Supply-chain intelligence · directional estimates
Data-center optics supply chain
A bottom-up map from compound-semiconductor wafers and lasers through DSPs, fiber attach, transceivers, switches and hyperscale AI clusters—built to show where revenue, cost and operating leverage actually accrue.
How to read this: percentages are analyst estimates for a representative AI/data-center dollar, not reported segment disclosures. They are intended to map economic exposure and flow-through, not to forecast any single company with precision. Hardware mix, integration, geography, inventory timing and customer concentration can move the ranges materially.
1. Executive map: where the optics dollar goes
Power & materialsInP/GaAs, silicon, wafers, epitaxy, specialty glass
~8–14% of system cost
~8–14% of system cost
Photonics devicesLasers, modulators, photodiodes, VCSELs, driver ICs
~18–28%
~18–28%
PackagingCo-packaging, fiber attach, lenses, isolators, thermal
~12–20%
~12–20%
TransceiversDSP, TIA, gearbox, PCB, housing, test
~30–45%
~30–45%
System integrationSwitch/accelerator assembly, optics qualification, spares
~10–20%
~10–20%
$100
Illustrative optical interconnect dollar at the data-center system level
30–45%
Transceiver / active-module share of optical BOM
15–30%
Typical gross margin range: scaled module makers
2–4×
Potential revenue multiplier from 800G→1.6T port upgrades
2. Layer-by-layer supply chain and economics
| Layer | What is supplied | Representative players / exposure | Share of optical BOM | Cost / margin logic | Flow-through signal |
|---|---|---|---|---|---|
| Substrates & epitaxy | InP/GaAs wafers, epitaxial growth, silicon photonics wafers, SiN, specialty glass | Sumitomo Electric, Freiberger, IQE, IntelliEPI, Shin-Etsu / Coherent ecosystem | 8–14% | Material-heavy; yield, wafer size and epitaxy utilization drive cost. Scarcity can create strong pricing. | Lead-time / yield wafer starts, epi utilization, substrate pricing |
| Light sources | EML/DML lasers, CW lasers, VCSELs, pump lasers, gain chips | Coherent, Lumentum, Broadcom, Innolight, Mitsubishi Electric, II-VI heritage | 12–20% | High R&D and qualification content; laser yield and burn-in are key. Mix shifts to EML/CPO can change economics. | Laser ASP 800G/1.6T mix, wavelength count, yield |
| Modulation & detection | Modulators, photodiodes, balanced receivers, TIAs, drivers | Broadcom, MACOM, Coherent, Lumentum, Marvell ecosystem | 8–14% | Analog performance and process know-how support high value-add; defect rates can be painful. | Content / ASP baud rate, lane count, PAM4 adoption |
| DSP & electrical engine | PAM4 DSP, gearbox, retimers, FEC, driver/TIA electronics | Broadcom, Marvell, Credo, MaxLinear, MACOM | 15–25% | Semiconductor gross margins are high, but leading-edge silicon and packaging consume significant cost. DSP may be bypassed in some CPO architectures. | Semis leverage DSP ASP × port count; CPO displacement risk |
| Passive optical assembly | Fiber arrays, lenses, isolators, mux/demux, optical benches, connectors | Senko, US Conec, Corning, Fujikura, Sumitomo Electric, Molex | 12–20% | Precision assembly and automation matter more than raw material. Fiber attach is a bottleneck at scale. | Capacity active ports, attach yield, automation capex |
| Module manufacturing | Transceiver PCB, housing, thermal design, firmware, calibration, test | Coherent, Lumentum, Fabrinet, Innolight, Eoptolink, Hisense, Source Photonics | 20–30% | Assembly/test is labor- and capex-intensive; contract manufacturers capture volume with lower margin. | Volume / mix shipments × ASP; qualification and RMA |
| Switch / system integration | Optical ports in switches, NICs, accelerators; service, spares and qualification | Arista, Cisco, Nvidia, Broadcom, Celestica, Wiwynn, Dell | 10–20% of system value | System vendors capture integration and software value; optics may be pass-through or bundled. | Ports switch ports × optics per port × attach rate |
3. Revenue and cost waterfall
| Representative 800G/1.6T module dollar | Revenue share | Cost share |
|---|---|---|
| Optical components & lasers | 20–30% | 25–35% |
| DSP / driver / receiver electronics | 20–30% | 20–28% |
| Packaging, fiber attach & passives | 15–22% | 15–22% |
| Assembly, calibration & test | 15–22% | 12–18% |
| R&D, warranty, overhead & profit | 18–30% | — |
A module maker's revenue is not equal to industry value-add: components can be purchased and resold, while contract manufacturers report revenue with thinner margins.
4. What drives earnings flow-through?
- Bandwidth transition: 400G→800G→1.6T increases dollars per port, but content may be split across more lanes and components.
- AI scale-out: GPU clusters require dense east-west bandwidth; switch and NIC port count is often the cleanest demand proxy.
- Co-packaged optics: can move value from pluggable modules to switch/package suppliers; lower power and reach are traded against serviceability.
- Capacity: fiber attach, EML lasers, advanced packaging and test—not just wafer capacity—can constrain shipments.
- Customer concentration: hyperscaler qualification can produce step-ups, then sharp digestion when deployments pause.
5. KPI dashboard for investors
Upstream
- InP/GaAs wafer starts and epi utilization
- Laser/EML yield and backlog
- Fiber-array and connector lead times
- Photonic packaging capex
Midstream
- 800G/1.6T mix and module ASP
- Gross margin after component inflation
- Transceiver qualification wins
- RMA / field failure rates
Downstream
- Hyperscaler capex and AI cluster networking
- Switch port shipments and attach rates
- Ethernet vs InfiniBand / proprietary fabrics
- Power-per-bit and rack density
6. Key risks and caveats
Optical revenue can grow while profit does not if module ASPs fall faster than volumes rise, if lasers or DSPs are supply-constrained, or if customers dual-source. The most important distinction is content growth per port versus total port growth. CPO, linear-drive optics, silicon photonics and optical I/O are plausible architecture changes, but adoption timing and supplier share are uncertain.
Reference set: company annual reports and earnings materials from Broadcom, Coherent, Lumentum, Marvell, Credo, Fabrinet, Arista, Cisco and Nvidia; OIF / IEEE Ethernet roadmaps; public hyperscaler capex disclosures.
3. Named company-to-company network — optics
The table is intentionally transaction-oriented. “High” means the commercial path is directly visible in filings, product qualification, or a clearly documented supplier/customer relationship; “Medium” means ecosystem or channel inference rather than a disclosed bilateral contract.
| Company | Buys from | Sells to | Product / role | Economic flow-through | Evidence |
|---|---|---|---|---|---|
| Corning | Silica, dopants, energy suppliers | CommScope, OFS, Corning optical connectivity | Fiber, specialty glass | Fiber content rises with port count; pricing and utilization drive margin | High |
| Fujikura | Glass preforms, resins, equipment | NTT, Sumitomo, module/connector makers | Optical fiber, ribbon, harnesses | Fiber volume is a capacity signal; modest material share, high qualification | High |
| Sumitomo Electric | InP/GaAs wafers, chemicals | Broadcom, Lumentum, module integrators | Lasers, photodiodes, modules | Captures component ASP and scarcity premium | High |
| IQE | GaAs/InP substrates, gases | Coherent, Lumentum, Broadcom, VCSEL makers | Epitaxial wafers | Epi utilization flows to revenue before module shipment | High |
| Coherent | IQE/epi, wafers, MACOM/Broadcom ICs | Fabrinet, hyperscaler-qualified module buyers | Lasers, EMLs, transceivers | High component content; mix and yield determine gross margin | High |
| Lumentum | InP wafers, epi, DSP/driver ICs | Cisco, Ciena, Nokia, cloud module channels | Lasers, coherent optics, modules | Telecom + datacom mix; cloud qualification unlocks volume | High |
| Broadcom | TSMC wafers, substrates, EDA/IP | Arista, Cisco, Nvidia, Coherent and module OEMs | DSPs, switch ASICs, drivers | High-margin silicon; every port carries content | High |
| Marvell | TSMC, packaging, EDA | Cloud OEMs, optical module makers, switch vendors | PAM4 DSP, electro-optics, interconnect | DSP ASP × port transition; design wins precede volume | High |
| Credo | TSMC, OSAT, substrates | Arista, cloud ODMs, module makers | AECs, retimers, optical DSPs | Content shifts from copper to active electrical/optical links | High |
| MACOM | Foundries, InP/GaAs epi, packaging | Coherent, Lumentum, transceiver OEMs | RF/analog, laser drivers, optical ICs | Analog content grows with lane speed; pricing sensitive | High |
| Mitsubishi Electric | Compound wafers, packaging materials | Telecom and datacom module OEMs | EML/DML lasers, optical devices | Qualified laser supply can constrain module output | High |
| Hamamatsu | Semiconductor wafers, glass, packaging | Instrument and network equipment OEMs | Photodiodes and detectors | Detector content tracks lane count and performance | High |
| Innolight | Lasers, DSPs, passives, PCB | Nvidia, Arista, hyperscaler ODMs | 800G/1.6T transceivers | Direct AI-networking volume; BOM inflation passes through ASP | Medium |
| Eoptolink | Lasers, DSPs, optics, PCB | Nvidia, Cisco, Arista, ODMs | Datacom transceivers | Shipment volume × ASP; customer concentration matters | Medium |
| Source Photonics | Lasers, DSPs, optical components | Carrier, cloud and equipment OEMs | Optical transceivers | Qualification and mix determine margin; channel inventory risk | High |
| Hisense Broadband | Lasers, DSPs, optical assemblies | Cloud, telecom and equipment OEMs | Pluggable transceivers | Chinese supply base; volume and price competition | Medium |
| Fabrinet | Components from Coherent/Lumentum and IC vendors | Coherent, Lumentum, Cisco, optical OEMs | Contract optical/electronic manufacturing | Revenue tracks customer outsourcing; margin is conversion not component markup | High |
| Celestica | Optical components, PCBs, ASICs | Arista, Nvidia, cloud/telecom OEMs | ODM manufacturing and integration | AI networking ramp flows through manufacturing utilization | High |
| Jabil | Optics, PCBs, semiconductors | Cloud, networking and equipment OEMs | Contract manufacturing | Low-margin volume leverage; inventory timing important | High |
| Sanmina | Optics, boards, connectors | Networking and communications OEMs | Optical/electronic assemblies | Manufacturing revenue follows customer programs | High |
| Molex | Fiber, plastics, copper, metals | Nvidia, hyperscalers, switch/accelerator OEMs | Connectors, cages, fiber assemblies | Rack density increases connector dollars per system | High |
| Amphenol | Metals, polymers, cable and fiber | Dell, Nvidia, Arista, cloud ODMs | High-speed connectors and interconnect | High-margin connector content per rack and port | High |
| TE Connectivity | Metals, plastics, cable, optical components | Cloud, servers, switches, OEMs | Connectors and high-speed interconnect | Port density and power delivery expand content | High |
| US Conec | Ceramics, polymers, fiber | Molex, Amphenol, hyperscale integrators | MPO/MTP connectors | Connector attach is a direct rack-build proxy | High |
| Senko | Fiber, ferrules, polymers | Module makers, cloud and telecom installers | Fiber connectors and assemblies | Precision assembly bottlenecks can delay shipment | High |
| Ayar Labs | Foundry wafers, lasers, packaging | AI accelerator and switch designers | Optical I/O chiplets | Emerging architecture; design-win signal, not mature revenue | Medium |
| Lightmatter | Foundry wafers, packaging, lasers | AI system and switch OEMs | Photonic interconnect / compute products | Longer design cycle; potential DSP/SerDes displacement | Medium |
| Nvidia | TSMC wafers, HBM, Broadcom/Marvell ecosystem, optics | Dell, Supermicro, HPE, AWS, Meta, Microsoft | GPU systems, networking, optical-linked clusters | Optics content is downstream of GPU cluster scale-out | High |
| Arista | Broadcom ASICs, optics from Coherent/Innolight, connectors | Microsoft, Meta, cloud and enterprise buyers | Ethernet switches | Port shipments × optics attach × switch ASP | High |
| Cisco | Broadcom/Marvell ASICs, optics, PCBs | Cloud, enterprise, telecom operators | Switches, routers, coherent systems | Optics bundled in systems; service/software changes margin | High |
| Dell | Nvidia/AMD GPUs, HBM systems, optics, ODM assemblies | Enterprise, cloud, sovereign AI buyers | AI servers and networking | Memory/optics pass through system BOM; demand proxy is backlog | High |
| Supermicro | Nvidia/AMD, HBM, optics, ODM/board suppliers | Cloud and enterprise AI customers | AI servers and racks | Rack integration exposes full networking and optics BOM | High |
| Wiwynn | Nvidia/ASIC systems, optics, boards | Meta, Microsoft, Google and cloud customers | Cloud server/ODm racks | Direct cloud build exposure; low-margin throughput | High |
| Quanta | Processors, memory, optics, boards | Cloud OEMs and hyperscalers | Servers and networking ODM | Rack build and customer capex flow-through | High |
| Wistron | Processors, memory, optics, boards | Cloud and enterprise OEMs | Server/AI manufacturing | Volume and utilization, not component ASP | High |
| Accton | Broadcom ASICs, optics, boards | Arista, cloud/white-box buyers | Switch ODM | Switch port and optics attach signal | High |
| AWS | Nvidia/AMD, Arista/Broadcom ecosystem, ODM racks | Enterprises and developers | Cloud compute/network services | Capex converts to internal optical demand; utilization determines returns | High |
| Microsoft | Nvidia/AMD, Arista/Cisco, optical modules, ODMs | Azure customers | Cloud AI clusters | AI capex is demand trigger; procurement may be multi-tier | High |
| Meta | Nvidia/AMD, Arista/Broadcom, optical suppliers | Users/advertisers | AI and hyperscale network | Large scale-out optical port demand | High |
| TPU supply chain, optical modules, Broadcom/Arista ecosystem | Cloud and products | AI clusters and networking | Custom ASICs change optics and switch mix | High | |
| Oracle | Nvidia/AMD, servers, optics and networking | Cloud customers | OCI AI infrastructure | Cluster deployment drives downstream optics consumption | High |
| Ciena | Coherent/Lumentum, DSPs, fiber and boards | Telecom carriers and cloud interconnect | Optical transport systems | Long-haul cycle differs from datacenter pluggables | High |
| Nokia | Coherent/Lumentum, optics, ASICs | Telecom operators | Optical transport and routing | Carrier capex and coherent wavelength content | High |
| Juniper | Broadcom ASICs, optics, boards | Cloud and enterprise | Routing and switching | Acquisition/portfolio mix affects optics attach | High |
| Astera Labs | TSMC, packaging, connectors | Nvidia/AMD systems, cloud OEMs | PCIe/CXL connectivity and retimers | Adjacent to optics; rack bandwidth raises content | High |
| Amphenol PCD | Amphenol materials and assemblies | Server/network OEMs | High-speed and power connectors | Rack density and thermal design drive content | High |
End-to-end transaction chains
AI cluster network:
TSMC → Broadcom (switch ASIC/DSP ecosystem) → Arista/Cisco/Accton → Celestica/Fabrinet → Coherent/Lumentum/Innolight/Eoptolink → Nvidia/Dell/Supermicro racks → Microsoft/Meta/AWS/Google. Demand flows backward from ports and cluster build; component qualification and fiber attach can gate shipments.
TSMC → Broadcom (switch ASIC/DSP ecosystem) → Arista/Cisco/Accton → Celestica/Fabrinet → Coherent/Lumentum/Innolight/Eoptolink → Nvidia/Dell/Supermicro racks → Microsoft/Meta/AWS/Google. Demand flows backward from ports and cluster build; component qualification and fiber attach can gate shipments.
Laser path:
Silicon/InP substrate suppliers → IQE / Sumitomo Electric → Coherent / Lumentum / Mitsubishi Electric → Fabrinet / Jabil → optical module OEMs → Arista/Cisco/Nvidia systems → hyperscalers.
Silicon/InP substrate suppliers → IQE / Sumitomo Electric → Coherent / Lumentum / Mitsubishi Electric → Fabrinet / Jabil → optical module OEMs → Arista/Cisco/Nvidia systems → hyperscalers.
Connector path:
Corning/Fujikura fiber → Senko/US Conec → Molex/Amphenol/TE → switch and server ODMs → rack integrators. Port density increases connector, harness and fiber dollars even when transceiver ASP falls.
Corning/Fujikura fiber → Senko/US Conec → Molex/Amphenol/TE → switch and server ODMs → rack integrators. Port density increases connector, harness and fiber dollars even when transceiver ASP falls.
CPO / optical-I/O path:
TSMC → Ayar Labs / Lightmatter and package partners → Nvidia/Broadcom/custom ASIC designers → advanced AI systems. This is a potential architecture shift, with lower maturity and higher inference uncertainty than pluggables.
TSMC → Ayar Labs / Lightmatter and package partners → Nvidia/Broadcom/custom ASIC designers → advanced AI systems. This is a potential architecture shift, with lower maturity and higher inference uncertainty than pluggables.
Prepared by Singularity Research · Estimates reflect public filings, industry structure and best judgement. Percentages may not sum perfectly because of rounding and channel overlap.