Optical I/O for AI Chips Market Size & Share 2026-2035

Optical I/O for AI Chips Market Size

The global optical I/O for AI chips market was valued at USD 1.1 billion in 2025 and is estimated at USD 1.5 billion in 2026. It is projected to reach USD 14 billion by 2035, advancing at a 27.8% CAGR from 2026 to 2035.

The expansion reflects a change in where the network becomes the limiting resource in AI systems: as accelerator clusters add more devices, the electrical distance between compute, switching silicon, and front-panel optics consumes a growing share of the power and signal-integrity budget. Optical I/O addresses that constraint through a spectrum of architectures, from linear pluggables at the rack edge to engines located beside, or within, an ASIC package.

The near-term revenue base remains weighted toward linear pluggable optics (LPO), which offer a relatively serviceable path to 800G and 1.6T links. The longer-duration growth pool is co-packaged optics (CPO), near-package optics (NPO), and in-package optical I/O (OIO), where shorter electrical paths can reduce the energy burden of moving data. Broadcom introduced its 51.2-Tbps CPO switch platform in March 2024, anchoring the shift from laboratory demonstrations toward commercial switch architectures ^{[1]Broadcom, "51.2-Tbps Co-Packaged Optics Ethernet Switch Platform," investors.broadcom.com.}. The commercial question is therefore no longer simply whether optics replace copper in AI fabrics; it is which integration tier provides enough power relief to justify the accompanying qualification, repair, and packaging complexity.

Silicon photonics is the principal volume platform because it can combine photonic functions with semiconductor manufacturing disciplines, although it still depends on lasers and other active optical components that may sit outside the silicon die or be integrated heterogeneously. The addressable market includes silicon photonics (SiPh), indium phosphide (InP), thin-film lithium niobate (TFLN), heterogeneous/hybrid platforms, and other photonic materials across AI training, AI inference, data center networking and switching, and HPC. Demand is concentrated among hyperscalers and cloud providers, whose scale makes port-level power savings economically material, while enterprise and sovereign AI projects broaden the buyer base over the forecast period ^{[2]Nature Electronics, "Co-Packaged Optics for High-Performance Computing and Artificial Intelligence," nature.com.}.

This sequencing favors suppliers that can support multiple integration tiers rather than bet on a single form factor. It also means that revenue growth will not map cleanly to unit shipments: deeply integrated solutions carry more photonic and packaging content per deployed link, whereas LPO retains substantial volume where fleet operators prioritize interchangeable modules and supply diversity. The forecast therefore reflects a layered migration, with pluggables funding current-scale deployments while co-packaged and in-package designs address the increasingly severe bandwidth-per-watt constraint.

Coverage spans 2022-2035, with 2025 as the base year and 2026-2035 as the forecast period. The assessment covers product type, technology, application, and end-user segmentation, alongside North America, Europe, Asia Pacific, Latin America, and the Middle East & Africa.

Key Drivers

AI-cluster scale raises optical content per compute deployment

AI training fabrics require dense, low-latency links among accelerators, switches, storage, and control layers. As port speeds rise, electrical loss and equalization requirements become harder to manage at the front panel. This makes optics a capacity-enabling input rather than a peripheral component, particularly in scale-out networks where a large cluster multiplies the number of high-speed links. OIF’s 2025 interoperability demonstration of 112G, 224G, and 448G common electrical I/O signals illustrates the industry’s preparation for higher lane speeds ^{[3]OIF, "448G, 224G and 112G CEI Interoperability Demonstration," oiforum.com.}.

CPO commercializes a power-led switch architecture

By locating optical engines adjacent to a switch ASIC, CPO removes much of the long electrical path between switching silicon and detachable transceivers. Broadcom’s 51.2-Tbps platform integrates eight 6.4-Tbps optical engines and reports a 70% power reduction relative to conventional pluggable approaches. That differential has particular relevance in AI halls where network power adds to already high accelerator and cooling loads. The benefit is strongest where a buyer can standardize an architecture across a large installed fleet and absorb the resulting co-design process.

Domestic photonics investment strengthens supply-side capacity

The U.S. Department of Commerce announced CHIPS incentives for Infinera in January 2025, including support for indium phosphide fabrication and advanced packaging ^{[4]NIST and U.S. Department of Commerce, "CHIPS Incentives Award for Infinera," nist.gov.}. It separately announced preliminary terms with Coherent for expansion of InP wafer fabrication in Sherman, Texas. These measures matter because InP lasers and related active components remain critical inputs even when silicon photonics is used for passive routing and modulation. Manufacturing support can reduce an input bottleneck, but it does not eliminate the lengthy qualification requirements of high-speed optical components.

Standards move high-speed optics from bespoke design toward repeatable procurement

Common electrical interfaces and interoperable optical specifications determine whether buyers can source across vendors without rewriting a switch or accelerator design. OIF’s 800ZR implementation agreement and updates to 400ZR provide one example of how interface agreements support scalable high-capacity connectivity. For AI chips, higher-speed electrical I/O work has a similar commercial effect: it gives component developers a clearer target for 1.6T and beyond, while allowing system manufacturers to separate performance qualification from a single supplier relationship.

Photonic integration improves the economics of bandwidth density

Higher lane rates require tighter coordination among drivers, modulators, photodetectors, packaging, and thermal management. Marvell demonstrated a 200G 3D silicon photonics engine in 2024, showing the importance of integrating optical and electronic functions for accelerated infrastructure ^{[5]Marvell Technology, "200G 3D Silicon Photonics Engine," investor.marvell.com.}. Technology development is not only about maximum bit rate; lower-loss routing, lower drive voltage, and more compact optical engines can determine whether an architecture fits a constrained power and package budget.

Key Restraints

Advanced optical packaging carries a cost and yield penalty

CPO, NPO, and OIO concentrate several delicate functions in a small physical envelope: laser coupling, photonic die attachment, high-speed electrical interfaces, thermal paths, and fiber management. A module can be replaced at the front panel; a failed co-packaged optical engine may require a more involved service action. This changes qualification thresholds and makes early adoption most viable for operators with high port volumes, in-house systems engineering, and an explicit power-constrained deployment case ^{[6]Semiconductor Engineering, "Co-Packaged Optics in AI Data Centers," semiengineering.com.}.

Ecosystem maturity lags the most ambitious integration roadmaps

The industry has demonstrated common electrical interfaces, but mechanical, optical, thermal, test, and field-service practices remain architecture-specific at the advanced integration tiers. The consequence is a split market. LPO can scale where buyers need rapid deployment and vendor choice, while CPO and OIO need coordinated participation by ASIC suppliers, photonics vendors, foundries, contract manufacturers, and cloud operators. That coordination requirement can defer revenue even when the theoretical energy advantage is clear.

The supply chain implication is consequential. Incentives for InP capacity can ease availability of active optical components, yet they cannot substitute for design-for-test and serviceability discipline. Vendors that package photonic performance with credible manufacturing and lifecycle support are likely to move ahead of suppliers whose differentiation is confined to laboratory bandwidth. This creates room for LPO to remain commercially relevant even as CPO grows faster.

Optical I/O for AI Chips Market Segment Analysis

By Product Type

LPO is valued at USD 696.98 million in 2025, USD 932.22 million in 2026, and USD 4,914.54 million in 2035, representing a 20.29% CAGR. Its position reflects its compatibility with operationally familiar pluggable architectures. CPO rises from USD 113.66 million in 2025 to USD 186.44 million in 2026 and USD 3,931.63 million in 2035 at a 40.32% CAGR. Its growth is tied to switch-side power density and the evidence of commercial deployment from platform suppliers.

NPO advances from USD 88.22 million to USD 138.94 million and USD 2,527.48 million, at 38.03%, offering a compromise between electrical reach reduction and separable optical-engine serviceability. OIO grows from USD 70.97 million to USD 111.04 million and USD 1,965.82 million at 37.62%, but requires accelerator packaging architectures that can allocate optical interfaces near compute. OBO records USD 132.89 million, USD 175.16 million, and USD 702.08 million, respectively, at 16.68%; it serves a transitional role where board-level proximity is useful but full co-packaging is premature.

By Technology

SiPh expands from USD 611.30 million in 2025 to USD 875.22 million in 2026 and USD 9,548.25 million by 2035, at 30.41% CAGR. It provides the principal route to volume photonic integration, but its reliance on laser sources makes hybrid supply chains persistent. InP rises from USD 265.78 million to USD 350.33 million and USD 1,404.15 million, at 16.68%, retaining a critical role in active optical functions and benefiting from domestic capacity investment.

TFLN moves from USD 145.62 million to USD 198.91 million and USD 1,404.15 million at 24.25%; its commercial relevance depends on translating modulator performance into manufacturable wafer processes. Heterogeneous/hybrid integration increases from USD 53.72 million to USD 83.13 million and USD 1,404.15 million, the fastest technology growth at 36.90%, because it combines the strengths of silicon, III-V materials, and other photonic layers. Other platforms progress from USD 26.30 million to USD 36.22 million and USD 280.83 million, at 25.55%.

By Application

AI training totals USD 588.68 million in 2025, USD 815.84 million in 2026, and USD 6,739.94 million in 2035, at 26.44% CAGR. Large synchronized training fabrics place a premium on high-radix switching and low-latency connectivity, making this the main early market for 800G and advanced switch optics. AI inference grows from USD 154.95 million to USD 229.20 million and USD 3,089.14 million at 33.51%, as production serving disperses demand across more facilities and network tiers. Data center networking and switching rises from USD 259.00 million to USD 363.39 million and USD 3,369.97 million, at 28.08%; it includes the broader east-west and campus connectivity that supports AI operations. HPC and scientific computing reaches USD 100.09 million, USD 135.38 million, and USD 842.49 million, at 22.52%, where procurement cycles are typically more qualification-led.

By End User

Hyperscalers and cloud providers represent USD 752.68 million in 2025, USD 1,053.35 million in 2026, and USD 9,548.25 million in 2035, at 27.75% CAGR. Their concentrated port volumes make them the first economic adopters of CPO. Enterprise data centers expand from USD 99.81 million to USD 152.01 million and USD 2,387.06 million, at 35.80%; the higher rate reflects a smaller base and the spread of private AI infrastructure, although integration choices will remain more conservative. HPC and research institutions advance from USD 190.01 million to USD 251.76 million and USD 1,123.32 million at 18.08%. Government and defense rises from USD 60.23 million to USD 86.69 million and USD 982.91 million, at 30.97%, supported by secure-compute and domestic-supply priorities evidenced by photonics manufacturing programs.

Application and end-user patterns reinforce that distinction. Training clusters and hyperscalers are the logical proving ground because they concentrate traffic, procurement authority, and power cost. The faster inference and enterprise growth rates point to a later broadening of demand, but they do not imply identical product choices. Inference operators can value modularity and operational resilience more highly than the maximum energy reduction available from an accelerator-adjacent optical engine.

Optical I/O for AI Chips Market Regional Analysis

North America

North America is valued at USD 637.88 million in 2025, USD 871.66 million in 2026, and USD 6,178.28 million in 2035, at 24.31% CAGR. The United States accounts for USD 602.52 million in 2025, USD 822.94 million in 2026, and USD 5,807.42 million in 2035, at 24.25%. Canada records USD 35.36 million in 2025 and USD 370.86 million in 2035, at 25.30%. The region combines hyperscaler demand, switch and photonics design leadership, and policy support for active optical-component manufacturing. This co-location enables faster CPO qualification but also concentrates purchasing power among a limited set of buyers.

Europe

Europe progresses from USD 339.30 million in 2025 to USD 489.27 million in 2026 and USD 5,616.62 million in 2035, at 31.15% CAGR. Germany grows from USD 75.21 million to USD 1,234.77 million by 2035 at 31.04%; the United Kingdom rises from USD 78.34 million to USD 1,459.28 million at 32.70%; and France advances from USD 73.65 million to USD 1,122.52 million at 30.07%. Spain reaches USD 392.88 million from USD 34.48 million at 26.26%, Italy reaches USD 449.01 million from USD 35.26 million at 27.71%, and Rest of Europe reaches USD 902.03 million from USD 41.57 million at 34.69%. Russia moves from USD 0.78 million to USD 56.13 million at 49.50%, reflecting a low starting point. European growth is associated with local AI infrastructure and photonics capability; the commercial challenge is converting research and policy strengths into qualified, high-volume supply.

Asia Pacific

Asia Pacific increases from USD 84.82 million in 2025 to USD 122.32 million in 2026 and USD 1,404.15 million in 2035, at 31.15% CAGR. China rises from USD 33.17 million to USD 589.82 million at 32.09%; Japan progresses from USD 17.35 million to USD 280.88 million at 30.86%; South Korea increases from USD 14.65 million to USD 224.75 million at 30.16%; and Taiwan advances from USD 10.03 million to USD 168.63 million at 31.36%. Australia grows from USD 2.70 million to USD 56.13 million at 34.13%, and Rest of APAC rises from USD 6.93 million to USD 83.94 million at 27.06%. The region’s role is dual: it is a demand market and a component/manufacturing base. That position creates opportunities for local supply chains, while technology-access conditions and advanced packaging availability can shape which integration tiers scale first.

Latin America

Latin America advances from USD 22.05 million in 2025 to USD 30.88 million in 2026 and USD 280.83 million in 2035, at 27.80% CAGR. Brazil rises from USD 8.12 million to USD 123.36 million at 30.34%; Mexico from USD 5.22 million to USD 90.27 million at 32.22%; Argentina from USD 1.73 million to USD 28.08 million at 31.31%; and Rest of Latin America from USD 6.99 million to USD 39.12 million at 16.65%. Deployment is more likely to follow cloud and colocation expansion than domestic optical-engine production, favoring interoperable, serviceable architectures in the early years.

Middle East & Africa

MEA rises from USD 18.66 million in 2025 to USD 29.69 million in 2026 and USD 561.66 million in 2035, at 38.64% CAGR. Saudi Arabia advances from USD 3.93 million to USD 196.58 million at 46.68%; the UAE from USD 3.26 million to USD 179.53 million at 48.24%; South Africa from USD 1.64 million to USD 56.17 million at 40.64%; and Rest of MEA from USD 9.82 million to USD 129.38 million at 26.47%. These rates are sensitive to the execution of sovereign and hyperscale infrastructure plans. Where projects proceed, they can create concentrated demand for high-speed optical networking; where timelines slip, a small initial base can magnify the effect on realized growth.

MEA’s high forecast rate should be read as a timing-sensitive infrastructure opportunity rather than evidence of a mature local optical supply chain. Latin America is similarly demand-led, although at a lower scale. Across both regions, LPO and other modular options provide a more plausible initial route than deeply integrated CPO, because they limit field-service exposure while network operators establish high-speed AI capacity. Regional winners will be determined as much by qualification, logistics, and local support as by a photonic engine’s headline bandwidth.

Optical I/O for AI Chips Market Share & Competitive Landscape

Marvell Technology holds 68% of market share, followed by Coherent Corp at 12%, Lumentum Holdings at 7%, Ayar Labs at 2.5%, and other suppliers at 7%. The distribution reflects the importance of a portfolio that spans optical and electrical semiconductor content, photonic engines, lasers, and system relationships. Marvell’s announced acquisition of Celestial AI in December 2025 is particularly significant because it brings in-package optical fabric capabilities into a company already active in data infrastructure ^{[7]Marvell Technology, "Agreement to Acquire Celestial AI," investor.marvell.com.}.

Marvell Technology, Coherent Corp, Lumentum Holdings, Broadcom Inc., and Ayar Labs form the global key-player group. Broadcom’s 2024 CPO switch launch establishes it as a reference supplier for switch-side integration. Coherent’s planned InP expansion supported through CHIPS incentives reinforces the importance of laser and component supply. Lumentum reported fiscal 2025 results that highlight its cloud and networking exposure, while Ayar Labs’ USD 155 million financing in December 2024 supports its effort to commercialize optical I/O chiplets for AI infrastructure ^{[8]Ayar Labs, "USD 155 Million Financing for Optical I/O," ayarlabs.com.}.

North American participants include Intel Corporation, Ranovus Inc., MACOM Technology Solutions, Lightmatter, Celestial AI, Ciena Corporation, and POET Technologies Inc. Intel has demonstrated an optical compute interconnect chiplet co-packaged with a CPU. Ranovus participates in integrated optical-engine development; MACOM supplies high-speed analog connectivity components and announced 1.6T-focused solutions in 2025. Lightmatter is developing photonic interconnect products for AI systems. Ciena addresses coherent networking at campus and data-center interconnect layers, while POET introduced an 800G optical module for AI networks in March 2024.

In Asia Pacific, Sumitomo Electric Industries, Hamamatsu Photonics, Furukawa Electric Co., NewPhotonics, and PhotonicX AI Pte. Ltd. extend the supplier base across fiber, optical components, semiconductor devices, and AI interconnect development. Scintil Photonics and Salience Labs represent the European group, with emphasis on silicon-photonics and optical-switching innovation. Nexus Photonics is the designated niche/disruptor participant, focused on advanced photonic integration. Their competitive relevance depends on converting differentiated integration approaches into a manufacturable, qualified supply position rather than competing only on laboratory performance.

Competitive advantage is likely to remain layered. Incumbents with access to high-speed DSPs, photonic engines, laser supply, and major cloud accounts can control early CPO programs. Specialists can still capture value where they solve a discrete integration bottleneck, such as laser coupling, low-power modulation, optical I/O chiplets, or testability. The principal strategic risk is that a technically superior photonic architecture may fail to scale if it lacks a foundry, packaging, and field-support route acceptable to hyperscale buyers.

Recent Industry Developments

• March 2024: Broadcom introduced its 51.2-Tbps CPO Ethernet switch platform, combining eight 6.4-Tbps optical engines with a switch architecture for scalable AI systems.
• March 2024: Marvell demonstrated a 200G 3D silicon photonics engine for accelerated infrastructure.
• March 2024: POET Technologies announced entry into the AI optical-module market with an 800G 2×FR4 OSFP module based on its Optical Interposer platform.
• December 2024: Ayar Labs announced USD 155 million in funding, including investment from AMD, Intel Capital, and NVIDIA, to advance optical I/O for AI infrastructure.
• January 2025: The U.S. Department of Commerce announced CHIPS incentives for Infinera supporting indium phosphide fabrication and advanced packaging.
• March 2025: MACOM announced high-performance solutions for 1.6T applications.
• December 2025: Marvell announced an agreement to acquire Celestial AI, accelerating its scale-up connectivity capabilities for next-generation data centers.