The LYTE ETF Is Not a Photonics Trade. It Is a Leveraged Bet on the AI Compute Bottleneck.

The consensus is wrong because it treats optical interconnect as a byproduct of the AI trade. It is not a byproduct. It is the bottleneck. The AI optical module market is forecast to grow 57% this year to $26 billion. That number has been repeated so often that it now sounds like a law of physics. It is not physics. It is an assumption about cloud capital expenditure in 2025 and 2026, and it deserves the same forensic suspicion I applied to whitepapers during the 2017 ICO boom. Most of those projects died because they ignored an input that could not be replaced. Photonics has the same property. Sell-side commentary frames optical as one more AI beneficiary. That misses the structural mismatch. The bottleneck is not assembly; it is upstream chip capacity. Expanding high-speed laser chip production takes 12 to 18 months. Module assembly can scale in three to six months. When those two speeds diverge, pricing power moves to the slowest layer. That is exactly where Roundhill Photonics and Optical ETF (LYTE) positions its largest weights. Context LYTE owns a concentrated slice of the optical supply chain. Lumentum and Coherent anchor the high-end compound semiconductor layer. Zhongji Innolight, Eoptolink, and TFC dominate module assembly, silicon photonics integration, and passive optics. The top five holdings account for more than 67% of the fund. This is not broad sector beta. It is a directed bet on one pathway: AI accelerator to optical transceiver. Read as a supply chain map, the portfolio tells a clearer story. The American names own indium phosphide epitaxy, laser design, and waveguide IP. The Chinese names own scale, yield, and delivery speed. One side controls the chip generation. The other side controls the production ramp. That is not a geopolitical accident. It is a hedged expression of a bifurcated industry. There is a hidden signal in the selection. Innolight and Eoptolink are leading silicon photonics adopters. TFC makes the fiber array components that silicon photonics modules require. The ETF is therefore not just betting on optical demand; it is betting that silicon photonics replaces conventional EML lasers as the mainstream architecture. I would put that probability at roughly seven out of ten. The silicon light engine may not be the cheapest option today, but it scales into the data center ecosystem more gracefully than discrete lasers. Code is law, but capital decides who writes it. In photonics, the real code is the epitaxial recipe and the waveguide layout. Core Analysis The photonics stack leads to three conclusions that matter for positioning. First, process geometry is the wrong lens. Optical chips do not obey the logic-chip shrink religion. High-speed lasers and detectors are built on InP and GaAs with feature sizes in the 0.13 to 0.5 micron range. Silicon photonics uses mature CMOS nodes, typically 45 to 130nm. The small DSP in every high-speed module, however, sits on 7nm or 5nm foundry capacity from TSMC, supplied mostly by Broadcom and Marvell. That creates a hidden leverage point. The optical periphery is exposed to advanced logic constraints through a single low-cost component. If foundry access is weaponized, module delivery speed will not save anyone. Second, the technology gap between Chinese and American players is not uniform. At the module level, Innolight and Eoptolink are at global parity. They ship 800G in volume and are developing 1.6T in parallel. At the laser chip level, the Chinese ecosystem remains roughly one to two product generations behind Lumentum and Coherent. Using the EML progression as a ruler, the gap looks like 25G to 50G to 100G to 200G. At the co-packaged optics level, or CPO, the gap is two to three years. The market often describes this as one Chinese lag story. It is actually a layered map of bottlenecks, and each layer has a different margin structure. Third, profit pools are not proportional to value share. High-end optical chips capture 30 to 40% of industry value. Module assembly captures 40 to 50%. Passive components capture 10 to 15%. Yet the margin order is different. Lumentum and Coherent earn 45 to 60% gross margin on high-end chips. Innolight and Eoptolink earn 30 to 35% at the module layer. TFC earns more than 40% gross margin and more than 25% net margin on passive components. Investors dismiss passives because the products sound boring. That is a mistake. High reliability, high consistency, and manufacturing scale form a quiet moat. I learned this during my due diligence years: the least glamorous layer of a complex system often has the most durable pricing power. The demand side strengthens the core thesis. A single H100 GPU typically requires eight optical modules. A GB200 NVL72 scale rack drives hundreds. This explains why module makers are running at 85 to 95% utilization. Capacity is the constraint, not demand. The shortage is concentrated in three places: high-speed EML and CW lasers, DSP allocation on advanced foundry nodes, and precision coupling equipment. The market sees sold-out module orders and assumes the winner is the module maker. The more important insight is that the upstream chip bottleneck is expanding more slowly. An ETF that gives 15% plus weights to Lumentum and Coherent is effectively harvesting that bottleneck premium. There is a cyclical layer too. Optical modules follow a predictable cost-down curve of 15 to 25% per year. 800G transceivers entered at roughly $1,800 to $2,500 and settled to $800 to $1,200. 1.6T will enter near $2,500 to $4,000. Price erosion is not the problem, because product mix upgrades offset it. The margin question is not quarterly ASP pressure. It is whether 1.6T qualification arrives at hyperscalers on time. That is the signal I am watching in the second half of 2025. Capacity and capital spending tell the same story from a different angle. Module makers spend only 10 to 15% of revenue on capex. Chip makers spend 15 to 25%. Equipment lead times are six to twelve months, not ASML levels. MOCVD epitaxy tools have Chinese alternatives from AMEC and Naura. This matters. It means the industry can respond to demand faster than logic, but it also means the barrier to entry is lower at the module layer. The true moat is not in the cleanroom. It is in customer qualification cycles and supply-chain relationships. A module that passes a hyperscaler 1.6T qualification cannot be replaced in a quarter. R&D spending tells another layer of the story. Coherent and Lumentum each spend $500 to $800 million annually. Chinese module leaders spend $100 to $200 million. At first glance, that looks like an overwhelming American advantage. It is not. The two ecosystems are playing different games. The American names invest in materials science and laser physics. The Chinese names invest in application engineering and rapid iteration. Both can be rational strategies. The risk is when investors assume that R&D dollars map linearly to market share. They do not. The interaction between chip physics and system execution is the variable that matters. Contrarian Angle The popular story is a geopolitical battle between American chip design and Chinese manufacturing. The uncomfortable truth is that the unhedged risks sit elsewhere. Consider gallium and germanium. China restricted exports of these materials in 2023. InP and GaAs epitaxy depend on gallium. If those restrictions escalate, the tax lands directly on the highest-margin layer of the chain: Lumentum and Coherent. At the same time, Japan controls most of the high-purity InP substrate supply. Chinese substrate alternatives remain in the validation stage. The market treats input materials as a footnote. In a complex system, the silent monopoly is usually the one that is not in the headline. The second contrarian point is about Thailand. Media coverage calls it tariff avoidance. It is something bigger. Zhongji Innolight, Eoptolink, and TFC have all invested in Thai manufacturing capacity. Once North American orders can be fulfilled from Thailand, a Chinese optical company stops being a Chinese exporter and becomes a multinational manufacturer. That changes its exposure to export controls, its customer conversation, and potentially its valuation multiple. The geopolitical risk premium embedded in Chinese optical stocks may be overstated. The third point is the uncomfortable one. LYTE concentration is not a flaw if cloud capex plans are real. It is a flaw if the 57% growth number is a forecast rather than a hard budget line. Leading cloud providers have guided more than $300 billion of capital expenditure for 2025. Optical interconnect is roughly 5 to 8% of AI server bill of materials. The math is compelling. The execution risk is enormous. The ETF is a clear-eyed expression of that execution risk. History does not repeat, but it rhymes. In 2022, the market learned that Terra algorithmic stability was not a law of science. In 2025, the optical buy-side must learn that CSP capital plans are not contractual obligations. Export control analysis reinforces the point. Optical chips are not yet in the same restricted category as advanced logic. No EUV tools are required. No 16nm line is at stake. The critical exposure is the DSP and its access to leading-edge foundry capacity. If Washington restricts DSP exports to China, module makers face a short-term production pause. But there is a reciprocal cost: the United States would lose access to the largest module manufacturing capacity in the world. That is why I rate full decoupling at around 15% probability, partial DSP tightening at 40%, and the status quo at 45%. The baseline world is not a clean war. It is a negotiated gray zone in which both sides need each other more than they admit. The structure of LYTE is itself a hedge. If the United States wins the photonics race, Lumentum and Coherent capture the value. If China scale and execution win, Innolight, Eoptolink, and TFC dominate the module layer. The fund monetizes whichever side of the Pacific holds pricing power. That is intellectually neat, but it does not hedge the common factor: hyperscaler capital expenditure. If the AI buildout pauses, all five holdings fall together. This is the single most important risk in the ETF, and it is not diversified away by geography. It is only diversified by time. Takeaway Volatility is the fee for admission to the future. That sentence has guided my position sizing since the 2022 collapse, and it applies to LYTE. This is not a diversified optical fund; it is a disciplined view. The view works if 1.6T volumes arrive on schedule, CPO remains a 2027 story, and hyperscaler capex stays rigid. It fails if AI infrastructure spending pauses, if DSP supply is weaponized, or if hyperscalers vertically integrate and squeeze the module layer. This cycle is not about predicting the direction. It is about positioning ahead of consensus. The next signal is not the price of optical modules. It is the qualification timing of 1.6T at each hyperscaler. Volume follows qualification, not conference calls. And risk is not what you don't know. It is what you know that turns out to be wrong. Watch the qualification data. Everything else is beta.