The data suggests we have been auditing the wrong layer.
Roundhill's LYTE ETF landed in August with a portfolio that reads like a modem-era joint venture: Lumentum at roughly 15%, Coherent at 15%, Zhongji Innolight at 14%, Eoptolink at 14%, and Tianfu Communication at less than 8%. This is not a thematic basket. It is a weighted map of the global optical module supply chain. And the anomaly is not valuation. The anomaly is that America's financial infrastructure has issued a product that explicitly refuses to decouple from China's manufacturing base.
Here is the twist nobody in the crypto commentary room is talking about. A US-listed ETF has encoded the exact interdependence that Washington's policy engine spends every legislative cycle trying to unwind. Lumentum and Coherent represent American light-chip incumbency. Zhongji Innolight and Eoptolink represent Chinese module scalability. Tianfu represents the passive components that everyone takes for granted. The architecture of the ETF reveals the true intent: this is not a bet on one country beating the other. It is a bet on a commodity chain that cannot be separated without raising the price of every AI cluster on Earth.
For anyone who spent the last cycle inside rollup disputes, the pattern is familiar. The settlement layer is not the layers you argue about. It is the layer you ignore until it fails.
Context: The Physical Layer Two
The optical module stack is the physical Layer 2 of AI infrastructure. It inherits demand from the GPU layer above, but it has its own execution logic: light in, light out, measured in photons per watt, and priced in transceivers per rack.
LYTE is a dilute play on that logic. Its holdings span three nodes of a very specific value chain. Lumentum and Coherent sit upstream, where InP quantum wells, EML/DFB laser diodes, coherent optical engines, and silicon-photonics dies are designed and fabricated. Zhongji Innolight and Eoptolink sit midstream, where transceiver modules are assembled, coupled, tested, and shipped into QSFP-DD and OSFP packages. Tianfu Communication sits in the upstream passive layer: ceramic ferrules, lens arrays, isolators, and the kind of optical components that never get a keynote slide but stop a supply chain cold when missing.
The market context is equally specific. AI data centers now contribute an estimated 40% to 50% of optical module revenue, growing at triple-digit annual rates. Traditional cloud remains a slow-growth 30% slice. Telecom is a low-single-digit laggard. The entire narrative is one product cycle: 800G modules today, 1.6T modules tomorrow, and co-packaged optics waiting in the wings.
The global optical module market's CAGR is projected to jump from the old 5% to 8% range to 15% to 20% through 2028. That is a structural re-rating, not a cyclical hiccup. But structural re-ratings are exactly when the market stops reading footnotes.
Core: The Real Cost Curve Is Yield, Not Lithography
Tracing the yield curve back to the InP substrate reveals where the industry's true constraint lives.
The process node debate is a trap. Optical modules do not run on 3nm or 5nm. The industry runs on two parallel geometries. The electrical DSP layer is built by Broadcom and Marvell on 5nm or 7nm at TSMC. The optical chip layer is built on 130nm to 45nm silicon photonics, or on InP/GaAs material systems that do not care about Moore's Law. The bottleneck is not lithography. The bottleneck is material physics.
Current volume production looks like this: 100G EML and DFB lasers are mature. 200G EML is in the painful production ramp. Silicon photonics is mostly a 45nm to 130nm game. DSPs are already at leading-edge geometry. This asymmetry should remind any protocol architect of a modular blockchain: the execution layer is running on frontier silicon, while the data-availability layer is still fighting defect density.
The unit of throughput is not compute. It is coupling.
Every 800G module requires precision alignment between a laser chip, a lens, and a single-mode fiber. Misalignment by a micron means insertion loss. Insertion loss means retest. Retest means yield decay. The industry's dirty secret is that high-speed EML laser yield, especially 200G per-channel, still sits in the 50% to 70% range. The leaderboard is defined by who can keep that yield above 60% in mass production, not by who can draw a prettier block diagram.
Yield is verification. In semiconductor economics, verification is the only currency that matters.
I learned this lesson in 2017 while auditing Uniswap v1's transferFrom logic. I spent four nights tracing a simple library pattern and found a 12% gas reduction using unchecked arithmetic. The merged pull request saved a meaningful fraction of early DEX gas costs. The technical lesson was not about arithmetic. It was about locating the exact opcode where a trivial inefficiency becomes systemic. The optical module industry has the same shape. The most ignored parameter is not line rate. It is the coupling yield between a laser diode and a glass waveguide. Every failed die is not a defect. It is a gas payment.
The China-versus-America distinction follows the same logic. In module assembly, Chinese manufacturers are not catching up. They are already at parity or ahead. Zhongji Innolight controls an estimated 30% to 40% of the high-speed datacom module market. Eoptolink is a first-tier supplier. The Chinese edge is not secret technology. It is manufacturing speed, automation, and a brutal tolerance for thin margins.
Upstream, the picture inverts. Lumentum and Coherent hold an estimated combined 30% to 40% share of the EML/DFB laser chip market. Their R&D intensity is higher: roughly 15% to 20% of revenue, versus 5% to 10% for the Chinese module leaders. The gap is not one year. In high-end coherent optical chips and 200G EML, the American incumbents still lead by one to two generations. That translates into roughly two to three years of material advantage.
The hidden information in the LYTE weights is that the ETF designers are implicitly treating these two groups as complements, not rivals. That is a defensible reading. The chain needs Lumentum's lasers and Zhongji's assembly capacity. It does not need them to compete with each other in every segment.
The DSP Dependency Is The Oracle Feed Problem
If there is one structural risk that reminds me of DeFi oracle latency, it is the DSP dependency.
A high-speed optical module is useless without a digital signal processor to encode and decode the light. The high-end DSP market is effectively a duopoly: Broadcom and Marvell. Their DSPs run on leading-edge 5nm TSMC process nodes. Chinese module makers do not control that design. They do not control that geometry. And they most certainly do not control the export license.
The import dependence for high-end DSPs is brutal. Domestic Chinese DSP supply for 800G-class modules is still likely below 10%. EML and DFB laser chips above 100G are better, with Chinese supplier penetration perhaps in the 20% to 30% range. InP substrates remain heavily dependent on Japanese and American suppliers. EDA tools are dominated by Synopsys and Cadence. Every node on the chain has a different dependency level, and none of them are near full autonomy.
This is the technical version of oracle feed latency. In DeFi, a slow price feed creates a window for arbitrage. In optical infrastructure, a constrained DSP supply creates a window for geopolitical arbitrage. The party that controls the DSP controls the pace at which every Chinese module maker can ship.
The industry has responded with the same logic as every decentralized platform: inventory hoarding and domestic substitution. Chinese national funds are likely to push into optical chips and high-speed DSPs over the next three to five years. The likely path is a mid-rate breakout first, followed by a slower assault on 200G-plus parts. But that path is measured in years, not quarters.
Capacity And The Marginal Cost Trap
The capacity picture is strong but fragile. Industry utilization is running at an estimated 80% to 95%. Chinese leaders are close to full capacity on 800G lines. Capex intensity is 10% to 20% of revenue, below the weight of semiconductor fabs but rising.
Equipment delivery is faster than the chip world. A large module line can go from tool installation to production in three to six months. Depreciation schedules are five to seven years. The marginal drag on gross margin from new capacity is roughly one to three percentage points. As long as utilization stays above 60%, depreciation is covered.
The fragility is not today's utilization. It is the second derivative. If cloud capital expenditure pauses, the same capacity that looks like a growth moat becomes a margin anchor. Module prices for a given speed decline 10% to 20% per year. New products, such as 800G and eventually 1.6T, reset the price curve. But the window of pricing power is short. The industry has been here before: every new speed grade brings a temporary margin expansion, followed by a Chinese price war that compresses everyone toward the same spread.
The demand side is where the bullish case rests. One large AI training cluster can consume tens of thousands of 800G modules. The migration to 1.6T, starting in 2025, should lift average selling prices and reset margins again. The inventory cycle is currently mild, with low channel inventory and tight lead times. This is the healthiest phase of the cycle. It is also the phase where investors confuse a high tide with a trading edge.
Financial Engineering: This ETF Is A Beta Sweep
A closer look at the financials confirms the ETF is not designed to isolate alpha. It is designed to sweep beta across a high-demand industry.
Gross margins cluster in the 30% to 35% range for the major module makers, with Tianfu Communication standing apart in the 45% to 50% range because passive components are a high-margin toll booth. R&D spending is modest for the Chinese names by global standards, but their return on capital is strong. Zhongji Innolight's ROE is estimated in the 25% to 30% range. Chinese ROIC levels are above 20%, meaningfully above a WACC of 8% to 10%. These are genuine value creators in the short run.
The US incumbents look worse on this metric. Lumentum and Coherent carry legacy business lines and lower ROIC, somewhere in the 5% to 8% range. Their bull case is nonlinear: if coherent optics and TFLN-based modulators become core to 1.6T-plus systems, their material science advantage becomes a royalty on the whole chain.
Valuation is the weak point. Chinese leaders trade at 30 to 50 times trailing earnings. Forward multiples are equally stretched. EV/EBITDA in the 15 to 25x range is fully loaded with AI optimism. This is not a 2020 entry price. This is a crowded trade. The LYTE ETF as a package gives the buyer a diversified position across three segments, but it does not give the buyer a margin of safety.
Threat Model: The Blind Spots No ETF Can Hedge
Contrary to the prevailing narrative that export controls are the main risk, I would argue the deeper risk is architectural substitution.
The most obvious threat is the supply-chain one. Zhongji Innolight, Eoptolink, and Tianfu are not currently on the US BIS Entity List. Optical modules themselves are not explicitly restricted. But the DSPs inside those modules are precisely the kind of high-end semiconductors that export rules are designed to catch. If the US restricts Broadcom and Marvell from selling 5nm-class DSPs into Chinese module supply chains, the Chinese module makers face a short-term delivery crisis. US-based Lumentum and Coherent would be relative beneficiaries. But the global AI buildout would slow, which would eventually hurt every holding in the ETF.
The second-order geopolitical risk runs in the opposite direction. China controls the vast majority of global optical module assembly, and it has already shown a willingness to restrict exports of gallium and germanium. Those elements are inputs into compound semiconductors. If China weaponizes module assembly the same way it weaponized rare-earth processing, the cost curve of every Western data center shifts upward. Decoupling is not a one-way trade.
The third blind spot is customer concentration. The top five customers of a large module maker can account for more than 60% of revenue. Google and Meta alone may exceed 40% for some names. The cloud giants do not need a fiery breakup. They simply need to guide next year's capex lower and the entire LYTE thesis loses its anchor. This is not a malicious scenario. It is a normal cloud capex cycle.
The fourth and most underestimated blind spot is technology substitution. CPO, or co-packaged optics, threatens to move the optical engine off the pluggable module and onto the switch package itself. LPO, or linear-drive optics, attempts to strip the DSP out of the module entirely. If CPO reaches cost parity by 2026 or 2027, then the current generation of QSFP-DD and OSFP capacity becomes transitional infrastructure. The same pattern destroyed modem companies two decades ago. The architecture of the industry changes faster than its depreciation schedule.
A threat model that ignores these risks is not a threat model. It is a marketing brochure.
The Hidden Thesis Beneath The Weightings
The LYTE ETF's weightings are more informative than its name. Lumentum and Coherent are in the portfolio for their upstream chip physics. Zhongji Innolight and Eoptolink are in the portfolio for the exact reason that makes them hard to short: they assemble the modules that make AI clusters physically possible. Tianfu is in the portfolio because passive components are the most quietly predictable margin in the chain.
The balance between the two blocks is close: roughly 30% US weight against roughly 34% to 37% Chinese weight. That is not an accident. An issuer that wanted pure American exposure could have left out the Chinese names entirely. The fact that it did not is a declaration that the financial market still views this as a coupled system. Political decoupling and financial coupling are diverging. In the long run, one of those signals is lying.
Based on my experience tracing fraud-proof dispute windows in optimistic rollups, I can tell you that the safest systems are those where the challenge period is longer than the economic volatility of the underlying asset. The optical module cycle has no such challenge period. The bull case is priced as if AI capex is infinite. The bear case is that the ETF is a leveraged bet on the calendar delivery rhythm of hyperscale data centers.
Takeaway: The Real Settlement Layer Is A Photon
The question this ETF does not answer is what happens in 2027, when co-packaged optics quietly move the optical engine from a replaceable module onto the switch substrate. The standalone optical module will not disappear overnight. But if the historical pattern of networking holds, it will migrate downmarket and become the copper cable of the next decade.
For now, the investment thesis is simple: AI compute needs light. Even in the most conservative cloud capex scenario, 800G modules are the toll road to every large-scale GPU cluster. But the margin of safety at 30 to 50 times earnings is thin. The next big trade is not buying the ETF. The next big trade is identifying which companies in the chain own the material science that survives the transition to co-packaging. Verification is the only currency that matters, and the market has not yet verified that the current module makers will be the ones to capture the CPO margin.
The ETFs have done their job. They have made the supply chain visible. But visibility is not edge. In 2026, the world will learn whether the optical module industry was the settlement layer of AI infrastructure, or just another pluggable block in a consensus mechanism that history is about to upgrade.