The AI Fabric Is Going Back to the Switchboard

For a decade the fastest network in any data center was a stack of packet switches, each converting light to electrons, inspecting headers and converting back. Google broke with that model years ago, wiring its TPU pods through racks of micro-mirror optical circuit switches. In 2026 the rest of the industry has stopped treating that as a curiosity. Nvidia has written OCS into its scale-up roadmap, Lumentum is about to post its first nine-figure OCS quarter, and a 20-vendor supply chain is forming around a device that does nothing more sophisticated than point one fiber at another. That simplicity changes what the fiber plant inside an AI cluster has to look like.

What Google proved

The reference deployment is still Google's Apollo program. Its first-generation Palomar switch is a 136-by-136 port, strictly non-blocking 3D MEMS device: two arrays of tilting silicon mirrors steer any input fiber to any output fiber, with no buffers, no serializers and no awareness of what the light carries. Google has used Apollo both in the spine of its Jupiter network, where it credited the switches with cutting capex by roughly 30 percent and power by around 40 percent, and as the backbone of every TPU generation since 2015. The payoff is topology: a 3D torus of TPUs is cheap and fast but rigid, and the OCS layer lets Google carve a pod into arbitrary slices, route around a failed chip and rewire the mesh between jobs without touching a patch panel.

Ironwood, the seventh-generation TPU now deploying in Google Cloud, is the largest expression of that idea. Cubes of 64 chips are stitched through OCS into pods of 9,216 chips sharing 1.77 petabytes of directly addressable HBM. Reconfiguration takes milliseconds, glacial by electronic standards and perfectly adequate for a training job that runs for weeks on a traffic pattern it knows in advance.