A navigation aid for placing any optical-networking deal. Written 2026-05-30. The whole field solves one problem: moving bits between chips fast without burning too much power. Almost every component is a converter (electrons ↔ light), a decider (where does this bit go?), or the road between them.
The converter chain (one bit, GPU A → GPU B)
Data starts electrical on a chip; to send it far you put it on light and read it back.
- Laser — the light source. Steady blank carrier (flashlight held on).
- Modulator — impresses data onto the beam by flicking it billions of times/sec (the hand flicking the flashlight into Morse). The materials race (TFLN, BTO, silicon — Lithium Niobate (TFLN)) is just ways to flick faster/cheaper/cooler.
- Photodetector — converts light back to electrical at the far end (reads the Morse).
- Transceiver — does both directions: TRANS-mitter (laser+modulator) + re-CEIVER (detector) + DSP. The translator at each fiber end. Each conversion (O-E-O, optical-electrical-optical) burns power — the villain of the whole story.
- Pluggable — a transceiver as a removable faceplate module (a dongle). Most of the market.
- Co-Packaged Optics (CPO) — moves the transceiver next to the switch chip (mm not cm of electrical trace) to cut power. “CPO vs pluggables” = where you put the converter.
- Fiber — the road.
The deciders (where OCS fits)
The hierarchy
chip → server → rack → pod → datacenter. (chip = desk, rack = office, pod = floor, datacenter = building; transceivers = phones on desks, switches = mailrooms on floors, OCS = the re-wireable riser.)
- Chip/package: electrical + emerging optical I/O (Ayar, Lightmatter).
- Scale-up: GPUs within a pod, fast/tight, mostly electrical (NVLink, copper). See Optical Vs Electrical Scale Up.
- Scale-out: pods across the building; top-of-rack → spine → core; where big packet switches live and where OCS competes as a reconfigurable patch panel.
Where deals sit (the placement question)
AI broke the old model: thousands of GPUs in lockstep, so electronic switches hit power/bandwidth walls and every conversion hurts. Two layers, two responses:
The one question that places any deal: is it a converter (link layer) or a decider (switch layer), and at which rung of the hierarchy? That fixes who it competes with and which thesis it tests. For the switch-layer speed axis (why “MEMS is too slow”), see the spectrum table in Optical Circuit Switching (OCS).