 2026.

Poet Technologies (POET US) is the most speculative name in our list, but its potential still makes it a worthwhile inclusion. The bet is that they succeed in commercializing a solution to a critical problem in the photonics supply chain: scalable photonic packaging at wafer level. The problem today is that photonic components like lasers and modulators are essentially wired together one at a time. The process is slow, expensive, and fundamentally doesn’t scale. This is a major bottleneck and a key reason behind transceiver supply falling short of demand, even as upstream capacity expands.

POET’s answer is the Optical Interposer, a wafer-level platform that integrates electronic and photonic components onto a single chip. In practice, this means components snap into place on the interposer at the wafer level rather than being individually positioned. If it works at scale, it dramatically reduces assembly cost, shrinks the form factor, lowers power consumption, and enables a potential step change in production volumes. Poet Technologies (POET US) If wafer-level passive assembly becomes the standard, every transceiver maker in the supply chain would need access to an interposer platform, either by licensing POET’s technology, partnering with them, or developing their own.

The partnership list suggests the industry is taking the possibility seriously. Still, the execution risk is significant, and the big integrated CPO players have the resources to potentially solve the packaging problem internally rather than rely on a third-party. The window to prove itself is also narrowing. By 2027, the CPO supply chain will likely consolidate around a handful of proven suppliers, and either POET has a seat at that table or it doesn’t. The company has over $300M in cash and minimal debt, though that cash came from roughly $400M in equity raises in the past twelve months.

At the risk of further dilution, we still think this is a reasonable bet for investors with the risk tolerance for a binary outcome. To put it all on one page (or if you’ve simply scrolled to the bottom, which we understand), we have created the following basket for this photonic transformation. Again, we believe there is more upside here than in the more traditional optical names that have already priced much of this ramp. We are overweighting Himax given our relative conviction on AAPL/NVDA upside combined with its modest current valuation. We underweight POET given its more speculative positioning.

4) The Basket You can see the basket on the Citrindex portal here. Depending on the market environment, we would consider shorting AAOI against this basket or monitoring some of the names mentioned in our previous optics coverage as shorts, given significant copper exposure. We’re also less constructive on the winners thus far, which we view as being priced relatively for perfection in a tough macro environment. However, that doesn’t mean we’re bearish on the overall optics ramp. We just view the basket above as a much better (and more asymmetric) way to play it from here.

We’ll keep this updated as things progress. Additionally, here’s the compiled performance of the names mentioned in our optics coverage from Summer 2024 - this should give a better understanding of what has worked (and what hasn’t) in the space since we originally began covering jpeg) Interconnects & Connectivity Basket: Individual Name Performance Since August 2024 Source: Citrini Research, Bloomberg