ECOC 2025 (51st European Conference on Optical Communication): Key Takeaways
Updated: 20 hours ago
Introduction
ECOC 2025 showcased how rapidly optical communications is evolving in the AI era. With the rise of generative AI and massive GPU clusters, data centers now demand far more bandwidth, energy efficiency and reliability, putting optical modules and network architecture at the center of the industry's attention. From breaking through the 400G/lane and 448G technology thresholds, to the architectural choice between pluggable, LPO and CPO, to the role of OCS and coherent technology in AI networks, the ecosystem is diverse and both competitive and collaborative. Vendors and standards bodies (Omia, Cignal AI, OIF, IEEE, Meta, Huawei, Coherent and others) each offered views on market size, technical bottlenecks and industry roadmaps, together painting a full picture of where optical communications is heading over the next 5–10 years. This report focuses on five themes: AI as the driving force, 400G/lane and 448G electrical interfaces, diversified modules and packaging, energy efficiency and reliability, and exploration of network architectures, and systematically organizes the market forecasts and technology roadmaps.
Five Key Takeaways
1. AI Is the Biggest Driver of Optical Communications
Almost every talk (Omia, Cignal AI, PhotoniX AI, Source Photonics) pointed to the same conclusion: AI investment and GPU-cluster demand are driving explosive growth in the optical module market. (Omia / Cignal AI / PhotoniX AI / Source Photonics)
CAPEX has reached hundreds of billions of US dollars, and AI will become the main growth engine for optical modules. (Omia / Cignal AI)
2. 400G/Lane and 448G Are the Industry's Key Thresholds
400G/lane → 1.6T / 3.2T is the inevitable path; components face >100 GHz modulation/receiver bandwidth challenges. (Huawei / Synopsys / Source Photonics)
The 448G electrical interface is evolving in parallel; the choice of FEC and PAM4 vs. PAM6/8 determines latency and power. (Synopsys / IEEE)
3. Modules and Packaging: Multiple Paths Coexist
Pluggable: still the mainstream, with great flexibility in deployment and maintenance. (Eoptolink / Cisco / Meta)
LPO: more power-efficient than retimed pluggables; energy savings already visible from 800G to 1.6T. (Meta / Cisco)
CPO: best efficiency and density, but thermal management, reliability and serviceability are the biggest challenges. (Meta / Broadcom+Corning)
The future landscape will be a hybrid coexistence of Pluggable + LPO + CPO. (Omia / Cisco)
4. Energy Efficiency and Reliability Become Core KPIs
Hyperscaler targets: <4 pJ/bit and >2 Tbps/mm shoreline density. (OIF)
GPU clusters are extremely sensitive to latency and BER; retransmissions significantly slow down training. (OIF / Meta)
System reliability must reach ASIC level or higher so optics do not become the bottleneck. (Meta / Broadcom+Corning)
5. Diverging Explorations of Future Network Architectures
Pragmatists: chiplet-based designs, SiPh + EML/MRM hybrids, incremental evolution. (PhotoniX AI / Huawei / Coherent)
Disruptors: all-photonic AI networks that attempt to eliminate electrical packet switching. (Oriole Networks)
Standardization: IEEE and OIF are accelerating Ethernet/Coherent iterations; interoperability and collaboration are key to deployment. (IEEE / OIF / Marvell)
ECOC 2025 Optical Communications Market Overview
1. Datacom (Data Center Optical Modules)
2023→2024: industry grew ~43% YoY; in Q1 2025 Datacom accounted for 63% of revenue, rising to 65% in Q2; trailing-twelve-month revenue went from $6B→$12B, doubling in two years. (Omia)
High-speed module shipments: 2024 ~22.5 million units → 2025 ~39 million units. (Cignal AI)
1.6T modules: **~20,000 units** shipped in H1 2025, with **~500,000 units** forecast for the full year. (Panel: Cisco / China Telecom / Broadcom / Terahop)
Market value: 2025 ~$15B → 2029 ~$24B. (Cignal AI)
Price target: high-speed module pricing is approaching $0.5/Gbps, with two rounds of price negotiation already in 2025. (Cignal AI)
3.2T timing: not until 2029 will it enter real volume production and take over from 1.6T. (Cignal AI / Huawei)
2. Telecom / Coherent (incl. ZR / ZR+)
Market size: 2025 ~$5B → 2029 ~$8B. (Cignal AI)
Deployment progress:
400ZR is already commercial. (OIF)
800ZR became the workhorse in 2024–2025. (OIF / Marvell)
1600ZR targets spec completion in 2026, with adoption starting in 2027–2028. (OIF / Marvell)
3. OCS (Optical Circuit Switching)
Google's deployment benefits: power 40%, cost 30%, throughput +30%. (Google, cited in an OIF session)
Market size:
2029 ~$1.6B. (Cignal AI, conservative estimate)
2030 ~$2B. (Coherent)
4. Technology Options and Market Segmentation
Scale-out (across racks / across data centers):
Currently dominated by 800G pluggables, transitioning to 1.6T pluggables in 2026–2028. (Cignal AI / Omia)
At 10 km, ZR fills the gap. (Marvell / OIF)
Scale-up (in-rack, 1–20 m):
2024–2025 is still copper-dominated, with reach shrinking to 1–2 m. (Synopsys / OIF)
Optics begin to take over in 2026–2027 (LPO / CPO / chiplet optics / 1060 nm SM-VCSEL + MCF). (Meta / PhotoniX AI / Furukawa / Tokyo Institute of Science)
The 2–40 km gray zone:
Coherent Lite is seen as a candidate for campus DCI and OCS fabrics. (Marvell / OIF)
Market Direction (Key Recap)
Datacom is the main battlefield: market grows $15B→$24B over 2025–2029, units 22.5M→39M. (Cignal AI)
1.6T is the current hot spot: ~20,000 units already shipped in 2025, ~500,000 units for the full year. (Panel)
3.2T only becomes mainstream in 2029. (Cignal AI / Huawei)
The coherent market is smaller but stable: 2025 ~$5B → 2029 ~$8B, centered on 800ZR and 1600ZR. (Cignal AI / OIF / Marvell)
OCS is a new growth segment: $1.6–2B by 2029–2030, driven by energy efficiency and topology simplification. (Cignal AI / Coherent)
Why ZR Coherent Is Becoming More Important (incl. Market)
Technology and Architecture Drivers
AI DCI demand: >95% of DC-to-DC distances are ≤120 km; IMDD reach falls short, and 400/800ZR fits exactly. (Cignal AI / OIF)
IMDD distance bottleneck: each doubling of speed cuts reach roughly fourfold, gradually confining it to ≤10 km; ZR covers 10–120 km+. (Marvell)
Mature standards: 400ZR (2020) → 800ZR (2024) has a complete ecosystem; 1600ZR in 2027–2028 will be adopted, driving IP-over-DWDM. (OIF)
Serviceability / reliability: pluggable coherent + DSP/FEC keeps errors in check, meeting the high-reliability needs of AI workloads. (Marvell / OIF)
Market Size (incl. ZR)
2025 ~$5B → 2029 ~$8B. (Cignal AI)
Growth in 2024–2026 is mainly driven by 800ZR; 1600ZR in 2027–2028 brings a new wave of replacing traditional transponders. (OIF / Cignal AI / Marvell)
Scale-Up vs. Scale-Out: Requirements, Timing, and Market Size
1. Definitions and Requirements
Scale-up (in-rack GPU interconnect)
Range: 1–20 m; mostly GPU-to-GPU within the same rack/row.
Requirements: ultra-low latency, high density, strong thermal capacity; copper reach drops rapidly as speeds rise. (Synopsys / OIF)
Technology forms: CPO, LPO, short-reach IMDD, SiPh/VCSEL chiplets.
Scale-out (across racks / across pods)
Range: tens to hundreds of meters, extending to DCI (10–120 km).
Requirements: long-reach optics are a must, serviceability first, massive bandwidth. (Cignal AI / OIF)
Technology forms: 800G/1.6T pluggables, ZR/ZR+, with some exploration of CPO.
2. Timing
Scale-up
2024–2026: copper dominates, but reach and power are constrained. (Synopsys / OIF)
~2026–2027: next-gen GPU pods drive optical scale-up (CPO / short-reach optical modules). (Meta / Cisco)
~2028+: hybrid optoelectronic architectures; super GPU pods require optical interconnect.
Scale-out
2025: 800G pluggables are mainstream; 1.6T begins qualification. (Cignal AI / Omia)
2026–2028: 1.6T deploys at scale; DCI is led by 800/1600ZR. (OIF / Marvell)
2029–2030: 3.2T goes commercial; AI networks dominate data center infrastructure. (Cignal AI / Huawei)
3. Market Size Differences
Scale-up: an emerging market (SAM expansion), with bandwidth demand 100× that of a traditional CPU bus. (PhotoniX AI)
Scale-out: an existing market (Datacom → AI fabric); 2024–2025 is the inflection point with 800G as mainstream; before 2030 Datacom will be fully dominated by AI networks. (Omia)
Datacom modules ~$23B in 2025 and ~$30B in 2030, mainly from scale-out; scale-up grows faster and its share keeps rising. (Omia / Cignal AI)
Scale-Up: Technology Options and Progress
1. Copper (DAC / ACC / AEC)
Nearly 100% today; but at 200G→400G→800G/lane, reach has already shrunk to 1–2 m. (Synopsys / OIF)
2. LPO / LRO
Saves ~20–30% power versus retimed pluggables; good serviceability, but the interoperability ecosystem is still catching up. (Meta / Cisco)
3. CPO / NPO
Saves ~65% power versus pluggables and another ~35% versus LPO; but serviceability, volume manufacturing and thermal challenges are significant; >15 million hours of lifetime data accumulated. (Meta)
The 102.4T CPO system targets ≥1 Tb/s/mm shoreline density. (Broadcom+Corning)
4. CPC / On-board / Chiplet Optics
Uses ultra-short copper or on-board optics to balance power and serviceability; the chiplet form factor adds flexibility. (PhotoniX AI / Cisco)
5. Materials and Modulation
SiPh + EML/MZM (>100 GHz), MRM (low power but needs thermal control/linearization), TFLN (ultra-wide bandwidth but cost/integration unresolved). (Huawei / Source Photonics / fibeReality / Coherent)
Timeline
2024–2025: copper still mainstream, LPO ramps up. (Synopsys / Meta / Cisco)
2026–2027: CPO/NPO pilot commercial deployments. (Meta / Broadcom)
2028+: hybrid / chiplet approaches may become mainstream. (PhotoniX AI)
VCSEL (Its Role in Scale-Up)
1. Positioning
Short reach with high efficiency, low cost and high channel density; traditional 850 nm multimode is limited to <100 m.
2. New Developments
1060 nm single-mode VCSEL + multicore fiber (MCF):
Tokyo Institute of Science: 2.88 Tbps (16×180 Gbps) @500 m. (Tokyo Institute of Science)
Furukawa: 50 Gbaud×16ch CPO, 2 km transmission, 3.95 fJ/bit, more efficient than many SiPh CPO solutions. (Furukawa)
3. Technical Challenges
Reaching 400G/lane still requires bandwidth and reliability breakthroughs; long-term high-temperature lifetime needs validation.
4. Comparison with Other Technologies
Vs. EML/SiPh: advantages in cost and efficiency, but slightly weaker reach/bandwidth.
Vs. MRM/TFLN: better suited to short-reach, large-scale integration; the latter are better for high bandwidth.
In 1–20 m scale-up scenarios it is already an attractive low-power path. (Furukawa / Tokyo Institute of Science)
Why OCS Is Needed
1. Bottlenecks of Traditional Electrical Switching
Multiple O-E-O layers bring high power, high latency and high cost; unstable queuing delay hurts synchronized AI collectives. (OIF / Google)
2. Characteristics of AI Networks
Long-lived, high-volume traffic (elephant flows) + tight synchronization: any slow link drags down the whole job.
3. Efficiency and Cost Advantages
Google's measured results: power 40%, cost 30%, throughput +30%. (Google, cited in an OIF session)
4. Architectural Simplification and Reliability
Flatter topology, fewer active components, smaller failure domains; synergy with the all-photonic network vision. (Oriole Networks)
5. Status and Challenges
Google's OCS is deployed at scale, but reconfiguration takes hours/days; startups are attempting ns-level optical switching, which requires matching software and cost structures. (Google / Oriole Networks)
Modulator Technology Comparison: Pros, Cons and Progress (Quick Look)
1. SiPh-MZM (Mach-Zehnder Modulator)
Pros: mature and reliable, the workhorse for 400/800G, CMOS-compatible. (Huawei)
Cons: large footprint, relatively high drive voltage and power, high cost.
Progress: already the mainstay for 200G/lane; >110 GHz demonstrated, can support 400G/lane (packaging is a major challenge). (Huawei)
2. MRM (Micro-Ring Modulator)
Pros: small size, high density, low power; suited to CPO / chiplets. (fibeReality / PhotoniX AI)
Cons: temperature-sensitive and needs heater tuning, ~40 GHz bandwidth, insufficient linearity.
Progress: 200G PAM4 demonstrated; reaching 400G/lane requires breakthroughs in thermal control / linearization.
3. EML (Electro-Absorption Modulated Laser)
Pros: compact single chip, high bandwidth (~100 GHz), suited to 2–10 km. (Huawei / Source Photonics)
Cons: higher cost, limited thermal stability, relatively low output power.
Progress: the mainstay for 400G/800G LR; supports 448G/lane and 1.6T/3.2T FR/LR. (Huawei / Source Photonics)
4. VCSEL (Vertical-Cavity Surface-Emitting Laser)
Pros: lowest cost, excellent efficiency (fJ/bit class), deployed in volume for short reach.
Cons: ~40 GHz bandwidth, multimode interference, traditional reach <100 m.
Progress: 1060 nm single-mode + MCF can reach 2 km and 2.88 Tbps; 3.95 fJ/bit efficiency is strong, but 400G/lane and lifetime breakthroughs are still needed. (Furukawa / Tokyo Institute of Science)
5. TFLN (Thin-Film Lithium Niobate)
Pros: >100 GHz, low Vπ, high linearity; suited to advanced modulation formats. (Coherent / fibeReality)
Cons: high cost, limited integration with Si CMOS, yield challenges.
Progress: >100 GHz demonstrated with potential for 400G/lane and beyond, but not yet in volume production.
6. InP (Indium Phosphide)
Pros: high bandwidth, can be integrated with SOAs.
Cons: high cost, limited integration, worse power than SiPh.
Progress: stable in coherent; in Datacom it is gradually being replaced by SiPh/EML, and is better suited to long-haul telecom. (Marvell / OIF)
Energy Efficiency and Thermal Management
1. Why It Has Become the Top Metric
In AI factories, optical interconnect's share of energy consumption is rising to 20–30%; if not reduced, it will eat into the power available to GPUs. (OIF / Meta)
Hyperscaler KPIs: <4 pJ/bit and >2 Tbps/mm. (OIF)
2. Sources of Module Power Consumption
DSP (5–15 W for coherent), driver/TIA (200–400 mW per channel), lasers/pumps, TEC/cooling (up to 25–30%), MRM heaters, etc. (Meta / Lumentum / Source Photonics)
3. Technical Levers for Saving Power
Architecture: LPO (~30% savings vs. retimed), CPO (~65% savings vs. pluggable, another ~35% vs. LPO). (Meta)
Component integration: 4-in-1 pumps sharing one TEC (25%+ cooling power savings), DGE+OCM integration, EML differential drive at ~1 V. (Lumentum / Source Photonics)
Modulation/FEC: short-reach IMDD is more power-efficient; higher-order PAM saves bandwidth but increases FEC/DSP power and latency. (IEEE / Synopsys)
4. Thermal Management
CPO requires liquid or immersion cooling; 102.4T CPO systems need co-design of thermal management and fiber routing. (Broadcom+Corning)
Air cooling is nearing its limits; liquid cooling (direct-to-chip) is becoming mainstream, and immersion cooling is under evaluation. (Cisco / Meta / Google)
5. Industry Perspectives
Meta: CPO testing shows **~65% energy savings**.
OIF: presented an energy-efficiency roadmap and targets.
Cisco: advocates "thermal design first."
Broadcom+Corning: demonstrated a systematic 102.4T CPO system design.
Summary
Energy efficiency is the new KPI (<4 pJ/bit, >2 Tbps/mm).
The biggest consumers are DSP / pumps / TEC.
Solutions: LPO / CPO, pump integration, differential drive, SiPh integration.
Thermal management evolves: air cooling → liquid cooling → immersion cooling.
Reliability must be ≥ ASIC level, or it drags down overall efficiency.
Conclusion
ECOC 2025 clearly mapped the transformation of the optical communications industry: AI is driving demand at an unprecedented pace, Datacom has become the core battlefield, 1.6T and 3.2T optical modules are taking over in succession, while ZR coherent and OCS fill the needs for longer reach and better efficiency. Short-reach scale-up is shifting from copper to optics, mid-to-long-reach scale-out relies on pluggables and coherent, and the overall market will keep expanding through 2025–2030. The various modulator technologies (MZM, MRM, EML, VCSEL, TFLN, InP) and module architectures (pluggable, LPO, CPO, chiplet) will not play out as a single-winner elimination race, but will coexist and evolve according to application. Ultimately, energy efficiency and reliability will be the core KPIs for judging whether a technology gets deployed, while standardization and supply-chain maturity will determine the pace of market adoption.
In short, ECOC 2025 told us this: optical communications in the AI era is no longer just about chasing speed — it is about finding the best balance among power, latency, reliability and cost. This industry restructuring will define the foundation of future data centers and AI factories.
Comments