Google Insights: Charting the Next Frontier of Optical Switching

In this article, Google explores device technologies for next-generation optical circuit switches (OCS), with a clear focus on data center networks and machine-learning supercomputers. Key device parameters—such as insertion loss, crosstalk, port count, reconfiguration time, and polarization sensitivity—directly shape system-level performance and reliability.

Introduction

Large-scale systems depend on networks to move information from sources to destinations through switching fabrics. Today, most hyperscale data networks are built around electronic packet switches (EPS) arranged in a fixed Clos topology. While these networks can support arbitrary communication patterns, they do not scale gracefully across critical system metrics like cost, latency, and reconfigurability.

These well-known scaling limits motivated early research into optical circuit switches, which can dynamically reshape network topology to better match prevailing communication patterns. That foundational work has since led to real-world deployment of OCS in large data center networks and machine learning systems. Optical switching has become a core enabler of networks that are high-performance, cost-effective, and reconfigurable. This article reviews the commercial OCS landscape and outlines the device-technology directions that may define the next generation of optical switching.

Background

Electronic packet switches queue packets in shared memory and make local routing decisions based on packet headers, forwarding each packet to the appropriate output port. End-to-end connectivity is formed by multi-hop paths that traverse multiple switches. Because routing decisions occur on a per-packet basis, packets from the same source to the same destination can experience different delays—an outcome that becomes especially problematic for synchronized machine learning workloads.

Optical circuit switches take a different approach. They establish an end-to-end optical “circuit” between an input port and an output port. Traffic stays in the optical domain as it passes through the switch, and packets are routed along a preconfigured path rather than being individually inspected and forwarded based on headers. As a result, all packets traverse the same optical path and experience the same latency—an ideal property for tightly synchronized training jobs.

Many OCS designs are also largely rate-insensitive, allowing a switch to remain useful across multiple generations of optical transceivers operating at different data rates. That simplicity comes with a tradeoff: optical switching generally requires centralized control. At large scale, building the control plane can demand as much—or more—engineering effort than building the switching hardware itself.

Future Optical Switching Technologies

Google highlights several device technology routes for commercial and research OCS platforms, emphasizing that real-world performance depends on whether switching is implemented in the spatial domain or the wavelength domain, and whether the device is realized as a three-dimensional free-space architecture or a two-dimensional planar architecture.

There is currently no single switching technology that is optimal across all applications and all performance metrics. For today’s large-scale deployments, optical switches are typically designed around three practical priorities: very high port counts, low insertion loss, and low return loss.

3D Free-Space Switching

A widely deployed commercial approach uses custom MEMS (micro-electromechanical systems) mirrors. In this design, high-reflectivity micro-mirrors—fabricated using deep reactive ion etching—are steered by high-voltage signals that actuate comb-drive structures around each mirror. Each mirror can rotate around two axes, and pairs of such devices can create an arbitrary 3D optical path from any input port to any output port.

Google notes that MEMS-based OCS has delivered meaningful cost advantages in large-scale data center networks. It has also improved availability and performance when used in TPU superpod-style machine learning systems, where consistent latency and high bisection bandwidth are critical.

Beyond mechanical MEMS mirrors, Google points to emerging non-mechanical options for 3D free-space switching—such as two-dimensional digital liquid crystal (DLC) pixel arrays. These devices leverage polarization to digitally steer beam direction. By building a folded cascade from N binary stages, a switch with 2^N ports can be realized, enabling large port counts through modular composition.

2D Planar Switching Devices

In contrast to 3D free-space architectures, many research efforts in 2D planar switching use waveguide crossbar matrices: N waveguides in each direction, with a binary switching element placed at each of the N² intersections to control optical routing.

A large portion of this research is driven by silicon photonics (SiP), designed for compatibility with standard CMOS manufacturing. In principle, SiP planar switches can offer lower cost per port, faster switching speed, tighter integration with electronics, and potentially higher reliability due to lower drive voltages than many 3D free-space systems.

However, these advantages have not yet been fully realized in mass-produced systems. The most persistent challenges include high loss during fiber coupling and within the switching fabric, as well as limited port counts. Some drawbacks—especially insertion loss—are difficult to avoid across many planar architectures.

1. Interference-Based Devices

Interference-based planar switches have been studied extensively. Two representative devices are:

Mach–Zehnder interferometers (MZIs), which create switching states via single-pass interference.

Microring resonators, which create switching states via multi-pass interference in a resonant cavity.

Resonator-based devices can achieve lower drive voltages, but they typically have narrower bandwidth and can be more complex to control.

Both families rely on changing refractive index to produce constructive or destructive interference. Common tuning methods include thermal tuning and electro-optic effects, where an applied electric field changes refractive index. The induced index change is wavelength-dependent, which constrains usable bandwidth.

Thermal tuning is relatively slow (microseconds rather than nanoseconds) and demands careful control to avoid thermal crosstalk across closely packed devices. For interference-based switches at scale, key challenges include reducing overall loss, enabling polarization-diverse designs, and managing increasing crosstalk as the number of cascaded elements—and thus port count—grows.

2. Heterogeneously Integrated Devices

Google highlights a fast-growing application that intensifies switching requirements: photonic quantum computing. In this context, optical switches are used to generate initial computational resources and to perform feed-forward operations across stages of a quantum workflow. Because these stages can depend on prior measurement outcomes, overall system speed can become limited by switch reconfiguration time. Quantum photonics also imposes exceptionally strict constraints on loss and crosstalk.

To address these demands, researchers are exploring high-speed switches based on heterogeneous integration—combining thin-film materials with strong electro-optic effects onto foundry silicon photonics. This approach can enable lower drive voltage and higher speed. Micro-transfer-printing methods are also being developed to support scalable heterogeneous integration. Even so, the core limitations of interference-based devices still apply, and the field must also solve the manufacturability and packaging challenges required for practical deployment.

3. Silicon Photonics MEMS Devices

MEMS is also being explored in silicon photonics planar switching. In this architecture, input and output fiber-array units connect into a 2D waveguide crossbar, and MEMS-actuated couplers at each intersection direct light into one of two paths. The MEMS photonic integrated circuit is then integrated with a CMOS control chip.

Compared with the analog MEMS used in free-space systems, binary MEMS couplers can switch up to 1000× faster and have demonstrated relatively large port counts in research prototypes. Still, the packaging challenge remains formidable—especially achieving low-loss fiber-to-waveguide connections across 2N interfaces, a pain point shared broadly across silicon photonics switching approaches.

4. Wavelength-Switching Devices

Another route is wavelength switching, built from tunable lasers, passive arrayed waveguide gratings (AWGs), and tunable filters. Compared with spatial switching approaches, tunable lasers can be more expensive and power-hungry, while passive optics may introduce higher loss and operate over fixed wavelength bands. These characteristics tend to limit port count and constrain the per-port usable bandwidth.

Conclusion

As optical circuit switching matures commercially, research activity around next-generation OCS device technologies is accelerating. As application demands diversify—from hyperscale data centers to AI supercomputers and emerging quantum systems—some of today’s research-stage device technologies are expected to transition into future compute and networking platforms at production scale.

Appendix: The Origins of Google’s OCS Effort

Google’s optical switching journey began years ago with a quiet but ambitious transformation of its data centers—replacing conventional network infrastructure through a deeply integrated, internal approach long envisioned by the networking community.

At the center is a program often referred to as “Project Apollo,” built on a simple principle: replace electronics with optics wherever possible, and use optical circuit switches instead of traditional electronic switches. In interviews toward the end of 2023, Google’s systems and services infrastructure leadership described why this shift matters.

Keeping Data in the Optical Domain

A fundamental inefficiency in data center communication comes from bridging two domains. Computation occurs electronically, so server-side data starts in the electrical domain. Yet transporting data is often faster and more efficient in the optical domain.

Traditional network topologies repeatedly convert signals between electrical and optical forms—hop by hop. Each conversion increases cost and power draw, while per-packet electronic processing adds latency.

OCS changes the equation by keeping traffic in the optical domain for as long as possible. Using micro-mirrors to redirect beams, the system forms direct optical cross-connects from a source to a destination port, reducing the frequency of electrical switching.

Google argues this approach reduces latency by avoiding unnecessary in-datacenter processing, and cuts power because the primary ongoing consumption is holding mirror positions—typically far lower than the power profile of large EPS systems.

How the Switching Hardware Works

In Apollo-style switches, fiber bundles bring light into the system, where multiple silicon components reflect beams using MEMS micro-mirror arrays. Each mirror can be rapidly reoriented to redirect an optical signal to a different fiber in the output bundle.

In one described design, each array contains 176 micro-mirrors, though only 136 may be used to improve yield. That yields 136² possible input-output combinations between two mirror components, totaling 18,496 connection possibilities.

Google reports a maximum power draw around 108 watts for the optical switching system—often lower in typical operation—compared with roughly 3000 watts for comparable EPS-based systems. Over the past several years, the company has deployed thousands of such OCS systems, describing it as the largest OCS deployment of its kind.

Custom Engineering at Scale

Delivering this system required extensive customization—not just in switch design, but in manufacturing and production tooling. Google describes building custom test, alignment, and assembly stations for MEMS mirrors, fiber collimators, optical chips, and subcomponents. It also developed automated alignment tools capable of placing lens arrays with sub-micron precision.

The company also designed and deployed custom optical components such as transceivers and circulators. Circulators help route light directionally between ports, and Google notes these designs can cut fiber counts roughly in half compared with prior approaches.

On the transceiver side, Google pursued low-cost WDM transceivers spanning multiple generations of optical interconnect speeds (40, 100, 200, 400GbE). A key design goal was balancing cost, power, and modulation format requirements while tolerating the additional insertion loss introduced by reflective optical switching.

Software-Defined Control: Orion

A broader piece of the vision is Orion, Google’s software-defined networking layer. Orion predated Apollo, meaning Google already operated a logically centralized control plane. That foundation made it feasible—though still difficult—to extend traffic engineering and routing control to manage direct-connect optical topologies and perform real-time traffic-aware mirror reconfiguration.

Engineering Challenges: Reconfiguration Time

One of Apollo’s key challenges is reconfiguration time. Clos networks built from EPS can effectively interconnect ports continuously with rapid, per-packet decisions. Optical switching is different: changing a direct connection requires mirrors to physically reconfigure, which can take seconds—far slower than electronic packet forwarding.

Google’s strategy is to reduce how often reconfiguration is needed by exploiting stable, long-lived communication patterns at scale. It describes traffic behavior in terms of “superblocks”—aggregations of roughly 1 to 2000 servers. While traffic volumes between superblocks are not perfectly constant, they tend to exhibit stability that can be leveraged to keep data on optical paths for long durations, and reconfigure only when patterns shift significantly.

This also enables dynamic infrastructure allocation. If additional electronic packet switching capacity is needed, Google suggests it can temporarily assign an unused “superblock” as a backbone element, effectively recruiting capacity without requiring data synchronization—since the block is not itself a traffic source.

Why Google Believes It’s Worth It

Google’s thesis is that optical switching can become a building-level infrastructure primitive because photons are agnostic to how bits are encoded. In practice, that means a single OCS deployment can remain useful across multiple transceiver generations—from 10Gbps upward to 40, 200, 400, 800Gbps and beyond—without replacing the optical switching fabric.

This creates operational flexibility: upgrades can happen at Google’s pace rather than forcing disruptive, full-fabric redesigns. Google argues this reduces downtime risk, lowers capital expenditure because EPS refresh cycles are avoided, and can cut network-related costs significantly—while also delivering substantial energy savings.

Looking forward, Google is aiming for OCS systems with higher port counts, lower insertion loss, and faster reconfiguration. The upside is large: Google notes that the internal bisection bandwidth of modern data centers can rival that of the broader internet—meaning even incremental efficiency gains translate into enormous savings in money, energy, and latency.