An optical circuit switch moves light, not packets. Inside an OCS, mirrors or other photonic elements connect an input port directly to an output port in the optical domain — no electrical conversion, no packet buffering, no serialization chips in the data path. A circuit stays set until the switch is reconfigured, which is the opposite of an electrical packet switch, which rewrites every packet through silicon on its way through. In an AI data center, that difference turns into power and latency: every electrical switch hop replaced by an analog optical path removes the SerDes, buffer memory, and forwarding silicon that hop would otherwise carry.
We have published the backdrop that makes this question urgent: data center AI networking will surge to nearly $20B in 2025, on a published path we have traced in public — exceeding $300B this decade and approaching $1T early next decade as scale-up, scale-out, and scale-across fabrics all expand. The fastest-growing traffic inside those fabrics is accelerator-to-accelerator, and every hop it takes through electrical switch silicon costs power the cluster would rather spend on compute. Our analysis of networking, power efficiency, and flexible architectures in AI data center networks made the general case; OCS is the aggressive version of it, removing hops rather than making them cheaper.
The proof point is in production. Google’s engineering account of its Jupiter data center network describes optical circuit switching and wavelength division multiplexing operating at datacenter scale, with software-defined control setting the circuits and reconfiguring the topology as traffic and maintenance demand. The lesson from that deployment shapes everything we expect next: OCS succeeds where the map of the fabric changes on the scale of hours, days, or maintenance windows — not where flows change packet by packet. The switch sets analog paths; the software decides where they go; the packets ride without ever touching an electrical forwarding stage in that layer.
Rack-scale systems are pulling in the same direction from a different angle. NVIDIA’s Vera Rubin Ultra NVL576 connects 576 Rubin Ultra GPUs into a single scale-up domain over a two-layer all-to-all NVLink topology, carrying the fabric over co-packaged optics. OCS and co-packaged scale-up fabrics are siblings, not the same device: both move work into the optical domain and strip electrical hops out of the path, but a co-packaged switch still forwards flows through silicon, while an OCS sets analog paths between ports. We expect both to grow on the same trendline — more optics, fewer electrical conversions — and we think buyers should evaluate them as complementary fabric instruments, not competing products.
OCS is also not a general replacement for the packet switch. Bursty, unpredictable traffic needs buffering and per-flow forwarding; a mirror path serves a fixed map. The technology fits fabrics with stable or periodically re-wired topologies, and it should not be confused with the line-system optics of data center interconnect, which solves a different problem between buildings rather than inside them. Inside the data center, the honest comparison is hop by hop: what does each electrical forwarding stage cost in power, latency, and capital, and which of those hops can an analog path remove without breaking the workload’s flexibility.
From the transceiver and switch side, where our research sits, OCS changes the port arithmetic rather than shrinking it. Optical paths still need transceivers, connectors, and fiber, and we forecast scale-up optics adding $10B+ by 2030 alongside scale-up switching above $30B, with AI scale-out Ethernet revenue surging to over $100B by 2030 — inside a total data center networking market we forecast to reach $200B by 2032 from slightly over $50B in 2026. OCS will compete for a slice of the switching spend while adding optical path spend of its own, and we forecast the OCS market to exceed $10B by 2030 — a number we will break out by architecture and vendor in our quarterly trackers as deployment broadens beyond the first mover.
650 Group researches several of the technologies we covered in this blog, including Data Center, Co-Location & Cloud and Enterprise, and more.