China’s 416-Mirror MEMS Chip Could Lower the Barrier to Optical Circuit Switching

Suzhou Chipsens Technology’s new “public-edition” chip is a standardized MEMS micromirror array designed for optical circuit switches (OCS). The strategic importance is component-level: instead of requiring every Chinese system maker to develop a large mirror array from scratch, a common qualified device could make domestic OCS designs easier to build, customize and scale.

The launch is therefore best understood as a supply-chain milestone for AI networking—not as the arrival of a finished 400×400 optical switch.

What Chipsens launched

The chip uses electrostatic, dual-axis MEMS actuation and integrates 416 independently movable silicon micromirrors on a single die. The number describes the mirror elements; it should not be read as a 416×416 port matrix. Reported package dimensions are 33.6 × 20.85 × 0.66 millimeters. 23

Reported specifications include:

• X-axis mechanical deflection of ±6.5°
• Y-axis mechanical deflection of ±5.5°
• Infrared reflectivity of at least 96%
• Response time below 10 milliseconds
• More than one billion switching cycles
• Volume die yield above 90%
• A stated ability to support 400×400-port-class OCS products 23

These figures come from company and industry reports. They indicate the product’s claimed capability, but they are not a substitute for independently published benchmark results.

How optical circuit switching works

An electronic packet switch receives data, processes routing information, and forwards packets through an electrical switching fabric. An optical circuit switch instead establishes a physical light path between two endpoints. Once the path is configured, data can travel through the switch without being repeatedly converted from optical to electrical signals and back again.

In a MEMS OCS, control software selects a connection and electrostatically driven mirrors tilt to redirect a beam from an input fiber toward the correct output fiber. The resulting circuit is dedicated until the network reconfigures it.

That makes OCS particularly useful for traffic with three characteristics:

1. High bandwidth: large AI workloads move substantial volumes of data between accelerator groups.
2. Predictable communication patterns: training and other coordinated workloads can create long-lived flows that justify a dedicated circuit.
3. Tolerance for millisecond-scale changes: OCS is generally used to reconfigure topology at the workload or scheduling level, not to make a new decision for every packet.

The technology is attractive because it is data-rate and wavelength agnostic, offers low latency, and avoids much of the switching power associated with electronic processing.

OCS does not eliminate electronic networking. Short-lived, bursty or irregular traffic still benefits from packet switching, while an optical layer requires topology management, scheduling, monitoring and fault recovery.

Why AI clusters are adopting the idea

As accelerator systems grow, fixed point-to-point cabling becomes increasingly difficult to manage. A reconfigurable optical layer can change which groups of processors have direct connections, potentially reducing electrical switching layers and the number of conversion points in the network.

Google’s Ironwood architecture provides a prominent example. Google describes OCS as the network used to connect TPU cubes beyond a single cube, while its documentation explains that optical links and OCS faults can be bypassed to improve the availability of TPU slices. 1721

Google’s published Ironwood material describes systems scaling to 9,216 chips in a superpod. That does not mean every AI cluster will use the same design, but it shows why optical topology control becomes more valuable as the number of interconnected accelerators rises. 21

Power claims require similar care. Industry material has cited approximately 40% network-power savings for particular OCS architectures, but that is an architecture-level result, not a guaranteed saving from installing Chipsens’ micromirror chip alone.

Why a standardized chip matters in China

OCS systems combine several difficult elements: mirror arrays, optical engines, fiber alignment, packaging, drive electronics, control software and system-level reliability. A standard micromirror component can allow equipment companies to concentrate on those integration layers rather than independently funding the most specialized part of the optical engine.

That could lower the barrier to customized OCS products for Chinese data centers, backbone networks and large AI-compute clusters. It could also reduce reliance on imported core components, although the scale of that effect will depend on customer qualification, sustained yield, system reliability and actual deployment volumes.

Chipsens says its team began working on MEMS-array switching in 2019. Reports then describe a large-array tape-out in 2025, use in a 320×320-channel OCS system, and the standardized product launch in August 2026. 124

The company also reports vertical capabilities spanning micromirror design, wafer processing, packaging and testing. A separate financing report says Chipsens completed a Pre-A round to accelerate MEMS micromirror-array development and AI-data-center optical interconnect applications. 67

Where MEMS fits among OCS technologies

MEMS mirrors are one of several approaches to optical circuit switching. Other routes include:

• Liquid-crystal switching: can offer strong stability and low optical loss, but is generally slower to reconfigure.
• Piezoelectric or direct beam steering: moves an optical element rapidly, with trade-offs in control complexity, scale and cost.
• Robotic fiber switching: physically aligns fibers and can offer very low loss, but mechanical movement is slower.
• Silicon-photonic or guided-wave switching: can be compact and fast, but scaling port count, optical loss and control remains an engineering challenge.

MEMS is widely viewed as the leading approach for many high-radix OCS designs because it combines high port density, relatively low insertion loss, broadband operation and millisecond-class reconfiguration. Academic and industry comparisons identify 320×320-class MEMS designs as a practical large-scale option, while market research expects MEMS to remain the dominant technology category in the near term.

The frequently repeated claim that MEMS holds more than 70% of the OCS market should be treated cautiously. Market share varies with the market definition, year and analyst methodology. The available evidence supports MEMS as a leading or dominant approach, but not a universally verified 70% figure.

The roadmap—and the limits of the current claim

Chipsens’ reported current product range includes 300×300 and 320×320-class arrays, while the company has described a plan to advance toward 1,024×1,024 ports. That is a roadmap, not evidence that a 1,024×1,024 product is currently shipping or qualified. 79

Likewise, proposals for OCS layers serving clusters with 100,000 or more accelerator cards describe a possible system direction. Reaching that scale would require more than a larger mirror array: it would depend on optical packaging, control-plane software, scheduling, fault handling, cabling, thermal design and coordination with electronic switches. Google’s own documentation illustrates the importance of rerouting and availability mechanisms alongside the optical fabric. 1721

The most defensible conclusion is narrower but significant: Chipsens has reported a standardized 416-mirror component with specifications intended to support domestic 400×400-class OCS designs. If those performance, yield and reliability claims hold through broader customer qualification, the chip could make Chinese OCS system development more accessible and strengthen local supply of a critical AI-interconnect component.

The bigger promises—large energy reductions, 100,000-card optical fabrics and 1,024×1,024 switching—remain system-level opportunities or future plans rather than results established by this launch alone.