Optical Circuit Switching Returns
Why AI clusters are reviving a transparent switch with no packet logic: how an optical circuit switch works, where its insertion loss lands in the link budget, the millisecond-versus-packet tradeoff it forces, and the traffic patterns that make it pay in a training fabric.
Scope
This article explains the optical circuit switch (OCS) and why it has moved from a niche to a standard element of AI datacenter fabrics: what it does, how MEMS-based switches are built, how its transparency changes the link budget, the tradeoff it forces against packet switching, and the traffic conditions under which it saves power and money. Background: optical circuit switching, insertion loss, and the module-architecture context in Pluggable, LRO, LPO, CPO.
Two ways to move a packet across a fabric
An electrical packet switch reads every packet, looks up its destination, and forwards it, converting optics to electronics and back at each hop (an O-E-O crossing), and it does this continuously whether or not the topology ever changes. That flexibility is the right tool for bursty, unpredictable traffic, and it is why datacenter networks were built from it. It also costs power at every hop and every conversion, and at AI-cluster scale those hops multiply.
An optical circuit switch does something older and simpler: it makes a physical light path from an input port to an output port and leaves it there. There is no packet logic, no lookup, no O-E-O conversion; photons enter one fiber and leave another. The switch does not know or care what modulation, bit rate, or wavelength passes through it, because it never looks. It is a reconfigurable patch panel operated by mirrors rather than hands.
How a MEMS optical circuit switch works
The dominant implementation steers light with micro-electromechanical mirrors. Each input fiber is collimated into a free-space beam that strikes a tiltable MEMS mirror; the mirror aims the beam across a small free-space gap to a second mirror, which couples it into the chosen output fiber. Setting a connection means tilting two mirrors to the angles that link the desired ports, a three-dimensional beam-steering problem solved by an internal control loop. Commercial units are large: Lumentum's R300 offers up to 300 by 300 ports, with the R64 a smaller 64-port member of the same family.
The consequences of this construction are the whole reason OCS is attractive. Once the mirrors are set they hold position with almost no static power, so a connected switch draws close to nothing, in contrast to a packet switch that burns power forwarding every frame. The insertion loss is low, on the order of 1.5 dB or less for a good MEMS switch, because the light only reflects off two mirrors and recouples. And because the path is transparent, one switch serves any current or future line rate; a fabric built on OCS does not need replacing when the transceivers move from 800G to 1.6T.
Where the loss lands
Transparency has a direct link-budget consequence that is easy to miss: because the switch never regenerates the signal, its insertion loss adds directly to the optical path it sits in, alongside the fiber, connectors, and any mux. A packet switch resets the budget at every hop by receiving and retransmitting; an OCS does not, so a 1.5 dB switch in the path is 1.5 dB the transceiver at the far end must have in hand. Two or three cascaded OCS stages, each with its connectors, can consume a meaningful fraction of a short-reach budget, which is why OCS fabrics are designed with the switch loss budgeted from the start rather than added as an afterthought. The Datacenter Link Budget Explorer treats exactly this kind of transparent element as a fixed loss between transmitter and receiver.
Milliseconds, not packets
The cost of removing the packet logic is that an OCS cannot make per-packet decisions. Reconfiguring a MEMS switch means physically tilting mirrors, which takes milliseconds, an eternity next to the nanoseconds a packet switch spends per frame. An OCS therefore cannot do statistical multiplexing, the trick that lets a packet network oversubscribe a link and share it among bursty flows. It switches circuits, not packets, and it is useful only where the traffic pattern holds still long enough that a circuit set once is used many times before it needs to change.
That condition sounds restrictive until you look at what an AI training cluster actually does. The communication pattern of a training job is structured and repetitive: the same groups of accelerators exchange the same gradients in the same pattern for hours. A topology matched to that pattern, set once at job launch and held, extracts almost all the benefit of a full packet fabric at a fraction of the power, and reconfigures in milliseconds when the job or the failure map changes. The mismatch between millisecond switching and microsecond traffic that rules OCS out of general networking is simply not present.
Why now
Optical circuit switching is not new. Google reported running its Jupiter datacenter fabric on MEMS optical circuit switches at scale in 2022, replacing the top layer of electrical spine switches with OCS and reconfiguring the topology to match demand. What changed by 2026 is that the argument stopped being one hyperscaler's in-house advantage and became an industry position: the Open Compute Project published a white paper on optical circuit switching for AI in April 2026, and merchant OCS products (Lumentum's R-series, MEMS units from Eoptolink and others) put the capability on the market rather than behind one company's walls. The driver is the same one pushing optics into every other layer of the AI datacenter: at cluster scale, the power and cost of O-E-O conversion in the network became large enough that eliminating a whole switching tier is worth the loss of per-packet flexibility.
The honest boundary: OCS does not replace packet switching, it removes a layer of it where the traffic is predictable. The access and aggregation layers, where traffic is bursty, stay electrical. OCS wins the slowly-changing core of a fabric whose job is known in advance, which for an AI training cluster is most of it.
References: Poutievski et al., "Jupiter evolving: transforming Google's datacenter network via optical circuit switches and software-defined networking," ACM SIGCOMM (2022); Open Compute Project, "Optical Circuit Switching for AI" white paper (April 2026); Lumentum optical circuit switch product materials (R300, R64), 2025-2026. The optical circuit switching entry gives the device summary, and Pluggable, LRO, LPO, CPO covers the module layer it connects.