Does optical circuit switching replace packet switches in AI clusters?

Not in current designs. OCS is layered alongside electrical packet switching: it handles the long-lived, predictable collective-communication paths, while electrical switches retain bursty traffic and topology changes. The hybrid pattern is what production deployments actually use.

How long does an OCS reconfiguration take?

Orders of magnitude slower than a packet switch converging. Circuit reconfiguration runs from seconds to minutes depending on the switch technology and the verification steps around it. That latency is acceptable for job-level topology changes and unacceptable for per-flow decisions.

Why do AI workloads suit circuits at all?

Collective communication in training - all-reduce, all-to-all between fixed process groups - produces traffic between the same endpoint sets for the lifetime of a job. When the traffic map is stable for days, setting up dedicated paths stops being overhead and starts being efficiency.

What does OCS do to the optics bill of materials?

It trades some switching optics for path optics. Circuits pass through the OCS, so fibers and connectors multiply, and every path's insertion loss budget now includes the switch element. Expect more fiber infrastructure and stricter connector hygiene, not fewer transceivers overall.

Is OCS relevant below AI scale?

Generally no. Below the scale where job-level topology stability exists, the reconfiguration latency and per-path cost outweigh the power and latency savings. The economics only work once a fabric hosts large, long-running, topology-aware jobs.

Optical Circuit Switches in AI Clusters: An Engineering View

An OCS moves no packets and reads no headers. That is precisely why it fits the one workload whose traffic map is stable for weeks at a time.

AI Datacenter Optics · Published 2026-10-27 · 5 min read

The strongest argument for optical circuit switching in AI clusters is not optical at all. It is a traffic observation: training collectives talk between the same sets of endpoints for the whole life of a job. An all-reduce ring does not renegotiate its membership; a pipeline-parallel stage pair does not wander. When the communication map is stable for days, the packet switch's core competency - adapting to unpredictable flows - is being paid for and barely used.

That observation is not new. What changed is scale making it economic: at thousands of accelerators per job, the power, latency and capital of an electrical switching tier justify revisiting a design in which those tiers handle only what genuinely needs them.

What an OCS actually does

An optical circuit switch takes fibers in, redirects them, and sends them out. No packet parsing, no buffering, no congestion control, no table lookups at line rate. Modern implementations mostly use arrays of steerable mirrors or wavelength-selective elements; the engineering properties that matter for fabric design are the same in either family.

Attribute	Electrical packet switch	Optical circuit switch
Forwards	Packets, per hop	Continuous light paths
Buffers / queues	Yes	None
Reconfiguration	Per flow, microseconds	Per circuit, seconds to minutes
Per-port power	High (SERDes + buffering)	Near zero in the switch itself
Latency added	Switch + queueing	Propagation only
Congestion handling	Rich (ECN, scheduling)	None - avoidance must come from topology
Failure granularity	Port	Path or port array

The last row is the one operators underestimate. When a circuit is misconfigured or a mirror array degrades, the blast radius is a set of paths, not a queue, and the symptom in the job is a stall in a collective - which the training framework experiences as a hung rank rather than a packet loss event.

The traffic argument, stated honestly

Circuit switching wins only where circuits are set up once and used heavily. Training meets that bar for the collectives; it does not meet it for everything:

Stable for the job's lifetime: gradient all-reduce rings, tensor-parallel links between fixed ranks, pipeline stage handoffs. These are circuit-shaped: high volume, fixed endpoints, long duration.
Bursty and unpredictable: data loading, checkpointing to storage, debugging traffic, any elastic scaling event. These need electrical switching and will keep it.

Production designs therefore layer the two: an electrical tier for the irregular traffic, an OCS tier carrying the heavy, stable collectives - sometimes as dedicated paths that bypass the electrical tier entirely, sometimes as a slowly reconfigurable topology layer that reshapes the network between jobs. The AI cluster interconnect design article covers the rail-optimized electrical baseline that this layer sits on top of.

The consequences for optics nobody budgets early

Adding an OCS tier changes the optical plant more than the switch vendor's slides suggest.

Fiber count explodes

Every circuit is a dedicated fiber path through the switch. A fabric that used to aggregate many flows onto one switch-to-switch trunk now runs individual paths per collective edge. Structured cabling, MPO hygiene and labeling standards stop being back-office concerns and become availability factors - the MPO polarity rules matter an order of magnitude more when a single mispolarized trunk silently inverts a path mapping.

Insertion loss joins the budget

Every path now includes the OCS element, typically adding a few dB depending on technology and port count. That loss comes out of the transceiver's budget, which constrains reach class and sometimes forces a higher-power PMD for paths that are physically shorter than the optics assume. Run the numbers per path - the habit from link budget work transfers directly, with the OCS loss as one more line item.

Connector quality becomes a reliability metric

A packet switch isolates a dirty connector to one link. A circuit switch passes light straight through, so marginal connectors on circuit paths degrade many endpoints' collective performance at once, and the symptom appears as training throughput variance that no one can attribute. Connector inspection cycles move from "annual best practice" to "job-level operational requirement".

The transceiver mix shifts toward short reach

Circuits mostly connect nearby rows within the same machine room, which favors parallel short-reach PMDs and, in some designs, directattach-style economics over long-reach optics. The power saved at the switching tier partially reappears as more short-reach transceiver ports; the net depends on topology and should be modeled, not assumed.

Failure modes and operations

Operating an OCS tier introduces new runbooks:

Path verification. After any reconfiguration, paths must be verified optically (power levels per path) before the job layer assumes them. The verification step, not the mirror movement, dominates reconfiguration time.
Topology drift discipline. The circuit map is now part of the cluster's state. It needs version control, diff review and rollback - the same rigor as a routing policy, because that is what it is.
Failure semantics. Training frameworks must learn to distinguish "path down, reconfigure" from "rank slow, restart". That integration work is real engineering effort and belongs in the project plan from day one.
Mirror degradation. Port arrays age. A periodic per-path optical power sweep catches drifting ports before jobs land on them.

Where electrical still wins, plainly

For most clusters being built in 2026, the answer is "almost everywhere":

Below a few thousand accelerators, the tier's fixed cost has no payback surface.
Workloads with unstable communication patterns (frequent job churn, heterogeneous models, elastic inference) never hold circuits long enough to amortize them.
Multi-tenant fabrics cannot promise a circuit map to tenants who arrive and leave hourly.
Teams without topology-aware job schedulers get nothing from circuits; the payoff requires software that can exploit a stable map.

The realistic adoption path is incremental: keep the electrical spine, add OCS for one specific collective pattern (typically the widest all-reduce tier), measure, and expand only if the operational machinery - path verification, topology versioning, scheduler integration - actually worked the first time. Clusters that skip the operational groundwork and buy the switch first reliably discover that the optics were never the hard part.

Sources

Jouppi et al., "TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning" (arXiv) - the published large-scale deployment of OCS in a production ML system, including the performance and availability rationale.
Ethernet Alliance - multi-vendor demonstrations covering switching and optics technology roadmaps relevant to fabric design choices.
OIF - interoperability work on the coherent and short-reach interface families used in cluster fabrics.
IEEE 802.3df Task Force - the 800G PMD baseline that the electrical tiers in hybrid designs are built from.

OCS implementations differ widely in switching technology, port counts and loss characteristics. Figures here are planning-level guidance; validate against a specific switch and topology model before committing a design.

FAQ

Does optical circuit switching replace packet switches in AI clusters?: Not in current designs. OCS is layered alongside electrical packet switching: it handles the long-lived, predictable collective-communication paths, while electrical switches retain bursty traffic and topology changes. The hybrid pattern is what production deployments actually use.
How long does an OCS reconfiguration take?: Orders of magnitude slower than a packet switch converging. Circuit reconfiguration runs from seconds to minutes depending on the switch technology and the verification steps around it. That latency is acceptable for job-level topology changes and unacceptable for per-flow decisions.
Why do AI workloads suit circuits at all?: Collective communication in training - all-reduce, all-to-all between fixed process groups - produces traffic between the same endpoint sets for the lifetime of a job. When the traffic map is stable for days, setting up dedicated paths stops being overhead and starts being efficiency.
What does OCS do to the optics bill of materials?: It trades some switching optics for path optics. Circuits pass through the OCS, so fibers and connectors multiply, and every path's insertion loss budget now includes the switch element. Expect more fiber infrastructure and stricter connector hygiene, not fewer transceivers overall.
Is OCS relevant below AI scale?: Generally no. Below the scale where job-level topology stability exists, the reconfiguration latency and per-path cost outweigh the power and latency savings. The economics only work once a fabric hosts large, long-running, topology-aware jobs.