Industry giants are making significant investments, making the transition from electrical to optical networks in data centers an imperative. At last week's Appliance & Electronics World Expo (AWE) in China, a new achievement in reconfiguring intelligent computing network architecture with Optical Circuit Switching (OCS) was officially launched. Shanghai Electric, in collaboration with Xizhi Technology, Biren Technology, and ZTE, unveiled the "Optical Leap Super Node 128-Card Commercial Edition," achieving long-term stable training performance and significantly enhancing model training efficiency while reducing transmission latency by over 90% compared to traditional electrical switching.
The super node centers around Xizhi Technology's world-first silicon photonic OCS optical switching chip, initially introduced at last July's World Artificial Intelligence Conference (WAIC). In just over half a year, it has progressed from proof-of-concept to practical commercial deployment, marking China's first end-to-end OCS batch deployment solution. Additionally, the solution has been successfully adapted for several domestic large-scale models, including StepFun and DeepSeek, demonstrating its potential as a cost-effective, high-efficiency computing solution.
Amid growing challenges in power consumption and latency for large model training with traditional electrical switching networks, OCS technology—with its low power usage, minimal delay, and flexible scheduling—is emerging as a key enabler to break through computational efficiency ceilings.
01. Traditional Electrical Circuit Switching Networks Grapple with Power Issues: How OCS Becomes the "Game-Changer" in the AI Era
It must be acknowledged that traditional electrical packet switching remains dominant in data centers today, primarily due to its technological maturity and well-established ecosystem. In principle, when exchanging data, electrical packet switches first convert optical signals from fiber links into electrical signals via optical modules. The data packets are then parsed and forwarded by switching chips before being converted back into optical signals for transmission. This entire process relies on "optical-electrical-optical" (O-E-O) conversion and electronic switching chips for data processing.
This can be likened to a busy railway freight station. When data-laden trains arrive, they must stop, unload cargo, sort and regroup items onto another train before departing. Each stop and the continuous operation of handling equipment contribute to higher latency and energy consumption. For instance, a typical 400G optical module consumes about 10W, with energy usage around 25 picojoules per bit. When combined with the power consumption of SerDes chips inside switches, total energy consumption can easily exceed 30 pJ/bit. While more advanced 800G modules offer better efficiency, their per-bit energy consumption remains in the 15–20 pJ range. In hyperscale data centers, this energy usage accumulates rapidly, increasing operational costs and imposing significant cooling and design pressures.
This is where OCS proves its value. OCS can alter the propagation path of optical signals at the physical layer. When data streams reach a switching node, they bypass the complex switching process of electrical systems; instead, optical signals are directly guided to target links, establishing an end-to-end optical pathway. This is akin to deploying an intelligent railway switch system that allows trains to change tracks without stopping, maintaining high-speed throughput. Although maintaining this "switch" system requires energy, its consumption is several orders of magnitude lower than O-E-O conversion, theoretically reaching femtojoule per bit levels.
Beyond energy efficiency and latency, OCS offers advantages in bandwidth, reliability, and compatibility. It is not constrained by electronic switch port rates or SerDes speeds, enabling easier support for high-bandwidth data transmission. It also operates independently of specific data transmission protocols, seamlessly integrating with interconnect protocols from different vendors to eliminate ecosystem lock-in risks. Moreover, OCS can rapidly re-establish optical paths at the physical layer. If a link or device fails, the network can reconfigure optical pathways within seconds to bypass the issue, enhancing overall stability and fault tolerance.
02. Divergent OCS Technology Paths: Chinese Players Explore Practical Solutions
While OCS advantages are clear, its high technical barriers and interdisciplinary nature have led various manufacturers and research institutions to explore different paths based on their expertise and industrial ecosystems. Some continue with traditional MEMS optical switches from optical communications, others pursue highly integrated optical switching via silicon photonic waveguides, and some employ optical control technologies like liquid crystals or piezoelectric ceramics.
Among these, MEMS-based OCS is relatively mature. This approach uses micro-mirror arrays to alter light beam directions, enabling port-to-port optical path reconfiguration. Alphabet is the only company globally to achieve large-scale mass production and deployment of OCS, with about a decade of R&D experience using MEMS technology. Combined with Alphabet's full-stack software, this solution greatly enhances data exchange efficiency in TPU training clusters.
MEMS-based solutions offer large port scales and stable optical performance but face limitations such as millisecond-level switching speeds due to mechanical structures, larger device sizes, higher system packaging complexity, and elevated costs. Another notable path is silicon photonic OCS, which uses silicon-based waveguides to route and switch optical signals on-chip. Unlike MEMS, silicon photonics has no moving parts, allowing switching speeds in microseconds or even nanoseconds. Startups like Xizhi Technology are advancing this field, with Xizhi developing the world's first silicon photonic OCS optical switching chip. Related research papers have been accepted by SIGCOMM 2025, a top-tier international conference on communication networks.
Silicon photonic OCS can leverage mature CMOS manufacturing processes for mass production, offering better integration and potential cost advantages. However, challenges remain in controlling optical loss and managing thermal crosstalk. Beyond MEMS and silicon photonics, other OCS technologies include liquid crystal-based path control and piezoelectric ceramic-driven optical structure adjustments, each providing different trade-offs in speed, stability, or cost.
Overall, the OCS field is in a phase of parallel exploration across multiple paths, with various technologies balancing port scale, switching speed, energy efficiency, and manufacturability. No single dominant approach has emerged, and future applications may see coexistence of multiple technologies. Given China's current development stage and technological environment, silicon photonics—benefiting from mature semiconductor manufacturing—holds practical advantages in industry chain coordination and scalable production, making it a key focus for companies like Xizhi.
03. Hardware Deployment First, Software Ecosystem Follows: OCS Is No Longer Optional
In discussions with Xizhi Technology CEO Shen Yichen, it was revealed that over the past six months, Xizhi has driven commercial adoption of optical interconnect and switching super nodes through simultaneous advancements in both hardware and software. The R&D process involved significant engineering challenges, such as initial instability in optical signal connections and less-than-expected smooth transmission switching. To address these, Xizhi collaborated with GPU and server manufacturers, forming large technical teams to achieve stable deployment through extensive optimization.
Additionally, Xizhi developed supporting software for optical switching and worked with ecosystem partners to enhance the software environment. The industry has largely reached consensus on the "copper to optical" transition. By the end of 2025, market research firm Cignal AI released a report highlighting that as hyperscale data centers and AI computing centers deploy, OCS is shifting from validation to large-scale pilot and commercial use, expanding beyond Alphabet to more vendors and applications. The total OCS market is projected to reach at least $2.5 billion by 2029.
Leading players like NVIDIA are voting with action. This month, NVIDIA invested $2 billion in two OCS-focused companies, Lumentum (MEMS path) and Coherent (liquid crystal path), signing multi-billion-dollar purchase commitments and securing future production capacity priority, signaling long-term confidence in OCS. Regarding NVIDIA's moves, Shen Yichen noted that Xizhi predicted 2–3 years ago that 30% of future data center chips would be optical, and NVIDIA would fully embrace optical chips. NVIDIA's investments are driven by anticipated rapid volume increases in optical chips, necessitating early supply chain preparation.
On technology path choices, Shen maintained an open perspective, stating that while different OCS approaches vary in specific metrics, they share 80–90% core functionality. The current priority is turning optical switching clusters from concept to reality, deploying large-scale optical interconnect and switching clusters. Hardware deployment must precede software optimization; only with physical optical switching clusters can software and systems be tailored for optimal performance. Transitioning from "electrical switching" to "optical switching" represents the most critical software change, with subsequent optimizations offering marginal gains. Shen analogized, "Rather than debating which motor technology to use, it's better to get an electric vehicle on the road first. Once the vehicle is operational, supporting infrastructure will naturally develop."
04. Conclusion: Data Center Networks Shift from Electrical to Optical as Chinese Solutions Gain Technological Edge
OCS technology is reshaping AI computing infrastructure unprecedentedly, delivering qualitative leaps in energy efficiency, latency, and bandwidth while providing robust support for training, inference, and broad adoption of next-generation large-scale models. In this accelerated adoption phase, Chinese companies like Xizhi Technology are pioneering a self-sufficient, high-performance, and high-efficiency domestic computing path through innovations such as silicon photonic OCS chips and the Optical Leap Super Node. Shen Yichen revealed that beyond supporting thousand- to ten-thousand-card training clusters, Xizhi plans to expand the 128-card commercial super node for inference applications. At this year's WAIC, Xizhi will unveil its next-generation optical switching cluster方案.
Comments