Silicon cannot emit light — indium phosphide is the chokepoint CPO relocates but cannot remove
inp-is-the-laser-chokepoint · conviction medium · status open · horizon 2026-2028 (substrate balance called for 2H28) · as of 2026-08-03
The bear case on laser vendors is that co-packaged optics collapses laser content: a sell-side BoM puts 91% of silicon photonics' cost advantage in replacing a $160 EML with a $16 CW laser. The physics says the laser cannot be removed at all — silicon's indirect bandgap cannot generate light, so every mass-produced SiPh platform integrates an InP source. The open question is not whether lasers survive but whether FEWER, HARDER, SCARCER lasers is a bigger or smaller revenue line, and no source in the desk's corpus states both terms.
Robust to undisclosed shares. 1 derived input under this thesis; redrawing every supply weight the industry does not publish moves none of them by more than 25%. Computed from evidence at most 17 days old (oldest input: lumentum).
Exhibits
Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.
Series available as data/inp-is-the-laser-chokepoint.csv
Exhibit 2What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.
The chokepoint is physical and sits upstream of every optical architecture
86% if the 3 gates are independent, 90% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Silicon Photonics at 0.90 — Guides, splits and modulates light; cannot generate it. The bound on how far silicon absorbs the laser.
The narrow point is the substrate, and it is export-gated — a policy variable, not a capacity one
64% if the 3 gates are independent, 75% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: Visual Photonics Epitaxy (VPEC) at 0.75 — Merchant MOCVD epitaxy between substrate and laser fab; added to make the chain traceable. Reported expanding InP MOCVD 60 to 64 units.
The second term is now measured, and it points down: laser revenue per unit of bandwidth falls
61% if the 3 gates are independent, 80% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: External laser source (ELS) at 0.80 — The architecture producing the reduction — one high-power CW feeding eight photonic engines instead of a laser per lane per module.
Therefore the position is a TRANSITION trade on scarcity, not an ownership trade on content
63% if the 3 gates are independent, 70% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.
Weakest link: External laser source (ELS) at 0.70 — The CPO architecture that keeps lasers in the BoM. ~8% of Rubin Ultra scale-up content — present, but a small slice.
The variant
Consensus
Co-packaged optics and silicon photonics commoditise the optics bill of materials. SiPh carries a 32% BoM advantage at 1.6T and a 20% price advantage while earning a HIGHER gross margin, so the merchant laser vendors are being disintermediated by a semiconductor process — the familiar pattern where a discrete component gets absorbed into silicon.
Variant
The absorption is bounded by physics, not by economics. Silicon has an INDIRECT bandgap and cannot efficiently convert electricity into photons; indium phosphide's direct ~1.34 eV bandgap can. Every mass-produced silicon-photonics platform — NVIDIA, Broadcom, Marvell, Cisco — integrates an InP laser somewhere in the package, and CPO changes the laser's position and mounting rather than removing it from the BoM. Meanwhile the CPO-grade parts are HIGH-POWER CW lasers and external-laser-source modules that are harder to manufacture than the EMLs inside pluggables, Lumentum runs five InP fabs and still ships >30% below demand, and NVIDIA committed $2bn across BOTH suppliers in March 2026 with purchase guarantees and capacity rights. You do not buy capacity rights in a component whose value is collapsing.
Differentiator
The desk went and got the second term, which neither side of the public argument states. NVIDIA discloses 18 ELSFP modules for the entire 144-port Quantum-X switch, each feeding eight 1.6Tb/s engines, and claims ~4x fewer lasers per unit bandwidth. An 800G pluggable runs 4-8 lasers per port; CPO at 4x fewer runs 1-2. Stack that on the sell-side BoM EML $160 vs CW $16 and laser revenue per port goes from $640-1,280 to $16-32. Break-even needs a CPO-grade CW at 4-8x an EML, and harder-to-manufacture is not 4-8x. Even ~5.7x port growth does not offset it. So the desk holds BOTH sides precisely: the chokepoint is structural and physics-bounded, the current shortage is real enough that NVIDIA paid $2bn for capacity rights, AND the terminal laser revenue line is smaller. That is a TRANSITION trade, not an ownership trade, and almost nobody frames it that way.
Falsifiers
claim: CPO-grade CW lasers price high enough to defend revenue · criterion: A disclosed or teardown-derived ASP for an ELSFP-class high-power CW laser above $320 — more than 2x an EML — which would materially narrow the per-port revenue collapse computed here · horizon: 2028-06-30 · settles: refuted
claim: Laser revenue per unit of optical bandwidth falls despite scarcity · criterion: Combined Lumentum + Coherent datacom laser revenue grows slower than deployed 800G-equivalent optical ports over any four consecutive quarters through 2028 · horizon: 2028-06-30 · settles: confirmed
claim: Silicon absorbs the light source after all · criterion: Any mass-produced silicon-photonics platform ships at volume with NO indium-phosphide laser in the package — a monolithic silicon or germanium-on-silicon emitter qualified for datacom · horizon: 2028-12-31 · settles: refuted
claim: The substrate constraint resolves and takes the scarcity premium with it · criterion: Chinese InP substrate export licensing is lifted or routinely granted AND Lumentum reports shipping to demand, before the 2H28 balance point the sell-side calls · horizon: 2028-06-30 · settles: refuted
claim: SCOPE — structure right, no consequence · criterion: LITE and COHR track the optical complex within benchmark dispersion through the horizon regardless of how the laser-content question resolves · horizon: 2028-12-31 · settles: refuted
Open questions
CLOSED 2026-08-03, with three caveats that bound it. NVIDIA discloses 18 ELSFP modules for a 144-port Quantum-X switch and claims ~4x fewer lasers per unit bandwidth. CAVEAT ONE: an ELSFP is a MODULE containing multiple CW dies, so 18 modules is not 18 lasers — the usable figure is the 4x-per-bandwidth claim, which is NVIDIA's own marketing number and is not independently verified. CAVEAT TWO: the $16 CW price comes from a sell-side BoM and may describe a lower-power part than an ELSFP-grade high-power CW feeding eight engines; if the real part is 2-3x that, the collapse is smaller but still a collapse. CAVEAT THREE: ELS modules are PLUGGABLE and field-replaceable, so they may carry replacement revenue a soldered laser would not — the desk has no failure-rate or replacement-cycle figure, and that is the residual unknown.
The 4-8 lasers per 800G pluggable figure is a lane-count INFERENCE (4x200G or 8x100G), not a disclosed BoM. It is the least-sourced number in the chain and the conclusion is proportionally sensitive to it.
Two sources point opposite ways on substrate tightness within days — a CEO describing supply tighter than DRAM/NAND, and a channel note reporting VPEC 3Q revenue +23% QoQ on IMPROVED substrate supply. Not strictly contradictory (a laser fab can be short while its input eases) but the desk must not hold a shortage premium and an easing thesis simultaneously.
High-power CW and ELS modules are described as harder to manufacture than EMLs. If true that supports price mix, but the desk has no ASP for a CPO-grade CW laser — only the $16 figure from a BoM table that may describe a different part.
Reasoning chain
The chokepoint is physical and sits upstream of every optical architectureVALID
premises
Indium phosphide (InP)0.95 strong
REACTIVE from derived:indium-phosphide.demand_pull (today 69.0, band 27.59-110.37 -> p 0.8-0.98).
why
Pull through the InP layer is the scarcity term. It is deliberately wired to the TRANSITION half of the thesis — the half that says the shortage is real now — not to the terminal-revenue half, which resolves on CPO penetration and cannot be read from a demand cell. Anchors are centred on today's reading so this makes the premise LIVE without re-rating the thesis. Direct ~1.34 eV bandgap; the material every datacenter laser is built on. Physics, not a market view.
Wired 2026-08-03 to close a structural gap — the desk was arguing about laser value with no node connecting substrate to laser maker.
Silicon Photonics0.90 strong
Guides, splits and modulates light; cannot generate it.
why
The bound on how far silicon absorbs the laser.
Silicon photonics is a semiconductor business for the waveguides and a NON-semiconductor business for the light. That asymmetry is physical and does not resolve with process maturity, which is why the laser layer keeps reappearing as a constraint however packaging evolves.
The narrow point is the substrate, and it is export-gated — a policy variable, not a capacity oneVALID
premises
AXT (Tongmei)0.85 strong
Chinese InP/GaAs substrate supply, sitting under Chinese export licensing.
The substrate step, the narrowest point in the chain.
Visual Photonics Epitaxy (VPEC)0.75 strong
Merchant MOCVD epitaxy between substrate and laser fab; added to make the chain traceable.
why
Reported expanding InP MOCVD 60 to 64 units.
Substrate -> epitaxy -> laser chip -> module. Controls apply at the substrate, so laser supply is licensable rather than buildable — a different risk class, and not diversifiable by adding fabs.
The second term is now measured, and it points down: laser revenue per unit of bandwidth fallsVALID
premises
Co-packaged optics0.85 strong
NVIDIA Quantum-X Photonics: 144 ports at 800Gb/s, 115.2 Tb/s, 24 optical engines on TSMC COUPE, and only 18 ELSFP external-laser modules for the whole switch.
External laser source (ELS)0.80 strong
The architecture producing the reduction — one high-power CW feeding eight photonic engines instead of a laser per lane per module.
The pluggable baseline: 4-8 lasers per 800G port depending on lane rate.
NVIDIA claims ~4x fewer lasers per unit bandwidth. At 4-8 lasers/port for pluggables and 1-2 for CPO, combined with EML $160 vs CW $16, laser revenue per port falls from $640-1,280 to $16-32. Holding revenue flat needs a CPO-grade CW at 4-8x an EML price, which harder-to-manufacture does not plausibly reach. Port growth of ~5.7x over the roadmap does not offset it.
Therefore the position is a TRANSITION trade on scarcity, not an ownership trade on contentVALID
premises
External laser source (ELS)0.70 moderate
8%
The CPO architecture that keeps lasers in the BoM.
why
~8% of Rubin Ultra scale-up content — present, but a small slice.
ELS is how light reaches a co-packaged engine without sitting on the heat source.
Lumentum Holdings0.90 strong
30%
Five InP fabs, shipping >30% below demand per its own CEO; NVIDIA capacity commitment.
Fewer lasers per port at higher difficulty and under supply scarcity. Whether that is a larger or smaller revenue line requires units AND price, and only price is in evidence. The desk holds the structure and declines the direction.
Sources
InP direct ~1.34 eV bandgap vs silicon indirect; all mass-produced SiPh platforms integrate an InP laser; CPO relocates rather than removes it; five InP fabs shipping >30% below demand; NVIDIA $2bn across Lumentum and Coherent, March 2026, with purchase guarantees and capacity rights — linkacc 2026-08-03