Chapter 11: Photonics, Optical I/O & Co-Packaged Optics

11.1 Overview

Electricity cannot keep up with AI. As GPU clusters scale from thousands to hundreds of thousands of accelerators, the copper traces and traditional optical modules that carry data between them hit fundamental physical limits: signal degradation over distance, power consumption per bit, and connector density per faceplate. Photonics solves these problems by converting electrical signals into light, which travels faster, farther, and with less energy loss than electrons through copper.

This chapter covers the companies that manufacture optical transceivers, laser components, silicon photonics devices, and the emerging co-packaged optics (CPO) technology that integrates photonics directly onto switch and GPU substrates. It sits above the networking silicon layer (Chapter 10), which designs the switch ASICs and NICs that generate the electrical signals, and connects to the connectors and cables layer (Chapter 12), which provides the physical fiber optic infrastructure.

The optical transceiver market was valued at approximately $14.7-15.7 billion in 2025, with LightCounting estimating actual sales of optical transceivers and related products at $23.8 billion including ancillary components ¹²³. The market is growing at 13-17% CAGR depending on the source, but the AI-relevant segment (800G and above datacom transceivers) is growing far faster. The transition from 800G to 1.6T transceivers began shipping in volume in late 2025, and 3.2T transceivers are on the development roadmap for 2027-2028.

The supply chain has three tiers. At the component level, companies like Coherent, Lumentum, Broadcom, and MACOM make the lasers (EMLs, VCSELs, CW lasers), photodetectors, and optical DSPs that go into transceiver modules. At the module level, companies like Innolight (now TeraHop), Coherent, Eoptolink, and Cisco assemble complete transceiver modules. At the contract manufacturing level, Fabrinet is the dominant outsourced assembler, producing transceivers for clients including Lumentum, NVIDIA, Cisco, and AWS.

NVIDIA’s $4 billion investment in Coherent and Lumentum ($2 billion each), announced in March 2026, is the single most significant capital commitment in the photonics supply chain. The deals include multi-billion-dollar purchase commitments, future capacity access rights, and support for new US-based fabrication facilities. NVIDIA is explicitly investing in co-packaged optics, which integrates optical engines directly alongside switch ASICs, eliminating the electrical-to-optical conversion bottleneck at the transceiver. NVIDIA’s Quantum-X Photonics InfiniBand switches and Spectrum-X Photonics Ethernet switches with CPO are expected to ship in 2026 ⁴⁵.

Co-packaged optics represents the next inflection point for data center interconnects, comparable to the transition from pluggable to on-board optics. By embedding photonics on the switch substrate, CPO reduces power consumption by up to 40%, increases bandwidth density, and enables scaling AI clusters to millions of GPUs. However, CPO requires fundamental changes to switch board design, thermal management, and supply chain logistics. Volume CPO deployments are projected for 2027-2028 for scale-out networks, though early products are shipping in 2026 ⁶.

Several private startups operate at the frontier of optical I/O. Ayar Labs develops optical I/O chiplets using TSMC’s processes. Lightmatter builds photonic interconnect fabric (Passage) and photonic computing chips (Envise). Celestial AI was acquired by Marvell for $3.25 billion in late 2025, validating the strategic value of silicon photonics in the data center ⁷. These companies represent the next generation of optical interconnect technology, which could eventually replace electrical signaling even within a single chip package.

11.2 Market Sizing & Growth

Optical transceiver market (total): Valued at $14.7-15.7 billion in 2025 depending on the source, with projections ranging from $25 billion by 2029 (MarketsandMarkets, 13% CAGR) to $46 billion by 2034 (Fortune Business Insights, 17% CAGR). LightCounting’s direct survey of vendors found $23.8 billion in actual 2025 sales when including related optical products ¹²³.

800G and above segment: 800G transceivers are the current volume standard for AI clusters. Cignal AI projects 800G optics shipments grew 60% in 2025. The 1.6T transition is underway, with Coherent, Lumentum, and Innolight all shipping 1.6T transceivers in late 2025 and ramping through 2026 ⁸.

Co-packaged optics: CPO is expected to ramp on both InfiniBand and Ethernet platforms starting in 2026-2027. Dell’Oro Group lists CPO as a key 2026 trend. NVIDIA’s $4 billion investment in Coherent and Lumentum is explicitly targeting CPO laser components ⁴⁶.

Coherent: FY2025 (ending June 2025) revenue was a record $5.81 billion, up 23% year-over-year. Networking segment revenue surged 49% to $3.42 billion. Data center revenue grew 61% for the full year. Q1 FY2026 revenue was $1.6 billion (+17% YoY). The company began shipping 1.6T transceivers and its first optical circuit switch (OCS) in Q4 FY2025. Coherent estimates OCS will add $2 billion to its addressable market by 2030 ⁸⁹¹⁰.

Lumentum: Q2 FY2026 (ending December 2025) revenue was $665.5 million, up 65.5% year-over-year. Q3 FY2026 revenue reached a record $808.4 million, up 90.1% year-over-year. Cloud & Networking represents approximately 88% of revenue. The company holds an estimated 50-60% share of EML lasers and is the only supplier shipping 200G-per-lane EML devices at volume, the key component for 1.6T transceivers. NVIDIA’s $2 billion investment includes funding for a new US fabrication facility ¹¹¹²¹³.

Fabrinet: FY2025 revenue was a record $3.42 billion, up 19% year-over-year. Optical communications represented 76% of sales. Q1 FY2026 (ending September 2025) revenue broke the $1 billion quarterly threshold for the first time, reaching $1.13 billion (+36% YoY). Fabrinet manufactures transceivers for Lumentum, NVIDIA, Cisco, and AWS among others ¹⁴¹⁵.

Innolight (TeraHop): Revenue of $5.27 billion in 2025, making it the largest optical transceiver vendor by revenue. Innolight holds over 50% share of NVIDIA’s optical module procurement. The company was first to complete 1.6T testing with NVIDIA and is expected to retain at least 50% of the global 1.6T market ¹⁶.

MACOM Technology: Supplies drivers, TIAs, and analog components for optical modules. MACOM is the sole supplier of drivers/TIAs for NVIDIA’s in-house transceiver production (via Fabrinet). The company is a critical second-tier component supplier in the optical chain ¹⁶.

Total optical serviceable addressable market (SAM): Coherent disclosed in its Q1 FY2026 earnings that it estimates the total optical SAM at $70 billion by 2030, comprising $50 billion in existing markets (pluggable transceivers, telecom, industrial) and $20 billion in emerging AI-driven applications (CPO, optical circuit switching, advanced coherent) ¹⁰. Goldman Sachs Global Investment Research (April 2026) projects the broader optical networking addressable market will reach approximately $154 billion by 2027-2028, a ninefold increase from the $15 billion base, with inter-rack AI connections alone representing a new $75 billion market. Goldman also projects silicon photonics penetration rising from 6% in Q1 2024 to 45% by Q4 2028 ²⁴. The Coherent figure ($70B) represents a single company’s view of its own addressable market; the Goldman figure ($154B) represents the full ecosystem including networking equipment, fiber, and infrastructure. Both figures confirm the order-of-magnitude expansion underway.

CPO market sizing: extreme divergence. Analyst projections for the co-packaged optics market by 2030 range from $2 billion (Mordor Intelligence) to $8.1 billion (Yole Group) to $130 billion (DigiTimes) ²⁵. This 65x spread in estimates for the same market in the same year reflects fundamental uncertainty about CPO adoption timelines. LightCounting projects pluggable transceivers will retain 80%+ market share through 2030 ³, suggesting CPO remains a niche rather than a wholesale replacement in that timeframe. The report uses the conservative end of these ranges; if CPO adoption accelerates faster than projected, the InP laser demand multiplier (Section 11.5) and future bottleneck timeline (Section 20.4b) would need to be pulled forward by 12-24 months.

NVIDIA optical supply chain commitments: NVIDIA has committed up to $7.7B to secure optical supply: $2B each in Coherent and Lumentum (March 2026) ⁴⁵, plus $500M in Corning with rights to purchase up to $3.2B in stock (May 6, 2026) ¹⁸²². Corning is building 3 new US optical manufacturing plants, hiring 3,000+, and increasing US optical capacity by 10x. Meta signed a $6B anchor deal. These investments signal that optical connectivity is the binding constraint on AI cluster scaling at current GPU densities.

Relative to the $725B capex framework (see Section 1.2), optical transceivers, fiber, and connectors represent approximately 4-7% of total AI infrastructure spending ($30-55B), placing this layer in the mid-range by dollar value but disproportionately high in bottleneck severity due to sole-source dependencies (Lumentum EMLs, Corning fiber).

11.3 Supply Chain Flowchart

PHOTONICS & OPTICAL I/O SUPPLY CHAIN
    |
    |---> UPSTREAM: InP SUBSTRATES & EPITAXY (Tier 0: Materials)
    |    |
    |    |-- InP SUBSTRATES (fewer than 5 volume suppliers globally)
    |    |    AXT Inc. (AXTI): ~35% global share, doubling capacity 2026
    |    |    Sumitomo Electric (5802.T): ~30% share, vertically integrated
    |    |    JX Advanced Metals (ENEOS sub.): investing 1.5B JPY in capacity
    |    |    Coherent: brought 6-inch InP to full production FY2025-2026
    |    |
    |    |-- InP EPITAXY
    |    |    IQE (IQE.L): InP epi wafer foundry, 18-24 month qualification
    |    |    Coherent, Lumentum: internal epi growth
    |    |
    |    |-- MOCVD EQUIPMENT (grows the InP epitaxial layers)
    |    |    AIXTRON (AIXA): ~40-50% MOCVD share (dominant outside China)
    |    |         G10-AsP enables high-volume InP PIC production
    |    |    Veeco (VECO): ~12% MOCVD share, MBE systems
    |    |    Nynomic/LayTec (M7U): monopoly in-situ metrology on MOCVD
    |    |         Sensors embedded in every MOCVD reactor
    |    |
    |    +-- SILICON PHOTONICS FOUNDRIES
    |         GlobalFoundries (GFS): largest SiPh foundry post-AMF acquisition
    |              SCALE CPO platform (May 2026), targeting $1B+ SiPh by 2028
    |         Tower Semiconductor (TSEM): SiPh foundry for Innolight/NVIDIA
    |              CPO foundry technology announced Nov 2025
    |         Intel Silicon Photonics: 21.5% SiPh share, licensing model
    |         X-FAB (XFAB): European SiPh capacity (200mm)
    |              |
    |              v
    |---> OPTICAL COMPONENT MANUFACTURERS (Tier 1: Lasers, Detectors, DSPs)
    |    |
    |    |-- LASER SOURCES
    |    |    Coherent: EMLs, VCSELs, CW lasers, silicon photonics
    |    |         InP capacity tripled YoY, moving to 6-inch wafers
    |    |    Lumentum: EMLs (50-60% share), VCSELs, CW lasers for CPO
    |    |         200G/lane EML sole supplier at volume
    |    |    Broadcom: silicon photonics integration for CPO
    |    |    II-VI legacy --> now part of Coherent (merged 2023)
    |    |    Sivers Semiconductors: InP-based photonic ICs (smaller)
    |    |
    |    |-- OPTICAL DSPs / DRIVERS / TIAs
    |    |    Marvell: PAM4 DSPs, coherent DSPs (market leader)
    |    |    Broadcom: PAM4/coherent DSPs, SerDes
    |    |    MACOM: drivers, TIAs (sole supplier for NVIDIA in-house)
    |    |    Semtech (SMTC): drivers, TIAs, signal conditioning
    |    |         Content per module scales from ~$8 (800G) to ~$80 (3.2T)
    |    |    MaxLinear: PAM4 DSPs (smaller share)
    |    |
    |    +-- SILICON PHOTONICS PLATFORMS
    |         Intel Silicon Photonics (integrated PIC platform)
    |         Coherent (internal SiPh for 1.6T transceivers)
    |         GlobalFoundries (SiPh foundry services, SCALE CPO platform)
    |         Tower Semiconductor (SiPh foundry for Innolight, others)
    |              |
    |              v
    |---> OPTICAL TRANSCEIVER MODULE MAKERS (Tier 2: Complete modules)
    |    Innolight/TeraHop: ~50%+ of NVIDIA procurement, largest by revenue
    |    Coherent: vertically integrated (components + modules)
    |    Eoptolink: growing Chinese supplier, entering NVIDIA supply chain
    |    Cisco: via Acacia (acquired 2021) and Luxtera acquisitions
    |    Lumentum: via Cloud Light acquisition (2023), vertically integrated
    |    Broadcom: silicon photonics-based transceivers
    |    Applied Optoelectronics (AAOI): 400G/800G modules
    |    Hisense Broadband: Chinese volume manufacturer
    |         |
    |         v
    |---> CONTRACT MANUFACTURING (Tier 3: Assembly & test)
    |    Fabrinet: dominant (~20%+ of outsourced optical assembly)
    |         $3.42B FY2025 revenue; assembles for Lumentum, NVIDIA, Cisco, AWS
    |    Jabil (JBL): emerging second source, own 1.6T OSFP transceiver
    |         $500M+ SE US manufacturing investment for AI DC infrastructure
    |         |
    |         v
    |---> OPTICAL TEST & MEASUREMENT (Tier 3b: Production gating)
    |    Keysight (KEYS): 800G/1.6T/CPO test platforms, DCA-M oscilloscopes
    |    Viavi (VIAV): production test, TestCenter D2 1.6T appliance
    |    Anritsu (6754.T): BERTWave for 10G-1.6T transceiver eval
    |    Teradyne/Quantifi (TER): PIC automated test (acquired Q2 2025)
    |         |
    |         v
    |---> OPTICAL NETWORKING SYSTEMS (Tier 4: DCI transport)
    |    Nokia (NOK): optical transport, InP PICs (ex-Infinera, $2.3B Feb 2025)
    |         AI & Cloud optical revenue +49% in Q1 2026
    |    Ciena (CIEN): DCI coherent transport, WaveLogic 6
    |         |
    |         v
    |---> CO-PACKAGED OPTICS (CPO) -- emerging, 2026-2028 volume ramp
    |    NVIDIA: Quantum-X Photonics (IB) + Spectrum-X Photonics (Ethernet)
    |         Uses Lumentum lasers + Coherent silicon photonics
    |    Broadcom: Tomahawk CPO (Humboldt, Bailly, Davisson series)
    |    Marvell: Celestial AI acquisition ($3.25B) for photonic fabric
    |    Ayar Labs: optical I/O chiplets (TSMC process) -- Private
    |    Lightmatter: Passage photonic interconnect fabric -- Private
    |         |
    |         v
    +---> OPTICAL CIRCUIT SWITCHING (OCS) -- emerging
         Coherent: liquid-crystal OCS (first revenue Q4 FY2025)
         Lumentum: MEMS-based OCS (R64 switch)
         Google: internal OCS deployment (Project Apollo)
         Polatis (acquired by Huber+Suhner): beam-steering OCS

11.4 Key Companies

Company	Ticker	Exchange	Approx. Mkt Cap	Role in Buildout	Key Metric
Coherent Corp	COHR	NYSE	~$65.6B	Vertically integrated: EMLs, VCSELs, CW lasers, SiPh, transceivers, OCS	FY2025 revenue $5.81B (+23% YoY); NVIDIA $2B investment; stock +87% YTD
Lumentum Holdings	LITE	NASDAQ	~$7.2B	EML lasers (50-60% share), VCSELs, CW lasers for CPO, transceivers, OCS	Q3 FY2026 revenue $808M (+90% YoY); NVIDIA $2B investment; stock +156% YTD
Fabrinet	FN	NYSE	~$25.0B	Dominant contract manufacturer for optical transceivers and assemblies	FY2025 revenue $3.42B (+19%); Q1 FY2026 $1.13B (+36%)
Innolight (TeraHop)	300308	Shenzhen	~$136B	Largest optical transceiver vendor globally; >50% of NVIDIA procurement	2025 revenue $5.27B; first to ship 1.6T to NVIDIA
Marvell Technology	MRVL	NASDAQ	~$149B	PAM4/coherent DSPs; acquired Celestial AI ($3.25B) for photonic fabric	Electro-optics revenue $819M in Q3 FY2026 (+2.6x YoY)
Broadcom	AVGO	NASDAQ	~$2.0T	SiPh integration, CPO (Tomahawk CPO series), PAM4 DSPs	CPO versions of Tomahawk shipping since 2022
NVIDIA	NVDA	NASDAQ	~$5.2T	CPO switches (Quantum-X/Spectrum-X Photonics); $4B invested in optics	Spectrum-X Photonics Ethernet CPO switches in 2026
Cisco Systems	CSCO	NASDAQ	~$381B	Transceivers (via Acacia/Luxtera), SiPh, 800G LPO technology	~1/3 of AI orders are optics; owns Acacia coherent DSPs
MACOM Technology	MTSI	NASDAQ	~$27.5B	Drivers, TIAs, analog ICs for optical modules; sole supplier for NVIDIA in-house	Critical component supplier; GaAs and InP capability
Ciena	CIEN	NYSE	~$77.5B	DCI coherent optical transport; WaveLogic 6 technology	FY2025 revenue $4.77B (+19%); #1 in DCI globally
Eoptolink	300502	Shenzhen	~$75.6B	High-speed Ethernet transceivers; entering NVIDIA supply chain	2024 revenue $1.2B (+175%); 800G leader in MCF
Applied Optoelectronics	AAOI	NASDAQ	~$12.0B	400G/800G transceiver modules for data centers	Received multi-million-dollar 800G order from NA hyperscaler
Nokia	NOK	NYSE	~$70.0B	Optical networking post-Infinera ($2.3B, Feb 2025); InP PIC fab	AI & Cloud optical revenue +49% Q1 2026; combined optical ~$3.6B
Corning	GLW	NYSE	~$161B	Optical fiber (sold out through 2026); NVIDIA $500M + $3.2B rights (May 2026)	10x US optical capacity expansion; 3 new plants; Meta $6B anchor
Keysight Technologies	KEYS	NYSE	~$62.0B	Dominant optical/photonic test equipment; 800G/1.6T/CPO platforms	FY2025 revenue $5.38B (+8%); strong SiPh testing growth
GlobalFoundries	GFS	NASDAQ	~$40.6B	Largest SiPh foundry (post-AMF acquisition Nov 2025); SCALE CPO platform	SiPh revenue doubled 2025, targeting $1B+ by 2028
Tower Semiconductor	TSEM	NASDAQ	~$25.1B	SiPh foundry for Innolight (NVIDIA’s #1 optical supplier); CPO foundry	Expanded Innolight collaboration Mar 2025; 1.6T SiPh scaling; stock ~3x from 2025 lows
Semtech	SMTC	NASDAQ	~$11.3B	Laser drivers, TIAs, signal conditioning; content scales 10x at 3.2T	FY2026 revenue ~$1.05B; Signal Integrity >$300M
AXT Inc	AXTI	NASDAQ	~$7.6B	~35% global InP substrate production; pure-play III-V substrate company	InP backlog $60M to $100M+; doubling capacity 2026
Viavi Solutions	VIAV	NASDAQ	~$12.0B	Optical production test equipment; 1.6T test appliance	Q2 FY2026 revenue $369M (+36%); NSE driven by DC demand
AIXTRON	AIXA	Frankfurt	~$5.6B	~40-50% MOCVD equipment share (dominant outside China); InP PIC enabler	G10-AsP system for high-volume InP production
Jabil	JBL	NYSE	~$37.5B	Emerging second-source optical assembly; own 1.6T OSFP transceiver	$500M+ SE US manufacturing; Ottawa photonic packaging
IQE	IQE	LSE AIM	~$580M	InP and GaAs epitaxial wafer foundry; multiple Tier 1 AI/hyperscale wins	Launching 6-inch foundry platform for SiPh; 18-24 month qual cycles
Nynomic (LayTec)	M7U	Frankfurt	~EUR 150M	Monopoly in-situ metrology on MOCVD reactors; sensors inside every InP tool	LayTec sensors embedded in AIXTRON/Veeco MOCVD equipment
POET Technologies	POET	TSX-V/NASDAQ	~$1.7B	Optical interposer platform for photonic integration	Early-stage; optical interposer could reduce CPO costs
Sivers Semiconductors	SIVE	Nasdaq Stockholm	~$1.2B	InP-based photonic ICs; optical engines for data centers	Small-cap component supplier; InP foundry services
Ayar Labs	Private	Private	~$3.8B (Series E, Mar 2026)	Optical I/O chiplets on TSMC process; in-package photonics	$500M Series E; UCIe-compliant TeraPHY; 2026 production
Lightmatter	Private	Private	~$4.4B (last round)	Passage photonic interconnect fabric; Envise photonic compute	Passage M1000: 114 Tbps; UALink member

11.5 Bottleneck Analysis

EML laser supply for 1.6T transceivers (SEVERE): Lumentum holds an estimated 50-60% share of EML lasers and is the only supplier shipping 200G-per-lane EMLs at volume. These are the critical component enabling 1.6T transceivers. Lumentum’s indium phosphide wafer fab capacity is fully allocated. CEO Michael Hurlston quantified the gap on the Q3 FY2026 call: “We are significantly undershipping demand…supply-demand imbalance is probably even higher than we reported…somewhere greater than 30%.” All capacity is committed under long-term agreements: “even as we increase the capacity, everything is spoken for” ²⁷. The company is investing in a new US fabrication facility (partly funded by NVIDIA’s $2 billion), but new fab capacity takes 18-24 months to come online. Until additional EML capacity ramps, the 1.6T transition is constrained by laser supply ¹¹¹².

Indium phosphide (InP) wafer and laser supply chain (SEVERE, future Tier 1 candidate): Industry estimates suggest global InP wafer demand reached approximately 2 million units in 2025, against production capacity of roughly 600,000 units, implying a structural supply gap of approximately 70%. Coherent management confirmed InP as the binding constraint on their Q1 FY2026 call: “Given the exceptionally strong demand environment and industry-wide constraints in indium phosphide, capacity expansion remains a highest priority” ²⁶. The gap is being addressed through capacity expansion (AXT, Coherent moving to 6-inch with yields already surpassing legacy 3-inch lines) but will take 2-3 years to close. The full InP dependency chain has four steps, each with concentrated suppliers:

InP substrates: AXT Inc. (AXTI, NASDAQ; MACOM is a major customer and supply chain partner, not the parent company) holds ~35% of global InP substrate production from its Tongmei Beijing facility (the world’s largest III-V wafer fab by the company’s claim, ~300,000 sq ft). AXT is doubling capacity through 2026 and currently produces 2-inch, 3-inch, and 4-inch InP wafers. The industry-wide transition to 6-inch InP wafers is underway: Coherent brought 6-inch InP into full production in FY2025-2026, while AXT pioneered the format historically but currently ships primarily 4-inch at volume. Other substrate suppliers: Sumitomo Electric (vertically integrated from ingot to finished fiber), JX Advanced Metals. Total global InP substrate suppliers number fewer than five at volume.
InP epitaxial wafers: IQE (LSE AIM, ~$580M) supplies InP epi wafers to multiple laser vendors. Coherent and Lumentum also grow epi internally. Qualification cycles for new epi suppliers are 18-24 months.
Laser dies (EML, CW, DFB): Lumentum (50-60% EML share), Coherent (vertically integrated, tripled InP capacity YoY in FY2025, transitioning to 6-inch wafers), Sumitomo Electric. For CPO external light sources, CW DFB laser arrays are the critical component, distinct from the EMLs used in pluggable transceivers. Each CPO-equipped switch needs 8-16 CW laser channels, multiplying per-rack laser demand versus pluggable architectures.
Laser driver ICs: Every CW laser array requires a driver IC from companies like Semtech or MaxLinear. This is a secondary chokepoint in the same module that receives almost no analytical attention.

Shared constraint with telecom: Lumentum and Coherent also supply EDFA (erbium-doped fiber amplifier) pump lasers for submarine and long-haul fiber networks. These compete for the same InP manufacturing capacity as data center CPO lasers. The demand model for InP should include both data center and telecom/submarine demand sources; most analysis treats them separately.

As CPO adoption scales (projected 2028-2030 for volume), the InP laser supply chain transitions from a Tier 2 bottleneck (current, RPN 128) toward potential Tier 1 severity. The 4-inch to 6-inch InP wafer transition, analogous to the painful 200mm-to-300mm silicon wafer transition, will determine whether laser cost comes down fast enough for CPO economics to work at scale. InP wafer yields remain sub-30%, structurally worse than silicon.

Fabrinet as contract manufacturing concentration (MODERATE-HIGH): Fabrinet (FN, NYSE, ~$25.0B) is the dominant outsourced precision optical assembly company, essentially filling the role TSMC plays for chip fabrication. Q1 FY2026 revenue reached $1.13B (+36% YoY), its first billion-dollar quarter ¹⁵. FY2025 revenue was $3.42B ¹⁴. NVIDIA accounts for 27.6% of revenue (down from 35.1% in FY2024 as the customer base diversifies); Cisco is 18.2%. Fabrinet is building “Building 10” in Thailand (2M sq ft, completion end of 2026) to double 1.6T transceiver assembly capacity, adding approximately $3B in revenue capacity ¹⁵. Competitors exist (Benchmark Electronics, Celestica), but Fabrinet’s moat is operational excellence and customer qualification stickiness, not structural monopoly. Switching contract manufacturers requires 6-12 months of requalification.

Optical DSP concentration (MODERATE-HIGH): Marvell and Broadcom dominate the PAM4 and coherent DSP market for optical transceivers. The DSP is the intelligence inside each transceiver, and it accounts for roughly 50% of module power consumption. NVIDIA is developing its own optical DSPs for in-house transceivers (produced via Fabrinet), using MACOM for analog components. If NVIDIA’s internal program scales, it could reduce dependence on Marvell/Broadcom DSPs but also compress the addressable market for third-party DSP vendors ¹⁶.

Co-packaged optics integration complexity (MODERATE, emerging): CPO requires changes to switch board design, thermal management (lasers are temperature-sensitive), and testing methodology. The existing switch ecosystem (Arista, Cisco, white-box ODMs) is optimized for pluggable transceivers. Transitioning to CPO means these system vendors must redesign their products, which creates adoption friction. The 2026-2028 timeline for volume CPO reflects this integration complexity, not just silicon readiness ⁶.

MOCVD equipment and metrology for InP expansion (MODERATE-HIGH, second-order): Expanding InP laser capacity requires MOCVD reactors. AIXTRON (AIXA, Frankfurt, ~$5.6B) holds approximately 40-50% of the global MOCVD market (dominant outside China; Chinese competitor AMEC is gaining share domestically). Every MOCVD reactor requires in-situ metrology sensors from Nynomic’s LayTec subsidiary (M7U, Frankfurt, ~EUR 150M), which holds a monopoly position in this niche. The dependency chain mirrors the Trumpf-Zeiss-ASML pattern: LayTec metrology enables AIXTRON equipment, which enables InP epitaxy, which enables laser production. If InP capacity is the binding constraint on CPO adoption, AIXTRON and LayTec are the equipment bottlenecks constraining the capacity expansion itself. Equipment lead times for MOCVD systems are 12-18 months.

Silicon photonics foundry concentration (MODERATE-HIGH): The SiPh chips inside optical transceivers are manufactured by a small number of foundries. Tower Semiconductor (TSEM, NASDAQ, ~$25.1B) is the SiPh foundry partner for Innolight, which supplies over 50% of NVIDIA’s optical modules. GlobalFoundries (GFS, NASDAQ, ~$40.6B) acquired Advanced Micro Foundry in November 2025 to become the largest pure-play SiPh foundry, launching its SCALE CPO platform in May 2026 and targeting $1B+ in SiPh revenue by 2028 ¹⁹. If Tower cannot scale SiPh capacity fast enough, Innolight’s 1.6T production is constrained regardless of laser or DSP supply. This is a hidden dependency that receives minimal analyst attention because Tower’s role is buried inside Innolight’s vertically-presented modules.

Optical test equipment as production gating factor (MODERATE): Every optical transceiver and photonic IC must pass production test before shipment. As the industry scales from millions to tens of millions of 1.6T units annually, test equipment throughput becomes a capacity constraint. Keysight Technologies (KEYS, NYSE, ~$62.0B) dominates optical/photonic test with DCA-M sampling oscilloscopes and FlexOTO software for CPO high-lane-count testing. Viavi Solutions (VIAV, NASDAQ, ~$12.0B) provides production-line test appliances. Teradyne acquired Quantifi Photonics in Q2 2025 specifically for PIC automated test. Test throughput can gate production ramp speed even when component supply is adequate.

Chinese supplier geopolitical risk (MODERATE): Innolight/TeraHop and Eoptolink, both Chinese companies, together represent a majority of global optical transceiver shipments. Innolight alone holds over 50% of NVIDIA’s optical module procurement. If US-China trade restrictions expanded to cover optical transceivers, the supply chain would face severe disruption. To date, optical modules have not been subject to the same export controls as semiconductor equipment or advanced chips, but the risk is non-trivial given the strategic importance of the technology ¹⁶.

11.6 Risks

Linear pluggable optics (LPO) disrupts traditional transceivers: LPO technology removes the DSP from the transceiver module, relying instead on the host switch’s DSP. This reduces transceiver power consumption by 50% or more (from 7-9W to 2-4W for 400G) and cuts latency by up to 90%. If LPO gains widespread adoption, it commoditizes the transceiver module and shifts value to the switch ASIC vendors (Broadcom, NVIDIA). Transceiver vendors like Coherent and Innolight would see lower ASPs. Cisco is actively pushing 800G LPO technology ¹⁷.

CPO cannibalization of pluggable transceivers: If CPO scales as projected, it integrates optical engines directly onto switch substrates, reducing or eliminating the need for separate pluggable transceiver modules. This is a structural threat to pure-play transceiver vendors. Coherent and Lumentum are hedging by supplying the CW lasers and photonic ICs for CPO, but their transceiver assembly revenue could decline. Innolight and Eoptolink, which are primarily module assemblers without deep component integration, are most exposed ⁶.

Chinese pricing pressure: Chinese optical transceiver vendors price 20-25% below Western incumbents. Innolight and Eoptolink have achieved excellent quality and reliability, and they are rapidly gaining share outside China as well. This creates structural margin pressure for Western vendors. Coherent and Lumentum are responding with vertical integration (owning more of the component stack) and differentiation through advanced technologies like OCS and CPO lasers, but the pricing gap persists ³.

Overcapacity risk in 2027: LightCounting warns that optical component and transceiver production capacity is catching up with demand, and this could lead to more competition and sharper price declines by the end of 2026. The optical industry has a history of boom-bust cycles: periods of shortage drive capacity investment, which then leads to oversupply and margin compression. If AI capex slows while new capacity comes online, the photonics supply chain could experience a painful correction ³.

First principles check: Is the optical bottleneck real? Yes. Copper signaling doubles in power consumption per bit for each halving of reach at a given data rate. At 200G per lane (the 1.6T transceiver standard), copper can carry signals only a few centimeters before requiring signal conditioning (retimers). Beyond the edge of a server, optical fiber is the only viable medium for high-bandwidth data transfer. The physics is unambiguous: as AI clusters scale, the proportion of total system cost and power consumed by interconnects grows, and optical solutions are the only path to managing that growth.