Imagine changing which parts of a data center are physically connected without asking someone to move a single patch cable. A controller sends instructions. Tiny mirrors tilt. Light takes a different path.
That is the intriguing job of an optical circuit switch, or OCS. It can look like another anonymous rack-mounted box, yet its role is unusual: it changes the optical connections on which the rest of the network depends.
If an Ethernet switch is a traffic director, an OCS is closer to a road junction whose connections can be rearranged. The distinction has helped reshape networks carrying cloud workloads and large-scale machine learning.
What Is an Optical Circuit Switch?
An optical circuit switch establishes configurable connections between optical input and output ports. Once a circuit is established, the optical signal travels along that connection until the configuration changes.
The device does not need to inspect every Ethernet frame or IP packet to decide where that signal goes. In the all-optical switching path, it redirects light without converting the carried data into electrical form and back again.
Think of a programmable fiber patch panel, with software replacing much of the manual connection work. The Mission Apollo paper describes how Google brought optical circuit switching into large-scale production data-center networking.
The endpoints still need suitable transmitters and receivers. The control electronics also consume power. “All-optical” describes the data path through the switching fabric; it does not mean the complete network runs without electricity.
How Tiny Mirrors Change a Network
In the MEMS implementation described in Google’s explanation of Jupiter, two arrays of movable mirrors steer light between fibers. MEMS stands for micro-electro-mechanical systems: microscopic mechanical structures controlled electrically.
A simplified sequence looks like this:
- Light from an input fiber enters the optical assembly.
- Controlled mirror angles steer it through the selected optical path.
- The light couples into the chosen output fiber, which carries it onward.
The mirrors do not read a destination address. Software has already selected the connection. Many packets can pass through that established optical circuit without the OCS making a fresh packet-level decision for each one.
MEMS is an implementation choice, not the definition of all optical switching. Here, the focus is the fiber-port circuit switch and the MEMS designs that make its operation particularly memorable.
Two Very Different Times Inside One Box
Here is a detail worth pausing over: passing through an established connection and changing that connection are different operations.
The 2022 CALIENT S320 datasheet lists:
| Specification | Published value | What it describes |
|---|---|---|
| Latency | Less than 30 nanoseconds | Transit through the optical switch. |
| Channel setup time | Less than 50 milliseconds | Establishing a channel configuration. |
These are specifications for that product and document, not universal OCS figures. The important lesson is that millisecond-scale setup does not mean every passing packet suffers a millisecond-scale switching delay.
Conversely, low transit latency does not make reconfiguration instantaneous. The network must handle the affected connection becoming temporarily unavailable, and link recovery can add time beyond the optical switching operation itself.
That is why the interesting engineering extends beyond the mirrors.
Why Reconfigure the Connections at All?
Consider four groups of servers connected through a fabric with several optical links between groups. A new workload generates heavy communication between groups A and B, while other links carry less traffic.
An orchestrated design may reassign some optical connections to give A and B more direct capacity. It has not invented extra bandwidth: those ports and links are being reassigned from somewhere else. The controller must preserve enough connectivity and capacity for the remaining traffic.
There is also a less glamorous, highly valuable use: adding capacity or replacing equipment without rebuilding the entire fabric in one disruptive operation.
Google’s 2024 account of Jupiter’s evolution describes optical circuit switching and software-defined control working together to support incremental upgrades and multiple hardware generations. The benefit is a network whose interconnections can evolve alongside the machines using it.
If you have ever labeled both ends of hundreds of cables, “software-controlled topology” probably sounds less abstract already.
This Is Already More Than a Laboratory Idea
Two published deployments show why the equipment deserves attention:
Jupiter data-center networks. In their SIGCOMM 2022 Jupiter Evolving paper, Google’s authors reported a 30% reduction in capital expenditure and a 41% reduction in power over the network evolution they studied. Optical circuit switching, software-defined control, topology changes, and operational automation were part of that redesign. Those are system-level results, not savings guaranteed by purchasing one OCS.
TPU v4 supercomputers. The TPU v4 research paper describes optical circuit switches reconfiguring the interconnect topology of a 4,096-chip machine-learning system. This makes physical connectivity a resource that can be adapted, rather than treating it as a permanently fixed installation decision.
The cloud-network and accelerator-interconnect examples serve different architectural roles. They demonstrate the same broader idea: control over optical connections can improve how a large system is assembled, operated, and used.
Does This Mean Packet Switching Is Going Away?
No. An optical circuit can carry packet traffic. The circuit determines an optical connection; packet switches and endpoints still perform their own forwarding and communication functions.
| Equipment | Main decision in this comparison |
|---|---|
| Ethernet packet switch | Which output should carry this frame? |
| Optical circuit switch | Which optical ports should be connected? |
| Network controller | Which connections and packet paths should the network configure? |
In Jupiter, optical connections link aggregation blocks that still contain packet switches. Optical and packet switching operate together at different layers. Our packet-switching article explains the sharing of link capacity that continues above those physical connections.
This also separates OCS from the idea of reserving an end-to-end telephone circuit for each individual conversation. One optical connection can carry traffic from many applications.
The Constraints That Make the Design Interesting
Optical compatibility still matters. A transparent optical path does not translate incompatible signaling formats or make mismatched endpoints communicate. Wavelength range, fiber type, insertion loss, and the available optical power budget remain engineering constraints.
Reconfiguration must earn its cost. Moving a connection for every short transaction would be a very different proposition from adapting a fabric for a sustained demand pattern or a maintenance operation. The useful timescale depends on the device, workload, and control system.
The controller has a difficult job. It must coordinate topology changes with traffic placement, routing, capacity, and failure handling. In Google’s production design, software-defined networking is integral to the result, rather than an optional dashboard added afterward.
An OCS does not become a firewall. Optical port selection is not packet inspection, user authentication, or application filtering. Those functions must be provided elsewhere in the architecture.
The Revolutionary Part Is the Extra Degree of Freedom
Faster links let a network move more data. An optical circuit switch adds another possibility: changing which optical connections exist, under software control.
That will not make an OCS a useful replacement for a small office’s Ethernet switch. But at data-center scale, it changes the questions an architect can ask. Instead of designing around a fixed web of connections, the system can make parts of that web configurable.
Sometimes the most consequential piece of network equipment is the one that quietly changes what all the other equipment can connect to.