Market and Technology Trends
Co-Packaged Optics for Data Centers 2026
CPO redefines AI interconnect economics, growing to $112B by 2031 as value shifts to packaging and lasers.
The data center interconnect industry faces its biggest structural shift in 40 years, as AI clusters push optics onto the package - co-packaged optics (CPO) reshaping a market that scales toward ~$112B by 2031.
- Completely new market model: This edition introduces a completely new market model covering switches and optical engines for scale-out and scale-up, plus optical engines for XPU packages. Demand is built bottom-up: every GPU and AI-ASIC package shipped each year carries a per-package intensity of bandwidth and optical engines (OEs), summed across chip generations for a granular, defensible forecast. NVIDIA coefficients are derived from published rack parameters, while non-NVIDIA volumes are built from architecture archetypes, and units multiplied by ASP drive revenue.
- Two new BoM analyses: two brand-new bill-of-materials analyses covering the CPO switch and the optical engine, alongside a dedicated teardown of the ELSFP module. Each analysis breaks down component costs and suppliers, giving customers a clear view of where cost and value sit inside next-generation CPO systems.
- Structural update: A refreshed structural chapter delivers qualitative analysis of scale-in, scale-up, and scale-out architectures, clarifying how each approach addresses bandwidth, power, and thermal constraints. This helps customers benchmark competing system architectures and anticipate which topology will win in different deployment scenarios.
- Detailed roadmap to market: A newly detailed roadmap maps technology and application timing across the CPO landscape, identifying who ships what and when. Customers can use this roadmap to time their own product launches and investment decisions against the pace of the broader ecosystem.
- Updated ecosystem mapping: The ecosystem mapping has been substantially expanded with finer granularity on ASIC/XPU vendors, fiber array unit (FAU) suppliers, light source providers, and test and assembly partners. This gives customers a more precise view of where every player sits in the CPO supply chain.
- Value migration: A new analysis tracks where value migrates as optics move onto the package, highlighting which players stand to gain and which risk losing share as the industry shifts from pluggable to co-packaged optics. This is essential reading for any company repositioning its roadmap around CPO.
- Strategic capital movement: This edition tracks strategic capital movement from recent fundraising rounds and investments across the CPO value chain, revealing which companies and technologies are attracting the most investor attention. Customers can use this intelligence to benchmark their own strategic positioning.
- CPO ecosystem capital flows: A dedicated new section examines the CPO ecosystem's fundraising wave and acquisition wave side by side, showing exactly where capital is flowing across the supply chain. It is a must-read for investors and strategy teams tracking consolidation and competitive dynamics in real time.
- CPO reliability: A new chapter dedicated to CPO reliability addresses one of the industry's most pressing adoption concerns, covering the technical and qualification challenges that must be solved before co-packaged optics can scale in production. This gives customers the risk context needed to plan deployment timelines with confidence.
- Extended laser sources chapter: The laser sources chapter has been significantly extended, providing deeper analysis of the light source technologies and suppliers competing to serve CPO architectures. This expanded coverage helps customers evaluate sourcing options as the laser landscape evolves alongside CPO adoption.
- Fiber Array Units (FAUs) for CPO: A brand new chapter delivers qualitative and industrial analysis of Fiber Array Units (FAUs) for CPO, covering the manufacturing processes, suppliers, and technical challenges unique to this critical component. This gives customers a comprehensive view of a part of the value chain that has received little dedicated coverage until now.
Objectives of this report
- Provide a detailed analysis and comprehensive view of co-packaged optics (CPO) for data centers.
- Provide the context of why DC operators explore CPO technology
- The explosive growth of AI, particularly large language models (LLMs) and generative AI as a major market driver, and other trends in data centers impacting the optical module technology, market, and industry
- Provide detailed market forecasts from 2021 – 2031, split by applications – scale-out and scale-up:
- CPO switches
- CPO optical engines
- ELS modules
- Review the supply chain
- Industrial ecosystem, supply chain of selected players
- Who truly stands to benefit from CPO transformative approach?
- Where capital is going?
- Examine technology approaches for CPO
- Applications: scale-out networking & scale-up AI/ML high-performance computing
- CPO technology architectures: optical engines for scale-out and scale-up
- Roadmaps for CPO adoption
- Competitive landscape - Technical parameters
- CPO solutions comparison
- CPO market leaders & merchant silicon
- Optical I/O specialists (start-ups)
- Discuss CPO challenges and compare with pluggable optics
CPO Goes Mainstream: Scale-Up Bet and the Reshaping of AI Interconnect
CPO has moved from concept to commercial reality faster and bigger than the 2023 outlook assumed, driven by NVIDIA and a maturing ecosystem. Total CPO optical engine revenue inflects sharply between 2026 and 2031 (~228% CAGR), reflecting hyperscaler capex near $500B in 2025 alone.
Scale up is the killer application: copper NVLink's ~2m reach cap, against per GPU bandwidth doubling toward 14.4 Tb/s, makes CPO the only credible path to larger GPU world sizes. Scale up switch OE grows to ~214% CAGR, while scale up compute (XPU) OE also expands significantly, together ~94% of the market and the largest line in the forecast.
Scale out remains incremental. Pluggables stay "good enough" for 0.5–2km reaches, and since networking is only ~9% of cluster power, scale out switch OE plateaus at a modest level by 2031 (~127% CAGR), a market test that matures the supply chain rather than the ultimate volume driver.
The opportunity is front loaded with certainty but extends well beyond 2030, continuing to climb through 2031, with scale up switches and compute remaining the dominant engine. NVIDIA's scale up CPO ramp pace remains the key signpost to watch
Optics Moves Onto the Package: The CPO Architecture Transition in AI Compute
Four architectures, FRO, LRO, LPO and CPO, form a single progression shifting signal processing from module to ASIC, trading module cost for lower power and latency; FRO stays essential for long reach, LPO leads 2025–2027 intra datacenter volume, and CPO sits at the endpoint.
CPO wins on density and efficiency: copper is capped near 10m at 200G/lane, while CPO reaches ~3.2 pJ/bit versus ~10 pJ/bit for copper, an ~80% energy cut, as switch/XPU bandwidth climbs from 51.2T toward ~204T.
Packaging and the laser are pivotal: TSMC's COUPE (micro ring modulators, SoIC bonding) is becoming the central enabler, Broadcom is migrating its own CPO line onto it, since fan out packaging can't scale past ~100 Gbps/lane, while remote laser choice (CW DFB, WDM, combs) drives system BOM and reliability.
Coupling is the gating HVM issue: fiber array units at micron tolerances are the binding constraint, though Meta/Broadcom's Bailly field data (zero errors to 4M port hours, ~2.6M hour MTBF) is encouraging.
Photonic interposers, Lightmatter's Passage M1000 and Celestial AI's OMIB, and the UCIe chiplet standard point toward optical I/O as compute's next frontier.
How CPO Is Reshaping the Optical Interconnect Supply Chain
Capital is industrializing: over $500B has flowed into silicon photonics/CPO M&A and funding between late 2025 and mid 2026, with 10+ incumbents buying into the optical value chain, shifting from validation toward lasers, packaging, fiber attach and test, while hyperscaler linked capital reinforces vertical integration. Value is relocating structurally toward the optical engine, advanced packaging and light sources, with coupling as the new bottleneck, while DSP/retimer vendors and merchant pluggable suppliers lose share as optics move on package.
Foundry and packaging are the key lock in: TSMC's COUPE (SoIC hybrid bonding) is hard to replicate, and system OEMs, NVIDIA, Broadcom, AMD, capture rack scale value. Four business models are emerging: vertically integrated system companies, chip makers adding CPO, design start ups partnering with hyperscalers, and integrators; NVIDIA is the most integrated, controlling design through TSMC COUPE fabrication.
Specialists own coupling, assembly and test: Corning (fiber), FOCI/TFC/Senko (FAU coupling), Foxconn/FIT (lasers), Fabrinet/ShunSin/Micas/Celestica (assembly) and Keysight/Ficontec/Teradyne/Chroma (test) are critical since CPO can't be field swapped like pluggables.
Glossary
Definitions
Identity Card
Methodologies & definition
Report objectives
Scope of this report
Companies cited
Historical forecast comparison
What we got right, what we got wrong
About the authors
3-Page summary
Executive summary
Context
Market forecast
- CPO market: Scale-out
- CPO market: Scale-up
- CPO market: Scale-in
Market trends
- Scaling connectivity for AI infrastructures
Application trends
Industry & supply chain
- Supply Chain
- Fundraising and M&A
Technology trends
- CPO architectures
- Fully retimed DSP (FRO) vs LRO vs LPO vs CPO
- Copper vs Optics
- Reliability
- Benchmark analysis
- UCIe
- CPO competitive landscape
- Lasers for CPO
- Fiber Array Units (FAUs)
- Bill of Materials analysis of CPO
Conclusion (Outlook)
- Summary
- Outlook: Future challengers to CPO
- Outlook
YG related presentation
3M, Accelink, Advanced Micro Foundry (AMF), Advantest, AFR, AIM Photonics, AIO Core, Alchip, Alibaba Cloud, All Ring Tech, Alphawave, Amazon (AWS), AMD, Amkor Technology, Amphenol, AOI (Advanced Optoelectronic Inc.), Ardentec, Arista Networks, ASE Holdings, A*STAR, AsteraLabs, AttoTude, Avicena, AXT, Ayar Labs, Baidu, BizLink, Broadcom, Broadex Technologies, Browave Corporation, ByteDance, Cambricon, Casela Technologies, CEA LETI, Celestial AI, Celestica, ChipMOS, Chiral Photonics, Chroma ATE, Ciena, CIG, Cisco, Coherent, Corning, Credo, Dell Technologies, Delta, DenseLight, DoGain, Enablence, Enlightra, Ennostar, Eoptolink, Eribel Systems, Ericsson, EVERBRIGHT, Fabrinet, Fiberhome, Ficontec, Finisar, FIT, FOCI, FormFactor, Foxconn, Fraunhofer, Fujikura, Fujitsu, FuriosaAI, Furukawa Electric, Genuine Optics, GlobalFoundries, Google Cloud, Graphcore, GUC (Global Unichip Corp), H3C, Hakusan, HANA Micron, Hengtong Group, Hewlett Packard Enterprise, Hisense Broadband, Huawei, Hyperlume, IBM, imec, IMECAS, Industrial Technology Research Institute (ITRI), Innolight, Innolume, Innovium, Intel, IntelliEpi, Inventec, IQE, Jabil, JCET, Juniper, Keysight, KYEC, LandMark, Lenovo, Ligent, Lightmatter, LioniX, Lumentum, Luminar, LuxshareICT, MACOM, Macrochip, Marvell, Maxim Integrated, MaxLinear, Mediatek, Meta, Micas, Microchip, Microsoft Azure, Mitsubishi Electric, MixxTech, Molex, multiLane, Nanosys, National Semiconductor Translation and Innovation Centre (NSTIC), NcodiN, Nepes, Neptec Optical Solutions, NewPhotonics, Nokia, NRC-CNRC, NTT, Nubis Communications, Nvidia, O-Net, OpenAI, OpenLight, Oracle, PETRA, Phix, Pilot Photonics, POET Technologies, Prysmian, QCT (Quanta Cloud Technology), QD Laser, Quanta Computer, Quintessent, Radiall, Ragile Networks, Rain Tree Photonics, Ranovus, Rebellions, Ruijie Networks, Samsung, Samtec, Scintil Photonics, SemiAnalysis, Semtech, SENKO, Shijia, ShinEtsu, Shinko, ShunSin Technology, Sigmui Technology, Sigurd, SilTerra, Sivers Photonics, SMIC, SOITEC, Source Photonics, SPIL, STATSChipPAC, Sumitomo Electric, TE Connectivity, Tencent, Tenstorrent, Teradyne, Terahop, Teramount, Tesla, TFC Optical Communication, Thorlabs, Three-circle Group, Tower Semiconductor, TrueLight, T&S Communications, TSMC, UMEC, United Microelectronics (UMC), US Conec, UTAC, VLINK, VPEC, Winstek, Wistron, Wiwynn, X-FAB, Xscape Photonics, Xsight Labs, X (Twitter), Yamaichi, YOFC, ZTE and more.