In-Depth Analysis of OCS: Optical-Layer Direct-Connect Switching Technology

In new data center scenarios characterized by generative AI and large-scale parallel computing, network traffic is shifting from a traditional “north-south” (also known as frontend network) orientation to an “east-west” (backend network) orientation, which is characterized by high-frequency inter-node communication. This shift places higher demands on network bandwidth, latency, and topological flexibility. Traditional packet-switched networks have gradually revealed bottlenecks in congestion, power consumption, and scalability when handling large-scale collective computations (such as All-Reduce).

Against this backdrop, OCS (Optical Direct Connect) has begun to attract attention as a switching technology based on direct optical connections. By establishing on-demand end-to-end optical paths at the physical layer, OCS bypasses intermediate packet processing, achieving ultra-low latency, non-blocking bandwidth, and superior energy efficiency, thereby providing a new architectural alternative for AI training clusters and high-performance computing networks.

Against this backdrop, Optical Circuit Switch (OCS), a switching technology based on direct optical layer connections, has begun to attract attention. By establishing on-demand end-to-end optical paths at the physical layer, OCS can bypass intermediate packet processing, achieving ultra-low latency, non-blocking bandwidth, and better energy efficiency, providing a new architectural complement for AI training clusters and high-performance computing networks.

What is OCS?

An OCS (Optical Circuit Switch) is an all-optical switching device that operates at the physical optical layer. Its core function is to establish direct optical paths between different fiber optic ports on demand, enabling direct routing and interconnection of optical signals. Unlike traditional switching architectures, OCS maintains the optical signal form throughout the switching process, without undergoing “optical-electrical-optical (O-E-O) conversion.” Instead, it changes the light propagation path through optical means (such as microelectromechanical systems micromirror arrays) to switch the optical signal at the input port to the target output port.

Essentially, OCS is more akin to a “connection-reconfiguration device at the optical fiber layer,” functioning similarly to dynamically adjusting the connection relationships of optical links within a network, rather than parsing and forwarding data packets. This endows it with characteristics such as ultra-low latency, protocol independence, and high-bandwidth transparent transmission.

It should be noted that OCS is not a simple replacement for traditional electrical switches but rather serves as a complementary technology. In current data center networks, electrical switches still handle core functions such as traffic scheduling and protocol processing, while OCS is used to provide efficient direct connection paths in high-bandwidth, fixed, or semi-fixed communication scenarios. Particularly in large-scale collective communication scenarios such as AI training, OCS can significantly improve network transmission efficiency and resource utilization.

NADDOD’s 800G N9520–64OC intelligent computing switch, specifically designed for high-end AIGC intelligent computing scenarios, is a core electrical layer supporting device for the OCS optical layer network. As an 800G electrical switch product, the N9520–64OC undertakes core responsibilities that the OCS optical layer cannot cover, such as protocol parsing, fine-grained traffic scheduling, service logic forwarding, and low-latency tightly coupled communication. It forms a collaborative architecture with OCS of “high-bandwidth non-blocking direct connection at the optical layer + fine-grained intelligent management and control at the electrical layer,” becoming the core hub of large-scale AI cluster networks.

The Advantages of OCS

As a switching technology based on the physical optical layer, OCS offers significant advantages in terms of energy efficiency, bandwidth utilization, and network architecture flexibility, making it particularly well-suited for AI and high-performance computing scenarios that demand high bandwidth and low latency.

• Energy Efficiency and Latency: By eliminating “optical-electrical-optical (O-E-O) conversion” and packet processing, OCS significantly reduces system power consumption and processing overhead. Unlike traditional electrical switches, which require signal conversion and forwarding decisions, OCS performs path switching solely at the optical layer. This results in lower single-hop latency and reduced jitter, delivering stable and predictable transmission performance — a critical factor for inter-process communication in large-scale AI clusters.
• Bandwidth Capacity and Rate Adaptation: OCS is not constrained by the fixed rate limits of switching chips but directly leverages the physical transmission capacity of optical fibers to achieve true rate transparency. Whether it is 400G, 800G, or future 1.6T and higher speeds, OCS can support them without replacing equipment; only the optical modules at both ends need to be upgraded. This feature not only improves link utilization efficiency but also significantly extends the equipment lifecycle, reducing hardware replacement and operational costs for data centers during generational upgrades.
• At the network architecture level: OCS offers excellent reconfigurability. By integrating with Software-Defined Networking (SDN), it can dynamically adjust optical path connections on demand and flexibly construct different logical topologies, thereby improving cluster resource utilization and enhancing the network’s resilience in the event of failures. Additionally, OCS supports on-demand scaling and incremental deployment, allowing for gradual expansion based on business scale and avoiding large-scale one-time investments.

Overall, by delivering efficient and transparent connectivity at the optical layer, OCS provides data centers with a low-power, low-latency, and future-proof supplementary network infrastructure solution.

Limitations of OCS

While OCS offers significant advantages in optical layer switching, its practical deployment still faces certain limitations and requires rational evaluation based on specific scenarios.

• Handover Latency: OCS typically relies on mechanical structures (such as MEMS micromirrors) for optical path adjustment, resulting in millisecond-level reconfiguration latency, significantly higher than the nanosecond-level forwarding latency of electrical switches. Therefore, OCS is more suitable for scenarios with relatively stable connections and less suitable for services requiring real-time handover or extremely low latency response.
• Cost: OCS is currently still in a highly customized stage, resulting in relatively high overall prices. For example, a 300-port OCS device typically costs between $100,000 and $120,000 in bulk purchases, representing a significant initial investment.
• Technology and Ecosystem Perspective: The successful deployment of OCS depends on supporting optical devices and system capabilities, including customized optical modules (such as integrated circulators and wavelength division multiplexers) and dedicated scheduling and control software. Currently, the related industry chain is not yet fully scaled up, and the overall ecosystem maturity still needs improvement.
• Application Scope: OCS is more suitable for network layer scenarios with relatively stable traffic patterns and low path adjustment frequency. It is usually used as a supplement to existing electrical switching networks rather than a general alternative.

OCS Core Technology Solution

Currently, OCS is mainly developing along four technical routes, including implementation schemes based on MEMS, liquid crystals, piezoelectric ceramics, and silicon optical waveguides.

MEMS (Micro-Electro-Mechanical Systems) Technology

MEMS is one of the most mature and widely adopted technological approaches in current OCS systems. Its core principle involves constructing an array of micro-mirrors on a silicon wafer; by adjusting the angles of these micro-mirrors, the propagation direction of the incident light is altered, thereby precisely routing the optical signal to a designated output port to achieve optical path switching.

In practical implementation, MEMS solutions typically employ two sets of dynamically adjustable micromirror arrays to control the optical path: the input optical signal is collimated into a parallel beam, which is first reflected by the first set of micromirrors to a target position, then undergoes secondary reflection and correction by the second set of micromirrors, and finally coupled into the corresponding output fiber. This method of optical path routing in three-dimensional space ensures high control precision and stability.

Based on their structural design, MEMS optical switches are primarily classified into two categories: 2D MEMS and 3D MEMS.

• 2D MEMS are based on a planar structure and use electrostatic actuation of a mirror to switch the optical path in a fixed direction; they feature a relatively simple structure and are easier to manufacture.
• 3D MEMS, on the other hand, allow the micromirror to rotate freely along two axes, enabling more flexible optical path control while offering superior performance in terms of size, power consumption, insertion loss, and crosstalk.

Overall, MEMS technology has advantages in reliability, stability, and engineering maturity, and can better meet the needs of data centers for large-scale, high-speed optical connections, thus becoming the mainstream implementation solution for OCS.

Digital Liquid Crystal (DLC) Technology

DLC is a non-mechanical light-switching solution. Its core principle involves using an applied electric field to adjust the orientation of liquid crystal molecules, thereby altering the polarization state and propagation path of light to achieve the reconfiguration of optical channels. Compared to solutions that rely on mechanical motion, DLC switches optical paths through purely electrical control, eliminating the need for moving parts.

In practical implementation, the input optical signal first passes through a birefringent device to adjust its polarization state. When no voltage is applied, the optical signal propagates along a predetermined path; when voltage is applied, the orientation of the liquid crystal molecules changes, thereby altering the polarization direction of the light, which is then guided to the target output port by a polarization splitter. By cascading liquid crystal delayers with birefringent structures, flexible routing between multiple ports can be achieved, with current scales supporting up to hundreds of ports.

This solution demonstrates good performance in terms of reliability, service life, and drive power consumption, while also offering certain cost advantages. However, its optical switching time is typically on the order of hundreds of milliseconds, which is significantly slower than solutions such as MEMS. Therefore, it is more suitable for scenarios with low switching frequency requirements, such as network redundancy and backup applications.

Piezoelectric Ceramic Technology

Optical switching solutions based on piezoelectric ceramics (also known as Direct Beam Steering, DBS) utilize the material’s electromechanical coupling effect to generate micrometer-level precise deformation when a voltage is applied, thereby driving optical components to switch the light path.

The basic structure consists of an array of fiber collimators, a two-dimensional piezoelectric actuator, and a position control unit. The fiber collimators are mounted on the piezoelectric actuator, arranged in a two-dimensional array, and positioned opposite another array. When the piezoelectric ceramics undergo displacement or angular tilt under the influence of voltage, the beam’s emission direction can be precisely adjusted to align the input with the target output port, thereby establishing the optical path connection.

This solution features fast response speed, low insertion loss, and high control precision, making it suitable for applications with certain requirements for switching performance and optical efficiency.

Silicon Photonic Waveguide Technology

Silicon photonic waveguide solutions are a type of optical switching technology based on chip-level integration. By constructing a photonic waveguide matrix on a silicon-based chip, they enable the transmission and switching of optical signals along predefined paths. Its core principle involves utilizing the thermo-optic effect or plasmonic dispersion effects (such as carrier injection/depletion) to alter the refractive index of the waveguide material, thereby dynamically controlling the optical path to achieve switching functionality. Typical implementations include structures based on Mach-Zehnder interferometers (MZIs).

Unlike mechanical solutions such as MEMS, silicon photowavguides require no moving parts and can directly modulate optical paths via electrical signals, offering faster response times. Their switching times typically reach the microsecond range or even shorter, making them more suitable for scenarios with frequent traffic fluctuations, such as AI scale-up and memory pooling applications.

In terms of engineering implementation, this solution offers high integration and excellent compatibility with CMOS processes. It can be mass-produced using existing semiconductor manufacturing systems and supports monolithic integration of switching matrices and driver circuits, demonstrating strong scalability potential.

However, silicon photonic waveguide solutions currently face certain challenges, including relatively high insertion loss, as well as crosstalk and reliability issues in multi-channel, large-scale arrays. Furthermore, the technology as a whole is still in the development stage and has not yet been deployed on a large scale commercially.

OCS Application Scenarios

OCS is not suitable for all network environments; its application has clearly defined hierarchical positioning and usage requirements, and is primarily focused on scenarios with high bandwidth demands and relatively stable traffic patterns.

• In AI training clusters: OCS is currently a relatively mature application area that has been widely implemented at scale. By connecting large-scale clusters of AI acceleration chips (such as TPUs) and building topological structures like 3D Torus, OCS can support efficient collaborative training among thousands to tens of thousands of chips, thereby improving overall computing resource utilization.
• Data Center Backbone Network Layer (Spine Layer): OCS can be used to replace some traditional electrical switches, primarily handling large-scale “elephant traffic” with relatively fixed paths. Such scenarios typically do not require frequent link switching, making them well-suited for OCS to provide high-bandwidth, low-power direct connectivity.
• Data Center Interconnect (DCI) Scenarios: OCS can serve as an optical cross-connect device, connecting different data centers or campus networks to provide flexible, high-bandwidth interconnection channels that meet cross-region data transmission needs.

Additionally, OCS can be used for network redundancy and backup. When the primary optical switch fails, OCS can rapidly switch optical paths to restore service, thereby enhancing the overall reliability of the network.

It is important to note that OCS is not suitable for network scenarios with frequently fluctuating traffic and extremely high real-time requirements, such as southbound networks involving tightly coupled communication between GPUs and TPUs. Such scenarios demand extremely low switching latency, and the millisecond-level switching speed of OCS, which relies on mechanical structures, struggles to meet these requirements.

Addressing the limitations of OCS in millisecond-level switching latency, low-latency tightly coupled communication, and dynamic traffic scheduling, the N9520–64OC fills these gaps. The N9520–64OC deeply optimizes RDMA lossless networks, supports PFC/ECN flow control and MMU memory tuning, and constructs an end-to-end low-latency RoCE network. Its nanosecond-level forwarding latency fully meets the requirements of tightly coupled GPU/TPU communication. Simultaneously, it features Nadod RALB remote dynamic load balancing and ENLB end-network collaborative load balancing, resolving the equal-cost multipath conflict problem of traditional ECMP. This enables efficient scheduling of bursty, dynamic, and fine-grained traffic that OCS cannot handle.

OCS Market Trends

From a market development perspective, OCS is currently still in the early adoption phase, with demand concentrated in the short term, but its long-term growth potential is gradually becoming apparent.

In the short term, the OCS market size is fairly certain, with the primary driver being Google’s deployment needs for its TPU clusters. Global demand in 2026 is projected to range between 18,000 and 30,000 units: a conservative estimate is approximately 18,000 units (of which MEMS solutions account for about 15,000 units and LCD solutions for about 3,000 units); under an optimistic scenario, total demand could reach 20,000 to 30,000 units, with Google accounting for approximately 18,000 to 20,000 units, Oracle for about 2,000 units, and the remainder consisting of trial purchases by other cloud providers. Based on mainstream 300-port devices with a unit price of approximately $100,000 to $120,000, the corresponding market size would be approximately $1.8 billion to $3.6 billion.

In the medium to long term (through around 2030), market expectations are highly divergent. On the one hand, a cautious view holds that Google will continue to dominate the OCS market, while investments by other cloud providers (such as Microsoft, Meta, and Oracle) will remain primarily in the testing and validation phases, making large-scale commercial deployment unlikely in the short term. The global cumulative market size is projected to reach approximately 50,000 to 80,000 units by 2027/2028. On the other hand, the industry has also signaled a degree of optimism. For instance, vendors have raised their long-term projections for the OCS market, with the total addressable market (TAM) now estimated at over $2 billion. Meanwhile, application scenarios are gradually expanding from internal interconnections within AI clusters to include spine-layer replacements and data center interconnect (DCI) solutions.

From an industry and ecosystem perspective, current key OCS vendors include Google, Huber+Suhner, Coherent, Calient, Polatis, and iPronics. In 2025, the Open Compute Project (OCP) established an OCS working group to promote the open-source development of all-optical switching technology, which is expected to accelerate the maturity of the industry chain and market penetration. Meanwhile, in addition to Google, leading companies such as Microsoft, Meta, and NVIDIA are also advancing their related technology strategies. For example, Google has been deploying OCS at scale in its AI clusters since TPU v4 and plans to further evolve the technology by integrating it with memory pooling architectures in the future; NVIDIA may incorporate OCS into its next-generation network architecture to enable topology reconfiguration; and companies like Meta are also continuing to increase their investments in data center interconnect (DCI) infrastructure.

Overall, although OCS still faces challenges in terms of cost, insertion loss, and reconfiguration latency, its long-term development potential in the field of optical interconnect deserves continued attention as application scenarios expand and the ecosystem gradually matures.

Conclusion

Overall, as a layer-optical switching technology, OCS has demonstrated clear value in the evolution of AI-driven data center networks: by providing high-bandwidth, low-power, and reconfigurable optical connectivity, it effectively addresses the shortcomings of traditional electrical switching networks in large-scale collective communication scenarios. Although it still faces certain limitations in terms of switching latency, cost, and ecosystem maturity, with ongoing technological advancements and the gradual expansion of application scenarios, OCS is expected to continue penetrating areas such as AI clusters, data center backbones, and DCI, becoming a vital component of future data center network architectures.

The NADDOD N9520–64OC is a flagship product that exemplifies the synergistic evolution of OCS and electrical switching, providing a mature “optical-electrical synergy” solution for AI-driven next-generation data center networks. It not only supports SDN (Software-Defined Networking) but also seamlessly integrates with OCS’s SDN control platform, enabling unified scheduling and coordinated reconfiguration of optical and electrical paths; but it also features comprehensive optimizations for AIGC scenarios. From 800G high-bandwidth transmission to support for clusters with tens of thousands of cards, and from lossless RDMA networks to AI-powered intelligent operations and maintenance, it is deeply integrated with OCS’s optical layer technology. This resolves the congestion and power consumption bottlenecks of traditional electrical switching networks, establishing itself as the core network backbone for large-scale AIGC intelligent computing clusters and high-performance computing data centers.