The “switch” layer, distinct from the “link” layer (Optical Interconnect, Co-Packaged Optics).
Mechanism
OCS establishes a dedicated lightpath between input and output ports and holds it for the duration of a flow, reconfiguring the topology relatively slowly. It switches whole wavelengths/fibers, not individual packets. Reconfiguration time is set by the switching medium (MEMS tilt, LCOS phase, free-space diffraction) and is typically µs-to-ms class — fine for stable, long-lived flows and topology reconfiguration, too slow for per-packet routing. The complement, switching at packet granularity, is Optical Packet Switching (OPS).
The discriminating spec for any optical switch is switching speed. Per Zigzag Networks: ns-class competes with NVLink-class fabric; µs/ms competes only with incumbent OCS for slow topology reconfiguration.
Where it sits (three distinct layers)
- Scale-up OCS / OPS — the switch fabric (this page). Replaces electrical top-of-rack/spine with light for MoE all-to-all and collective ops.
- In-package optical I/O — chip-to-chip (Ayar Labs, Lightmatter Passage, Celestial AI). Not switching.
- Fabric-replacement protocols — NVLink, Broadcom Tomahawk. Electrical-native.
Incumbent landscape
Switching-speed spectrum (the discriminating axis)
The live competitive bet in OCS is which medium delivers fast (sub-µs) switching at high radix and foundry-manufacturable cost — where Neye Ai’s silicon-photonic MEMS currently looks strongest.
The convergence trap and the enabling-layer wedge
The whole venture cohort is building the same thing — a reconfigurable optical switch fabric — and differentiating only on the actuation medium (3D-MEMS, silicon-photonic MEMS, Mach-Zehnder / micro-ring, active metasurface, LCOS, free-space diffraction). That is a sustaining-innovation arms race on one axis (ports × switching speed × power), and the zero-success record (no standalone OCS company has ever reached venture scale) says that race ends in strategic acquisition, not a standalone winner. The re-synchronisation reframe above sharpens it: if the binding constraint is timing rather than the switch medium, a better switch is the wrong thing to build.
So the orthogonal, earlier-stage bets are the layers the whole race needs regardless of which switch medium or hyperscaler wins:
The robust version of the bet: back the layer that wins regardless of which switch wins. #1 (source) is safest and already in-portfolio; #2 (sync) is the most contrarian and the clearest “sell to everyone.” Weakest links held honestly: the wavelength-routing flip has been academically promising for 15+ years without displacing switches (so bet the source layer on “every fabric needs sources,” with the flip as upside, not as the thesis); and the sync primitive risks being built in-house by hyperscalers (the merchant-vs-captive test).
Companies using
Connected ideas
Frontier (open questions)
- The pattern where no standalone OCS company has ever reached venture scale (they get acquired or stay small): broken by 2030, or holds?
- ns-class switching: does it arrive, and from which medium?
- Medium winner for scale-up AI fabrics (MEMS / LCOS / free-space diffractive / soliton-microcomb)?
- Does the architecture flip to wavelength routing (passive AWGR + nanosecond-tunable lasers), moving value from the switch to the light source?
- Is the backable-early wedge the switch at all, or the enabling layer (tunable-laser/comb source; merchant burst-mode sync primitive) the whole cohort must buy?