Lumentum's latest earnings call wasn't about lasers. It was a confession. The company can't make enough of them. Inventory is bone dry. EML laser chip capacity is stretched to the limit, and every AI data center on Earth โ from Google's clusters in Iowa to GPU farms quietly rising in Eastern Europe โ is waiting on these tiny pieces of indium phosphide that convert electrical signals into pulses of light. Meanwhile, a new ticker started trading on NYSE Arca. LYTE. Roundhill's optical module ETF. A single basket holding American chipmakers and Chinese module manufacturers in roughly equal weight. In a bear market for digital assets, I've spent the last week not parsing on-chain flows โ but tracing the physical supply chain that makes all digital data possible. The signal is loud. And nearly every crypto analyst is missing it. The same week, a protocol lost 40% of its liquidity providers. Two narratives colliding: the ethereal world of digital assets deflating, while the industrial layer of the AI complex tightens like a drum.
The optical module industry is the plumbing of the digital age. Every byte that travels across the internet โ every transaction pushed to a validator set, every shard of a Layer2 state commitment, every AI model inference โ moves as light through fiber optic cables, terminated by transceivers that convert photons to electrons and back again. Yet this industry doesn't behave like Silicon Valley. It's a hybrid: part semiconductor physics, part precision manufacturing, part commodity production.
The chain splits into three distinct power centers. Upstream, photonic chip companies like Lumentum and Coherent design and fabricate the laser chips โ EML and DFB lasers at 100G/200G per lane, built on indium phosphide (InP) and gallium arsenide (GaAs) substrates. This is material science, yield-challenged and unforgiving. A 200G single-channel EML laser's yield sits between 50% and 70% at industry average. The physics doesn't scale the way transistors do. You can't shrink an InP laser the way TSMC shrinks a logic die; the material itself is the constraint. Photonic chip fabs use contact lithography and DUV steppers, not EUV. Moore's Law simply doesn't apply in the same timeline.
In parallel sits the electrical layer. Every module needs a DSP โ a digital signal processor designed by Broadcom or Marvell, fabbed at 5nm and 7nm by TSMC. This is the true choke point. The DSP duopoly controls the brains of every high-speed optical module. And Chinese module makers โ despite their manufacturing strength โ have less than 10% domestic DSP substitution. Place an export control on high-speed DSPs, and a Chinese module supply chain stalls within two quarters.
Then there's the module assembly layer. Zhongji Innolight and Eoptolink โ the Chinese giants โ control roughly 30-40% of the 800G data-comm module market. Their R&D budgets, at 5-10% of revenue, are far smaller than Lumentum's 20% or Coherent's 15%. But their output per R&D dollar is exponentially higher. This isn't a technology gap. It's a business-model gap. The Chinese firms excel at manufacturing intelligence: supply chain integration, packaging, yield management, and high-velocity iteration. I haven't mentioned Tianfu Communication yet โ China's leading passive optical component maker, with gross margins around 45-50%, higher than any active module player. Passive components โ lenses, filters, ceramic ferrules โ don't get the glory. They also don't get commoditized as brutally. Tianfu's inclusion in LYTE is the quiet confirmation that the market understands the full stack.
Now let's dig into what LYTE's construction actually tells us. The construction details matter, so let's walk through them. The fund weights Lumentum and Coherent at roughly 30% combined, against roughly 36.7% for Zhongji Innolight, Eoptolink, and Tianfu Communication. Roughly balanced โ a deliberate hedge. This is a financial instrument that bets on both sides of the Pacific simultaneously. It's not a clean thematic ETF. It's a diplomatic agreement expressed as a product. And if you map the dependencies, the basket looks less like a portfolio and more like a hostage treaty. The Chinese module makers need American DSPs โ a single export-control ruling on those chips could halt production in a quarter. The American chip makers need Chinese manufacturing scale to meet AI delivery timelines; no other country can stand up 800G module capacity this fast. Lumentum and Coherent could theoretically supply everything themselves, but their cost structures would never satisfy hyperscaler pricing expectations. And if China's gallium and germanium export restrictions โ already live since 2023 โ tighten further, those same American chip makers face an input-cost shock. The ETF isn't just diversified exposure to AI infrastructure. It's a cross-collateralized bet that political decoupling remains rhetorical. That's the hidden thesis in plain sight.
Let me push on the technical layer, because that's where the real constraints live. During my Prague Protocol audit in 2017, I learned something that stuck: the most dangerous vulnerabilities hide in the parts of a system nobody reads. In DeFi, that's the token contract's edge cases. In the optical industry, it's the yield curve. High-speed EML lasers have a 50-70% yield at best. That means for every wafer of laser chips, a third to half are scrap. Which means the industry's effective capacity ceiling is far lower than its nominal capacity. Which means the market's ability to respond to demand is structurally constrained. The market prices the narrative of AI growth, but the physical substrate moves at the speed of semiconductor material science. s fragmented logic. โ supply can't be conjured; it must be grown, wafer by wafer, with patience the narrative never has.
The demand side is real. I've been tracking hyperscaler capex since the 2022 crash, and the trajectory is unlike anything I've seen. AI training clusters consume thousands of 800G transceivers each; a single large cluster can require 10,000+ modules. Industry capacity utilization sits at 80-95%, with Chinese leaders near full output. Inventory across the chain is normal to low. This is a genuine supply squeeze โ not a paper narrative. But โ and this is where my DeFi instincts kick in โ supply squeezes are exactly where narrative error becomes most expensive. Remember the 2020 DeFi Summer. The 'money lego' narrative assumed composability would compound value infinitely. It was true โ until it wasn't. The flaw wasn't in the smart contracts; it was in the assumption that liquidity is sticky and user behavior rational. In the optical module world, the equivalent assumption is: 'Hyperscaler capex will never slow.' That assumption deserves scrutiny.
Look at buyer concentration. The top five customers of most module makers account for over 60% of revenue. Google, Microsoft, Meta, Amazon. They dictate terms. They push annual price reductions of 10-20% on the same speed grade. The only way module makers maintain ASPs is to ride the technology treadmill: introduce 800G, harvest margins, then move to 1.6T, then 3.2T. Each generation demands new capex, maturing yields, and fresh risk. This is structurally similar to crypto mining โ a hardware-dependent, capital-intensive business where profitability depends on a single external price and on flawless timing of hardware refreshes. The winners in mining were never the cheapest miners; they were the ones who could access capital at the lowest cost and time their equipment purchases perfectly. The same logic applies to optical modules.
Here's a detail most analyses skip. The capex-to-revenue ratio for module makers has climbed from single digits to 10-20% over the past three years. New production lines, cleanroom expansion, and automation in China are absorbing tens of billions of RMB. The depreciation schedules โ typically five to seven years โ mean gross margins will absorb a one-to-three percentage point drag for the foreseeable future. That's tolerable when utilization stays above 80%, and brutal if demand dips to 60%. Anyone modeling this industry as a stable-margin compounder is ignoring the capital cycle. This is cyclical manufacturing dressed in AI hype.
The next act is already in motion. 1.6T modules โ doubling the 800G data rate โ are entering sampling. The ASP uplift is significant, which is why Lumentum and Coherent are pouring hundreds of millions into InP wafer capacity, and why Chinese module makers are standing up new production lines. The 1.6T cycle, expected to inflect in late 2025 and into 2026, will separate the players who timed their capex correctly from those who over-built for 800G just as its ASP curve bends downward. The capacity decisions made today will determine who survives the next industry trough.
Now consider the substitution question, because it determines the decade's power balance. Chinese DSP development is progressing, but from a staggeringly low base โ under 10% domestic substitution today. Industry consensus suggests mid-rate DSPs could see domestic breakthroughs within three to five years, but 200G-class electrical interfaces lag further. If China's DSP program catches up faster than expected, the entire dependency structure I've described flips. The American chipmakers lose their leverage, and the Chinese module makers become vertically integrated monopolists. That's the scenario LYTE's balanced weighting implicitly dismisses. It's the market's way of saying: 'we don't believe the Chinese DSP story yet.' That disbelief may be the single best contrarian trade in the whole narrative.
The valuation picture complicates the thesis further. Chinese module leaders trade at 30-50x trailing earnings, while Lumentum and Coherent sit at more modest multiples. But look at return on invested capital: Zhongji Innolight posts ROIC above 20%, while the American photonic giants hover around 5-8% โ barely above their cost of capital. The market is paying a massive premium for Chinese manufacturing efficiency. In a bear market, that premium is exactly what contracts first. I can't help but see the parallel to Layer2 fragmentation in crypto โ dozens of networks, but the same small user base, liquidity sliced into ever-thinner fragments. In the optical space, we have dozens of startups chasing the same four hyperscaler customers, with value concentrating into the few players who own both scale and vertical integration.
But here's the genuine insight underneath all of this. The optical module chain is the closest thing to a 'chain abstraction' in the physical world. s fragmented logic. โ the industry's fragmentation is also its resilience. Because the chain spans multiple geopolitical blocs, no single regulator can kill it. Because manufacturing and design are separated, the two sides are locked into mutual dependence. And because technology iterates in speed grades rather than architectural revolutions, incumbents have time to adapt. This is the opposite of crypto's L2 fragmentation problem, where value dissipates across incompatible silos. Here, cross-border fragmentation creates durable value by making decoupling prohibitively expensive.
Let me connect this to my own trajectory. In 2026, I launched a speculative research project on the convergence of AI agents and blockchain. The core question: can decentralized networks provide the compute, verification, and economic settlement for autonomous agents? I failed to build the product โ ENFP scope creep, as always. But the research led me to a counterintuitive conclusion: the binding constraint on agent economies isn't consensus algorithms or tokenomics. It's bandwidth. It's the physical layer โ the fiber, the transceivers, the lasers โ that moves data between compute nodes at the speed and cost that make distributed training and inference viable. Blockchain's contribution to the AI stack will remain trivial until the physical layer is solved. And the physical layer, right now, is solving itself through a supply chain binding Shenzhen and Silicon Valley in mutual necessity.
Which brings me to the contrarian read. The bear case isn't technology. It isn't even geopolitics, at least not in the obvious form. The bear case is narrative collapse. The launch of LYTE itself is a sentiment signal. When an asset class reaches the point of niche financialization โ when a sub-industry gets its own thematic ETF โ it often marks the peak of narrative enthusiasm. I saw it with NFT index products in 2021. I saw it with crypto mining ETFs at Bitcoin peaks. It's happened in every cycle I've studied โ from the 2017 ICO mania to the 2021 NFT summer. The operational thesis may still play out over a multi-year timeframe, but the ETF's existence tells you the story has become visible enough to package. And visible narratives are, by definition, late-stage narratives. s fragmented logic. โ the trade no longer belongs to the patient few; it belongs to whoever is willing to buy the ticker.
Consider the CPO risk too. Co-packaged optics โ integrating the optical engine directly onto the switch ASIC โ could erode the removable transceiver market by 2027. If CPO's yield and cost curves break through earlier than expected, the entire module assembly layer that the Chinese champions dominate becomes significantly less valuable. Incumbents are racing to adapt โ Zhongji Innolight has CPO research programs, Coherent is pushing silicon photonics โ but transitions like this are brutal for companies built around old form factors. This is the chain-abstraction moment of the optical world: a sudden architectural shift that rewrites the value chain. The historical analog in crypto is the move from proof-of-work to proof-of-stake on Ethereum. It took years. But when it happened, the entire mining hardware supply chain lost its raison d'รชtre. The same fate could await pluggable module assembly if CPO matures faster than expected.
And the final contrarian angle, the one that makes me most uncomfortable: the ETF blends competitors into a single 'AI infrastructure' story. But the two halves have opposite risk profiles. Lumentum and Coherent are upstream material scientists with pricing power but slow growth. The Chinese module makers are downstream manufacturers with hypergrowth but zero pricing power. One is a bond-like bet on chips. The other is an equity-like bet on AI capex volatility. They don't hedge each other. They compound each other's failure. If AI demand collapses, American chipmakers' volumes drop, and Chinese module makers' margins compress simultaneously. The diversification in LYTE, I suspect, is optical.
So what do we do with this information? For crypto analysts, the optical module chain is a leading indicator we've been ignoring. When hyperscaler capex shifts, when fiber orders tighten, when laser chip yields constrain supply โ those are signals for the health of the entire digital economy, including the DeFi and L2 ecosystems we track. I'm now treating LYTE as a macro sentinel. If the industrial physical layer stumbles, the digital layer will feel it first โ through energy prices, bandwidth costs, and data center availability. The physical world has a cruel sense of timing; nothing mints a laser chip out of thin air. I'm also watching the 1.6T inflection and CPO timelines as the two variables that could reshape the basket's thesis within eighteen months. If the 1.6T cycle goes smoothly, the Chinese manufacturers consolidate their lead. If CPO accelerates, the value migrates upstream to the photonic chip designers. Either way, the ETF becomes a referendum on which half of the physical layer actually owns the AI narrative. The question that keeps me up at night: have we built a digital economy whose narrative runs far ahead of its physical capacity? Or is the physical layer finally catching up to what the narratives promised? The answer, I suspect, will be written in light pulses moving through fiber. And for the first time, there's an ETF that lets us bet on it. Watch the fiber. The future will travel through it โ or it won't arrive at all.