Alibaba Cloud's Wang Peng: The Evolution of OCS Technology for AI Clusters

C114 News, July 23 (Yan Yi) — As AI large-model training scales from tens of thousands to hundreds of thousands of GPUs, data center networks are approaching the physical limits of bandwidth, power consumption and cost. Against this backdrop, optical circuit switching (OCS) has become an industry focus thanks to its low latency, protocol transparency, ultra-low power consumption and reconfigurability.

At the Optical Switching System Technology Session of the 2026 China Optical Communications High-Quality Development Forum held on July 23, Alibaba Cloud optical network architect Wang Peng said OCS is not intended to replace traditional electrical switching, but to evolve alongside it, optimizing interconnect efficiency in AI clusters for specific scenarios. Wang Peng provided an in-depth analysis covering application positioning, scenario value and commercialization conditions, clarifying the OCS technology roadmap, domestic progress and the path to industrial deployment.

Why Are High Hopes Pinned on OCS?

Wang Peng noted that the scaling of AI clusters faces "four walls": the bandwidth wall, the power wall, the cost wall and the reliability wall. On bandwidth, multiple categories of communication overhead account for more than 50% of large-model training; although per-GPU egress bandwidth is evolving toward 1.6T, it still falls short of demand. On power, electrical packet switching (EPS) and DSP optical modules account for 20%-30% of AI cluster power consumption, and in a 100,000-GPU cluster the interconnect alone consumes more than 10 MW. On cost, the two together make up more than 80% of total cost of ownership (TCO) for the network, leaving limited room for savings. On reliability, AI training is highly sensitive to link failures, and even a single point of failure can interrupt model training.

Against this backdrop, the industry expects OCS to deliver low-power, low-latency, reconfigurable interconnect capability for AI networks and to optimize system TCO at certain scales. OCS has four core physical characteristics: pass-through latency below 10 nanoseconds, effectively near zero; protocol transparency, supporting Ethernet, InfiniBand, NVLink and others; no SerDes or DSP components, cutting average power per port pair to below the watt level; and dynamic optical path topology adjustments on timescales from milliseconds to minutes. In essence, OCS is a reconfigurable optical circuit switching system.

Wang Peng stressed that the core requirements for OCS differ across the tiers of an AI cluster. It will not simply replace electrical switching, but will evolve in coordination with it, strengthening the reconfigurable optical switching capabilities of AI clusters and complementing electrical switching in bandwidth, power consumption and reconfigurability.

At the scale-up (intra-node/intra-rack) level, OCS competes with NVLink Switch, with its core value lying in low latency and expanding the interconnect domain, while optical path reconfiguration enables rapid fault-domain convergence. At the scale-out (cross-rack/cross-pod) level, its core value is reconfigurable topology; by adapting to high-availability AI resource strategies, it serves as a replacement or complement to the Spine layer. At the scale-across (cross-cluster/campus-level) level, its core value is elastic compute scheduling: optical paths can be adjusted on a minute-by-minute basis, and the technology offers unique advantages thanks to large port counts and highly reliable optical protection.

Four Technology Routes Compete and Coexist

Although OCS offers clear physical advantages, its engineering implementation involves multiple challenges in optics, precision manufacturing and other areas, with each technology path having its own strengths and weaknesses. Wang Peng said there are currently four main technology routes for OCS engineering.

Among them, MEMS (micromirror array) has the largest port count (up to 1000×1000) and the highest maturity, making it closest to large-scale deployment, but it carries risks of micromirror stiction and high-voltage driver failure. DBS (piezoelectric ceramic) offers the best reliability with no mechanical wear, but costs are relatively high and the core technology is held by a small number of overseas vendors. DLC (digital liquid crystal) has no moving mechanical parts and requires low drive voltage, but suffers from aging and temperature drift. SiPh (silicon photonics) achieves microsecond-level switching speeds; combined with SOA, insertion loss approaches 0, and it has the greatest long-term cost potential, but current port counts are small (32-128) and it remains at the prototype validation stage.

Judging from the OSC system cost structure, the four routes share some common features: core components and precision packaging alignment together account for about 60%-70% of BOM cost across all four paths, making them the main battleground for cost reduction in volume production. The MEMS and DBS routes also bear 15%-18% in high-voltage driver-related costs. In addition, software and testing together account for about 18%-24% of cost, currently constrained mainly by low efficiency and consistency, as well as the lack of unified standards.

On a per-port cost basis, taking 300×300 ports as an example, MEMS costs about US$350-500 per port, SiPh about US$200-500, DLC about US$500-800, and DBS as much as US$1500-2000.

On the domestic supply chain, Wang Peng said China has full control over optical components, LC connectors, system software and other segments; it has partial manufacturing capability in MEMS micromirrors, silicon photonics chips, DLC panels, precision packaging, drive ASICs, FPGAs, and test and calibration equipment, but yield and consistency still need improvement; DBS technology, because of the deep coupling of piezoelectric actuators, precision packaging and closed-loop servo circuits, is not yet within domestic core capabilities.

The feasible path, therefore, is to focus on the MEMS and silicon photonics routes where domestic technology is stronger, avoid reliance on DBS, and use complete-system solutions to drive localization of components. This requires cloud providers to lead with real business scenarios, equipment vendors to carry out integration and validation, and upstream component makers to keep improving performance, yield and consistency.

Toward Commercial Scale-Up

With the technology roadmap and localization path clarified, how to make OCS a reality becomes the key question. Wang Peng noted that OCS must pass through five stages to move from "usable" to "deployable, operable and ready for volume deployment":

First is scenario entry: select suitable PoC scenarios in AI clusters, validate benefits and boundaries, and define applicable use cases. Second is engineering: bring port scale, insertion loss and reliability to production-grade levels, and evolve in step with 800G/1.6T/3.2T optical interconnects. Third is coordinated adaptation: establish interface performance monitoring, certification and O&M systems to support 7×24-hour stable operation of AI clusters. Fourth is system integration: coordinate the OCS controller with compute scheduling and network control, enabling bidirectional awareness between network topology and training task status. Fifth is standardization: unify interfaces, alarms and management models to achieve "deployable, operable and ready for volume deployment."

Industry efforts currently focus on the first two stages, and the next 2-3 years will be a critical period. However, volume commercialization of OCS will still require four-way collaboration among cloud service providers, equipment/component vendors, module makers and standards organizations:

In this division of roles, cloud service providers are responsible for demand pull and scenario validation, setting deployment requirements for operability, monitorability and mass production; equipment and component vendors must advance the engineering maturity of MEMS, SiPh and DLC approaches to achieve high port counts, low insertion loss, low cost and high reliability; optical module makers must adapt to 800G/1.6T optical interconnect technology and link budgets, promoting compatibility across pluggable, NPO and CPO solutions; and standards and open-source organizations must drive the standardization of interfaces, alarms and management models, improve test methods, interoperability specifications and certification systems, and lower the barriers to cross-vendor integration.

Finally, Wang Peng concluded that OCS currently lacks precedents for large-scale deployment in China, and a gap remains between the laboratory and production environments. In terms of commercialization, component breakthroughs are only the starting point; system-level engineering, operations and maintenance systems, and ecosystem coordination together determine the pace of commercialization. End-user demand, system software and system-level engineering in China still need time to mature through practice, and uncertainties remain. End users must serve as the demand-side driver, working with all parts of the industry to move OCS from "usable" to "deployed at scale," building a solid technical foundation for AI cluster interconnects.