The Optical Ledger: What the LYTE ETF Reveals About AI’s Real Bottleneck

Beneath the surface of the current AI-crypto convergence, a mirror maze has formed. We assume the next bull market will be written in smart contracts, zero-knowledge proofs, and tokenised compute markets. But the most consequential ledger in 2025 is not on-chain. It is written in photons. Roundhill’s Photonics and Optical ETF, LYTE, is the rare instrument that exposes this to plain sight, and the more I study it, the more I suspect it is not a passive product at all. It is a thesis with a ticker. We are hunting for truth in a mirror maze of hype. The first reflection is simple: every artificial-intelligence token, every decentralised GPU network, every crypto project that promises to finance physical AI infrastructure, eventually has to ride on a physical layer. GPUs do not talk to each other by magic. They communicate through lasers, modulators, DSPs, and fibre optic transceivers. LYTE has converted that plumbing into a portfolio. Its five heaviest holdings—Lumentum, Coherent, Zhongji Innolight, Eoptolink, and TFC Communication—together account for more than 67% of the fund. That is not diversification. That is conviction. Context: The ETF as a Narrative Artefact LYTE is built to capture the photonics supply chain behind large-scale AI data centres. The fund does not hold Nvidia. It holds the companies that send Nvidia’s GPUs signals through the other side of the rack. Lumentum and Coherent control the high-end laser and photonic chip layer, while Zhongji Innolight and Eoptolink dominate the optical module layer. TFC Communication provides the passive components—fibre arrays, isolators, connectors—that make a module manufacturable at scale. The historical frame matters. During the 2017 ICO mania, I spent forty hours a week separating real teams from whitepaper theatre. I learned that the most durable investments were not the loudest tokens, but the ones anchored to a physical problem that could not be solved by a website. The same logic applies to the AI narrative in the mid-2020s. AI is not an abstraction. It is a set of computing clusters, and every cluster is a customer for photonics. LYTE is effectively a bet that the AI cluster build-out will continue long enough for optical components to become the new oil. The AI optical interconnect market is forecast to grow 57% to roughly 26 billion dollars in 2025, according to the data underlying this analysis. That figure is the hook of the LYTE narrative. But a forecast is a promise written in Excel. The ledger remembers what the heart forgets, so let us slow down and examine the physical and geopolitical structure behind that number. The Blockchain Connection Why should a crypto analyst care? Because digital asset markets are increasingly financialised claims on physical compute. Tokenised AI protocols, decentralised physical infrastructure networks, and even Bitcoin mining all require data-centre density and high-speed networking. The same optical modules that carry AI cluster traffic carry validator communication, exchange order flow, and the backhaul of every oracle. If the optical supply chain becomes a bottleneck, the digital asset industry feels it through higher costs and slower infrastructure build-out. This is not a metaphor; it is a physical dependency. In 2025, I worked with asset managers in Malaysia on a narrative risk assessment framework that tried to quantify how social sentiment influences institutional adoption. The conclusion was uncomfortable: the most powerful narratives are never purely digital. They are anchored to hardware. The AI narrative is anchored to GPUs and optical transceivers. The crypto narrative is anchored to miners, stakers, and data centres. LYTE sits in the middle of that hardware anchor, and that is why it deserves the attention of anyone who trades tokens, not just technology equities. Core: Seven Lenses on the Optical Stack Lens 1: The Process Node Myth We assume that technology leadership always means smaller transistors. In photonics, that assumption is the mirror maze’s first trap. Lumentum and Coherent’s optical chips are built on compound semiconductor processes—InP and GaAs—that would look primitive in logic terms, often at 0.13 to 0.5 micron linewidths. The most advanced silicon photonic chips use mature CMOS nodes between 45 and 130 nanometres. The only place where leading-edge logic appears in the optical module is the digital signal processor, the DSP, which uses 7 and 5 nanometre-class TSMC processes. This means the moat in optical chips is not measured in nanometres. It is measured in epitaxial precision, laser design, and yield on exotic substrates. The relevant manufacturing step is MOCVD growth on InP wafers, not EUV lithography. The industry frontier is defined by Lumentum, Coherent, Broadcom, and Marvell. Chinese module leaders such as Zhongji Innolight and Eoptolink are already world-class at integration and packaging, but they still rely on imported high-end chips for a meaningful share of their 800G modules. The next roadmap is the 800G to 1.6T transition. 1.6T modules began small-batch shipments in 2025, with volume expected in 2026. The longer-term technology shift is co-packaged optics, CPO, where the optical engine is placed directly beside the switch ASIC. CPO is expected to penetrate gradually between 2026 and 2027. The inclusion of Chinese silicon photonics champions in LYTE is therefore a quiet referendum: the ETF is implicitly betting that silicon photonics displaces discrete EML-based modules as the mainstream solution. TFC’s fibre array work is the critical glue for that architecture. The technical gap is not uniform. At the module level, Chinese companies are at the frontier—they generate the world’s volume of 800G and are co-developing 1.6T. At the photonic chip level, Chinese suppliers lag roughly one or two generations. The 200G EML laser, for example, is still a bottleneck where leading US suppliers achieve production yields around 60% to 70%. Chinese laser makers are one to two years behind. By 2027 or 2028, Chinese domestic suppliers may reach self-sufficiency for 100G-class EMLs, but the 200G EML and the CPO ecosystem will remain difficult to conquer quickly. The deeper implication is that process node mythology can mislead portfolio construction. If an investor believes that advanced nodes are the only moat, the optical sector looks boring. If the investor understands that MOCVD recipes and laser facet engineering are just as hard, the sector looks essential. LYTE’s heaviest positions encode the second view. Lens 2: Yield and the Margin Ledger Yield is the invisible hand inside every optical product. A 200G EML laser with 65% yield is a different business from the same laser at 45% yield, even if the performance spec is identical. This is why Lumentum and Coherent can command gross margins of 45% to 60% on high-end chips, while Chinese module assemblers live with 30% to 35% margins. The value chain is not fair. It is a mirror that rewards whoever controls the piece that can fail. The module layer is a sandwich. Optical module makers sell to hyperscale cloud providers and Nvidia, whose top five customers often account for 60% to 70% of revenue. They buy DSPs from Broadcom and Marvell, two suppliers that effectively own the high-speed DSP market. They buy high-end EML and continuous-wave lasers from Lumentum and Coherent. They then add engineering, rapid iteration, and manufacturing throughput. In normal industry conditions, that position would be structurally weak. In the current shortage, it has been profitable enough, but the bargaining power remains medium-to-low. The supply chain table paints a clear picture. High-speed EML and CW lasers: China’s dependence is 60% to 70%, with substitutes still ramping. DSPs: dependence above 90%, with only early-stage alternatives. InP substrates: more than 80% of supply comes from Japanese suppliers. Passive components: low dependence, with Chinese manufacturers already self-sufficient. Test equipment: high dependence on US instruments. The conclusion is simple: the optical interconnect industry is nowhere near as exposed as leading-edge logic, but it has its own narrow choke points. The ledger remembers what the heart forgets, and the choke points, not the revenue headlines, will decide who keeps the margin. One of the least appreciated details is the profit pool distribution. Photonic chips account for roughly 30% to 40% of the value in an optical interconnect. Module packaging and integration account for 40% to 50%. Passive components contribute about 10% to 15%, and test equipment makes up 5% to 10%. This means Lumentum and Coherent are operating in the highest-margin slice, while Zhongji Innolight and Eoptolink are operating in the highest-volume slice. The ETF holds both sides, but the two sides have different risk profiles. A technology change that simplifies module assembly could shift value upward toward chips and passive components. A technology change that commoditises lasers could shift value downward toward integration. LYTE is exposed to both directions, and that is both its elegance and its fragility. Lens 3: Capacity, Capex, and China Speed One of the most misunderstood aspects of the photonics industry is the speed of module capacity expansion. Unlike a wafer fab, which can take years to bring online and requires 80% utilisation to break even, an optical module line can be installed and qualified in three to six months. Equipment lead times are six to twelve months for MOCVD reactors and coupling stations, not three years for EUV. Industry utilisation has been running at 85% to 95%, and 800G supply remains tight. The expansion story is therefore not a story of excess, at least not yet. The capital intensity is moderate. Module makers spend 10% to 15% of revenue on capex, while photonic chip makers spend 15% to 25%. Depreciation schedules run five to eight years. Manufacturing break-even comes at 55% to 60% utilisation, much lower than the 80% threshold of advanced logic. That means optical module expansion is genuinely less risky than building a leading-edge fab. It also means the market can add supply quickly when demand signals appear, and that is precisely the contradiction hidden inside the LYTE basket. The four largest expansion efforts matter. Zhongji Innolight is expanding in Suzhou and Tongling, adding millions of new units per year. Eoptolink is expanding in Chengdu. Coherent is expanding high-speed EML and DFB capacity in Texas. Lumentum is expanding high-end laser capacity across multiple countries. The common theme is that module capacity is expanding faster than upstream photonic chip capacity. For the next two years, the intermediate bottleneck is not the module assemblers; it is the laser chip supply and the DSP capacity at TSMC. LYTE’s overweight position in Lumentum and Coherent is effectively a capture of this bottleneck’s pricing power. China speed is a structural weapon. The ability to move an entire production line and achieve high yields within months is not something US or European competitors can easily replicate. But speed can also be a liability, because it invites geopolitical reaction. When a country can scale the critical components of the AI supply chain faster than anyone else, it becomes part of the security conversation. I have seen this pattern before: a technological advantage that begins as an economic asset and ends as a political target. The capacity story will peak at a predictable moment. 2026 is the earliest point where the 1.6T ramp collides with an expanded installed base. If the CSPs sustain their capital expenditure plans, the collision will be soft and profitable. If they pause or trim, the industry will carry excess module assembly capacity. The module layer is lighter than a wafer fab, so the damage would be smaller, but it would still punish the most leveraged balance sheets. Because LYTE concentrates 67% of its assets in five names, the punishment would be visible in the ETF price almost immediately. Lens 4: Demand Visibility and the 300 Billion Dollar Question Demand for optical modules is now dominated by AI data centres. Based on the revenue structures of LYTE’s holdings, data-centre and AI interconnects contribute about 55% to 65% of revenue. Telecom accounts for 15% to 20%, industrial lasers 10% to 15%, consumer electronics 5% to 10%. The direction is unmistakable. Older telecom-driven growth cycles had a compound annual growth rate of 8% to 12%. The AI-driven cycle is running at 25% to 35%. The 57% growth forecast for the AI optical module market is consistent with a structural inflection, not a cyclical bounce. The arithmetic behind the demand is stark. An Nvidia H100 GPU usually requires around eight optical modules to connect to its cluster. A GB200 NVL72 rack can require roughly double that per GPU, and a full rack of 72 GPUs needs hundreds of optical interconnects. The sell-side story is straightforward: as GPU shipments grow, optical module shipments grow with them. In addition, inference workloads will create secondary demand for lower-speed 400G modules, which means the demand curve has a long tail, not a cliff. Inventory cycles in the module industry are currently healthy. Channel inventory sits at roughly four to six weeks. There is no obvious accumulation of excess 800G stock as of mid-2025. But the term healthy is relative. The industry has a long history of telecom inventory corrections, and the transition from 400G to 800G to 1.6T creates moments where old products become trapped in the channel. The difference this time is the strength of the AI capex cycle. The 300 billion dollar question is not whether the technology works. It works. The question is whether the world’s largest cloud providers will execute their 2025 capital expenditure plans without interruption. Microsoft, Google, Meta, and Amazon are expected to spend more than 300 billion dollars on AI infrastructure in 2025, with optical interconnect representing roughly 5% to 8% of that total. This is the foundation of the 57% growth forecast. If those capital plans survive economic turbulence, then LYTE has a multi-year runway. If they do not, the five-name concentration will amplify the downside. The price curve adds another layer. Optical modules follow a familiar cost-down curve of 15% to 25% per year. An early 800G module can sell for 1,800 to 2,500 dollars, then fall to 800 to 1,200 dollars as volumes mature. The 1.6T generation is expected to start around 2,500 to 4,000 dollars. This is not a price war. It is an engineered cost decline that allows module makers to maintain margins if volumes rise and product mix shifts upward. My assessment: gross margins will remain stable, but any delay in product transitions will upset the delicate balance. Demand is also becoming less cyclical and more secular. The installed base of AI accelerators is doubling, and the ratio of optical modules per GPU is rising. In earlier telecom cycles, demand was driven by carriers replacing equipment every few years. In this AI cycle, demand is driven by the build-out of massive, non-redundant computing clusters that need to move terabytes of data between every node. The replacement cycle is shorter, the urgency is higher, and the tolerance for downtime is lower. That is why the industry can sustain a higher baseline growth rate than its own history. Lens 5: Geopolitics as the Invisible Co-Pilot The optical supply chain is deeply entangled with the US-China technology conflict, yet it remains in a partial exemption zone. Lumentum and Coherent are American companies, and the Chinese module makers—Zhongji Innolight, Eoptolink, and TFC—were not on the US BIS Entity List as of August 2025. There are no systematic US export controls on optical transceivers. But the entanglement is real. The first point of exposure is the DSP. High-end 7 and 5 nanometre DSPs are controlled indirectly by US export policy because they rely on TSMC advanced logic. If the US tightens advanced-process access for Chinese chip designers, the Chinese optical module industry faces an upstream constraint, even if the module itself is not listed. The second point of exposure is the InP substrate, dominated by Japanese suppliers like Sumitomo and JX Metals. If Japan ever aligns with a stricter US containment policy, the Chinese high-speed optical chip roadmap is delayed. The third point is Chinese countermeasures. China controls export of gallium and germanium, key inputs for InP and GaAs photonic devices. Any tightening would raise costs for Lumentum and Coherent, creating a pressure valve that neither side can fully control. This is why Zhongji Innolight and Eoptolink have spent considerable effort building Thai manufacturing bases. Publicly, the rationale is customer demand and labour cost. Strategically, it is a pre-hedge against a scenario where the US restricts products with Chinese supply chains. The Thai plants turn a Chinese company into a multi-country supplier. Yet the classification risk remains. If a US regulator decides that effective control still lies with the Chinese parent, Thailand may not protect the listing from restrictions. LYTE itself is a geopolitical hedge. With Lumentum and Coherent contributing roughly 30% or more of the fund, and Chinese A-share names around 36.7%, the ETF offers double exposure to both sides of the trade. In the baseline status-quo scenario, the fund benefits from continued global interconnections. In a full decoupling scenario, the fund suffers because its Chinese holdings lose access to US customers while its US holdings face higher costs from gallium and germanium restrictions. In a partial-restriction scenario, Chinese module makers accelerate domestic substitution, but short-term costs rise. The probability-weighted outcome is not as safe as the name Photonics and Optical suggests. The cost of full optical decoupling would be severe. Global optical interconnect costs would rise 30% to 50%, and delivery cycles would stretch by six to twelve months. AI deployment timelines would slip, not stop. The mirror maze multiplies, but the physical dependency remains. That is why the optimal strategy for both Washington and Beijing is to leave optical modules out of the most restrictive export-control categories, even while fighting over every other layer of the semiconductor stack. Optical modules are too important to the AI ambitions of both powers to become a battlefield, but too strategically sensitive to remain permanently outside the game. Lens 6: The Competitive Mirror Maze The competitive structure of the optical world can be expressed in a single sentence: Chinese companies dominate the module, American companies dominate the chip, and the future depends on whether the chip or the module becomes the main point of value capture. In data-centre optical modules, Zhongji Innolight is the world leader with roughly 25% share, followed by Coherent around 15%. In the overall module market, Zhongji has about 18%, with Coherent around 12%. In high-speed photonic chips, Coherent leads with around 35% share, and Lumentum is second with about 25%. In passive components, TFC Communication holds 10% to 15%. Eoptolink is a top-five telecom module supplier. The market is a duopoly in disguise: one side for chips, one side for assembly. Research expenditure reveals the deeper asymmetry. Lumentum spends about 18% of revenue on R&D. Coherent spends 12% to 15%. Zhongji Innolight spends 5% to 7%. Eoptolink spends 4% to 6%. TFC spends 6% to 8%. In absolute terms, the US companies invest hundreds of millions more every year, focused on materials science, epitaxy, and laser architecture. The Chinese companies invest in application engineering, manufacturing automation, and rapid customer response. This is not a simple race. It is a division of labour, and the social contract of that division will eventually be tested. The corporate strategy divergence is visible in the numbers. Lumentum and Coherent are essentially IDMs, designing and manufacturing their own photonic chips, then selling them both as components and as integrated modules. Zhongji Innolight and Eoptolink are integration specialists, buying the best available chips from multiple sources and combining them into high-performance modules. The IDM model captures more margin per unit but carries more capital expenditure and technology risk. The integrator model captures less margin per unit but moves faster and can switch between competing suppliers. For a narrative hunter, the interesting question is which model becomes more valuable during a transition from 800G to 1.6T to CPO. The transition to 1.6T and CPO will redraw the map. In the CPO world, the optical engine moves closer to the switching ASIC, requiring advanced 2.5D or 3D packaging and close co-design with Broadcom and Nvidia. That is an ecosystem where the value pool shifts toward chip integration and advanced packaging. Chinese module makers can build CPO demos, but they face an ecosystem gap. The same gap exists for the DSP. By excluding Marvell and Acacia, LYTE has chosen a pure-photonics definition, but if the market shifts to co-packaged optical SoCs, the ETF may miss the most important value increment. This is the hidden risk inside the product design, and it deserves more attention. The competitive dashboard must also include the possibility of new entrants. Broadcom and Marvell are increasingly visible in the optical chip space. Nvidia has its own optics roadmap. The CSPs are exploring internal designs and direct supplier relationships. If the largest cloud buyers decide to vertically integrate optical module procurement or co-design, they could compress the margins of independent module makers. The LYTE portfolio would then face a structural compression that no amount of volume growth could offset. Lens 7: The Hidden Ledgers of ETF Construction An ETF is not a neutral container. Every index is a set of assumptions dressed as rules, and LYTE is no exception. The weighting methodology has already made several hidden choices that the average investor may not see. The first hidden ledger is the silicon photonics bet. The inclusion of Zhongji Innolight and Eoptolink, both of whom have pushed silicon photonics into mainstream 800G production, suggests that the fund believes silicon photonics will gradually replace traditional indium phosphide EML solutions in the majority of AI interconnect applications. TFC’s fibre array components are precisely the kind of passive interconnect that silicon photonic modules need. The fund is not merely betting on optical speed; it is betting on a particular technology path. The second hidden ledger is the definition of the investment universe. Acacia, which operates under Cisco, is not included. Marvell, with its combination of DSPs and silicon photonics, is not included. This reflects a definition of pure optical devices rather than optoelectronic SoCs. The exclusion reduces overlap with diversified semiconductor giants, but it also means LYTE will not fully capture the value of DSP-enabled intelligence inside the optical module. If the next generation of optical modules becomes increasingly digital, the ETF will be underweight the most important silicon component. The third hidden ledger is geopolitical hedging. The fund has more than 30% allocated to US optical chip leaders and around 36.7% allocated to Chinese A-share modules. From a neutral market-cap perspective, Chinese module makers already control a larger share of global optical module revenue than US suppliers. The fact that the fund gives almost equal weight to both sides suggests an intentional hedge, whether or not the index methodology admits it. This is a hedge that pays in a connected world, but it fails simultaneously in a decoupled world. The fourth hidden ledger is the assumption that AI capital expenditure is a durable force rather than a speculative cycle. Every demand forecast in the ETF prospectus is a derivative of CSP capex. If AI investment slows, the fund has no defensive sleeve. The five largest holdings all depend on the same final customer. The ETF may look diversified across eight different equity names, but it is actually one large leveraged trade on hyperscaler spending. Contrarian: The Blind Spot Is Not Technology Every article about optical infrastructure eventually celebrates the technology. The contrarian question is not whether photonics works; it is whether LYTE is a reflection of the technology or a reflection of the narrative. The answer is uncomfortably balanced. The ETF is a concentrated bet that AI capital expenditure will not stop, that silicon photonics will win, and that the US-China supply chain will remain sufficiently connected to allow American chips and Chinese modules to ship without interruption. All three assumptions are reasonable. None is guaranteed. The most dangerous blind spot is the demand narrative itself. The 57% growth number, the 300 billion dollar capex figure, the 1.6T roadmap, the CPO timeline—all of these are extrapolations from a period of extraordinary monetary and speculative intensity. The story of AI infrastructure has become indistinguishable from the story of AI hype. In the current cycle, any negative surprise in corporate earnings or any political disruption can reverse the flow of capital faster than the physical supply chain can adjust. The same psychology I observed in the 2017 ICO market and the 2021 NFT summer is present here, rewired for a new sector. The names are optical modules instead of tokens, but the emotional architecture is the same. Another blind spot is the entry point. The five largest LYTE holdings are already successful companies. The market has already rewarded them. The ETF offers exposure to the optical bottleneck, but at a point where the bottleneck narrative is widely known. The ledger remembers what the heart forgets. In 2022, after Terra and FTX, the market remembered that trust is an asset; in 2025, the market seems to have forgotten that physical capacity can be overbuilt, that customer concentration is a risk, and that geopolitical truces are temporary. The technology itself hints at a future disruption. CPO is not a small change in the packaging process; it is a relocation of value. When the optical engine moves onto the switch package, the traditional optical module assembly layer loses relevance. The module maker becomes a component supplier rather than a system integrator. The switch ASIC maker, Broadcom or Nvidia, becomes the centre of gravity. If CPO penetration accelerates after 2027, LYTE’s module-heavy weightings will underperform a portfolio that owned switch ASIC vendors and advanced packaging suppliers. The ETF may be on the wrong side of the next packaging transition. There is also a more uncomfortable risk: the demand forecast could be right, but the profit capture could be wrong. If the hyperscalers continue to expand AI clusters, they will need more optical modules. But they may also use their purchasing power to force lower prices, or they may design their own optical engines. The model of selling independent optical modules to a handful of giant customers works well during shortage, but the balance of power will shift once supply catches up. The long-term margin for module assembly is unlikely to stay at 30% to 35%. The competitive gravity of the semiconductor industry always moves toward consolidation and downward price pressure, and optics is not immune. Takeaway: The Next Narrative Is Bottleneck Ownership The next chapter of the AI-crypto story will not be written in token tickets. It will be written in the physical constraints of the photonic supply chain. I am not predicting a crash in LYTE, and I am not dismissing the incredible technical progress in silicon photonics. I am saying that the mirror maze has become crowded. The investor who stands in front of it should track InP substrate supply, MOCVD bookings, CoWoS capacity, and the inventory choices at the largest cloud providers. Those data points will tell you whether the 57% growth forecast is being confirmed or contradicted before the price does. The old signatures of my profession have not changed. We are hunting for truth in a mirror maze of hype, and the hunt now runs through lasers, epitaxy, and trade policy. The ledger remembers what the heart forgets. Value flows to the bottleneck, and the bottleneck always leaves a physical trace. Follow the trace.