Optical Circuit Switching: Why Google and Lightmatter Are Ditching Electrical Switches for AI

Most AI infrastructure talk fixates on the chip. But a GPU spends a surprising share of its life waiting on the network, and every electrical switch between two accelerators is a toll booth where light gets converted into electrons, parsed, and converted back into light. Optical circuit switching asks a rude question: what if the switch never touches the electrons at all?

Before you read: Our earlier piece on silicon photonics covered why copper is giving up inside AI racks. This article goes one layer up, to the switches that connect those racks and can rewire the whole cluster on demand. Silicon Photonics: Why AI is Forcing Data Centers to Abandon Copper

Table of contents

• The Toll Booth Problem
• Switching With Mirrors
• Why Training Clusters Love It
• Lightmatter Wants the Whole Data Path
• Fun facts
• Sources

The Toll Booth Problem

In a conventional packet-switched fabric, every hop is an electrical switch. The signal arrives as light, becomes electricity, gets its header read, is sent to an output port, and is converted back to light for the next hop. Each conversion burns power and adds latency. Multiply that by the thousands of hops a large training job touches and the toll adds up fast.

Switching With Mirrors

Google’s answer, laid out in its 2022 SIGCOMM paper “Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale,” is an optical circuit switch built around arrays of MEMS mirrors. Light enters on one fiber, hits a microscopic mirror that tilts to aim the beam, and leaves on whichever fiber the mirror points at. No packet parsing, no buffering, no conversion. The route is set by an angle.

Anyone who has spent a Friday night re-patching fiber in a cage knows the appeal. The mirror does the re-patching in software, and nobody has to hold a flashlight in their teeth while counting fiber colors. Honestly, it’s a strange sentence to write about the most advanced networking gear in a hyperscale facility: the core of the switch moves by tilting.

Why Training Clusters Love It

Training traffic is unusually predictable. A large model’s communication patterns repeat for days, so the topology can be configured once per job and held steady. The TPU v4 paper (ISCA 2023) describes 4,096-chip pods wired as a reconfigurable 3D torus, where optical switches carve out slices for different jobs and route around failed racks without a technician touching a cable. The authors put the optical switching layer at under 5 percent of system cost and under 3 percent of system power. That is a rare number in this industry: low enough to be boring, high enough to matter.

The catch is speed. Reconfiguration happens on a timescale of milliseconds, which is an eternity for traffic that shifts microsecond by microsecond. The approach works because training schedules are stable. Hand it an inference workload with bursty, unpredictable flows and the advantage evaporates. Google’s design doesn’t replace the electrical network; it sits beside it, taking the long, steady flows that dominate training.

Lightmatter Wants the Whole Data Path

Google is wiring its own pods. Lightmatter, the Boston startup that spent years on photonic computing before pivoting hard toward interconnect, wants optics everywhere data moves. Its Passage and Photonic Fabric products pitch light as the transport between processors and memory, not just the link between boxes. If that works, the optical switch becomes one piece of a fully photonic path instead of an isolated appliance.

The problem is that the pitch is considerably bigger than the public evidence. Independent benchmarks at cluster scale are thin, and “our chiplet moves data at X terabits per second” is a sentence that is much easier to write than to verify at volume. I’d want to see a deployment, not a slide, before calling the photonic fabric a done deal.

Fun facts

• According to the Mission Apollo paper, the mirror positions are checked and corrected with camera-based feedback loops. After all the photonics, one of the most reliable tools for aligning a beam turns out to be a camera watching the light.
• Because the torus is reconfigurable, a failed rack can be routed around in software. The job keeps running while a technician decides when to fix the hardware, and nobody has to unplug anything at 3 a.m.

The honest verdict: optical circuit switching is a clever answer to a specific question, namely how to keep a predictable training cluster wired the way the job needs. Until someone outside Google publishes numbers at comparable scale, treat the all-optical data center as a promising architecture, not a settled one. The electrons aren’t going quietly, but they are being asked to do less.

Sources

• Google, “Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale,” ACM SIGCOMM 2022
• Jouppi et al., “TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings,” ISCA 2023
• Lightmatter, product materials on Passage and Photonic Fabric (lightmatter.co)