Core Conclusions
AI optical communications is not an optional accessory bolted onto AI systems; it is the core capacity-expansion layer that lets an AI cluster move from single-box compute toward large-scale training, inference, distributed storage, and cross-campus interconnect. The more compute expansion depends on multi-machine cooperation, the more the network shifts from a "bandwidth add-on" into the "system bottleneck." That is the fundamental reason Dell'Oro now breaks out the AI back-end switch market as a separate line and expects it to exceed 100 billion dollars in scale by 2030.
The themes that have genuinely converted into orders and revenue rank, in order of priority, as: 800G pluggable optical modules, early-stage 1.6T ramp, AI switches and switch chips, EML/CW lasers, key optical components and high-end optical manufacturing. By contrast, CPO, NPO, and Optical I/O sit closer to mid-term architectural directions; their position in the industry is very high, but their direct contribution to most listed companies' financials over the next twelve months remains weaker than 800G/1.6T pluggables.
800G is "present tense," not a concept. LightCounting notes that the market for AI-cluster optical modules, LPO, and CPO grows from roughly 5 billion dollars in 2024 to over 10 billion dollars in 2026; Dell'Oro likewise points out that AI back-end network switch ports are already dominated by 800G and will migrate toward 1.6T in 2027.
1.6T is the "next real upgrade," and its logic comes from 200G/lane electrical and optical interfaces, 102.4T switch chips, and the high-radix demands of next-generation AI scale-out/scale-up networks. Marvell, Cisco, and Broadcom have all publicly demonstrated 1.6T-related DSP, switching, and system roadmaps, which shows this is not a distant vision but a product and design-in window for 2026-2027.
3.2T still leans toward "roadmap confirmed, revenue not yet here." Broadcom has made clear that its 200T AI-era roadmap will lay the foundation for future 3.2T optical modules and a 204.8T switching platform; Marvell extended its roadmap to 3.2T and beyond through the acquisition of Polariton. But over the next 12-24 months, the earnings leverage from 3.2T for listed companies shows up more as R&D positioning and valuation imagination than as large-scale revenue.
The most directly benefiting listed companies are not every firm that "tells an AI optical-communications story," but those that have already entered key customers' procurement lists, possess volume production and delivery capability, and have financial validation on their books. This group includes Arista, Cisco, Broadcom, Marvell, Coherent, Lumentum, Fabrinet, and Applied Optoelectronics, along with China's Innolight, Eoptolink, T&S Communications, and Accelink.
The strongest current evidence chain for customer and order certainty does not lie in vague "AI exposure" but in NVIDIA's official ecosystem and long-term partnership commitments: NVIDIA has publicly disclosed that its CPO/silicon-photonics ecosystem partners include TSMC, Coherent, Corning, Foxconn, Lumentum, and SENKO, and notes that its pluggable optics ecosystem is supported by Coherent, Eoptolink, Fabrinet, Innolight, and others; in addition, NVIDIA signed multibillion-dollar procurement and 2 billion dollar investment partnerships with Coherent and Lumentum respectively, and is working with Marvell on NVLink Fusion and silicon photonics.
The highest-margin layers over the long run remain the switch chips and system software stack together with the key semiconductors/lasers backed by proprietary IP; the layers with the greatest revenue leverage are the pluggable modules and upstream components. This means high margins and high leverage often do not sit in the same place: Arista and Broadcom earn high earnings quality, Coherent and Lumentum earn the scarcity of key materials and components, and Innolight and Eoptolink earn the volume-and-price leverage of AI generational upgrades.
What may genuinely tighten over the next 12-24 months is no longer generic "800G module capacity," but 200G/lane DSP, InP/EML/CW lasers, high-power lasers for CPO, silicon-photonics packaging and test, and high-density connectors/FAUs/MT ferrules. LightCounting has already noted that overall supply pressure on conventional Ethernet optical modules is easing in 2026, which means commodity categories will more easily face price competition; the real scarcity is migrating toward the higher-spec layers required by 1.6T/CPO.
What may face substitution over the medium-to-long term is not "all optical modules being eliminated," but a shift in where value sits along the chain. LPO/LRO will thin out the value of short-reach DSP modules, silicon photonics will compress the value of discrete modulators and the WDM portion, and CPO/NPO will migrate part of the traditional module-assembly value into switch chips and optical-engine packaging. But until serviceability, replaceability, thermal design, yield, and standardization are resolved, pluggable modules will remain the workhorse form factor for the next several years.
The rise of Ethernet in AI scale-out is already reality, not expectation. Dell'Oro notes that 2025 Ethernet AI back-end switch sales have exceeded InfiniBand by more than two times, with Amazon, Microsoft, Meta, Oracle, and xAI all adopting Ethernet; this directly benefits Arista, Cisco, and Broadcom, and also means optical-module demand will spread from a single NVIDIA ecosystem to the much broader Ethernet chain.
There is no public evidence that cloud-vendor capital spending has hit an "AI build-out peak." Microsoft's FY26 Q1 CapEx was 34.9 billion dollars, Alphabet's 2026 Q1 CapEx was 35.7 billion dollars with roughly 40% of it going to data centers and network equipment, Meta's 2026 Q1 CapEx was 19.84 billion dollars, and Alphabet also gave full-year 2026 CapEx guidance of 175 billion to 185 billion dollars; Oracle's FY26 Q3 OCI revenue grew 84% year over year and RPO grew 325% year over year. These disclosures mean the demand base for AI networking and optical interconnect remains strong.
The biggest risk is not that "AI optical communications has no demand," but that demand is real while the order of profit distribution changes: if 800G supply is ample and ASPs fall quickly while 1.6T/CPO fail to pick up the baton in time, midstream module assemblers may see revenue grow faster than profit; conversely, lasers, key components, switch ASICs, and the system-software layer hold up better on profit resilience.
Supply Chain Panorama
In the AI optical-communications supply chain, what is genuinely worth investing research time in is not "who touches optics," but who can turn GPU-cluster expansion directly into visible orders, recognized revenue, and released profit. The following walks through upstream, midstream, and downstream, and puts "revenue-recognition method, supply bottleneck, margin profile, and benefit intensity" into one view.
Upstream Materials and Chips
Supply-Chain Position Sub-Segment Core Products AI Demand Driver Revenue Recognition Main Customers Supply Bottleneck Margin Profile Representative Companies Listed/Private Benefit Intensity Investment Leverage Main Basis Materials InP/GaAs/SiPh materials InP laser epitaxy, GaAs, SiPh wafers 200G/lane, 1.6T/CPO upgrades Wafer/epi-wafer shipments Laser and optical-chip makers InP and high-end epitaxy, SiPh packaging validation High barriers, but capex-heavy Coherent, Lumentum, TSMC, GF Fotonix Listed/Private 4 4 Chips InP optical chips EML, CW DFB, DFB, some detectors Rising 800G/1.6T volumes Chip sales / long-term supply Module makers, CPO optical engines High-speed yield, test, validation Margins steadier than modules Coherent, Lumentum, Yuanjie Semiconductor, Sigtone Photonics Listed 5 5 Analog chips TIA/Driver TIA, Driver, Retimer Higher speeds and lower bit-error rates Chip sales Module makers, system vendors Process node and customer qualification Mid-to-high margins, IP-driven Broadcom, Marvell, MaxLinear Listed 3 4 DSP DSP chips 800G/1.6T PAM4/coherent DSP 100G/lane to 200G/lane migration Chip sales / reference designs Module makers, switch/router vendors 3nm, power, ecosystem compatibility High ASP, high technical barriers Marvell, Broadcom, Cisco Acacia Listed/Internal business 4 5 Optical Components, Optical Modules, and Packaging Manufacturing
Supply-Chain Position Sub-Segment Core Products AI Demand Driver Revenue Recognition Main Customers Supply Bottleneck Margin Profile Representative Companies Listed/Private Benefit Intensity Investment Leverage Main Basis Lasers EML/CW/UHP lasers 100G/200G EML, CW lasers, high-power lasers 800G/1.6T pluggables and CPO Chip/component shipments Module makers, NVIDIA/CPO ecosystem InP production lines, high-speed packaging High margins, strong pricing power Lumentum, Coherent Listed 5 5 Optical components AWG, FAU, isolators, couplers, MT ferrules Wavelength-splitting/coupling/connector parts High-density routing and SiPh/CPO packaging Component sales Module makers, packaging houses, switch vendors High-precision alignment, temperature tolerance, insertion loss Mid-to-high margins, leaders better positioned T&S Communications, Sigtone Photonics, SENKO, US Conec, Fujikura Listed/Private 4 4 Optical modules 400G/800G/1.6T/3.2T OSFP, QSFP-DD, DR/FR/LR/ZR AI scale-out expansion Module delivery and acceptance Cloud vendors, NVIDIA ecosystem, switch vendors 200G/lane solutions, DSP, lasers Greatest revenue leverage, but also greatest ASP pressure Innolight, Eoptolink, Accelink, AOI, Coherent Listed 5 5 Packaging manufacturing Module foundry / precision optical manufacturing OSA, TOSA/ROSA, optical-module assembly Shipment-volume surge Manufacturing-service revenue Major optical-module and laser makers Yield, automation, process know-how Margins depend on mix and capacity utilization Fabrinet, Foxconn, Celestica Listed 4 4 Silicon Photonics, CPO, and Network Systems
Supply-Chain Position Sub-Segment Core Products AI Demand Driver Revenue Recognition Main Customers Supply Bottleneck Margin Profile Representative Companies Listed/Private Benefit Intensity Investment Leverage Main Basis Silicon-photonics platforms Silicon Photonics SiPh modulators, integrated platforms, external laser sources Lower power, smaller footprint, support for CPO/NPO Chip/optical-engine/platform licensing Switch-chip and AI-accelerator makers Packaging, thermal design, laser coupling High barriers once mature, heavy early investment Intel SiPh, Marvell, Broadcom, Ayar Labs, Lightmatter, Celestial AI Listed/Private 4 4 CPO/NPO/OIO Co-packaged / near-packaged / optical I/O CPO switches, NPO optical engines, Optical I/O 1.6T/3.2T and power pressure System/component revenue NVIDIA, switch OEMs, cloud vendors Yield, serviceability, standardization High long-term value, limited near-term revenue NVIDIA, Broadcom, Marvell, Ayar Labs, Ranovus Listed/Private 3 5 Switch chips Ethernet/IB/scale-up ASIC 51.2T/102.4T/204.8T switch ASIC Port-speed upgrades and high radix Chip sales Switch vendors, ODMs, cloud vendors Advanced process, software-stack adaptation High margins, high barriers Broadcom, Marvell, NVIDIA, Cisco Listed 5 4 Switch systems AI data-center switches 800G/1.6T switches AI cluster scale-out/scale-up System revenue / services Amazon, Microsoft, Meta, Oracle, xAI, etc. Customer qualification, supply cadence Margins higher than modules Arista, Cisco, NVIDIA, Celestica, HPE/Juniper Listed 5 4 Fiber and connectors Fiber, connectors, cables High-density multi-fiber connectors, MPO/MMC, AOC/DAC Routing density and data-center expansion Cable/connector shipments Cloud vendors, DC builders, system makers High-density connectors and insertion loss Relatively steady, weaker leverage than modules Corning, Fujikura, Sumitomo Electric, Furukawa Listed 3 3 Demand and Technology Path
Why AI Clusters Keep Pushing Up High-Speed Optical-Communications Demand
The core of AI training and large-scale inference is not the peak compute of a single GPU, but whether parameters, gradients, KV cache, checkpoints, and samples can be exchanged at high utilization between GPUs, between GPUs and storage, between racks, and between campuses. Dell'Oro has explicitly divided this class of networking into three domains, scale-up, scale-out, and scale-across, and believes future AI back-end network spending will keep growing across all three; NVIDIA, in its Rubin and Spectrum-X roadmaps, ties "sustained high utilization" and "reduced training time/TCO" directly to network capability.
From a physical-layer view, the reason AI data centers need more optics rather than more copper is simple: higher speeds, longer distances, greater power sensitivity, and hotter racks. NVIDIA has publicly stated that its Spectrum-X Photonics and Quantum-X Photonics, by integrating optics into the switch-ASIC package, can deliver more than 5x improvement in network power efficiency and more than 10x improvement in resilience; Broadcom likewise defines its 102.4T Tomahawk 6 CPO switch as infrastructure that provides more stable links and higher energy efficiency for AI scale-up/scale-out.
Cloud-vendor results also show that networking is not edge spending. Alphabet directed roughly 40% of its 35.7 billion dollar CapEx in 2026 Q1 toward data centers and network equipment; Microsoft's 34.9 billion dollar capital spending in FY26 Q1 also included server and network-equipment refresh beyond GPUs/CPUs; Meta's 2026 Q1 capital spending reached 19.84 billion dollars; and the strong growth in Oracle's OCI revenue and RPO shows that demand for AI-infrastructure leasing is still expanding.
What Training, Inference, Scale-Up, Scale-Out, and Scale-Across Each Mean
What training fears most is global synchronization and large-scale all-reduce; what inference fears most is latency, elastic scaling, and cross-model/cross-cache scheduling. Training therefore places higher demands on bandwidth, latency jitter, packet-loss control, and network determinism, while inference is more affected by cost, power, and network-tiering design. Dell'Oro argues that clusters relying mainly on scale-out are no longer sufficient to carry the next phase of agentic/physical-AI demand, that scale-up fabric is emerging, and that Ethernet will gradually expand from scale-out into scale-up.
Scale-up: occurs more within a single chassis/rack domain or even inside a rack, pursuing extremely high-bandwidth coupling of GPUs and memory. It is still dominated by proprietary solutions such as NVLink, but UALink and Ethernet-based approaches are beginning to contest future leadership.
Scale-out: connects large numbers of accelerator nodes across racks and rows. It is the most direct source of 800G/1.6T optical-module demand and the strongest growth zone for Arista, Cisco, Broadcom, and NVIDIA switches.
Scale-across: data-center interconnect, leaning toward ZR/ZR+ and Jericho-class DCI/AI WAN. Broadcom's Jericho4 and Marvell's 1.6T ZR/ZR+ roadmaps map directly to this layer.
The Upgrade Logic of 800G, 1.6T, and 3.2T
According to Dell'Oro, most AI back-end network switch ports have already shifted to 800G, are expected to shift to 1.6T in 2027, and to move toward 3.2T in 2030. At 2026 OFC, Marvell defined 1.6T as a "new era" for end-to-end connectivity in AI data centers; Cisco has also released 1.6T OSFP optics aimed at AI scale-out; and Broadcom, through 102.4T CPO switching and 200G/lane components, explicitly extends its roadmap toward future 3.2T optical modules.
The math behind this is not complicated. Take a 51.2T switch as an example: if it uses 64 ports of 800G, the optical-module count grows roughly linearly with port count. Lumentum has given a very clear illustration with EML chips: a single 51.2T, 64×800G switch platform could correspond to 512 chips of 100G EML, and 100 such switches would need over 50,000 EML chips. This example is not an exact count for every architecture, but it is enough to show that as AI clusters scale from thousands to tens of thousands to hundreds of thousands of cards, demand for optical components and modules compounds along "port count × generational upgrade."
For the question of "roughly how many optical modules an AI server, an AI rack, or an AI cluster needs," public sources can hardly give a single fixed number that fits all architectures, because it depends on whether copper is used inside the rack, the TOR/EOR layout, the degree of liquid-cooling disaggregation, and whether scale-up stays inside the chassis. The more reasonable approach is to give a range judgment: a single training server's external high-speed links typically run from "almost no optics" to "as many as several 400G/800G optical ports"; a cross-rack leaf-spine design pushes a single AI rack's boundary to tens of high-speed optical ports; and when a cluster reaches tens of thousands of GPUs, module demand usually enters the thousands to tens of thousands of 800G/1.6T range. This is an inference based on switch-port specifications and AI-fabric structure, not a uniform BOM.
The Applicable Scenarios for LPO, LRO, DSP, and CPO
The economic core of LPO/LRO is to cut the power and latency contributed by DSP and retimers in short-reach, controllable link-quality scenarios, which makes them better suited to the interior of AI data centers, especially links where cabling and interoperability can be tightly controlled. LightCounting explicitly believes AI is accelerating the adoption of LPO/CPO; Eoptolink has publicly treated LPO/LRO, XPO, NPO, and CPO as one technology family and put volume-production deployment in place.
But DSP will not disappear quickly because of LPO. The reason is that as speeds rise from 100G/lane to 200G/lane, the requirements on link budget, signal integrity, and interoperability grow higher, and in longer-reach, more complex deployments that emphasize compatibility, DSP-equipped modules remain mainstream. Marvell's 1.6T Ara DSP platform and Broadcom's continued strengthening of its 200G/lane optical-component roadmap both show that DSP retains considerable vitality in the 1.6T era.
CPO is a different level of question altogether: it does not "replace one chip," but redraws the boundaries among switch chips, optical engines, connectors, serviceability, and power delivery. NVIDIA has placed CPO in its Quantum-X and Spectrum-X Photonics products and given a 2026 volume-production timeline; Broadcom has also begun volume shipments of its third-generation CPO switch. Over the next 12-24 months, however, CPO is more likely to first replace some of the most extreme, most dense, and most power-sensitive switch-layer connections rather than fully displace traditional pluggable optical modules.
Scenario Forecast
Dimension Conservative Base Aggressive Assumptions Cloud-vendor CapEx growth slows but does not turn negative; training clusters focus more on ROI Hyperscalers and neo-clouds keep expanding; Ethernet share keeps rising Agentic AI, inference networks, and cross-campus interconnect surge; CPO reaches volume faster GPU-cluster build cadence Expansion continues but pace slows Training + inference dual-engine drive Sustained ten-billion-dollar-scale network spend 800G penetration Continues to dominate 2026-2027 High-level ramp Still large, but migrates faster toward 1.6T 1.6T penetration Mainly samples and small-batch design-in Clear ramp around 2027 Becomes the new mainstream around 2027 CPO adoption Pilots / partial deployment Begins landing in high-end switch layers in 2026-2027 Accelerated penetration driven by NVIDIA/Broadcom Most-benefiting layer 800G modules, EML/CW, switches 1.6T modules, DSP, lasers, switches Silicon photonics, CPO, high-power lasers, optical engines, connectors Most-benefiting companies Innolight, Eoptolink, Lumentum, Coherent, Arista Arista, Cisco, Broadcom, Marvell, Coherent, Lumentum, Eoptolink Coherent, Lumentum, Marvell, Broadcom, Ayar Labs, T&S Communications, Fujikura Main risks Rapid 800G ASP decline, end of inventory restocking Design-switch delays, customer qualification slower than expected CPO yield/serviceability, supply bottlenecks slowing revenue conversion Basis Segment Breakdown
To keep this readable, the eighteen segments requested by the user are compressed below into a single "investment-attractiveness matrix," with each segment mapped to its segment logic, revenue conversion, supply-demand, price and margin, catalysts, and risks. Scores use a 5-point scale, where higher means more worth prioritizing for research over the next 12-24 months.
Segment Segment Logic How AI Demand Converts to Revenue Current Supply-Demand Price/Margin Trend 12-24 Month Catalysts Main Risks Score Directly Benefiting Companies Basis 800G optical modules Still the workhorse of AI scale-out Switch/NIC upgrades pull modules directly Strong demand, but overall supply tightness has eased ASP falling, volume offsets price Cloud-vendor expansion, Ethernet keeps taking share Price wars, inventory cycle 4.5 Innolight, Eoptolink, AOI, Accelink 1.6T optical modules 200G/lane generational upgrade Pulled by 102.4T switch ASIC and high radix Early ramp, qualification-driven, greater scarcity Better early ASP/margin Cisco/Marvell/Broadcom roadmaps land Slow design-in cadence, standards evolution 4.5 Eoptolink, Coherent, Marvell, Cisco 3.2T optical modules Maps to the 204.8T era Shows up first in R&D and design-in Far from volume Still theme-ahead-of-fundamentals 400G/lane and 204.8T chip launches Long-dated, slow conversion 2.5 Broadcom, Marvell, a few forward-leaning module makers LPO/LRO Short-reach power/latency reduction Internal network-BOM optimization for specific cloud customers Localized adoption, not whole-market dominance If volumes rise, could improve power value AI internal-fabric optimization Interoperability, link budget 3.5 Eoptolink, Eoptolink-class players, Broadcom-supporting DSP modules Still mainstream in the 1.6T era Chip + module dual revenue recognition 3nm/200G/lane scarcity ASP and margins relatively steadier 1.6T ramp, DCI upgrades LPO squeezes short-reach demand 4.0 Marvell, Cisco Acacia, Broadcom Silicon photonics Lower power, higher integration over the medium-to-long term Platform/engine/chip revenue Application expanding but long customer validation High margins once mature, heavy early investment NVIDIA/CPO and hyperscaler adoption External-laser and packaging challenges 4.0 Intel SiPh, Marvell, Broadcom, Ayar Labs, Lightmatter CPO One of the endgame directions for power and bandwidth density System-level substitution, value migrating to switch ASIC and optical engine Direction clear, revenue conversion still early Unstable early margins, big potential after volume NVIDIA/Broadcom volume-production window Serviceability, thermal design, yield 3.0 NVIDIA, Broadcom, Coherent, Lumentum, Ayar Labs EML lasers Core 800G/1.6T component Module ramp directly pulls chip shipments Still a high-barrier bottleneck layer Margins better than module assembly NVIDIA long-term orders, 1.6T ramp Customer concentration, alternative paths 4.5 Lumentum, Coherent VCSEL Short reach, NPO, some high-speed interconnect Depends on architectural changes Affected by roadmap divergence Mediocre margin leverage Broadcom NPO solutions Squeezed by EML/SiPh 3.0 Broadcom ecosystem, some Taiwanese makers InP optical chips The real bottleneck for high-speed light sources Revenue recognized through the laser/module chain High-speed yield and capacity matter Structurally high margins 1.6T, CPO light sources Capex, customer qualification 4.5 Coherent, Lumentum, Yuanjie Semiconductor, Sigtone Photonics TIA/Driver The pacer of speed upgrades Attached to module/connector upgrades Higher technical barrier but not the scarcest Mid-to-high margins 200G/lane proliferation Displaced by system integration or price competition 3.5 Broadcom, Marvell DSP chips Core of data-rate upgrades 1.6T, ZR/ZR+, AEC/optics ramp 3nm and 200G/lane are key High ASP and barriers Start of the 1.6T era Partly displaced by LPO/LRO in some scenarios 4.0 Marvell, Broadcom, Cisco Acacia Optical components AWG/FAU/isolators/couplers Driven jointly by modules and SiPh packaging Looks fragmented, actually critical Leaders hold stronger pricing power CPO/NPO packaging ramp Price pressure from integration 4.0 T&S Communications, Sigtone Photonics Optical-module foundry Enjoys scale effects during ramp OEM/IDM orders convert to manufacturing revenue Strongly pulled by leading customers Depends on mix and automation 1.6T, CPO packaging spillover Customer concentration 4.0 Fabrinet, Foxconn Test equipment Every generational upgrade requires investment Follows module/chip validation Steady demand but smaller volume Usually steady margins 200G/lane and CPO validation Cadence lags shipments 3.5 Anritsu, Keysight, Viavi Fiber connectors Hard demand for high-density routing DC new-build/expansion and refresh Steady growth, slower than modules Less ASP pressure than modules New connector standards like MMC/TMT Cyclical, not AI-exclusive 4.0 Fujikura, SENKO, Corning, Furukawa Switch chips The value hub of AI networking Port upgrades monetized directly via ASIC/SerDes One of the highest barriers One of the highest margins The 102.4T/204.8T era Technology iterating too fast 4.5 Broadcom, Marvell, NVIDIA, Cisco Data-center switches The most certain system layer in AI budgets Whole units recognized together with OS/management software Strong demand, concentrated customers Earnings quality better than modules Rising Ethernet share Customer concentration, supply-chain constraints 4.5 Arista, Cisco, NVIDIA, Celestica InfiniBand vs. Ethernet Determines who controls the flow of network orders Fabric choice directly dictates ecosystem procurement Ethernet share rising clearly More open Ethernet ecosystem UEC/MRC/Ethernet scale-up IB still has an edge in high-performance training Ethernet 4.5 / IB 3.5 Arista, Cisco, Broadcom, NVIDIA Who Benefits Most Directly, Who Is More Likely Just a Theme
The most direct beneficiaries are modules, lasers, core components, and AI switch systems, because these layers are most easily pulled by orders and most quickly convert to revenue. On the financial evidence: Arista has raised its FY2026 AI fabrics revenue target to roughly 3.5 billion dollars; Cisco landed 2.1 billion dollars of hyperscaler AI-infrastructure orders in a single FY26 Q2 quarter; Coherent and Lumentum have established long-term partnerships with NVIDIA backed by procurement commitments; and the results of Fabrinet and AOI already directly reflect AI optical-component/module demand.
By contrast, companies that merely tell a "silicon-photonics/CPO/Optical I/O story" but have not yet formed meaningful revenue, or companies whose main business is still traditional telecom/backbone/general connectivity with AI as only a secondary increment, are better placed in "medium-to-long-term observation" rather than the "short-to-medium-term earnings-conversion mainline." This group is not without value; it is that the public financial evidence is not yet sufficient to support a "high-certainty, high-profit-leverage" conclusion.
Company List and Tiering
Master Table of Key Listed Companies
The table below prioritizes companies that genuinely sit at the hub of AI optical-communications order flow and already have public evidence. For customers, AI revenue share, backlog, or real-time valuation that cannot be confirmed from public sources, it uniformly marks "undisclosed / needs further verification."
Company Ticker/Market Sub-Segment AI Benefit Path Main Customers/Relationships In Key Supply Chain? Public Financial Evidence Order/Capacity/Product Progress Overall Assessment Tier Main Basis Innolight 300308.SZ / A-share 800G/1.6T optical modules High-speed module shipments to North American and global data centers Specific customers undisclosed; Innolight's website states its products are widely deployed at top global data-center operators Yes; NVIDIA lists Innolight as a supporter of its pluggable optics ecosystem 2025 annual report says 800G and 1.6T demand will dominate over the next three years; media cite the report showing a high share of overseas revenue Continues to scale 800G/1.6T delivery, with forward-looking 3.2T positioning One of the module leaders with the strongest revenue leverage and most thorough customer validation A Eoptolink 300502.SZ / A-share 800G/1.6T/LPO/NPO/XPO optical modules Converts AI demand through 800G/1.6T/LPO product upgrades Focused on large cloud-service customers, names undisclosed Highly likely; products target AI clusters/cloud DCs; among the first in the industry to volume-produce 1.6T 2025 annual report summary calls 800G/1.6T the core growth engine; first to volume-produce and ship 1.6T, and volume-producing LPO Launched 1.6T DR4, 6.4T NPO, and 12.8T XPO in March 2026 One of the Chinese leaders with the highest certainty and most aggressive technology roadmap A T&S Communications 300394.SZ / A-share Optical components/FAU/lasers/optical engines As an upstream key-component supplier, benefits from module upgrades and CPO/SiPh packaging Customers undisclosed Indirectly yes 2025 revenue 5.163 billion yuan, up 58.79% year over year; net profit attributable to parent 2.017 billion yuan, up 50.15% year over year Has positioned multi-channel high-power lasers for CPO and 1.6T OSFP optical engines; Phase II expansion in Thailand Not the largest revenue leverage, but strong upstream-component scarcity and better earnings quality A Accelink 002281.SZ / A-share Datacom optical modules and components 800G products in volume production, benefiting from AI-compute-related module iteration Customers undisclosed Indirectly yes Annual report summary says it has successfully developed and volume-produced multiple high-speed low-power datacom modules, including 800G VR8/DR8/2xFR4/2xLR4 Continues to ramp R&D and volume-production readiness for AI-compute-related optical modules Clear beneficiary, but its order and ultra-high-speed share evidence is still somewhat weaker than Innolight/Eoptolink B Sigtone Photonics 688313.SH / A-share AWG, FAU, CW DFB, EML pre-research Upstream passive/active components benefiting from 1.6T and CPO packaging Customers undisclosed Indirectly yes Annual report discloses 1.6T AWG chips and components advancing toward customer validation and volume-production import, plus industrialization of CPO polarization-maintaining components and high-temperature-resistant FAU 3.2T pre-research, small-scale pilots; CW DFB customer validation advancing Strong bottleneck attributes, but scale and customer concentration need continued verification B Coherent COHR / US EML/CW lasers, optical components, modules, CPO AI data-center components and optical engines benefit directly Public NVIDIA partner Yes; NVIDIA CPO ecosystem and long-term partnership FY26 Q3 revenue 1.81 billion dollars, non-GAAP gross margin 39.6%; management says 800G and 1.6T are driving growth Signed a multibillion procurement commitment with NVIDIA and received a 2 billion dollar investment; Datacenter & Communications became a new core segment A genuine upstream-bottleneck leader A Lumentum LITE / US EML, pump lasers, CW/UHP lasers, modules Lasers and some modules benefit directly from AI optical interconnect Public NVIDIA partner Yes FY25 Q4 Cloud & Networking revenue 424.1 million dollars, up 66.5% year over year; FY26 Q3 revenue 808.4 million dollars Signed a multibillion procurement commitment with NVIDIA and received a 2 billion dollar investment; building a new InP laser plant in the US One of the most typical beneficiaries of the laser bottleneck A Fabrinet FN / US Optical manufacturing / module foundry Benefits from the ramp at leading module and laser makers OEM/IDM customers, not disclosed individually Yes; NVIDIA names it in its pluggable optics ecosystem FY25 optical-communications revenue accounted for 76.6%; FY26 Q3 revenue 1.214 billion dollars, up significantly year over year AI optical communications drove record Q3 revenue; manufacturing capability and yield are core One of the most certain beneficiaries in the manufacturing layer A Applied Optoelectronics AAOI / US 800G/1.6T modules Driven by the 800G ramp and hyperscaler demand Specific customers undisclosed Needs further verification, but hyperscaler exposure is clear 2026 Q1 revenue 151.1 million dollars, up 51% year over year; says 800G/1.6T customer engagement is strong 800G has already formed real revenue, 1.6T advancing High leverage, but customer concentration and sustainable profitability still need verification B Arista ANET / US AI switches / network OS Ethernet AI-fabric system revenue recognized directly Microsoft, Meta and other cloud titans plus specialty providers; AI fabrics is an explicitly disclosed target Yes 2026 Q1 revenue 2.709 billion dollars, up 35.1% year over year; FY26 AI fabrics revenue target of roughly 3.5 billion dollars 800G continues to ramp; XPO high-density liquid-cooled pluggable solution launched; purchase commitments of 8.9 billion dollars Combines revenue certainty, earnings quality, and a moat A Cisco CSCO / US AI switches, Silicon One, Acacia optics Hyperscaler AI-infra systems and optics provide a dual-engine drive Hyperscalers, not all specific customers disclosed Yes Landed 2.1 billion dollars of hyperscaler AI-infra orders in FY26 Q2; FY26 guidance for AI-infra revenue above 3 billion dollars Launched 1.6T OSFP optics, G300, and network-utilization optimization solutions; Acacia optics performing strongly Strong certainty but AI purity lower than Arista B Broadcom AVGO / US Switch chips, CPO, optical components, NIC Switch ASIC and Ethernet AI switch are the layer with the greatest operating leverage Cloud/hyperscale, OEM Yes FY25 Q4 AI semiconductor revenue up 74% year over year; management expects FY26 Q1 to reach 8.2 billion dollars Tomahawk 6 102.4T CPO is in volume shipment; complete 200G/lane VCSEL/EML/CWL/CPO and NPO roadmap One of the strongest long-term moats, but not a "pure optics" name A Marvell MRVL / US DSP, AEC, optical-interconnect platforms, silicon photonics 1.6T DSP and Photonic Fabric enter the core of AI interconnect Cloud vendors/OEM, full list not disclosed Yes; AI-ecosystem partnership with NVIDIA FY26 Q4 revenue 2.219 billion dollars; clear 1.6T Ara DSP, ZR/ZR+, and OCS roadmaps Acquisition of Polariton points to 3.2T and beyond; OCS demo with Lumentum; NVIDIA announced investment and silicon-photonics partnership Medium-to-high earnings conversion, migrating to higher-value layers A Corning GLW / US Fiber / connectivity AI data-center high-density routing, CPO ecosystem Equipment vendors / data-center builders Yes; NVIDIA CPO ecosystem partner AI-related revenue not separately disclosed Plays the optical-connectivity role in NVIDIA's silicon-photonics CPO supply chain Benefit is real but indirect, with weaker profit leverage than modules/components C Fujikura 5803.T / Japan High-density connectors / fiber routing AI data-center high-density connectors and optical routing Data-center construction and equipment support Indirectly yes The company focuses on showcasing optical routing and connectors for AI data-center networks; PRIZM TMT/MMC point to large-scale AI DCs Continued benefit from connector-standard evolution and data-center expansion Steady beneficiary, bottleneck-leaning but with mild leverage B Furukawa Electric 5801.T / Japan Fiber, MT ferrules, DFB lasers Data-center-related products drive growth in communications solutions DC operators, communications system vendors Indirectly yes Financial materials state data-center demand is increasing and that communications-solution profit is driven by data-center-related products Strengthening data-center product development, manufacturing, and sales systems Clear beneficiary but more of an "infrastructure-materials supplier" C Sumitomo Electric 5802.T / Japan Fiber and Data Center Solutions Benefits from DC fiber-routing demand Data-center builders Indirectly yes Discloses a Data Center Solutions business, but AI-related financials not separately broken out Mainly connectivity and routing solutions More of a steady allocation than a high-leverage AI optics name C Ciena CIEN / US DCI / backbone equipment Benefits from scale-across and DCI Cloud/carriers Needs further verification Evidence of direct AI financial leverage is weaker than the module/switch chain Leans toward long-haul and DCI Has a logic, but should not be treated on par with the AI back-end mainline C ZTE 000063.SZ/HK Communications equipment / data-center networking Benefits from broad AI infrastructure and Ethernet switching Carriers/enterprise-government/some data centers Needs further verification Public sources show no intuitive financial breakdown of the AI optical-module chain Leans toward communications equipment rather than pure AI optical interconnect Related but not a purest beneficiary C Important Private Companies and Primary-Market Opportunities
Company Region Sub-Direction Core Products Customers/Partners Funding/Valuation Relationship to Listed Companies Investment Focus Main Risks Basis Ayar Labs US CPO / Optical I/O Optical I/O and CPO solutions Backed by NVIDIA and others Closed a 500 million dollar Series E in 2026 at a 3.75 billion dollar valuation, 870 million dollars in total funding Potential coopetition with NVIDIA/Broadcom/Marvell and others If CPO reaches large-scale volume, one of the most core primary assets Commercialization cadence, supply and test scaling Celestial AI US Photonic Fabric Chip-to-chip and memory-to-memory optical interconnect AMD Ventures and other investors Raised 250 million dollars in 2025, over 515 million dollars in total funding Forms potential coopetition with Marvell/Broadcom/AMD If scale-up/memory interconnect goes optical, its position is excellent Still in early platform-import stage Lightmatter US Silicon photonics / photonic interconnect Passage photonic interposer, chiplet GlobalFoundries manufactures part of its products 2024 Series D of 400 million dollars at a 4.4 billion dollar valuation May compete head-to-head with SiPh/CPO leaders Worth tracking for its leap from demo to volume Revenue conversion still early Ranovus Canada Optical I/O / NPO/CPO Odin Optical I/O core and high-speed interconnect Partnership with Cerebras, DARPA projects Latest valuation not disclosed Deeply tied to system/AI-chip makers Leans toward medium-to-long-term technical positioning Commercial scale and volume cadence Intel Silicon Photonics US Silicon photonics Silicon-photonics platforms and modules Large cloud vendors, OEM Large-company internal business Competes with Broadcom/Marvell/newcomers High industry maturity, strong volume-production experience Low financial transparency Cisco Acacia US coherent DSP/optics Acacia coherent optics, 1.6T-related Hyperscaler / Cisco systems Internal business Competes with Marvell/Broadcom DCI and AI scale-across value AI purity lower than the pluggable mainline Broadcom optical-interconnect business US CPO/NPO/VCSEL/EML Tomahawk 6 CPO, NPO, 200G/lane optics OEM, cloud vendors Internal business The core of value-chain reallocation If CPO penetration accelerates, one of the most benefiting large players Not a pure optics business; valuation anchor must be separated NVIDIA Photonics US CPO / silicon photonics Quantum-X / Spectrum-X Photonics Partners include Coherent, Corning, Lumentum, SENKO, and others Internal business Holds roadmap-defining power over the whole supply chain Not a standalone investment name, but sets the industry's cadence Adoption speed and ecosystem coordination Company Tiering and Investment Priority
Tier A: Core Direct Beneficiaries Innolight, Eoptolink, T&S Communications, Coherent, Lumentum, Fabrinet, Arista, Broadcom, Marvell. Their shared trait is that they either sell 800G/1.6T modules and key components directly, or sell AI switch systems/switch chips directly, and already have public orders, revenue targets, partnership commitments, or significant results to show for it.
Tier B: Clear Beneficiaries, but Carrying Cycle/Customer-Concentration/Valuation Risk Accelink, Sigtone Photonics, AOI, Cisco, Fujikura. Their AI logic all holds, but either customer and order disclosure is not complete enough, or AI revenue is not yet a high enough share of total revenue, or valuation and supply-cycle volatility is larger.
Tier C: Possible Long-Term Beneficiaries, but Mediocre Near-Term Financial Leverage Corning, Furukawa Electric, Sumitomo Electric, Ciena, ZTE. They lean more toward fiber, connectivity, DCI, and communications equipment; AI demand can transmit to them, but the path is longer and the profit leverage weaker.
Tier D: Theme Stronger Than Financial Evidence Pure CPO/Optical I/O theme-mapping names, or companies that map AI optical communications onto general communications equipment. This category does not mean the companies are weak; it means that at this stage, public sources are not sufficient to prove their AI optical communications has meaningfully converted into orders, revenue, and profit. Most primary-market companies also temporarily belong here.
Valuation, Competition, and Risk
Competitive Landscape and Customer Relationships
The division of labor between Chinese module makers and US component/system makers is already very clear. Chinese vendors hold a significant edge in the volume production, cost, and delivery speed of 800G/1.6T pluggable modules; US vendors hold stronger IP and customer stickiness in EML/CW lasers, switch chips, system software, CPO, and high-end DSP. NVIDIA's public ecosystem strings this chain together neatly: the pluggable optics ecosystem includes Eoptolink, Innolight, Fabrinet, and Coherent; the CPO/silicon-photonics ecosystem includes TSMC, Coherent, Corning, Foxconn, Lumentum, and SENKO.
Cloud vendors do not tend to procure from a single supplier. Dell'Oro argues that AI-cluster scale expansion and supply-chain constraints are pushing vendor diversity, which is why Ethernet is accelerating its expansion; that is also why Arista, Cisco, Celestica, and NVIDIA can all grow simultaneously in the AI back-end switch market. For optical modules, multi-supplier strategies are equally common, which benefits leading module makers in winning orders while also meaning ASPs and margins will not rise without limit.
The 800G-to-1.6T upgrade will not necessarily replicate the 400G-to-800G market-share path. The reason is that 1.6T relies more on 200G/lane DSP, EML/CW lasers, silicon photonics/high-end packaging, and switch-chip ecosystem coordination. Whoever has the deepest customer qualification, and whoever can soonest turn 200G/lane into high-yield volume production, has a better chance of prevailing in the next round of share reshuffling. Marvell, Cisco, Broadcom, Coherent, and Eoptolink all show strong positioning in public sources.
CPO will restructure the supply chain, but not eliminate pluggables across the board. The roadmaps of both NVIDIA and Broadcom show CPO has entered the volume-production stage, but NVIDIA also makes clear that its photonics will develop in parallel with pluggable optical transceiver technologies, the latter still supplied by Coherent, Eoptolink, Fabrinet, Innolight, and others. What is more likely over the coming years is a hybrid landscape of "CPO at the high-end switch layer, pluggables continuing at the broad-deployment layer."
Scoring Model and Total-Score Ranking
The scoring weights use the seven criteria suggested by the user: direct AI-demand exposure 25%, supply-chain position and pricing power 20%, technical barriers 15%, financial quality 15%, growth leverage 10%, valuation reasonableness 10%, and future catalysts 5%. Scores are on a 100-point scale, and the core logic is: first look at order conversion, then profit distribution, and only last at theme imagination.
Rank Company Total Score Ranking Logic Lead group Innolight 87 Directly benefits from 800G/1.6T, strong volume-delivery advantage, most direct AI-demand transmission, but price pressure needs continued tracking Lead group Eoptolink 86 One of the strongest on 1.6T, LPO, and NPO/XPO technical foresight, large upgrade leverage Lead group Arista 85 Clear AI fabrics revenue target, high financial quality, biggest beneficiary of the Ethernet trend Lead group Coherent 84 Core of the EML/CW and CPO ecosystem, NVIDIA long-term orders raise certainty Lead group Lumentum 83 Strong laser-bottleneck attributes, NVIDIA partnership and new US plant raise supply pricing power Lead group Broadcom 82 Strongest dominance in switch chips/CPO/200G/lane platforms, but "pure optics" leverage is more diffuse Strong group Marvell 80 Complete 1.6T DSP, silicon-photonics, and OCS roadmap, with the conversion window in 2026-2027 Strong group Fabrinet 79 The most stable beneficiary of manufacturing spillover, earnings quality better than ordinary foundry Strong group T&S Communications 78 Scarce upstream components, clear benefit path, but customer and order transparency weaker than US component leaders Strong group Cisco 76 Orders have appeared, but AI purity and growth leverage lag Arista Watch group AOI 72 High 800G earnings leverage, but customer concentration and profit sustainability still need verification Watch group Accelink 70 Real benefit, but needs to show ultra-high-speed product share gains more Watch group Sigtone Photonics 68 Sits in a possible bottleneck layer, but scale, customers, and volume cadence need continued verification Allocation group Fujikura 66 Long-term beneficiary of connectors and routing, steady rather than high-leverage Allocation group Corning 64 Important ecosystem position, but the path from AI optical communications to its profit is more indirect Valuation and Market-Expectation Analysis
Those that have already fairly fully priced in AI optical-communications expectations are usually the US system/component leaders with clear AI orders and high growth-stage valuations, such as Arista, Coherent, Lumentum, Marvell, and Broadcom. Especially as the market begins to capitalize the medium-to-long-term stories of 1.6T, CPO, and silicon photonics ahead of time, the margin for error in valuation declines.
Where an expectation gap still exists is often with companies already standing at a key-component or manufacturing bottleneck, but where market understanding still treats them as "ordinary parts suppliers," such as T&S Communications, Sigtone Photonics, and Fabrinet; their stories are not as sexy as the pure CPO theme, but they are more likely to convert profit first.
The classic "good company but too expensive" case is those that are high-certainty, high-quality, and already fully consensus-priced; the "cyclical-bottom reversal + AI demand" combination shows up more in leverage names like AOI, which were once hit by customer cycles but have re-attached to hyperscaler demand.
Risk Analysis
There are five categories of risk most in need of systematic tracking. The first is a slowdown in cloud-vendor AI CapEx. Current public data is still strong, but if token monetization, cloud-leasing prices, or capacity utilization fall short in the second half of 2026, network and optical-module spending will slow ahead of GPUs.
The second is module oversupply and ASP decline. LightCounting has explicitly noted that shortages of components like InP lasers are gradually easing, and that total 2026 Ethernet optical-module capacity could even exceed what customers actually need. This is especially unfavorable for the midstream module-assembly layer.
The third is misjudging the technology roadmap. If LPO/LRO penetration is slower than expected, the "de-DSP" narrative falls flat; if CPO penetration is faster than expected, the long-term valuation anchor of traditional pluggable-module makers could be revised down; and if scale-up continues to be dominated by proprietary interconnect over the long run, part of the Ethernet-expansion expectation will also prove excessive.
The fourth is customer concentration and export restrictions. Arista itself already emphasizes in its risk factors large-customer concentration, single/limited supply sources for key components, and the uncertainty of trade restrictions and tariffs; the same risk is more sensitive for Chinese module makers, except they usually do not disclose specific customers in public sources.
The fifth is valuation-bubble risk. When the market simultaneously prices in continued 800G ramp, early large-scale 1.6T ramp, rapid CPO landing, and silicon-photonics value-chain restructuring, any one of these slowing slightly could trigger a pullback in high-valuation companies. This point must be guarded against separately, especially for optical-communications leaders that are "already expensive with strong consensus."
Final Conclusions and Research List
The importance of AI optical communications within the AI supply chain can be summed up in one sentence: it is the key constraint layer that takes compute expansion from "buying more GPUs" to "actually connecting more GPUs efficiently." Without enough bandwidth, low enough latency, and high enough reliability in high-speed optical interconnect, the larger a GPU cluster grows, the more easily utilization slips and the worse the economics of training and inference become.
For the five sub-segments most worth watching going forward, I would narrow to these five mainlines: 800G/1.6T pluggable optical modules, EML/CW/InP lasers, AI switches and switch chips, key optical components and high-density connectors, and the silicon-photonics/CPO mid-term roadmap. If the goal is to find "revenue leverage, profit leverage, order certainty, and a long-term moat," these five mainlines are more effective than talking about "AI optical communications" in the abstract.
The ten listed companies most worth deeper research include: Innolight, Eoptolink, T&S Communications, Coherent, Lumentum, Fabrinet, Arista, Broadcom, Marvell, and Cisco. The first five lean more toward the "direct optical-benefit chain," and the last five lean more toward the "network-systems and high-value-chip layer"; only by viewing the two groups together can you balance revenue leverage and earnings quality at the same time.
The five private/internalized opportunities most worth continued tracking are: Ayar Labs, Celestial AI, Lightmatter, Ranovus, and Intel Silicon Photonics. What they jointly point to is not "next quarter's financials," but the long-term boundary restructuring of Optical I/O, silicon photonics, CPO, and chip-to-chip interconnect.
The three things the market most easily misunderstands are: first, CPO matters, but the most profitable layer in the near term is still 800G/1.6T pluggables and upstream lasers; second, industry demand growth does not mean every related company can make money, and what decides the winners is customer qualification, volume-production yield, and supply position; third, the rise of Ethernet in AI networking is not a question of "whether it happens," but of "how fast, how far, and how much share it takes."
The indicators most worth tracking over the next 6-12 months, in order, are: cloud-vendor CapEx and the OCI/AI-infra revenue measures, Arista/Cisco AI fabrics and AI-infra orders, Coherent/Lumentum laser and long-term-contract progress, ASP changes in 800G/1.6T modules, the cadence of 200G/lane DSP and 1.6T design-in, and the actual volume-production and customer-deployment cadence of NVIDIA/Broadcom CPO.
For the narrower and more research-efficient directions going forward, I recommend prioritizing: 1.6T optical modules, EML/CW lasers, silicon photonics and CPO, optical-module foundry, and AI data-center switch chips. These five directions make it easier to capture the real earnings leverage and expectation gaps of the next 12-24 months than studying the entire optical-communications chain in a generalized way.
This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.
Full report
Sign in to read the full report
Sign up free to unlock the full text, the Baillie growth scorecard, and full-text search.
Log in / Sign up free