Follow the Photons: Why the Optical Supply Chain Is the Real Settlement Layer of the AI-Crypto Trade

Hook: The 67% Concentration Problem

Ignore the ETF ticker. Look at the photons.

As of August 2025, the Roundhill Photonics and Optical ETF (LYTE) is trading as if the AI infrastructure trade cannot fail. Its five largest holdings — Lumentum, Coherent, Zhongji Innolight, Eoptolink, and Tianfu Communication — account for over 67% of the fund's total weight. The industry line puts AI optical module growth at 57% this year, reaching $26 billion. Impressive. Also irrelevant until you stress-test the construction behind the number.

Illusions dissolve under stress testing. I have performed this exercise before. In late 2017, as a junior quantitative researcher in Copenhagen, I ran Python scripts across the Ethereum mainnet to audit reserve claims for five ICO projects. Three of them held less than 5% of their advertised reserves in cold storage. The team divested immediately and missed an 80% drawdown. I bring that same forensic instinct to this analysis, because the LYTE basket is not a passive collection of semiconductor names. It is a concentrated bet on the physical layer that the AI-agent economy — including crypto's own compute-intensive ambitions — depends on. What follows is not an equity thesis. It is a structural audit of the optical interconnect chain, its bottlenecks, its geopolitical fault lines, and the fragility hidden inside a forecast that says everything grows forever.

Context: The ETF as a Geography of Dependencies

LYTE sits at a strange intersection that most crypto observers refuse to acknowledge: the gap between the promise of decentralized AI and the brutally centralized hardware that enables it. A GPU cluster with 72 NVIDIA GB200 accelerators needs several hundred optical modules for scale-up and scale-out interconnects. Each high-end GPU requires eight to sixteen optical transceivers, depending on cluster topology. Before a single proof is generated, before a single zk-rollup batch is verified, before any AI agent can autonomously transact on-chain, the photons have to move across the backplane.

The ETF's construction maps to this reality in five concentrated positions:

• Lumentum (U.S., roughly 15%+ of the fund): InP lasers, optical amplifiers, high-speed EML chips, some telecom heritage.
• Coherent (U.S., roughly 15%+): High-speed EML/CW lasers, IDM-level photonic chip capability, roughly 35% share of the high-speed optical chip market.
• Zhongji Innolight (China, via A-shares): The world's largest data-center optical module maker, approximately 25% share.
• Eoptolink (China): Second-tier module powerhouse, heavily exposed to North American CSP cloud demand.
• Tianfu Communication (China): Passive components — fiber arrays, isolators, connectors — with gross margins above 40%.

This is not a diversified portfolio. It is a deliberately compressed representation of one value chain: AI data-center interconnects. The fund skips Acacia, skips Marvell, skips Broadcom's optical divisions. What remains is an almost purist play on the optical physical layer.

In 2025, I led an economic modeling effort for AI-driven autonomous agents interacting with blockchain networks. I built simulations to predict how LLM-instructed agents would manipulate gas markets and oracle feeds. The work kept hitting the same external constraint: transaction settlement speed between agents was not limited by consensus design but by the physical latency of the data-center interconnect layer. That experience reframed how I read this ETF. The token layer is a narrative. The interconnect layer is a physical settlement system. When you evaluate infrastructure investments in crypto, you are implicitly evaluating this optical chain too.

Core: The Structural Audit of the Optical Layer

I. The Yield Surface of the Photonics Chain

In 2020, when DeFi Summer was peaking, I built dynamic models to separate organic TVL from liquidity-mining distortion across Aave, Compound, and Uniswap. The conclusion that mattered: short-term reward schedules were artificially inflating TVL by roughly 300%, and anyone who treated those yields as structural would be liquidated when incentives dried up. I applied the same lens here because the optical value chain has its own yield surface — a stacked sequence of gross margins that reveals where economic rent actually collects.

The numbers, estimated from 2024 fiscal data:

• High-speed optical chips (EML/CW lasers): 45–60% gross margin. This is the proof-of-work layer — the difficulty-adjusted energy of the entire system.
• Module assembly (800G/1.6T transceivers): 25–35% gross margin for Chinese leaders, 35–45% for U.S. incumbents with vertically integrated chips.
• Passive components (Tianfu's segment): 30–40% gross margin, but with scale effects that translate to net margins above 25%.

Follow the vector, not the hype. The vector points straight toward the upper layer. Value pools in this chain flow to whatever has the longest lead time, the harshest yield requirement, and the smallest set of qualified suppliers. That is the same dynamic I saw in DeFi: yield migrates to the real constraint, no matter how the marketing frames the ecosystem.

Lumentum and Coherent sit at the chip layer with pricing power because the bottleneck is structural. An EML wafer takes 12–18 months from expansion decision to volume output. The chain cannot accelerate it. Module makers such as Zhongji Innolight and Eoptolink can add lines in three to six months — a flexibility advantage that is genuinely impressive — but they remain price-takers on the chip layer. The gross margin gap between the module layer and the chip layer is the yield spread of this industry. It should be the first metric any defensively positioned investor watches.

II. The Bottleneck Map: Who Actually Holds the Leverage

Every blockchain has a validated set of constraints. Here, they are three:

The EML lockbox. Coherent holds roughly 35% of the high-speed optical chip market; Lumentum holds about 25%. Together they control the 100G and 200G EML production that powers 800G and next-generation 1.6T modules. Chinese chip vendors — Yuanjie Technology, Yunling Optoelectronics — are moving, but the yield gap is one to two years. At the 200G EML node, industry leaders achieve about 60–70% volume yield. Chinese fabs are still climbing that curve. This gap appears small until you realize that yield, not design, is the moat. A two-percentage-point yield difference at 200G EML is the difference between a profitable product and a marginal one.

The DSP bottleneck. Every high-end optical module requires a DSP — a digital signal processor that encodes and decodes the PAM4 signal. Over 90% of this supply comes from two U.S. companies: Broadcom and Marvell. These chips are fabricated on TSMC's 7nm and 5nm nodes. This is the same advanced-logic capacity that is already rationed across AI accelerators and networking silicon. If a geopolitical crisis lands on the DSP pin, Chinese module makers face a genuine production pause. The asymmetry is sharp: China assembles the modules, but the brain inside each module crosses the Pacific twice.

The InP substrate dependency. The indium-phosphide substrates that host high-speed lasers are dominated by Japanese suppliers — Sumitomo Electric and JX Metals control over 80% of the relevant market. Chinese providers such as Zhongke Jingdian are in early validation, not volume production. Meanwhile, upstream the gallium and germanium feed stocks for InP epitaxy are subject to Chinese export controls, which creates a rare two-sided vulnerability: Japan controls the substrate, but the element feedstock is under Beijing's administrative thumb. Both sides are exposed. Both sides pretend otherwise.

You can read this as a balance sheet or a war plan. The DSP is the capital controls of the AI era. The EML is its central bank. And this is the level of concentration hiding beneath an ETF that looks, on its face, like a diversified thematic product.

III. The Capacity Equation: China Speed Versus Physics

Utilization rates in global optical module manufacturing have been at 85–95% since mid-2024. 800G modules moved from scarce to tight, perpetually chasing hyperscaler demand. This feels like a shortage, but the more precise diagnosis is a structural mismatch between module capacity and chip capacity.

Module capacity is elastic. Equipment lead times are six to twelve months. A Chinese module factory can go from equipment move-in to volume production in three to six months — versus 12–24 months for a leading-edge logic fab. This speed is a structural advantage that U.S. and European competitors have not replicated, and after my 2022 audits of centralized exchange proof-of-reserves, I have learned to respect the operational edge of teams that can move physical gear quickly.

The same report paints the darker side: optical chip capacity is inelastic. MOCVD epitaxy tools take 12–18 months to deploy, and qualified process engineers do not scale like assembly-line workers. The industry is spending accordingly — Lumentum and Coherent have each committed hundreds of millions of dollars to high-speed laser expansion through 2026. Capital intensity is moderate by semiconductor standards (module makers spend 10–15% of revenue on capex, chip makers 15–25%), but the real constraint is not money. It is time.

My honest read of the delivery timeline: the 800G shortage resolves into balance by 2026, just as 1.6T volume ramps. The module makers will live through the squeeze, but their gross margins stay structurally capped by chip supply until 2027–2028, when Chinese EML vendors reach credible volume. The floor is a trap for the impatient — investors who buy the module names today expecting margin expansion will be waiting on the chip layer's timeline, not their own.

IV. The Demand Vector: When the Growth Number Is the Argument

The headline number — AI optical modules growing 57% to $26 billion — deserves deconstruction. It is not a forecast. It is a commitment schedule. The top U.S. cloud providers (Microsoft, Google, Meta, Amazon) are expected to direct more than $300 billion of capital expenditures in 2025, with the optical interconnect slice representing roughly 5–8% of that. The volume is not an estimate of end-user adoption; it is the consequence of already-committed, largely non-cancellable infrastructure spending.

The physics is unforgiving. An H100 GPU requires roughly eight optical modules in a typical scale-up deployment. GB200 NVL72 systems require sixteen per GPU when you account for the full NVLink spine. There is no software optimization that replaces a physical cable; there is no protocol upgrade that condenses the photon count. The compute-to-interconnect ratio is a thermodynamic fact, not an engineering choice.

The growth vector extends past training into inference. As inference workloads expand, demand flows down to 400G modules as a complement to 800G. The market is not choosing a winner at one speed. It is buying all speeds simultaneously because the installed base is heterogeneous. This is the difference between this cycle and the telecom cycles of 2022–2023. The 2022 correction was a genuine inventory overhang in a market with low structural growth. The current cycle is a structural shift — the CAGR of optical interconnects has moved from 8–12% in the telecom era toward 25–35% in the AI era. Volume without conviction is just noise; this volume has conviction embedded in purchase orders.

Still, I would flag one variable that most bullish reports underweight: the ROI question. Hyperscaler capex commitments are not strictly cancellable in 2025–2026, but the 2027 wave depends on whether the buildout actually produces returns. Every $300 billion of capex creates a balance-sheet expectation of future cash flow. If the AI margin story weakens, the optics chain feels it nine to twelve months later. This is not a prediction of a crash. It is a warning that the long-duration claims of the AI infrastructure trade have not been stress-tested outside of the boom.

V. The Geopolitical Hedge That Is Not a Hedge

The portfolio construction at LYTE deserves a closer look than most ETF analyses provide. Lumentum and Coherent are U.S.-domiciled, representing over 30% of the fund. Zhongji Innolight, Eoptolink, and Tianfu represent roughly 36.7% of the fund via Chinese listings. On its face, this looks like an even-footed allocation to both sides of the supply chain.

It is not.

The Chinese module makers generate the majority of their revenue from North American customers. Zhongji Innolight and Eoptolink both count Google, Meta, Microsoft, Amazon, and NVIDIA among their top accounts — their top five customers routinely exceed 60–70% of total revenue. The U.S. chip makers, meanwhile, rely on Chinese demand for modules even as they compete with those module makers in downstream markets. The fund is not a hedge. It is a double exposure to one single trade: North American AI capex sustained through the Chinese manufacturing industry.

Here is where the analysis gets uncomfortable. Based on my 2022 work auditing proof-of-reserves at centralized exchanges, I learned that counterparty concentration is the silent killer of medium-term returns. I identified solvency gaps at three major platforms before the FTX collapse. The lesson generalizes: when your counterparty risk is geographical, it does not appear on a balance sheet, and it cannot be diversified away by a ticker. The five holdings of this ETF are effectively five expressions of one counterparty — the U.S. hyperscaler oligopoly.

VI. The Geopolitical Scenarios: A Three-Branch Monte Carlo

I ran a scenario matrix on the geopolitical variables touching this chain, calibrated to the August 2025 policy environment. The output is sobering but not catastrophic.

Scenario A — Systemic decoupling (roughly 15% probability). The United States extends export controls to optical modules and high-speed optical chips. Chinese module exports to North America collapse. Global AI buildout slows as U.S. hyperscalers scramble for non-Chinese module capacity that does not yet exist. The cost of optical interconnects rises 30–50%, and delivery times extend by six to twelve months. The fund's thesis breaks in both directions simultaneously. This is the low-probability, high-impact tail, and anyone holding LYTE should acknowledge they are short the statesman gene in Washington and Beijing.

Scenario B — Targeted escalation (roughly 40% probability). The U.S. tightens access to DSPs and high-end lasers for Chinese module makers, as it has done with advanced logic. Chinese companies accelerate domestic substitution, passing over a 12-to-18-month cost and yield penalty before rebalancing. The fund experiences a sharp drawdown followed by a re-rating as the Chinese module suppliers recover with higher vertical integration. This is the scenario the market is not pricing because it looks like a contradiction: tighter controls eventually produce more self-sufficient Chinese suppliers.

The hidden accelerant is China's industrial policy. The third phase of the National Integrated Circuit Industry Investment Fund — roughly 344 billion yuan — explicitly includes photonics. Local governments add their own semiconductor funds. The Chinese response to export controls has historically been a full-throttle push for self-sufficiency. If I am forced to make a probability judgment, I would say the most likely surprise of the next 24 months is that Chinese silicon photonics design capability closes the gap faster than Western analysts expect.

Scenario C — Controlled friction (roughly 45% probability). The status quo persists. Optical technology remains in what I call the "regulatory exempt zone" of export controls — outside the core restrictions on logic chips, HBM, and advanced packaging. Under this scenario, the supply chain continues to operate with high efficiency, and the ETF trades on normal fundamentals. But even here, the gallium and germanium cards remain face-up on the table. China already demonstrated in 2023 that it can deploy these controls over key element feedstocks. Lumentum and Coherent both feel this exposure in their cost structure, which explains why gallium pricing has become a quiet margin topic in every earnings call.

VII. What I Would Cross-Examine at the Next Pitch Meeting

The statistical concentration of the portfolio, the bullish growth forecast, and the idyllic supply-demand balance all collapse into three questions that nobody asks aloud:

First, how durable is the price umbrella? 800G modules initially sold in the $1,800–$2,500 range and have already compressed toward $800–$1,200 as production scales. This 15–25% annual cost-down curve is the industry norm. The only way to counter it is product generation upgrades — 1.6T enters at $2,500–$4,000, resetting the margin clock for its two-to-three-year lifetime. The bull case for this ETF is, therefore, implicitly a bull case on the 3.2T roadmap arriving on time. Any delay in the next technology cycle converts today's tightness into tomorrow's margin squeeze.

Second, what happens when CSPs vertically integrate? Every major hyperscaler is evaluating internal optical development. Google has deployed in-house co-packaged optics research inside its data centers; NVIDIA is building a full-stack networking strategy that includes optical acquisition and partnership moves. If a Microsoft or Amazon decides to design modules in-house and tap directly into Lumentum and Coherent for chips, the entire addressable market for third-party module assemblers shrinks. This is the same pattern DeFi experienced when liquidity mining moved from third-party protocols back to the centralized venues — the intermediaries got crushed.

Third, what is the actual risk-free rate of this supply chain? Investors in a sideways crypto market often move into equity infrastructure products seeking security, but this portfolio is not risk-free in any regime. It carries technology risk (CPO could displace modular forms), geopolitical risk (the scenarios above), and concentration risk (five names, one thesis). I would frame the optimal allocation not as "safe AI exposure" but as "high-beta physics exposure" — a portfolio that ranks with ETH collateralization in terms of structural optimism and far below the safety of managed custody. In my experience, the promise of safety comes at precisely the moment when leverage is already embedded in the price.

Contrarian: The Decoupling Thesis Is Already a Fact, Not a Forecast

The mainstream framing of this sector is binary: either globalization survives in photonics or decoupling destroys the chain. I think both framings miss the point. The decoupling is already underway and it is happening inside the corporate structures of the very companies in this ETF.

The evidence patterns me as unusual. Zhongji Innolight and Eoptolink have both built production capacity in Thailand. Tianfu has followed. This is not hedging against a scenario; it is executing on an inevitability. The Chinese module leaders understand, better than any U.S. policy analyst, that "made in China" is becoming a liability for North American customers. The "China speed" advantage that I credited earlier is now being exported — Chinese companies are building multinational manufacturing footprints. They are transforming from Chinese exporters into global corporations with Chinese R&D cores. This is a more resilient model than either pure domestic production or U.S. reshoring.

Simultaneously, U.S. chip makers are increasing their dependency on Asian manufacturing and demand even as they execute expansion plans in Texas and across multiple countries. Coherent's Texas optical chip capacity, and Lumentum's multi-country expansion, are responses to a world in which "American" and "Chinese" are no longer clean supply chain categories. The truth is that both sides need each other to profit from the AI wave. The U.S. controls the DSP and high-speed chip layer; China controls packaging, passive components, and module assembly at cost levels no Western competitor can match. That interdependence is as durable as the gallium-in-InP loop that connects Japanese substrates, U.S. laser fabs, and Chinese assembly lines.

The counterintuitive conclusion is that this ETF's concentrated exposure to both sides of the chain is not a bug but a feature. It captures the messy globalized reality of the AI-era infrastructure trade. The bear case for the fund is not decoupling — it is a clean and total decoupling executed with no regard for the mutual damage it inflicts. The more probable path is a slow rearrangement where supply chains bifurcate into "China+1" and "US+1" versions, each with separate but overlapping optical architectures. In that world, holding Lumentum, Coherent, and Chinese module makers together is rational, not contradictory.

But here is the deeper trap I keep circling back to: the entire thesis assumes that AI capex remains a sovereign-grade commitment. The market has decided that $300 billion of annual hyperscaler capital spending is going to continue because it is now too big to cancel. That is exactly the kind of stable conviction that crypto natives recognize as the precursor to a liquidity event. When everyone expects a liquidity pool to go on forever, the smart trader triangulates the exit ramp.

Takeaway: Positioning for the Bottleneck, Not the Narrative

Let me state the investment conclusion as plainly as the analysis allows. The optical interconnect chain is the highest-conviction physical infrastructure play in the AI-crypto convergence narrative, and LYTE captures it with unusual purity. But purity is concentration, and concentration is risk — the ETF is exposed to three critical chokepoints (EML chips, DSP supply, and the InP substrate market) with whom the actual pricing power rests. A portfolio that includes Lumentum and Coherent is a portfolio that is long the bottleneck. A portfolio heavy in Chinese module makers is a portfolio long the manufacturing speed that remains unique to the Asian supply chain.

The optimal positioning, in my judgment, is to weight the chip layer over the module layer through 2026 — because the lead time asymmetry continues to favor whoever controls the slowest step. Watch the DSP allocation queues at TSMC as the early warning signal, and watch the Chinese EML yield improvements at 200G as the late-cycle signal. If the 57% growth figure proves soft, the chip layer will hold margins better than the assembly layer, which will absorb the first round of price declines.

I have been at this long enough to know when an industry is approaching its own refraction point. The shift from 800G to 1.6T to 3.2T is not a continuous upgrade cycle; it is a phase change. CPO — co-packaged optics — will reroute the value chain entirely by 2027, drawing optical engines into the same package as switching silicon. That transition will upend the module assemblers and reward whoever controls the packaging integration stack. The team that wins CPO is the team that wins the next cycle. I do not yet have a net conviction on who that is, but I know how to find out: follow the capital expenditure lines, follow the customer qualification programs, and do not trust the narrative until the revenue mix reflects it.

On-chain, the same discipline applies. I started this piece with a memory of auditing ICO reserves in 2017, and I will close with the version that matters for 2025: the AI-agent economy will not be constrained by consensus latency or gas prices, but by the photons running through this supply chain. The protocols will perform as their hardware allows. When I built my agent simulation this year, the binding constraint was never the smart contract logic. It was the interconnect fabric between the data centers. That is the vector that matters, everywhere — on-chain or off.

You do not need to buy the ETF to bet on the thesis. You need to know where the bottleneck lives, who controls it, and why the market keeps under-pricing the slowest step. The floor is a trap for the impatient. The bottleneck is a home for the ones who see the whole machine.

And the machine, this cycle, is built of light.