AI資料中心光路交換市場規模、佔有率和成長預測（2026-2032年）

2026年人工智慧資料中心光路交換市場規模估計為6.2億美元，預計2032年將達到35.8億美元。這意味著從2026年到2032年的複合年成長率為33.9%。

傳統資料中心網路透過多層電訊號分封交換來路由流量，從而建構連接。這種方式可以實現細粒度路由和統計復用，但隨著網路規模的擴大，需要高功率的交換晶片、收發器以及反覆的光電轉換。光路交換則提供了不同的功能。它在選定的端點之間建立專用光路，使高速資料訊號保持在光域內。由於光路交換本身不檢查資料包，因此對依賴延遲和頻寬的處理負載增加甚微。這非常適合那些持續時間足夠長，足以證明重新配置物理拓撲結構是合理的流量模式。谷歌在這方面擁有最令人信服的實際經驗。其 Jupiter 網路多年來一直採用光路交換技術，其雲端張量處理單元 (TPU) 系統也使用光路交換來連接加速器模組，並透過軟體繞過故障的光鏈路進行重新路由。 Ironwood 整合了晶片間互連、光路交換和資料中心網路，使系統能夠擴展到每個超級節點最多 9,216 個 TPU 晶片。谷歌也曾在其目前的TPU模組拓樸結構中提及光路交換技術。這顯示光路交換技術可以作為運作環境中的人工智慧互連層，而不僅限於實驗室或測試自動化領域。目前，商業市場正在蓬勃發展。 Lumentum提供基於64x64和300x300 MEMS（微電子機械系統）的平台，並採用基於SONiC的控制和gNMI管理。 Coherent提供從64x64到512x512的光路交換配置。 Huber+Suhner提供高達384x384的POLATIS系統，並計劃透過在波蘭開設專用生產工廠，將其產能至少提高五倍。 Molex憑藉其HiRadix平台進入市場，而Calient和Telescent則將成熟的光路交換和自動光纖交換技術引入人工智慧網路領域。

市場促進因素

• 隨著人工智慧叢集規模的擴大，固定網路拓撲的成本也隨之增加。大規模訓練和推理系統擴大跨越成千上萬個加速器。在這種規模下，如果鏈路發生故障或工作負載的通訊模式與實體網路不匹配，固定網路拓撲可能會導致昂貴的運算資源閒置。谷歌正在利用光路交換技術來提高切片可用性，該技術可以重新配置張量處理單元 (TPU) 之間的互連，並重新路由以繞過故障點。商業光路交換供應商正嘗試將類似的原理應用於圖形處理單元 (GPU) 和異質加速器叢集。軟體允許在不重新佈線資料中心或更換分組交換硬體的情況下重新定位實體光路。
• 電交換的電力消耗為光旁路技術的引入創造了空間。高頻寬電分組交換消耗大量電力，並且每次光纖鏈路連接到交換器時都需要進行光電轉換。光路交換可以繞過某些分組交換階段，用於不需要逐包路由的資料流，從而減少轉換和交換開銷。谷歌已報告稱，透過將光路交換整合到其 Jupiter 系統中，顯著電力消耗並降低了成本。同時，Lumentum、Coherent、Calient 和 HUBER+SUHNER 等商業供應商強調協議滲透性和交換路徑上的極低功耗。隨著加速器網路在人工智慧工廠總電力消耗中所佔比例越來越大，這一優勢顯得尤為重要。
• 光通道速率越高，協定透明交換的價值就越大。由於光路交換器路由的是光訊號，而不是處理特定的乙太網路或InfiniBand資料速率，因此即使連結從400G和800G升級到1.6T及更高速率，同一光交換器底盤也能繼續使用。 Calient和POLATIS將其平台定位為與位元率和協定無關，而Lumentum和Coherent則專門為人工智慧網路開發光路交換器。與依賴與連接埠速度和傳輸容量緊密相關的專用積體電路（ASIC）的分封交換相比，這可以延長設備的使用壽命。
• 軟體定義拓撲結構實現了可組合的人工智慧基礎設施。當運算、儲存和加速器被視為資源池而非永久連接的系統時，動態光交換的價值將進一步提升。 GoogleGKE的動態切片功能可以編配光路交換網路中張量處理單元（TPU）的實體重配置。同時，Drut Technologies和POLATIS也展示了光子原生資源配置。 Molex同樣將高基數光路交換定位為動態重配置人工智慧叢集拓樸結構的一種手段。隨著編配軟體的日益成熟，光路交換的應用範圍可能會從靜態容錯和自動化測試轉向工作負載感知的基礎設施配置。

市場限制因素

• 光路交換並不能完全取代所有流量流中的分組交換。許多微機電系統 (MEMS) 和機器人系統需要幾毫秒甚至更長的重配置時間，因此對於短時或突發性極強的流量，電分組交換架構仍然不可或缺。網路控制軟體必須決定何時建立電路、維持電路多長時間以及如何在重新配置期間安全地傳輸流量。隨著基數的增加，光損耗、連接埠數量、連接器密度和光纖管理都變得更加複雜。此外，由於光路交換器可以高度整合到實體層，因此運營商需要冗餘路徑。最後，由於大部分大規模運作經驗仍集中在超大規模資料中心業者大規模資料中心，因此要實現更廣泛的應用，商業平台需要證明其在大規模環境中的互通性、可靠性和易用性。

目錄

第1章：執行摘要

• 市場機會與主要發現
• OCS商業化計劃
• 主要收入來源

第2章 市場概覽

• 光路交換與分組交換的比較
• OCS在AI規模化和規模化架構中的作用
• 動態拓撲和災難復原
• OCS控制與編配
• 第三方OCS與內部開發的OCS實現方案的比較

第3章 市場規模及預測（2026-2032 年）

• 全球市場銷售
• 年增率分析
• OCS港口貨物運輸
• 按業務運營商和實施類型分類的銷售額

第4章 市場：透過技術轉換

• MEMS自由空間光路開關
• 機器人和實體光纖交叉連接
• 整合式光電光開關
• 其他新興的光開關技術

第5章 市場：透過 Switch Radix

• 最大 64×64
• 65x65 到 320x320
• 320x320 或更大
• 模組化多底盤架構

第6章 市場：依網路角色分類

• 人工智慧規模化架構
• 重建人工智慧橫向擴展脊柱
• 災後復原與韌性
• 資源分解和組合式基礎架構
• 測試、配置和自動化光纖管理

第7章 市場：依客戶類型分類

• 超大規模雲端供應商
• Neocloud 和 GPU 雲端供應商
• 主權人工智慧基礎設施
• 高效能運算
• 企業人工智慧基礎設施

第8章 區域市場

• 北美洲
  • 美國
  • 加拿大
• 歐洲
• 亞太地區
• 中東和世界其他地區

第9章 市場動態

• 促進因素
  • 擴展加速器集群
  • 資料包交換處理能力和 OEO 轉換開銷
  • 協定透明的頻寬擴展
  • 軟體定義的可組合人工智慧基礎設施
• 抑制因子
  • 電路重配置時間
  • 控制平面和調度複雜性
  • 高基數光損耗和光纖管理
  • 超大規模資料中心業者實施案例研究

第10章 競爭格局

• 價值鏈
• MEMS OCS供應商
• 機器人光交叉連接器器供應商
• 整合光電OCS開發公司
• AI網路平台與編配生態系統
• 產能和超大規模供應合約

第11章：公司簡介

• Lumentum
• Coherent
• HUBER+SUHNER / POLATIS
• Molex
• Calient
• Telescent
• Google
• Drut Technologies
• NVIDIA
• Broadcom
• Cisco
• Arista Networks
• Ciena
• Fujitsu
• NEC

第12章 近期趨勢

第13章附錄

The AI Data Center Optical Circuit Switching Market is estimated at USD 0.62 billion in 2026 and is projected to reach USD 3.58 billion by 2032, representing a CAGR of 33.9% during 2026-2032.

Traditional data-center networks build connectivity by passing traffic through several layers of electrical packet switches. This provides fine-grained routing and statistical multiplexing but requires high-power switch silicon, transceivers and repeated optical-electrical-optical conversion as networks scale. Optical circuit switching introduces a different function. It establishes a dedicated optical path between selected endpoints and leaves the high-speed data signal in the optical domain. The optical circuit switch itself does not inspect packets and therefore adds very little latency or bandwidth-dependent processing power. It is best suited to traffic patterns that persist long enough to justify reconfiguring the physical topology. Google provides the strongest production proof point. Its Jupiter network has used optical circuit switching for years, and Cloud tensor processing unit systems use optical circuit switching to connect accelerator cubes while software reroutes around failed optical links. Ironwood combines inter-chip interconnect, optical circuit switching and the data-center network into a system that scales to 9,216 tensor processing unit chips per superpod. Google also describes optical circuit switching as part of the current tensor processing unit pod topology. This shows that optical circuit switching can operate as a production AI interconnect layer rather than only a laboratory or test-automation technology. The merchant market is now broadening. Lumentum offers 64 by 64 and 300 by 300 micro-electromechanical systems based platforms with SONiC-based control and gNMI management. Coherent lists optical circuit switching configurations from 64 by 64 through 512 by 512. HUBER+SUHNER offers POLATIS systems up to 384 by 384 and opened a dedicated production site in Poland with plans to increase capacity at least fivefold. Molex is entering with a high-radix platform, while Calient and Telescent bring established optical circuit switching and automated-fiber-switching technologies into AI networking.

Market Drivers

• AI cluster scale increases the cost of fixed network topology. Large training and inference systems increasingly span thousands to tens of thousands of accelerators. At this scale, fixed network topology can leave expensive compute resources stranded when links fail or when workload communication patterns do not match the physical network. Google uses optical circuit switching to reconfigure tensor processing unit interconnects and route around failures, improving slice availability. Merchant optical circuit switching suppliers are positioning the same principle for graphics processing unit and heterogeneous accelerator clusters. Software can rearrange physical optical paths without recabling the data center or replacing packet-switch hardware.
• Electrical switching power creates room for optical bypass. High-bandwidth electrical packet switching consumes significant power and requires optical-electrical-optical conversion wherever fiber links terminate at a switch. Optical circuit switching can bypass selected packet-switch stages for flows that do not need per-packet routing, reducing conversion and switching overhead. Google reported substantial network power and cost benefits from integrating optical circuit switching into Jupiter, while commercial suppliers such as Lumentum, Coherent, Calient and HUBER+SUHNER emphasize protocol transparency and very low switching-path power. The benefit becomes more valuable as accelerator networks consume a larger share of total AI-factory power.
• Higher optical lane rates increase the value of protocol-transparent switching. Optical circuit switches route light rather than process a specific Ethernet or InfiniBand data rate, allowing the same optical switching chassis to remain useful as links migrate from 400G and 800G toward 1.6T and later generations. Calient and POLATIS position their platforms as bit-rate and protocol independent, while Lumentum and Coherent are developing optical circuit switching specifically around AI networking. This can extend equipment life compared with packet switches whose application-specific integrated circuit generation is tightly tied to port speed and forwarding capacity.
• Software-defined topology enables composable AI infrastructure. Dynamic optical switching becomes more valuable when compute, storage and accelerators are treated as pools rather than permanently wired systems. Google GKE dynamic slicing can orchestrate physical reconfiguration in the tensor processing unit optical circuit switching network, while Drut Technologies and POLATIS have demonstrated photonic-native resource composition. Molex similarly positions high-radix optical circuit switching as a way to dynamically reconfigure AI cluster topology. As orchestration software matures, optical circuit switching can move from static resilience and test automation toward workload-aware infrastructure composition.

Market Restraints

• Optical circuit switching does not replace packet switching across every traffic flow. Reconfiguration takes milliseconds or longer in many micro-electromechanical systems and robotic systems, so short-lived or highly bursty flows still require electrical packet fabrics. Network control software must decide when a circuit should be created, how long it should remain in place and how traffic is moved safely during reconfiguration. Optical loss, port count, connector density and fiber management become harder as radix increases. Operators also need redundant paths because an optical circuit switch can become a high-concentration physical-layer element. Finally, much of the largest production experience remains concentrated among hyperscalers, so broader adoption depends on merchant platforms proving interoperability, reliability and operational simplicity at scale.

Technology and Segment Insights

• Micro-electromechanical systems based free-space optical circuit switching represents the largest 2026 revenue pool. Lumentum, Coherent, Calient and HUBER+SUHNER all use mature optical beam-steering or micro-electromechanical systems approaches to deliver hundreds of ports with low optical loss and protocol-transparent operation. The technology benefits from field experience in telecom, test automation and hyperscale networks and is therefore the most commercially mature route for high-radix AI fabrics. Micro-electromechanical systems optical circuit switches form the largest commercial architecture in 2026, and 320 by 320 and larger switches form an important hyperscale deployment class during the forecast.
• Integrated-photonic optical switching is expected to grow fastest from a small base through 2032. Silicon-photonic and other planar approaches can potentially reduce size and switching time while integrating monitoring or wavelength functions more tightly. However, high radix, low loss, manufacturability and thermal stability remain significant engineering constraints. Integrated-photonic switching is the fastest-growing emerging technology from a small commercial base. Robotic fiber cross-connect systems form a separate niche where physical fiber paths can be reconfigured more slowly but at very high connector density, which is useful for planned topology changes, provisioning and large parallel-optics fabrics.
• By switch radix, the market spans up to 64 by 64, 65 by 65 to 320 by 320, above 320 by 320, and modular multi-chassis architectures. By network role, the market covers AI scale-up fabrics, AI scale-out and spine reconfiguration, failure recovery and resiliency, resource disaggregation and composable infrastructure, and test, provisioning and automated fiber management. By customer type, demand comes from hyperscale cloud providers, neocloud and graphics-processor-cloud operators, sovereign AI infrastructure, high-performance computing, and enterprise AI infrastructure. Optical circuit switching control and orchestration software form a recurring support layer covering topology programming, failover and workload-aware circuit control. Monitoring and optical circuit switching integration services form a project-based layer covering power monitoring, validation, deployment and lifecycle support.

Competitive and Strategic Outlook

• The merchant optical circuit switching market remains concentrated compared with Ethernet switching or optical transceivers. Lumentum, Coherent, HUBER+SUHNER with POLATIS, Molex, Calient and Telescent have the clearest commercial optical circuit switching or automated optical-switching portfolios. Their technologies differ in port count, switching mechanism, speed, insertion loss, control interfaces and target application. Micro-electromechanical systems dominate high-radix direct optical switching, while robotic cross-connects serve slower topology reconfiguration and physical provisioning use cases.
• Large network and AI platform companies form the surrounding ecosystem. Google is the most important reference operator because it has deployed optical circuit switching deeply in production tensor processing unit and Jupiter networks. NVIDIA, Broadcom, Cisco and Arista influence where optical circuit switching can complement packet fabrics as AI scale-up and scale-out topologies evolve. Drut Technologies is developing photonic-native composable infrastructure using POLATIS, while Ciena, Fujitsu and NEC bring optical-network control and switching expertise relevant to future data-center optical circuit switching integration. Competitive advantage increasingly depends on radix, optical loss, switching speed, reliability, software control, serviceability and manufacturing scale.
• Recent developments confirm accelerating commercialization. In September 2026, Coherent showcased optical circuit switching alongside co-packaged optics, near-packaged optics and next-generation AI data-center connectivity technologies at ECOC 2026. Lumentum exhibited its R300 optical circuit switch at ECOC 2026 as part of a broader scalable-AI infrastructure portfolio. In March 2026, Molex introduced a High-Radix Optical Circuit Switch Platform designed for large-scale reconfigurable AI cluster fabrics. Telescent launched a high-density robotic optical circuit switch for 1.6T DR4 and DR8 parallel-optics connections in graphics processing unit clusters. During 2026, Google expanded software-controlled optical circuit switching usage through dynamic tensor processing unit slicing and current tensor processing unit cluster architectures. In May 2026, Lumentum described micro-electromechanical systems based optical circuit switching as a strategic technology for reconfigurable AI data-center fabrics. Between 2025 and 2027, HUBER+SUHNER opened a dedicated POLATIS manufacturing site in Pisary, Poland, planning an at least fivefold capacity expansion over two years, which was followed by major multi-year hyperscale optical circuit switching orders.
• North America is the largest early market for AI data-center optical circuit switching because the United States combines hyperscale AI operators, cloud tensor processing unit deployments, large graphics processing unit clusters and a dense photonics supplier base. Google provides the most mature production example through Jupiter and tensor processing unit interconnect systems. Lumentum, Coherent, Molex, Calient and Telescent are all headquartered or maintain major operations in the region, giving United States operators access to both high-radix micro-electromechanical systems and emerging automated-fiber-switching platforms.
• The commercial opportunity is strongest where the cost of idle accelerators or redundant packet-switch layers is high enough to justify a programmable optical fabric. Training clusters, large inference domains and composable AI infrastructure meet that condition because network topology can materially affect accelerator utilization. Optical circuit switching is also attractive for operators that need rapid failover around fiber or switch-path faults without physically recabling thousands of links. The economics improve as optical lane rates increase because the same protocol-transparent optical circuit switch can switch successively higher amounts of bandwidth per port.
• North American deployments are likely to remain heterogeneous through 2032. Hyperscalers may continue to use proprietary or heavily customized systems, while neoclouds and enterprise-scale AI operators are more likely to adopt merchant optical circuit switching platforms integrated with Ethernet or InfiniBand packet fabrics. Control software and orchestration therefore become as important as the optical switch itself. Suppliers that expose standards-based application programming interfaces, telemetry and topology-management interfaces are better positioned to fit into multi-vendor AI fabrics.
• Asia Pacific is important through optical-component manufacturing, hyperscale buildout and large domestic AI infrastructure programs. Europe has a smaller deployment base but hosts HUBER+SUHNER and a strong photonics ecosystem, and the new POLATIS production expansion in Poland increases regional manufacturing capacity. Other regions are expected to adopt optical circuit switching selectively in sovereign AI and very large greenfield campuses where network architecture can be optimized from the initial design stage.

Conclusion

• The AI data center optical circuit switching market is positioned for rapid expansion, growing from USD 0.62 billion in 2026 to USD 3.58 billion by 2032 at a compound annual growth rate of 33.9 percent. Micro-electromechanical systems optical circuit switches form the largest commercial architecture in 2026, while integrated-photonic switching is the fastest-growing emerging technology from a small commercial base. AI resource composability increases the value of programmable physical-layer connectivity across clusters, and 320 by 320 and larger switches form an important hyperscale deployment class during the forecast. North America leads adoption through hyperscale AI operators and optical-networking suppliers, with Asia Pacific contributing through optical-component manufacturing and Europe through photonics and manufacturing capacity expansion. Reconfiguration time, control-plane complexity, high-radix optical loss and concentrated hyperscaler deployment experience remain the principal constraints. Success will depend on radix, optical loss, switching speed, reliability, software control, serviceability and manufacturing scale. Suppliers that combine mature optical hardware with standards-based orchestration interfaces are best positioned to capture value as optical circuit switching moves from hyperscale-specific deployments toward broader AI infrastructure adoption.

Key Benefits of this Report

• Insightful Analysis: Detailed market insights across regions, customer segments, policies, socio-economic factors, consumer preferences, and industry verticals.
• Competitive Landscape: Understand strategic moves by key players to identify optimal market entry approaches.
• Market Drivers and Future Trends: Assess major growth forces and emerging developments shaping the market.
• Actionable Recommendations: Support strategic decisions to unlock new revenue streams.
• Caters to a Wide Audience: Suitable for startups, research institutions, consultants, SMEs, and large enterprises.

What Businesses Use Our Reports For

• Industry and market insights, opportunity assessment, product demand forecasting, market entry strategy, geographical expansion, capital investment decisions, regulatory analysis, new product development, and competitive intelligence.

Report Coverage

• Historical data from 2021 to 2024, Base year 2025, and Forecast years from 2026 to 2032
• Growth opportunities, challenges, supply chain outlook, regulatory framework, and trend analysis
• Competitive positioning, strategies, and market share evaluation, and trade analysis
• Revenue growth and forecast assessment across segments and regions
• Company profiling including strategies, products, financials, and key developments

TABLE OF CONTENTS

1. Executive Summary

• 1.1. Market Opportunity and Key Findings
• 1.2. OCS Commercialization Timeline
• 1.3. Principal Revenue Pools

2. Market Overview

• 2.1. Optical Circuit Switching versus Packet Switching
• 2.2. OCS Role in AI Scale-Up and Scale-Out Fabrics
• 2.3. Dynamic Topology and Failure Recovery
• 2.4. OCS Control and Orchestration
• 2.5. Merchant versus Proprietary OCS Deployment

3. Market Size and Forecast, 2026-2032

• 3.1. Global Market Revenue
• 3.2. Annual Growth Analysis
• 3.3. OCS Port Shipments
• 3.4. Revenue by Merchant and Custom Deployment

4. Market by Switching Technology

• 4.1. MEMS Free-Space Optical Circuit Switches
• 4.2. Robotic / Physical Fiber Cross-Connects
• 4.3. Integrated-Photonic Optical Switches
• 4.4. Other Emerging Optical Switching Technologies

5. Market by Switch Radix

• 5.1. Up to 64x64
• 5.2. 65x65 to 320x320
• 5.3. Above 320x320
• 5.4. Modular Multi-Chassis Architectures

6. Market by Network Role

• 6.1. AI Scale-Up Fabrics
• 6.2. AI Scale-Out / Spine Reconfiguration
• 6.3. Failure Recovery and Resiliency
• 6.4. Resource Disaggregation and Composable Infrastructure
• 6.5. Test, Provisioning and Automated Fiber Management

7. Market by Customer Type

• 7.1. Hyperscale Cloud Providers
• 7.2. Neocloud and GPU-Cloud Operators
• 7.3. Sovereign AI Infrastructure
• 7.4. High-Performance Computing
• 7.5. Enterprise AI Infrastructure

8. Regional Market

• 8.1. North America
  • 8.1.1. United States
  • 8.1.2. Canada
• 8.2. Europe
• 8.3. Asia Pacific
• 8.4. Middle East and Rest of World

9. Market Dynamics

• 9.1. Drivers
  • 9.1.1. Growth in Accelerator Cluster Scale
  • 9.1.2. Packet-Switch Power and OEO Conversion Overhead
  • 9.1.3. Protocol-Transparent Bandwidth Scaling
  • 9.1.4. Software-Defined Composable AI Infrastructure
• 9.2. Restraints
  • 9.2.1. Circuit Reconfiguration Time
  • 9.2.2. Control-Plane and Scheduling Complexity
  • 9.2.3. High-Radix Optical Loss and Fiber Management
  • 9.2.4. Concentrated Hyperscaler Deployment Experience

10. Competitive Landscape

• 10.1. Value Chain
• 10.2. MEMS OCS Suppliers
• 10.3. Robotic Optical Cross-Connect Suppliers
• 10.4. Integrated-Photonic OCS Developers
• 10.5. AI Network Platform and Orchestration Ecosystem
• 10.6. Manufacturing Capacity and Hyperscale Supply Agreements

11. Company Profiles

• 11.1. Lumentum
• 11.2. Coherent
• 11.3. HUBER+SUHNER / POLATIS
• 11.4. Molex
• 11.5. Calient
• 11.6. Telescent
• 11.7. Google
• 11.8. Drut Technologies
• 11.9. NVIDIA
• 11.10. Broadcom
• 11.11. Cisco
• 11.12. Arista Networks
• 11.13. Ciena
• 11.14. Fujitsu
• 11.15. NEC

12. Recent Developments

13. Appendix