Market Overview
Size, architectures, and capital signals
- Brevintel analysis values the market at USD 438.6 million in 2025 ($439M), estimated at USD 544.6 million in 2026, and projected to USD 5.87 billion ($5.87B) by 2035—also cited as USD 5.9 billion—at approximately 30.3% CAGR from 2026 to 2035.
- Lightmatter led with over 21.4% market share in 2025; Lightmatter, Lightelligence, Celestial AI, Intel Corporation, and Ayar Labs collectively held 56.3%.
- Commercial activity concentrates on optical interconnect products that relieve bandwidth between GPUs, memory, and compute nodes while retaining conventional dies, versus native optical-compute that executes matrix operations in photonics.
- Lightmatter's October 2024 Series D raised USD 400 million and brought its reported valuation to USD 4.4 billion; Ayar Labs raised USD 155 million in December 2024 with AMD Ventures, Intel Capital, and NVIDIA among investors.
- The addressable stack includes photonic integrated circuits, modulators, laser sources, detectors, control electronics, optical interconnect substrates, in-package chiplets, and compilation software for photonic neural workloads.
- Silicon photonics remains the largest technology base via established semiconductor manufacturing; hybrid photonic-electronic designs place optical I/O or compute planes alongside CMOS logic.
- Proposed CHIPS incentives include up to USD 33 million for Coherent Corp.'s indium phosphide expansion and up to USD 93 million for Infinera's planned indium phosphide facility.
- Interconnect products have a clearer procurement path today; native optical compute carries greater upside where precision, calibration, and software integration prove out at production scale.
AI infrastructure power and bandwidth context
- The International Energy Agency projects global data center electricity demand of about 945 TWh by 2030, roughly double its 2024 level, as AI workloads expand.
- In the U.S., data centers consumed 176 TWh in 2023, and Lawrence Berkeley National Laboratory projects a range of 325–580 TWh by 2028.
- Optical neural network processors address part of this constraint by shifting bandwidth-intensive communication and selected linear-algebra operations into photonic architectures.
- Intel reported an integrated optical I/O chiplet supporting 64 channels at 32 Gbps in each direction while operating at 5 picojoules per bit.
- IBM's co-packaged optics prototype reported more than a fivefold power reduction versus mid-range electrical interconnects and a sixfold increase in fiber beachfront density.
- A 2025 Nature study demonstrated an integrated photonic accelerator with more than 16,000 photonic components, computing frequencies up to 1 GHz, and latency as low as 3 nanoseconds per cycle.
- Nature also reported fully forward mode training for optical neural networks in 2024, addressing training without conventional physical backward-propagation on optical hardware.
- Suppliers that package optics with familiar electronic interfaces can monetize near-term interconnect demand while optical-compute vendors prove accuracy, thermal stability, and model-compilation workflows.
Market Dynamics
Key drivers
- Rising demand for high-speed, energy-efficient AI processing makes interconnect energy a cluster operating-cost variable as demand approaches about 945 TWh by 2030.
- Expansion of AI and machine-learning workloads intensifies repeated matrix operations and accelerator-to-accelerator communication that photonic hardware can target.
- Fixed-weight inference has fewer real-time parameter-update needs and is a nearer-term photonic acceleration target than full training stacks.
- Foundry access is widening: GlobalFoundries offers production silicon photonics on 200-mm and 300-mm wafers with dense wavelength-division multiplexing, and TSMC positions silicon photonics in advanced packaging via its compact universal photonic engine.
- Electrical systems face limits in bandwidth density, signal reach, and power dissipation; co-packaged and in-package optical designs preserve software bases while attacking the costly interconnect layer.
- Integrated photonics progress expands viable architectures across silicon routing and compound-semiconductor active functions such as light generation, modulation, and nonvolatile optical-state control.
- Training creates a larger long-term opportunity because distributed models intensify communication, but it requires tighter coordination among optics, software frameworks, and cluster management.
- Brevintel analysis places neuromorphic optical computing near 34.2% CAGR, hybrid photonic-electronic integration near 32.6%, and training-optimized processors near 33.9%, versus about 28.9% for inference-optimized processors.
Key restraints
- High development and implementation costs span photonic PDKs, lasers, detectors, packages, fiber coupling, electronic control, and calibration software, expanding NRE and yield exposure at optical–electronic boundaries.
- Low-volume economics favor firms with capital, foundry partnerships, and customers willing to co-develop architectures; early interconnect revenue can finance longer optical-compute qualification cycles.
- Integration with existing GPUs, CPUs, ASICs, memory, and electrical standards adds optical alignment, thermal management, reliability testing, and serviceability requirements beyond chip-level performance.
- IBM reported polymer optical-waveguide co-packaged optics prototypes completing thermal cycling, humidity, and mechanical reliability stress testing—thresholds vendors must clear for deployment.
- Software constraints include compilation, quantization, drift compensation, and framework integration so ML teams avoid redesigning model-development workflows.
- Near-term demand favors products that improve bandwidth-per-watt while retaining familiar compute and software environments; hybrid architectures participate in current AI build-outs with lower precision and programming risk.
Segmentation Analysis
By technology
- Silicon photonics accounted for USD 233.7 million in 2025 and is projected to reach USD 2.92 billion by 2035 at approximately 29.3% CAGR, aided by wafer-level processing and packaging ecosystems.
- Hybrid photonic-electronic integration was valued at USD 114.4 million in 2025 and is projected to reach USD 1.83 billion by 2035 at approximately 32.6% CAGR.
- Lightmatter's Passage interconnect and Ayar Labs' TeraPHY optical I/O illustrate hybrid logic: photonics where electrical connectivity is under pressure while conventional compute chips remain in place.
- Free-space optics generated USD 60.7 million in 2025 and is projected to reach USD 710.3 million by 2035 at approximately 28.5% CAGR via diffractive elements, SLMs, lens arrays, or Fourier-domain processing.
- Emerging photonic materials were valued at USD 29.8 million in 2025 and are projected to reach USD 415.7 million by 2035 at approximately 30.8% CAGR, with indium phosphide important for laser gain not native to pure silicon.
- Free-space deployment remains sensitive to alignment, vibration, environment, and dynamic reconfiguration, favoring specialized image, scientific, and signal-processing uses over broad AI-server replacement.
- Silicon remains strong for passive routing and dense integration, but commercial optical neural networks still need robust sources, modulators, detectors, and weight stabilization.
- New modulators, phase-change materials, and hybrids must prove repeatable device behavior, compatible flows, and cost-effective integration before shifting mainstream architectures.
By computing paradigm
- Analog optical computing was valued at USD 233.1 million in 2025 and is projected to reach USD 2.68 billion by 2035 at approximately 28.3% CAGR via optical interference for linear matrix operations.
- Digital optical computing generated USD 122.0 million in 2025 and is projected to reach USD 1.67 billion by 2035 at approximately 30.6% CAGR, using discrete optical levels for noise tolerance and digital-error management compatibility.
- Neuromorphic optical computing was valued at USD 83.6 million in 2025 and is projected to reach USD 1.51 billion by 2035 at the highest paradigm CAGR of approximately 34.2%.
- Analog systems must control noise, thermal drift, mismatch, and conversion overhead; strongest uses tolerate quantization while cutting energy or latency.
- Neuromorphic designs target event-driven or spiking photonic behavior for low-power inference and sensor processing, gated by toolchains for mapping, calibration, and workflow integration.
- Brevintel analysis reads the 34.2% neuromorphic CAGR versus about 28.3% for analog as a software-gated path rather than a purely hardware-gated one.
By workload optimization
- Inference-optimized processors are the largest workload segment at USD 335.2 million in 2025, projected to reach USD 4.04 billion by 2035 at approximately 28.9% CAGR.
- Training-optimized processors are projected to rise from USD 103.4 million in 2025 to USD 1.82 billion by 2035 at approximately 33.9% CAGR.
- Fixed or infrequently updated inference models reduce continuous optical weight programming and clarify energy-per-inference comparisons.
- Optical interconnect products align with inference clusters by improving data movement without changing the fundamental model-execution environment.
- Training intensifies high-bandwidth links among distributed accelerators and needs synchronization, parameter updates, and software support.
- IBM has identified co-packaged optics as a possible route to shorten training time for a 70-billion-parameter large language model, contingent on system-level deployment at scale.
By application and end-use industry
- AI/ML acceleration is the largest application because neural workloads combine dense matrix operations with heavy movement among compute, memory, and network resources.
- Image recognition and computer vision can use parallel spatial processing in free-space and integrated optical systems under controlled industrial, surveillance, and sensor conditions.
- Natural language processing demand ties to large-model serving and distributed training; near-term photonics entry is more likely via interconnect and memory-fabric functions than full transformer replacement.
- Signal-processing opportunities arise where latency, bandwidth, and parallel transformation matter more than general programmability across communications, sensing, and analytics.
- Data centers and cloud providers dominate end use because cluster scale makes interconnect energy, bandwidth density, and cooling measurable procurement variables.
- Telecommunications can apply photonic processing to high-bandwidth switching and AI-enabled network management; automotive centers on sensor fusion, ADAS, and power-constrained onboard inference.
- Aerospace and defense value low-latency signal processing but face long procurement and environmental qualification; healthcare and life sciences remain selective for imaging and genomics acceleration.
- Edge computing and IoT create potential for low-power neuromorphic optical architectures where event-driven sensing and continuous inference constrain energy use.
Regional Outlook
North America
- North America was valued at USD 204.2 million in 2025, estimated at USD 254.6 million in 2026, and projected to reach USD 2.84 billion by 2035 at approximately 30.7% CAGR.
- The U.S. accounted for USD 175.5 million in 2025 and is projected to reach USD 2.46 billion by 2035 at approximately 30.9% CAGR.
- Canada contributed USD 28.7 million in 2025 and is projected to reach USD 372.5 million by 2035 at approximately 29.8% CAGR.
- Scale is underpinned by hyperscale AI infrastructure, semiconductor design activity, and venture-backed photonics companies.
- Lightmatter's 2024 financing and Ayar Labs' strategic investor group illustrate capital formation around photonic interconnect infrastructure.
- Proposed federal support for Coherent and Infinera indium phosphide expansions can improve supply-chain capacity for active optical components.
Europe
- Europe generated USD 114.4 million in 2025, estimated at USD 140.2 million in 2026, and projected to reach USD 1.33 billion by 2035 at approximately 28.4% CAGR.
- Germany is projected to rise from USD 31.7 million in 2025 to USD 418.8 million by 2035 at approximately 30.1% CAGR.
- The United Kingdom is projected to rise from USD 25.0 million to USD 309.8 million at approximately 29.2% CAGR; France from USD 18.0 million to USD 211.4 million at approximately 28.5% CAGR.
- Italy, Spain, the Netherlands, and the Rest of Europe are projected to grow from USD 10.7 million, USD 8.5 million, USD 7.0 million, and USD 13.6 million in 2025 to USD 115.8 million, USD 85.7 million, USD 66.3 million, and USD 123.9 million by 2035.
- Europe's position rests on semiconductor research, photonics design, and specialized computing developers, with commercial scale-up more fragmented across national markets than in North America.
Asia Pacific
- Asia Pacific was valued at USD 86.5 million in 2025, estimated at USD 108.8 million in 2026, and projected to reach USD 1.31 billion by 2035 at approximately 31.9% CAGR—the fastest major region.
- China is projected to expand from USD 37.4 million in 2025 to USD 599.5 million by 2035 at approximately 32.6% CAGR.
- Japan is projected to rise from USD 15.1 million to USD 226.0 million at approximately 31.7% CAGR; India from USD 12.8 million to USD 217.3 million at approximately 33.3% CAGR.
- South Korea, Australia, and the Rest of Asia Pacific are projected to reach USD 113.7 million, USD 50.7 million, and USD 107.2 million, respectively, by 2035.
- Growth ties to AI infrastructure investment, domestic semiconductor development, and pressure to cut power in dense computing deployments.
- Brevintel analysis places India at approximately 33.3% CAGR through 2035, ahead of China at approximately 32.6%, from a comparatively small 2025 base.
Latin America and Middle East & Africa
- Latin America was valued at USD 20.4 million in 2025, estimated at USD 24.8 million in 2026, and projected to reach USD 220.8 million by 2035 at approximately 27.5% CAGR.
- Brazil and Mexico are the region's largest markets, projected to reach USD 98.9 million and USD 65.1 million by 2035, respectively.
- Middle East and Africa was valued at USD 13.1 million in 2025, estimated at USD 16.3 million in 2026, and projected to reach USD 168.3 million by 2035 at approximately 29.7% CAGR.
- The UAE is projected to reach USD 67.8 million by 2035; Saudi Arabia USD 52.8 million; South Africa USD 28.3 million; Rest of Middle East and Africa USD 19.4 million.
- Regional opportunity ties to new AI and data-center infrastructure rather than replacing a large installed photonic-processing base, with cooling, energy, and import economics shaping adoption paths.
Competitive Landscape
Share structure and positioning
- The market was moderately concentrated in 2025: Lightmatter held approximately 21.4%, Lightelligence 11.4%, Celestial AI 9.6%, Intel 7.1%, and Ayar Labs 6.8%, together about 56.3%.
- Lightmatter's Passage photonic interconnect strategy addresses bandwidth between conventional AI chips while preserving the electronic compute substrate, backed by its USD 400 million Series D in October 2024.
- IBM's optical activity centers on co-packaged optics and system-level interconnect research, including a December 2024 prototype with reliability testing and reported power and density gains versus electrical approaches.
- Celestial AI's Photonic Fabric targets high-bandwidth connectivity among compute and memory; its USD 175 million Series C in March 2024 included strategic and institutional investors and disclosed hyperscale design-in activity.
- Intel reported shipment of more than 8 million photonic integrated circuits containing over 32 million integrated lasers, reinforcing its silicon photonics manufacturing and packaging role.
- Ayar Labs' TeraPHY optical I/O chiplet is positioned for in-package, high-bandwidth connectivity with strategic semiconductor and AI-infrastructure investor support.
- Advantage increasingly depends on converting photonic component performance into deployable systems via capital, foundry access, packaging partnerships, software integration, and customer qualification.
- In April 2026, the University of Southern California announced a photonic chip capable of surviving temperatures up to 1300°F (700°C); in March 2026 Nokia and Huawei each launched AI-era optical network solution suites.
- In 2025, Nature reported a large-scale integrated photonic accelerator with more than 16,000 photonic components, frequencies up to 1 GHz, and latency as low as 3 nanoseconds per cycle.
Global key players
- Lightmatter
- IBM Corporation
- Lightelligence
- Celestial AI
- Intel Corporation
- Ayar Labs
Regional players
- Luminous Computing
- Xanadu Quantum Technologies
- Optalysys
- iPronics
- OpenLight
Niche players
- PhotonicX AI
- SiEPIC
- Photon Bridge
- Mixx Technologies
- Neurophos