OCS is expected to become a key optical interconnect solution in the next stage of AI networks

Overview

This report focuses on optical circuit switches (OCS) in AI data center networking. OCS is a fully optical switching device that enables end-to-end optical networks through dynamic lightpath adjustment, reducing traditional opto-electrical-optical conversion. The report argues that as AI cluster scale expands and network power and latency become bottlenecks, OCS may move from Google TPU network proof-of-concept toward a broader deployment and testing stage among hyperscale cloud providers.

Core views

The core view is that OCS is not a substitute for all existing optical interconnect solutions; rather, it is complementary to CPO and pluggable transceivers at different network layers. OCS is suitable for dynamic, flexible routing and can be used both in TPU resource pooling scenarios in scale-up environments, in scale-out settings to replace part of Layer 1-2 electrical switching, and in scale-across interconnects across racks, clusters, and sites. If NVIDIA advances OCS in its next-generation Dragonfly/Feynman architecture, the industry may see an accelerated adoption inflection.

Analysis framework

The report combines technology-path comparison, application-scenario segmentation, TAM forecasting, supply-chain mapping, and vendor progress tracking. It first explains OCS advantages over traditional electrical switching in terms of low latency, low power, and high bandwidth, then compares four OCS approaches (MEMS, DLC, DLBS, and SiPh), and finally maps upstream precision optics, crystal materials, optical chips, FAU, and midstream systems/integration manufacturers.

Methodology notes

• Technology comparisonOCS technology route comparison

  Four OCS approaches: MEMS, DLC, DLBS, SiPh

  Compares maturity and applicable scenarios of different OCS approaches by port scale, per-port cost, insertion loss, switching time, crosstalk, and reliability.

• Market sizingTAM forecast

  OCS market size and CAGR

  Using Cignal AI estimates, the report estimates OCS market size rising from about USD 400 million in 2025 to over USD 2.5 billion in 2029, corresponding to roughly 58% CAGR over four years.

• Application scenariosScale-up / Scale-out / Scale-across network layering

  AI data center network hierarchy

  OCS use is divided into intra-cluster resource pooling, horizontal network expansion substituting part of electrical switching, and long-distance high-bandwidth interconnect across racks or clusters.

• Value chain analysisSupply-chain mapping

  Upstream components and midstream system integration

  Identifies key components such as MEMS mirrors, FAU, optical chips, lenses, and fiber arrays, and midstream system/module manufacturers to assess likely beneficiaries.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Coherent Corp (COHR.US)

  Represents the DLC approach and is involved in OCS development.

  Strengths

  The DLC approach has low-voltage and high-reliability characteristics, suitable for scenarios with stringent stability requirements.

  Weaknesses

  The report notes that liquid crystal approaches face greater challenges in volume production and yield ramp.

  Comparison

  Compared with MEMS, DLC is a non-mechanical approach; compared with SiPh, it has advantages in insertion loss and reliability characteristics.

  Risks

  OCS adoption timing, customer validation cycles, and mass-production yield remain uncertain.

• Lumentum Holdings Inc (LITE.US)

  One of the main developers of the MEMS approach.

  Strengths

  MEMS is currently the most mature OCS technology path, with a relatively simple optical path and a more stable supply chain.

  Weaknesses

  The table marks MEMS reliability as relatively lower, and single-port cost as higher.

  Comparison

  Together with Google, Lumentum represents the MEMS route, with maturity higher than the mass-production stage of DLC and SiPh.

  Risks

  Intensifying competition, client adoption timing, and technology-path displacement risk.

• NVIDIA Corp (NVDA.US)

  AI network architecture leader; discusses next-generation Dragonfly architecture and tracks OCS deployment in cross-rack and cross-cluster networks.

  Strengths

  Has ecosystem influence in defining AI cluster network architectures and advancing CPO+OCS combined solutions.

  Weaknesses

  The report does not provide details on NVIDIA's in-house OCS commercialization.

  Comparison

  Compared with optical communication component makers, NVIDIA is more of a system architecture and demand-pull force.

  Risks

  Uncertainty in architecture path, deployment timelines, and supply-chain choices.

• Suzhou TFC (300394.SS)

  A key component supplier that the report believes could benefit from OCS adoption.

  Strengths

  Involves critical OCS-related components such as lenses and fiber arrays, and is rated Buy in the report.

  Weaknesses

  The report does not detail the share of OCS revenue or customer validation progress.

  Comparison

  Compared with system makers, it is more likely to benefit from expansion in upstream component value capture.

  Risks

  Underperformance of premium optical module demand, tighter competition, slower product upgrades, and export-price/geopolitical risks.

• InnoLight

  A Chinese optical communication company and potential early beneficiary of OCS, whose subsidiary TeraHop is involved in SiPh-based OCS solutions.

  Strengths

  Displayed OCS products at OFC 2026 and has potential to expand from components to full OCS system solutions.

  Weaknesses

  The SiPh approach still faces high insertion loss, crosstalk, and multi-channel reliability challenges.

  Comparison

  Compared with Coherent/Lumentum, Chinese peers may enter via components and extend into full OCS manufacturing.

  Risks

  Customer validation, technology-route selection, yield in mass production, and export environment for offshore customers.

• Eoptolink / Accelink

  Chinese OCS product demonstration vendors entering full OCS manufacturing competition.

  Strengths

  Showed OCS-related products at OFC 2026, indicating rising participation by Chinese vendors.

  Weaknesses

  The report does not provide detailed information on orders, revenue, or mass-production progress.

  Comparison

  Represent Chinese optical communication players extending OCS system capability, similar to InnoLight.

  Risks

  Commercialization timing, cost control, customer qualification, and technology iteration risk.

Key data

• 2025 OCS market sizeabout USD400mnEstimated by Cignal AI; market in 2025 is led mainly by Google.
• 2029 OCS market size forecastabove USD2.5bnCignal AI forecasts a roughly 58% CAGR for 2025-2029.
• OCS latencyabout 10-100 nsReported to be about 1/100 of electrical switch latency.
• OCS power consumptionabout 1/5 of electrical switch power at the same bandwidthOnly the drive module consumes power, without needing switching chips and companion optical transceivers.
• Google Apollo OCS performance30% lower cost, 40% lower powerBased on the report’s description of Google’s MEMS OCS platform.
• Google research on scale-out40% lower power, 30% lower cost, 30% higher throughputReportedly citing Google study results.
• Network solution power comparisonPluggable 83 pJ/bit; Pluggable+OCS 50 pJ/bit; CPO 48 pJ/bit; CPO+OCS 31 pJ/bitComparison cited from NVIDIA in AIDC scenarios.
• CPO+OCS energy-saving effectabout 2.6x lower power than conventional pluggable solutionThe report states that CPO+OCS is the lowest-power combination in NVIDIA's cited design.

Impact & implications

If OCS shifts from being dominated by a single user (Google) to adoption by multiple cloud providers, value could expand from traditional optical transceivers to include OCS systems, optical components, MEMS arrays, FAU, lenses, optical chips, and system integration. Chinese optical communication players that can upgrade from component supply to full OCS products may improve their strategic position in AI data center network architectures.

Risks

• OCS insertion loss remains relatively high and may fail to meet standard optical module link budgets.
• Current OCS port density is still limited; connector density and performance may not match high-density wiring requirements of AI clusters.
• OCS failure modes, redundancy design, and hot-swap capability still need validation, potentially affecting high-availability requirements of AI clusters.
• Overall cost is still relatively high, which may constrain large-scale deployment.
• New routes such as DLC and SiPh face challenges in yield, insertion loss, crosstalk, and reliability during mass production.
• Adoption by hyperscale cloud providers is still in testing or early phases, and actual deployment speed may be slower than expected.
• Competition remains intense in optical module and optical component markets, and price pressure plus geopolitical factors may affect supplier profitability.

What to watch

• Deployment timelines by hyperscale providers such as Microsoft, META, and NVIDIA after small-batch OCS testing.
• Whether NVIDIA's next-generation Feynman architecture and Dragonfly topology adopt OCS.
• Progress in MEMS, DLC, DLBS, and SiPh routes regarding insertion loss, port density, switching time, and reliability.
• Customer validation, mass production, and order traction for OCS products from Chinese vendors such as InnoLight, Eoptolink, and Accelink.
• Actual power, cost, and throughput improvements of OCS plus CPO combinations in AI data centers.
• Demand-supply changes for key upstream components such as MEMS arrays, FAU, lenses, optical chips, and fiber arrays.