The Photon Is the Protocol: What LYTE's Concentrated Bet Reveals About AI's Physical Layer

Over the past seven days, I have been dissecting the Roundhill Photonics and Optical ETF (LYTE) — a fund that compresses an entire industry into five names. The top positions: Lumentum, Coherent, Zhongji Innolight, Eoptolink, and Tianfu Communication. Together they represent roughly 67% of the portfolio. This is not diversification. It is a thesis. And buried inside that thesis is a story about how the physical layer of artificial intelligence is being built, who controls the bottlenecks, and what happens when a promised demand curve bends. I spent three weeks in 2017 auditing the relayer architecture of a decentralized exchange instead of chasing an ICO token sale. That choice taught me a discipline I have carried into every analysis since: architecture determines outcomes long before prices reflect them. Trade flow follows structure; structure follows incentive; incentive follows physics. The same hierarchy applies to optical interconnect. To understand where the industry is heading, we follow the photons. LYTE is not a blockchain protocol, but its construction follows protocol logic. The ETF maps the full vertical of optical system design: upstream chip and material science (Lumentum and Coherent, whose InP and GaAs laser manufacturing defines the high end), midstream module integration (Zhongji Innolight and Eoptolink, assembling 800G and 1.6T transceivers for the world's hyperscalers), and downstream passive components (Tianfu Communication, whose fiber arrays and isolators quietly enable the entire system). Notably absent are the DSP designers — Broadcom, Marvell, Cisco's Acacia. The ETF excludes photonic-electronic SoCs entirely. That boundary is a statement: pure optical exposure, not integrated hybrid systems. It also means the fund forfeits the value accruing to the logic layer of optical interconnect, which may prove to be the more consequential omission. The process geometry is revealing. Optical chips run on compound semiconductor processes at 0.13 to 0.5 micron — unfathomably ancient by logic-chip standards, yet unforgiving in epitaxial precision. Silicon photonics runs on mature CMOS nodes anywhere from 45 to 130 nanometers. The DSPs inside every high-end module, by contrast, demand TSMC's 7nm or 5nm. Three layers, three vastly different physics regimes, three different scarcity characteristics. The economics follow accordingly. Optical chips capture 30–40% of the profit pool with gross margins of 45–60%. Module packaging captures 40–50% but at thinner 25–35% margins. Passive components deliver above 40% gross margins at scale because consistency becomes a moat in itself — Tianfu's profitability, at 40% gross and 25% net margins, exceeds most active module makers. Layer by layer, this mirrors a well-designed protocol stack: margin concentrates at the layer where scarcity is hardest to manufacture. This is where the popular narrative starts to crack. The market treats optical modules as the growth story, and for good reason — module leaders are shipping at 85–95% capacity utilization. But the module makers occupy what I call the "sandwich position." Upstream, they depend on Broadcom and Marvell for DSPs — a duopoly with over 90% share. Upstream again, they purchase high-speed EML lasers from Lumentum and Coherent, the same firms that partially compete at the system level. Downstream, their five largest customers — Microsoft, Google, Meta, Amazon, NVIDIA — command procurement leverage that keeps pricing disciplined. A 30–35% gross margin at Chinese module leaders reflects operational intensity, not structural bargaining power. The annual cost-down curve compounds the pressure: 25–35% price erosion per generation, from $1,800–2,500 per 800G unit at introduction down to $800–1,200 at maturity. Volume must double just to hold revenue flat. Based on my audit experience across decentralized systems, I have learned to find the point where consensus fails. In optical interconnect, that point is the laser chip. The industry is mid-transition from 800G to 1.6T, with 2025 seeing initial 1.6T shipments and 2026 expected to bring volume. But the binding constraint is not module assembly — capacity there scales in three to six months, a remarkably short cycle next to the 12–24 months of a logic fab. The constraint sits upstream, where 200G EML lasers hold yield rates of 60–70% at the best-run fabs. A five-point yield shift changes an entire quarter's margins. Chinese chip makers trail by one to two generations — the distance between 100G and 200G EML maturity — which is why Coherent commands roughly 35% and Lumentum another 25% of the high-speed optical chip market. They own the yield, and yield is the physical form of trust. The DSP layer is more fragile still. Over 90% of high-end PAM4 DSPs run on TSMC advanced nodes, placing optical interconnect in direct competition with AI accelerators for the same foundry capacity. Module revenue growth is therefore coupled to an external allocation decision made inside TSMC and NVIDIA — a dependency the market has not fully priced. A disruption in either the laser chip or the DSP layer stalls the entire AI buildout. Chinese substitution in DSP is a long horizon: the design tools, IP ecosystem, and advanced-node access form a triple constraint that will take longer than a decade to unwind. This is the quiet vulnerability of the entire stack, and it sits entirely outside the ETF's holdings. Module-layer dynamics tell a different story. Zhongji Innolight holds roughly 25% of the global datacom module market, and Chinese suppliers combined provide over half of the world's optical modules. Their R&D intensity runs 5–7% of revenue versus Lumentum's 18% — but their iteration speed and manufacturing elasticity have made them indispensable. When hyperscalers placed urgent 800G orders in 2024, only factories in Suzhou, Chengdu, and Thailand could scale within a quarter. "China speed" is a structural asymmetry that American competitors have not replicated, despite their chip-level depth. The winner in each optical generation is whoever reaches volume fastest, because the cost-down curve punishes laggards relentlessly. The geopolitical layer complicates every conclusion. Lumentum and Coherent are US-based; Zhongji Innolight, Eoptolink, and Tianfu are Chinese, though none sits on the BIS Entity List as of this writing. The ETF's split — over 30% to American incumbents, roughly 37% to Chinese leaders — reads as a deliberate geographic hedge, ethically ambiguous yet practically rational. But the hedge shares one point of failure: every layer depends on hyperscaler capex — projected to exceed $300 billion in 2025 across the top four CSPs — being executed as planned. The 57% projected growth in AI optical modules, reaching $26 billion, is a derived number. It follows corporate commitments, not protocol guarantees. If AI returns underwhelm for two consecutive quarters, every layer of this ETF falls in sync. There is no uncorrelated asset in the portfolio. The hidden variable is the Thailand shift. Zhongji Innolight, Eoptolink, and Tianfu have all established Thai manufacturing bases, repositioning themselves as multinational suppliers. The public narrative frames this as cost optimization. It is pre-emptive hedging against export controls and data-security procurement restrictions — the same logic that animated the TikTok divestment. If US regulators pressure hyperscalers to exclude perceived China-controlled supply chains, the Thai plants become the survival mechanism. The industry is building redundancy because the possibility of decoupling is now priced into corporate strategy, even where it has not yet materialized in regulation. This is a deeper transformation than any quarterly earnings release has acknowledged. The consensus view holds that Chinese module makers close the high-speed chip gap by 2027–2028 and achieve self-sufficiency in EML and CW lasers. My read diverges. Silicon photonics — integrating lasers and modulators onto CMOS-compatible silicon — represents a protocol-level inversion of industry power. As silicon photonics scales, value migrates from exotic InP epitaxy toward design integration and co-packaging collaboration with switch-chip vendors. The dominant players become those co-packaging optical engines with switching silicon (CPO), not merely selling discrete lasers. CPO will not reach meaningful commercial scale before 2027, but the directional signal is unambiguous — and it runs against the module-layer incumbents' current moat. Chinese leaders face a delicate trap: their module-layer dominance could be commoditized by the very transition that elevates Western chip and switching companies. Alternatively, their embedded relationships at hyperscalers could make them essential integration partners. The next 18 months will determine the scenario. Trust is not given; it is verified. The protocol remembers what the market forgets: that every exponential architecture sits on a physical substrate with yield rates, expansion cycles, and geopolitical fault lines. The firms that survive the next downturn will hold either proprietary chip design at the high end or manufacturing elasticity that absorbs demand shocks at speed. LYTE captures both sides of that divide — but only if its core assumption holds. Watch the yields at the 200G EML fabs. Watch the Thai factory utilization curves. Watch whether NVIDIA's GB200 ramp translates into optical orders at the pace the market has priced. The signals are visible from here. Stillness reveals the signal beneath the noise, and patience is the validator of true intent.