Co-packaged Optics Market Size & Share Analysis - Growth Trends and Forecast (2026 - 2031)

Global Co-packaged Optics Market Trends and Insights

Ramp-Up of 51.2 T Switch Silicon Requiring 1.6 T CPO

Second-generation 51.2 T switch ASICs integrate 64 ports of 800 G optics to sustain radix within a single rack unit.^{[1]Broadcom Inc., “Broadcom Tomahawk 5 Technology Brief,” broadcom.com} Broadcom’s Tomahawk 5-Bailly platform ships in volume, proving manufacturability at 100 G per lane. NVIDIA’s Quantum-X800 InfiniBand device delivers 144 ports of 800 G with only 4 dB insertion loss, a 63-fold eye-opening improvement over pluggables. Hyperscalers have pre-allocated foundry capacity through multi-year purchase commitments, effectively locking in demand. The architecture also lets operators collapse spine tiers, cutting capital expenditure per terabit. Sampling of 102.4 T silicon in late 2026 will pull 3.2 T and 6.4 T co-packaged interfaces into qualification cycles.

Hyperscale Data-Center Energy-Efficiency Mandates

The United States data-center power draw is projected to scale from 176 TWh in 2023 to as high as 580 TWh by 2028, with networking already near 23 TWh.^{[2]Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage 2023-2028,” lbl.gov} Co-packaged optics remove retimers and DSP stages, lowering per-port power by 30-40% at 800 G-plus speeds. Operators in power-constrained campuses now include “watts per terabit” in vendor scorecards, while EU directives embed efficiency metrics into tender documents. By reducing thermal load at the faceplate, co-packaged designs defer expensive electrical infrastructure upgrades. Regulatory pressure to curb carbon emissions is turning energy-efficient interconnects from an optimization lever into a procurement prerequisite.

Surge in AI/ML Cluster Bandwidth Demand

Training runs for trillion-parameter models saturate conventional three-tier Clos fabrics. NVIDIA’s DGX SuperPOD specifies 3.2 T of bi-directional bandwidth for every GPU, a figure that will double with next-generation accelerators.^{[3]NVIDIA Corporation, “Quantum-X800 Technical Architecture,” nvidia.com} Rail-optimized topologies become feasible when port density and power envelopes align, a balance enabled by co-packaged optics. Meta’s AI Research SuperCluster shaved 18% off training time for a 175-billion-parameter model after migrating from pluggable to 800 G co-packaged links. As real-time inference spreads to autonomous vehicles and interactive agents, latency budgets continue to tighten, boosting demand for integrated photonics.

Transition to 800 G - 1.6 T Pluggables Hitting Thermal Limits

OSFP and QSFP-DD modules cap heat dissipation near 18 W. Climbing to 1.6 T requires 200 G lanes, PAM4 modulation, and tighter FEC overhead, which raise bit-error rates beyond forward-error thresholds on moderate-reach links. Co-packaged engines distribute thermal load onto the ASIC heatsink, where vapor chambers and embedded heat pipes operate unconstrained by front-panel mechanics. Broadcom’s third-generation roadmap moves to 200 G lanes and charts a path toward 400 G lanes, a trajectory that pluggables cannot replicate without radical form-factor change.

Manufacturing Complexity and Heterogeneous-Integration Yield

Bonding III-V lasers, silicon-photonic waveguides, and CMOS logic within one package requires sub-micron alignment tolerance. Several foundries reported yields under 70% as of early 2026, inflating unit cost for all but the densest switches. TSMC’s early COUPE production targets 60-65% yield, implying 35-40% scrap. Advanced metrology, such as optical coherence tomography, is improving process control, yet the learning curve remains steeper than for electronic packaging. Until yields cross the high-80% threshold, enterprise and telco deployments will continue favoring legacy pluggables on a pure cost basis.

Interoperability and Standards Immaturity

The Open Compute Project released a mechanical spec in 2025, but wavelength grids, modulation formats, and FEC schemes still vary by customer. IEEE 802.3 groups expect to ratify co-packaged Ethernet at 800 G and 1.6 T only in late 2027. In the interim, every hyperscaler qualifies modules against bespoke test plans, limiting second-source viability. The Optical Internetworking Forum has proposed interoperability suites, yet participation is voluntary, and coverage is incomplete. Vendor-specific paths risk marooning early deployments on incompatible islands.

*Our updated forecasts treat driver/restraint impacts as directional, not additive. The revised impact forecasts reflect baseline growth, mix effects, and variable interactions.

Segment Analysis

By Data Rate: 6.4 T Interfaces Anchor Next-Generation Fabrics

In 2025, the 1.6 T segment captured a dominant 39.12% of the co-packaged optics market, while projections indicate that the segment of 6.4 T and above is set to expand at a robust CAGR of 36.69% through 2031. Early movers such as Tomahawk 5 and Quantum-X800 standardize on 2x800G breakouts or 16x100G lanes. Rapid sampling of 102.4 T silicon means 3.2 T and 6.4 T ports will dominate qualification roadmaps during 2026-2028. Operators with legacy 400 G fiber plants continue to use below-1.6 T links in brownfield upgrades, but new hyperscale builds prioritize port densities that only higher-rate optics enable.

TSMC’s mass-production COUPE flow supports up to 6.4 T per package by integrating modulators and detector arrays into a single interposer. Intel’s roadmap echoes this trajectory, aiming for 6.4 T by 2027 with monolithically integrated lasers. As AI cluster node counts balloon, network architects value bisection bandwidth over incremental cost, driving a secular mix-shift toward 6.4 T and above. Consequently, the 6.4 T cohort is forecast to capture an outsized share of the co-packaged optics market between 2028 and 2031.

Note: Segment shares of all individual segments available upon report purchase

By Component: Integrated Laser Sources Accelerate

Optical engines led the revenue stack at 43.76% in 2025, consolidating modulation, detection, and multiplexing on silicon. The laser-source category is set to post a 36.89% CAGR, the quickest among components, once heterogeneous III-V bonding matures. Coherent’s indium-phosphide designs target sub-500 mW per wavelength, unlocking new cost curves. Lumentum’s quantum-dot arrays improve wavelength uniformity, addressing reliability needs inside thermally turbulent switch enclosures.

Electrical IC content shrinks per port as SerDes functions co-locate within the switch die. Packaging and connector innovations, such as 64-fiber MPOs, raise panel density without sacrificing insertion loss. Miscellaneous passive components remain incremental plays. Over the forecast period, greater integration of lasers into the photonic stack is expected to lift the laser segment’s share of the overall co-packaged optics market.

By Integration Approach: Co-Packaged Dominates New Designs

Co-packaged architectures held a 55.67% share in 2025 and will widen their lead, growing at a 36.17% CAGR to 2031. By directly linking switch SerDes blocks to photonic modulators, engineers have shortened the electrical path length to under 1 mm, ensuring eye margins are maintained at 100 G lanes. NVIDIA's Quantum-X800 integrates 144 ports of 800G optics within its package, doing away with the requirement for a separate retimer power.

On-board optics remain serviceable for enterprise and telco workloads that value module swaps, yet their added latent heat and trace loss impose penalties at 200 G and beyond. Ayar Labs’ optical-interposer chiplet approach offers a middle ground, enabling mix-and-match integration while still qualifying as co-packaged. As yields rise and cost deltas narrow, co-packaged solutions will dominate incremental port additions, reinforcing their leadership in the co-packaged optics market.

By End-Use Application: HPC and AI/ML Clusters Accelerate

Hyperscale cloud data centers captured 62.34% of revenue in 2025, driven by the concentrated buying power of four dominant operators. However, HPC and AI/ML clusters are expected to expand at a 36.96% CAGR, the swiftest among applications, as frontier model training requires near-line-rate, all-to-all fabrics. Meta’s SuperCluster improved training time by 18% after shifting to co-packaged optics, validating the upside in performance.

Enterprise data centers are selectively adopting the technology, with a keen focus on latency-sensitive sectors like financial trading. While telco central offices represent a smaller segment, they are placing a premium on 5G core densification. Catering to this specialized need, Nokia has introduced its carrier-grade CPO platform. Furthermore, clusters that alleviate network bottlenecks not only address server underutilization but also position integrated photonics as a significant return on investment lever.

Geography Analysis

North America accounted for 47.83% of 2025 revenue, anchored by U.S. hyperscalers and CHIPS Act grants that subsidize domestic photonics pilot lines. In 2024-2025, Intel, Ayar Labs, and a consortium of universities garnered a combined USD 280 million for their silicon photonics R&D endeavors. While the funding has a cross-border dimension, it's limited: Canada is home to connector assembly plants, and Mexico specializes in back-end testing and finishing.

Asia Pacific is forecast to log a 36.91% CAGR through 2031, the quickest globally. TSMC’s COUPE ramp offers immediate volume in Taiwan, while Japanese and South Korean firms add connector and laser capacity. Chinese vendors innovate around export-control constraints by integrating hybrid components into mature nodes. India and Australia offer greenfield data center incentives, yet front-end wafer production remains concentrated in Taiwan and Japan.

Europe, the Middle East, and Africa account for the balance. The EU Chips Act backs photonics research at IMEC and Fraunhofer, but limited high-volume foundry capacity caps regional output. Germany, the United Kingdom, and France host niche optical component suppliers that have yet to meet hyperscale cost curves. Middle Eastern sovereign clouds drive pockets of demand, while African adoption stays embryonic.