Optical Circuit Switching (OCS): A Full Guide to Photonic Fabrics When Packet Switches Hit the Wall

Electrical packet switches still move most AI traffic. At spine scale and inside the densest accelerator pods, their power, hop count, and upgrade cadence are the constraint. Optical circuit switching (OCS) puts a programmable light path in that layer: no optical-electrical-optical conversion, no packet buffers, and no forced rate match when leaf optics jump from 400G to 800G to 1.6T.

This guide covers what OCS actually is, the MEMS / robotic / liquid-crystal / piezo / silicon-photonics trade space, Google's Jupiter and TPU deployments, where OCS sits next to CPO and GPU fabrics, and the honest limits that keep packet switches on the leaf.

Key Takeaways

• OCS steers light paths; it does not inspect or buffer packets.
• Google Jupiter cut network power ~41% and capex ~30% with OCS spines.
• MEMS dominates production radix; SiP wins on microsecond reconfig.
• Use cases today: spine replace, TPU topology, physical-layer failover.
• OCS is not a substitute for NVLink, InfiniBand, or leaf EPS.

Why packet spines are hitting the wall

AI clusters keep raising per-port speeds and east-west bandwidth. Every electrical spine hop regenerates the signal, buffers packets, and burns watts that scale with radix and optics fill. The Open Compute Project's April 2026 OCS white paper frames the problem bluntly: electrical packet-switch (EPS) fabrics struggle with congestion, power, and scalability as training and inference demand higher bandwidth, predictable latency, and large-scale synchronization.

OCS answers a narrower question than "replace Ethernet." It asks where a transparent photonic cross-connect beats another layer of store-and-forward silicon.

Here's the problem. Leaf switches still need packet decisions. Spines and accelerator interconnects often need stable, fat pipes that can be rewired in software when the job or the failure map changes. That is the OCS lane.

What OCS does (and does not do)

An optical circuit switch establishes a direct photonic connection between ports without O-E-O conversion. The path is bufferless and both protocol- and line-rate-agnostic. Latency is dominated by fiber and insertion loss, not queues. Jitter from packet buffering goes away because there is no packet buffer.

That also defines the downside.

OCS is circuit switching: typically one-to-one port mappings, not one-to-many packet fanout. It cannot enforce per-packet QoS, inspect headers, or emit packet-level telemetry. Insertion loss usually limits practical OCS hops to roughly one or two without high-power lasers or optical amplification. EPS fabrics regenerate at every hop; OCS does not.

Technology families: how the light actually gets steered

OCP's April 2026 survey groups production and near-production OCS into several mechanisms. Each trades radix, insertion loss, reconfiguration time, and latching behavior.

Robotic systems behave like automated patch panels: motors move fibers, paths latch without power, and reconfiguration is too slow for in-iteration AI traffic. MEMS mirrors (2D or 3D arrays) steer collimated beams in free space. Google's Apollo / Palomar platform, described in the Mission Apollo paper, is a manufacturable 136×136 3D-MEMS OCS with millisecond-scale switching, roughly 2 dB worst-case insertion loss, and bidirectional links via circulators that effectively double usable radix.

Liquid-crystal and piezo designs avoid or refine mechanical motion. Silicon-photonics switches route through on-chip waveguides with Mach-Zehnder interferometers or rings. They reconfigure fast, but insertion loss and crosstalk climb with port count unless semiconductor optical amplifiers compensate. Hybrid MEMS-plus-SiP designs try to keep both speed and loss under control.

That means technology choice is a control-plane decision as much as a BOM line. Slow, large-radix MEMS or robotic switches fit spine replacement and job-level topology. Fast SiP or metasurface switches matter when the research goal is in-workload reconfiguration.

Deployment pattern 1: replace the EPS spine

Google's Jupiter Evolving work (SIGCOMM 2022) is the clearest production proof for spine-layer OCS. Jupiter moved from a Clos spine of electrical packet switches to a direct-connect fabric among aggregation blocks, with a MEMS OCS datacenter interconnection layer and centralized SDN for traffic and topology engineering.

Google's published outcomes for that evolution include about 5× higher speed and capacity over the decade, roughly 30% lower capex, about 41% lower network power, incremental deployment without ripping prior generations, and average block-level path length around 1.4 (about 60% of traffic on a direct aggregation-block path). The Google Cloud systems blog on Jupiter reports similar order-of-magnitude gains versus prior Clos practice, including large cuts in downtime.

OCP restates the power math for a hypothetical large hall: sixty-four 16-slot spine chassis at ~30 kW each can approach ~1.9 MW at the spine alone when half-populated with coherent optics. Removing or collapsing that electrical tier is why OCS shows up first at the spine, not at every ToR.

Secondary benefit: the optical spine does not force a simultaneous rate upgrade. Leaf optics can move 400G → 800G → 1.6T while the OCS remains rate-agnostic. That is a different lifecycle from chassis EPS spines that must keep pace with leaf SerDes.

Co-Packaged Optics: A Full Guide to CPO, LPO, and Why AI Switches Are Absorbing the Transceiver

CPO versus LPO versus pluggables, Broadcom Bailly, NVIDIA Quantum-X and Spectrum-X, and an honest timeline for 1.6T and 3.2T.

Inside Deep Tech

Deployment pattern 2: reconfigurable AI pods (TPU Superpods)

Google's TPUv4 Superpod architecture puts OCS between hardwired Cubes rather than only at the campus spine. Each Cube holds 64 TPUs (16 servers × 4 TPUs). Sixty-four Cubes make a 4096-TPU Superpod. Cubes connect through three groups of sixteen 128-port MEMS OCS units (48 switches total), with software assigning topology per job.

OCP's summary of the TPU and Lightwave Fabrics literature highlights three payoffs:

• Topology match: job-specific fabrics improved large-language-model training performance by about 3.3× versus static configurations, with roughly 9% lower power than an EPS-based DCN fabric for the Superpod.
• Failure bypass: with 99.9% server availability, a static fabric supporting a 1024-TPU slice delivered only about 25% effective throughput; the reconfigurable Superpod reached about 75% for the same slice size.
• Subdivision: Cubes or Cube groups can be isolated for concurrent jobs and security domains.

Critical architectural limit: OCS sits between Cubes, not inside them. Intra-Cube links stay electrical. Non-surface TPUs still forward through other TPUs to reach the optical layer. The hard problem is the control plane that assigns Cubes, addresses, routes, and failovers for each job.

Deployment pattern 3: physical-layer resiliency

As clusters grow, switch and link failures are routine. Software reroutes on an EPS fabric often leave the job on a degraded graph. OCS can rewrite the physical graph: spare optical paths protect many components, and orchestration restores intended bandwidth instead of a permanent detour.

NVIDIA-linked research cited in the OCP white paper shows optically enabled switch resilience that adds OCS plus redundant electrical switches, with a control-plane sequence for recovery when spine or leaf EPS units fail. For operators, the economic claim is fewer full path duplicates and a shared spare optical pool instead of N+N electrical overbuild.

Where OCS sits next to CPO, silicon photonics, and GPU fabrics

Do not collapse these into one "optics for AI" slide.

Co-packaged optics moves the optical engine onto the switch or accelerator package to cut SerDes and pluggable power. Silicon photonics is the chip technology behind many of those engines and behind some OCS implementations. Active electrical cables keep short, retimed copper when optics are overkill inside the rack. NVLink, InfiniBand, and UALink are the packet or proprietary fabrics that still carry collective traffic and scale-up domains.

OCS is the reconfigurable photonic fabric layer between those domains. It does not replace NVLink inside a scale-up island. It does not replace InfiniBand or Ethernet leaf switching for multipath packet fabrics. It can sit above or beside them as a programmable optical crossbar for spines, pods, and failover.

See Inside Deep Tech's guides on co-packaged optics, silicon photonics, NVLink / InfiniBand / UALink, and active electrical cables for those adjacent BOM columns. Hall-level power density still drives liquid cooling whether the spine is EPS or OCS.

Silicon Photonics: A Full Guide to the Optical Chips Moving AI Data

SOI waveguides, modulators, Ge detectors, foundries, optical I/O vs CPO vs pluggables, and an honest timeline for GPU-package light.

Inside Deep Tech

Honest limits and who should wait

OCS is a poor fit when:

• Traffic is highly unpredictable and needs continuous per-packet multipathing.
• You need deep packet inspection, ACLs, or per-flow telemetry at the switch.
• Your optical budget cannot absorb even one OCS hop's insertion loss.
• The organization lacks SDN/ops maturity for topology engineering (Google's papers stress that the software is half the product).
• You confuse "optical" with "co-packaged": buying CPO switches does not give you OCS topology control.

Future research directions in the OCP paper (ACTINA / OCS-BCube for in-workload reconfiguration, MixNet for MoE expert-parallel all-to-alls) assume faster switches and tighter GPU/transceiver co-design. Those are 2025–2026 research architectures, not a second Jupiter you can buy off a price list tomorrow.

What to do Monday

If you run or buy AI clusters:

• Separate the questions: package optics (CPO/LPO), cable plant (DAC/AEC/AOC), packet fabric (Ethernet/IB/UALink/NVLink), and circuit fabric (OCS).
• Ask vendors for radix, worst-case insertion loss, latching behavior, reconfiguration time distribution, and dark-fiber switching support, not just "optical switch" marketing.
• Demand a control-plane story: who programs topology, how failures rematch circuits, and how jobs get isolated cubes or blocks.
• Benchmark against Google's published Jupiter and TPU numbers as existence proofs, not as your BOM. Your traffic matrix is not theirs.
• Read the OCP April 2026 OCS white paper and the Jupiter Evolving / Mission Apollo papers before greenlighting a spine rip-and-replace.

Packet switches are not dead. They are running out of cheap watts and upgrade headroom at the layers where light can stay light. OCS is how hyperscalers already exploit that gap. Everyone else should copy the architecture carefully, not the slogans.

FAQ