How Close Is Optical Circuit Switching (OCS) to Large-Scale Deployment?

As AI foundation models continue to expand at an unprecedented pace—growing by nearly 10× in parameter count each year—new computing paradigms such as test-time scaling are driving GPU clusters from hundreds of thousands to millions of accelerators. Traditional electrical switching architectures are rapidly approaching their physical limits in terms of power consumption, transmission distance, and port density.

Against this backdrop, Optical Circuit Switching (OCS) has evolved from an optional network optimization technology into one of the key enabling technologies for next-generation AI networking. Across the industry, there is growing consensus that OCS will become a fundamental building block for future AI infrastructure.

However, discussions across recent industry forums suggest that OCS is still in a transitional stage. While leading hyperscalers have begun large-scale validation and core technologies continue to mature, significant challenges remain—including physical scalability limits, system integration, and operational readiness. As a result, the industry's widely anticipated large-scale commercialization around 2028 still carries uncertainty. Adoption may progress more gradually than optimistic forecasts suggest, although the long-term strategic value of OCS is no longer in doubt.

Demand-Side Underlying Logic: AI Cluster Architecture Reconstruction — OCS Moves from the Background to a System-Level Core Driver
The core driving force behind the rise of the OCS industry is not the iteration of optical communication technology itself, but the underlying transformation of AI computing architecture. When cluster scale exceeds the 100,000-GPU level, the physical bottlenecks of electrical interconnects have become an explicit constraint on computing power growth. OCS is currently the only technology path that can systematically break through this bottleneck.

Electrical switching hits the physical ceiling; architectural upgrades are urgent
NVIDIA networking architecture experts have clearly pointed out that the exponential growth of AI model scale and the popularization of the test-time scaling paradigm are generating endless demand for GPU computing power and high-bandwidth networks. Cluster scale has expanded to the hundreds-of-thousands-of-GPUs level, and OCS has risen from a background peripheral component to a primary system-level architectural consideration. NTT has revealed the rigid boundaries of electrical interconnects from the physical layer: at single-link rates of 200 Gbps and above, the vertical scaling distance of traditional electrical interconnects (intra-machine and intra-rack GPU direct connections) is limited to within 1.5 meters, directly locking single-cluster scale at the hundreds-of-GPUs level and unable to match the expansion needs of AI computing power. The breakthrough path it proposes is a “CPO + OCS” combination: co-packaged optics (CPO) reduces bit energy consumption to below 5 pJ/bit, breaking through distance limitations; then decoupled OCS units construct flexible modular torus topologies, elevating vertical scaling scale to hundreds of GPUs and providing the foundation for flattened networking of large-scale clusters.

Scenario boundaries continue to extend: from horizontal scaling to vertical scaling penetration
Current mature OCS deployment scenarios are concentrated in the horizontal scaling domain, mainly used for fault tolerance, traffic bypass, and spine-layer switching. Through all-optical pass-through, they bypass failed nodes and improve network reliability and transmission efficiency. This value has been validated in deployments by leading customers such as Google’s TPU clusters. The industry’s next-stage goal is to push OCS into the vertical scaling domain (direct GPU-to-GPU interconnects), enabling flexible topology dynamic reconfiguration based on business needs. Once this scenario is broken through, OCS will upgrade from a “network auxiliary device” to a “core computing interconnect foundation,” with value and market space multiplying several times over—this is also the core direction of industry-wide technical breakthroughs.

Current Status of Technology Roadmap Industrialization: Dual-Track Parallel Iteration, Mature Route Scales First
OCS has not formed a single technology route dominating the landscape. Instead, it presents a dual-track parallel situation of “3D MEMS currently dominating commercial use, while 2D/silicon photonics tackles the next generation.” The two form differentiated complementarity targeting different port scales and application scenarios, rather than direct substitution.

3D MEMS: Absolute mainstay of current commercial use, advancing toward thousand-port scale
Optical switch solutions based on 3D MEMS micromirrors, with core advantages of complete transparency to transmission rates, mature processes, and controllable insertion loss, are the preferred choice for meeting high-radix (1,000+ ports) needs and the most concentrated technology route for vendor deployments at OFC 2026.

Industrialization progress
Overseas vendors such as Lumentum, DiCon, and Teraxion have achieved mass production of 300×300 to 360×360 port scales, with insertion loss controlled in the 1.5–3 dB range. DiCon has explicitly planned delivery of a 1024×1024 all-optical matrix switch in 2027, continuously pushing the upper limit of port scale. Domestic vendors including Accelink, Innolight, and Accelink Technology (and others) have also launched mass-production products at the 320×320 level, with performance metrics gradually catching up to overseas leaders.

Core challenges
Lumentum has pointed out that the industrialization bottleneck of 3D MEMS has shifted from laboratory performance to mass-production reliability and efficiency: first, long-term reliability verification of millions of switching cycles is required to meet data center service life requirements of over 10 years; second, high-throughput automated calibration at the scale of thousands of ports—traditional manual calibration is extremely inefficient and cannot support batch delivery. In addition, expanding port scale also brings physical constraints of increased physical volume and stepwise rising insertion loss.

2D MEMS and Silicon Photonics OCS: Next-generation technology directions focusing on high density and fast switching
New OCS solutions centered on 2D MEMS arrays and silicon photonics integration are the next-generation technologies being prioritized by the industry, emphasizing high integration, low power consumption, and fast switching speeds, targeting medium-to-small port high-density interconnect scenarios at the cabinet and rack levels.

Industrialization progress
nEye Systems’ 2D MEMS solution offers advantages in control system simplicity and reliability, has verified the potential to scale to 1000×1000 ports, with switching speeds reaching the sub-millisecond level, and is positioned as an automated “optical patch panel” for AI factories. iPronics’ silicon photonics OCS achieves ultra-high density of 256 ports per 1U, with single-port power consumption as low as 0.78 W, and programmable capabilities supporting integrated switching and signal processing.

Core challenges
The inherent shortcoming of 2D architectures is significant insertion loss accumulation: when the number of connections expands beyond 64, cascaded insertion loss rises sharply, greatly compressing the link power budget. Silicon photonics solutions still need breakthroughs in yield control and cost reduction at large port scales, remaining 2–3 years of technical iteration away from large-scale commercial use.

Ciena has summarized the industry’s system-level “wish list” for OCS: the ultimate goal is to achieve tens of thousands of non-blocking ports, microsecond-level switching speeds, and higher port density. Meanwhile, with the popularization of 800G and higher-rate links, the link budget has been compressed to 3–4 dB. Either low-loss dedicated transceivers must be developed, or coherent optics technology must be introduced to supplement the power budget—this is a common technical breakthrough direction across the industry.

Core Bottlenecks to Large-Scale Deployment: Four Barriers from Devices to Operations
Although the technical value of OCS has formed an industry consensus, large-scale popularization across the industry still requires crossing four core barriers in architecture, scenarios, systems, and operations—this is the core gap between market expectations and industrial reality.

Architectural physical limits: Inherent contradiction between port scale and performance
Whether following the 3D or 2D route, expansion of port scale faces rigid physical constraints:
· For 3D MEMS OCS: the larger the port radix, the higher the required micromirror angle control precision; insertion loss, equipment volume, and calibration difficulty rise synchronously. Costs and O&M complexity for devices above the thousand-port level grow nonlinearly.
· For 2D/silicon photonics OCS: limited by waveguide transmission loss and cascaded coupling loss, insertion loss accumulates rapidly once port count exceeds 64, directly consuming the link power budget and unable to support long-distance, high-rate transmission needs.

Vertical scaling dilemma: Core scenario value not yet released
Current mature applications of OCS remain limited to horizontal scaling and cross-connects, used for fault routing bypass and spine-layer traffic off""ing, and have not truly entered the core vertical scaling scenario of GPU-to-GPU interconnects. The essential reason is that direct inter-GPU interconnects impose extremely high requirements on latency jitter and link stability; current OCS switching speeds and optical path stability still cannot fully match the stringent standards of intra-machine interconnects. If the vertical scaling scenario cannot be broken through, the market space for OCS will remain limited to the network edge layer, with both value and growth ceiling significantly constrained.

Switching disconnection: Speed mismatch between optical and electrical layers
This is the most core system-level bottleneck facing dynamic OCS networking: the mirror switching speed of the optical switch itself can reach the microsecond level, but the optical module DSPs at both ends of the link require millisecond-level time to complete clock data recovery and re-locking of adaptive equalization filters. The order-of-magnitude speed difference between the optical and electrical layers causes millisecond-level link disconnections during OCS switching, making it unable to carry uninterrupted real-time computing services. It can only support relatively static topology configurations and cannot achieve true real-time dynamic reconfiguration. This problem cannot be solved by optical switch vendors alone; it requires coordinated optimization with optical module and DSP vendors and is a full-industry-chain systems engineering challenge.

Missing O&M systems: Hidden costs constitute a “maintenance crisis”
Large-scale deployment of high-radix OCS will bring concentrated layout of massive numbers of fiber connectors, and the cleanliness of fiber end-faces directly determines insertion loss levels and link reliability. Currently, the vast majority of data centers have not listed fiber cleaning and inspection as standardized O&M processes. Manual operations are low in efficiency and consistency; dust contamination leading to elevated insertion loss and frequent link failures constitutes hidden O&M costs and reliability risks. In addition, OCS management and control software has not yet formed unified standards and cannot seamlessly integrate into existing data center network orchestration systems. High customized development costs are also an important factor constraining adoption by small and medium customers.

Industry Breakthrough Paths: Scenario Adaptation and Full-Chain Collaboration to Explore Feasible Deployment Models
Facing the above bottlenecks, the industry has not remained solely at the technical breakthrough level. Instead, combining the unique characteristics of AI scenarios, it has explored multiple breakthrough paths with deployment feasibility, gradually advancing OCS from the laboratory to commercial use.

Hybrid networking + software-defined physical layer: Adapting to AI’s predictable traffic
The “software-defined physical layer” approach proposed by Cornell University researcher Rachee Singh has become the industry-recognized optimal deployment solution. Unlike the randomness of general cloud traffic, the traffic patterns of AI training and inference work""s are highly predictable—elephant flows such as model weight synchronization and All-Reduce account for an extremely high proportion, and traffic cycles are stable. Based on this characteristic, a hybrid networking model of “OCS + electrical switching” is adopted: high-bandwidth, long-duration elephant flows go through OCS all-optical pass-through, fully leveraging its low-latency and low-power advantages; small packets and random traffic are handled by traditional electrical switches, retaining flexibility. This solution does not require OCS to achieve millisecond-level real-time switching, thereby avoiding the shortcoming of switching disconnection while maximizing the value of optical switching, achieving the optimal balance between performance and cost. It is also the deployment model already validated by leading customers such as Google.

Cross-industry-chain collaboration: Breaking through system-level bottlenecks
Coordinated optimization across the full industry chain has become a consensus for addressing system-level problems such as switching disconnection and insufficient link budgets:
· Optical module and DSP vendors are developing dedicated chips supporting fast locking and burst mode, compressing re-locking time from the millisecond level to the microsecond level to match optical switch switching speeds.
· Coherent optics technology is beginning to sink into short-reach data center scenarios, offsetting OCS insertion loss through optical amplification gain, expanding the link power budget, and supporting networking at larger port scales.
· CPO and OCS fusion solutions continue to advance, further compressing latency and reducing power consumption through end-to-end all-optical architectures to adapt to the needs of vertical scaling scenarios.

Automated O&M and standardization: Lowering the threshold for popularization
Addressing O&M shortcomings, the industry is advancing from both tools and standards:
· Automated fiber cleaning, inspection, and calibration equipment is gradually maturing, capable of replacing manual operations to achieve batch, high-precision fiber end-face maintenance, significantly reducing O&M costs and human error.
· Source ecosystems and standardization processes are accelerating. For example, vendors such as Innolight are developing OCS operating systems based on the SONiC standard, promoting unification of management and control interfaces, lowering customer adaptation and customization costs, and laying the foundation for OCS to sink toward small and medium customers.

Industry Stage Assessment and Future Outlook
Current industrialization stage positioning
Overall, the OCS industry has crossed the “technical feasibility verification” stage and formally entered a transitional period of large-scale deployment by leading customers and full-industry ecosystem shortfall remediation. Leading vendors with vertical integration capabilities, such as Google, have achieved large-scale OCS deployment by leveraging self-developed complete technology stacks and rigorous O&M systems. However, for the vast majority of data centers and computing power operators, high costs, missing software ecosystems, and mismatched O&M systems remain difficult-to-overcome thresholds, and overall industry penetration remains low. The widely anticipated large-scale commercial use in 2028 still requires three core prerequisites to be met: yield breakthroughs and cost reductions for thousand-port-level 3D MEMS OCS, commercial availability of dedicated transceivers supporting fast locking, and basic maturity of automated O&M systems and standardized management/control interfaces.

Phased evolution path
· Short term (2026–2027): Batch volume by leading customers
The 3D MEMS route dominates the market, with applications concentrated in spine-layer switching and horizontal scaling scenarios of ultra-large-scale AI training clusters. Overseas leading cloud and AI vendors initiate batch procurement; domestic vendors follow with validation and small-batch deployment. Market scale grows rapidly but overall base remains limited; industry focus concentrates on hardware performance improvement and yield optimization.

· Medium term (2028–2030): Industry-wide large-scale penetration
System-level bottlenecks are gradually broken through; OCS begins penetrating GPU vertical scaling scenarios. 2D/silicon photonics solutions achieve popularization in cabinet-level and medium-to-small-scale clusters. Automated O&M tools and standardized ecosystems become basically mature; industry customers and intelligent computing centers begin batch adoption. The market enters an explosive growth stage, maintaining compound annual growth rates above 50%.

· Long term (after 2030): Architectural core status
OCS deeply integrates with CPO, silicon photonics, and chip-level optical interconnects, becoming the standard networking configuration for million-GPU-scale AI clusters. Optical interconnects thoroughly upgrade from supporting components of computing power to core components of computing architecture, with industrial value and market space achieving order-of-magnitude leaps.

Conclusion
The core industrial position of OCS in the AI era is no longer in doubt. Its large-scale takeoff is not a question of “whether,” but of “when.” Continuous iteration of device performance is only the foundation; full-industry-chain system collaboration, supporting improvement of O&M systems, and unified standards for software ecosystems are the keys determining the pace of industrial deployment. The next 3–5 years will be the critical window period for the OCS industry to move from “single-point breakthroughs” to a “closed-loop ecosystem.” Vendors with full-stack collaboration capabilities and deep binding to leading customers will occupy core positions in this round of industrial upgrading.