The market is treating artificial intelligence as a GPU story. The physical constraint is a light story. Every NVIDIA accelerator shipped this year required between eight and sixteen optical modules to communicate with its cluster — eight for the H100 generation, up to sixteen in GB200 systems — and as hyperscale clusters push past ten thousand accelerators, that dependency ratio keeps climbing. Yet the investment world has barely begun to read the supply chain that carries those signals.
Enter the Roundhill Photonics and Optical ETF (LYTE): five names spanning the optical interconnect chain — Lumentum and Coherent on the American side; Zhongji Innolight, Eoptolink, and TFC Communication on the Chinese side. Five holdings, roughly two-thirds of the fund, deliberately arranged. Most observers file this under "semiconductor ETF" and move on. That is a category error. This is a structural bet on AI's connective tissue, and the way these holdings are arranged reveals more about the next two years of AI infrastructure than any GPU roadmap.
LYTE holds no foundry, no logic design house, no memory maker. It is not a full-chain technology ETF; it is a key-node selection, deliberately skipping foundries and DSP design houses in favor of the optical layer's dual poles: chip design and module assembly. Upstream, Lumentum and Coherent fabricate high-speed optical chips — InP-based lasers, modulators, photodetectors — the highest-value layer in the stack, with gross margins of 45-60%. Midstream, Zhongji Innolight and Eoptolink assemble the 800G and 1.6T modules that dominate AI data centers, working at 30-35% gross margins while running at 85-95% utilization, a pace most semiconductor fabs cannot sustainably match. On the passive side, TFC Communication supplies fiber arrays, isolators, and connectors — the precision hardware every module vendor needs regardless of architecture.
The demand math is worth precision. The four largest US cloud providers are expected to spend more than $300 billion on capital expenditures in 2025, with optical interconnect consuming 5-8% of that total. Industry projections see AI optical modules growing 57% this year to $26 billion. The supply math is less appreciated: module capacity can be added within three to six months, while optical chip capacity — epitaxy, wafer processing, qualification — takes twelve to eighteen months. That asymmetry, not the GPU roadmap, is the investment thesis. The industry's physical limit over the next two years is not transistor count; it is the rate at which lasers can be grown, cleaved, and qualified.
When I audited the OpenYield protocol in 2020 and traced a critical reentrancy vulnerability in its flash loan module, the lesson was never about the code itself. It was about dependencies. A smart contract is only as sound as every external call it makes; a supply chain is only as sound as every supplier it cannot replace. Reading LYTE, I applied the same discipline. The dependency chain runs: InP substrates from Japan's Sumitomo and JX Metals; MOCVD epitaxy tools from a handful of global suppliers; DSP controllers from Broadcom and Marvell on Taiwanese advanced nodes; the lasers themselves from Coherent and Lumentum. Break any link, and AI infrastructure stalls.
The portfolio maps a split-brain supply chain. Chinese module makers lead the world in assembly and integration — Zhongji Innolight holds roughly a quarter of the data center module market as the global number one. American firms command the high-end chip layer — Coherent controls about 35% of the high-speed EML market, Lumentum another 25%. Neither jurisdiction can build a modern AI data center alone. The strategic position between Washington and Beijing is not decoupling; it is mutual hostage-taking. Each side holds an essential input the other cannot quickly replace.
High-end optical chips, not modules, are the binding constraint through 2026. The Chinese module leaders can scale output dramatically in under six months — they proved it with 800G. The lasers inside those modules cannot be scaled on the same timeline. Epitaxial growth is slow, consistency is harder, qualification is unforgiving. The Chinese ecosystem is closing the gap — measured along the 25G-to-200G EML evolution, the lag is one to two generations — with full self-sufficiency at the 100G EML tier expected around 2027-2028. The 200G tier and the co-packaged optics ecosystem remain American strengths for at least two to three more years.
The technology fork matters more than speed. The industry is splitting between traditional EML-based modules and silicon photonics, where modulation and detection are integrated onto CMOS-compatible wafers at mature 45-130nm processes. The Chinese names in this fund are among the most aggressive silicon photonics adopters in the world, and TFC's fiber arrays are precisely the component silicon photonics consumes in huge volumes. The hidden wager inside LYTE is that silicon photonics adoption accelerates faster than the incumbents' laser roadmap. If that happens, value shifts toward the integration layer — where the Chinese names sit — and pricing power migrates away from the high-margin chip monoliths.
The R&D comparison looks brutal at first glance. Lumentum spends roughly 18% of revenue on research; Coherent, 12-15%; the Chinese module makers, 4-7%. But the nature of the spending differs. US firms invest in fundamental materials physics — epitaxial structures and laser reliability. Chinese firms invest in application engineering and manufacturing iteration — the rapid yield learning that wins the assembly layer. At the module level, the competition is speed and cost. At the chip level, it is depth and patents. Both models succeed, in different layers. The ETF simply holds both sides of that split.
TFC Communication deserves the least glamorous and most profitable label in this fund. Its 40%+ gross margins and 25%+ net margins outperform most active module vendors. Passive components demand extreme consistency: a single misaligned fiber array can fail an entire 1.6T module's optical alignment test. Reliability at scale is a moat that needs no roadmap. In a gold rush, the company selling precision tools to every miner at every mine is the quietest equity on the board.
Pricing tells the rest of the margin story. Optical modules follow a 15-25% annual cost-down curve — 800G opened above $1,800 and has settled into the $800-1,200 band. The economics reset at each generation: 1.6T modules are entering at $2,500-4,000, and early-cycle margins on new products are the dominant earnings variable through 2026. This is not a price-war market; it is a technology-upgrade market. The winners are those who time the generation shift. Optical markets have broken companies before — the 2022-2023 telecom inventory bust is recent enough to remember — but the AI cycle carries structural demand visibility into 2027. "Structurally different" has been said at every cycle peak. Sometimes it is true.
No dependency audit is honest without listing the weakest links. DSP chips represent more than 90% import dependency for Chinese module makers, and every module still routes through Broadcom or Marvell parts on Taiwan's advanced nodes. InP substrates are over 80% supplied by Japan. China's gallium export controls, in place since 2023, add direct cost pressure to Lumentum and Coherent's epitaxy operations. And the next packaging frontier — co-packaged optics, where optical engines are co-packaged with switching silicon — depends on CoWoS-class advanced packaging, which is itself the industry's most constrained resource. Module plants reach manufacturing break-even at 55-60% utilization against roughly 80% for advanced logic fabs, so the sector absorbs overcapacity corrections more gracefully. That does not immunize it against the DSP bottleneck. The chain holds multiple single points of failure, and LYTE's concentration sits directly on top of every one of them.
Now the uncomfortable part. The 57% growth forecast embeds an assumption that US cloud capital expenditure commitments will be executed rigidly regardless of economic conditions or fading AI ROI enthusiasm. That word — commitment — is doing enormous work. If capex wobbles, the portfolio's concentration becomes its liability. I have watched this pattern before. In DeFi, the industry manufactured "liquidity fragmentation" as a problem warranting a wave of new products — a narrative doing the work of demand creation and, predictably, enriching the creators. The "deterministic growth" story around optical interconnect is similar in kind: a real tailwind elevated into a justification for concentration, and for dismissing an industry's cyclical history.
The geopolitical layer compounds the fragility. The Chinese module makers have quietly built Thai factories as a pre-hedge against export controls — a strategic move transforming them from Chinese exporters into multinational producers. But that hedge has limits. If Washington extends the TikTok logic to AI infrastructure supply chains — a "data security" exclusion on Chinese-controlled hardware — this portfolio faces structural repricing no matter where the assembly lines sit. LYTE is a dual-country bet designed to profit whichever side wins. Elegant in theory. Uncomfortable in practice, because in a real decoupling both sides lose.
Code is law, but humans are the protocol. Supply chains behave like protocols — precise, layered, and dependent on the humans who maintain them. And like any protocol, they can be forked.
The next 24 months will test whether the optical layer can scale faster than AI's appetite. Watch three signals: the 1.6T revenue mix at the Chinese module leaders; the ramp of the Thai production lines; and whether co-packaged optics moves from demonstration to deployment — remembering that advanced packaging capacity, not optical engineering, may govern that transition. We built trust in the chaos of supply-chain shocks, not despite it; the factories rising quietly in Texas and Thailand are the physical form of that trust. Education is the antidote to exploitation, and reading this ETF correctly — as a concentrated expression of AI's physical limits rather than a diversified technology fund — is the beginning of that education. Hold through the noise. The signal is in the light.