"Let there be light," and there was — an optical module. In an AI cluster the famous part is the GPU, but a GPU on its own is an island. To make thousands of them work as one machine, every chip has to talk to every other chip and to the switches between them, fast, and copper wire hits a physical wall well before AI bandwidths. So the signal leaves the electrical world and travels as light. The optical module is the device that makes that jump — the high-speed connector wired between GPU and GPU, and between GPU and switch. Without it, all that compute can't communicate, and the cluster is bottlenecked. That's why, in the map of AI infrastructure, optical modules are a layer of their own — the "blood vessels" of compute, sitting alongside the silicon "brain" and the physical "body" of structure and cooling.
How the light gets made
The mechanism is simple to picture and hard to build. It has three steps:
- Transmit. The electrical signal from the server hits a light-making chip — typically an EML(electro-absorption modulated laser) — which modulates it into a beam of light carrying the same information.
- Travel. That light runs down an optical fiber: fast, far, and with very little loss.
- Receive. At the far end a PD(photodetector) chip turns the light back into an electrical signal the next device can use.
The cleanest analogy is a lighthouse: it sends a message by switching its lamp on and off (Morse code). An optical module does the same thing, just unimaginably faster — encoding data into the light's on/off, its brightness, even its phase and polarization.
The routes racing to build it
There's no single winning design. To chase bandwidth, power and cost all at once, several technology routes are evolving in parallel — competing and combining:
| Route | Core idea | Where it stands |
|---|---|---|
| EML (traditional) | the EML chip both emits and modulates the light — mature and reliable | today's mainstream, esp. beyond 2 km; the chip supply is the bottleneck |
| Silicon photonics | build the optics in silicon — dense, cheap, low-power; needs an external CW laser for light | spreading fast; mainstream at 500 m / 2 km; 2025 was called the "year of silicon photonics" |
| LPO | linear-drive pluggable: drop the DSP chip and drive the laser directly — cuts power ~50% | a low-cost, low-power stopgap; "silicon photonics + LPO" is becoming the short-reach standard |
| CPO | co-packaged optics: put the optical engine right next to the switch chip; the signal travels mm, not cm | the long-term "endgame"; volume production not expected before ~2027 — for now, more concept than reality |
| Micro LED | use Micro LEDs as the emitter, "wide-and-slow" across many channels for bandwidth at low cost | emerging, early, but feasible; cost could fall below ¥50 a module |
| OCS | optical circuit switch: tiny MEMS mirrors switch data in the light domain, skipping the electrical conversion | core to Google's TPU clusters; not a module itself, but it reshapes how many modules you need |
Where the light gets stuck (and where the value is)
Follow the supply chain and the money pools exactly where the supply is tightest — the same lesson as the scarcity map:
- Optical chips (EML / CW / pump laser). The core bottleneck. Dominated by a few foreign giants (Lumentum, Coherent), with orders booked into 2028 and a supply gap over 30%.
- The optical engine. The value core — the assembled light-making heart — supplied by firms like T&S Communications (天孚) into module makers such as Fabrinet and Innolight (旭创).
- FAU (fiber array unit). The passive part that aligns the fiber to the optical chip. The precision bar is so high it's still almost entirely built by hand — a quiet, severe choke point.
- Faraday rotators / isolators. The component that stops stray light reflecting back into the laser. Effectively a Coherent monopoly, kept strategically tight, slow to expand.
Where the light goes to work
- Inside the data center (scale-out / scale-up). Linking GPUs to switches — the main use today, iterating fast from 800G toward 1.6T (gigabit to terabit per second).
- Between data centers (DCI). Connecting sites tens to thousands of kilometers apart. Demand is exploding in steps, and the products carry high prices and high margins — the world the Nokia / DCI piece is about.
- Inside the rack (CPO / NPO). The next-generation move Nvidia and Google are pushing — pulling the optics off the faceplate and in next to the chip.
The whole picture — one mind map
OPTICAL MODULE — turns electricity into light, and back again │ ├─ ROLE ............ the comms "nerves" between GPU "brains" │ GPU↔GPU, GPU↔switch; copper hits a wall, light doesn't │ ├─ HOW IT WORKS .... TX: EML chip modulates electricity → light │ → fiber (fast, far, low-loss) → │ RX: PD chip turns light back into electricity │ ├─ TECH ROUTES ..... EML ............. mature, mainstream, wins beyond 2 km │ Silicon photonic. dense / cheap / low-power (needs CW laser) │ LPO ............. drops the DSP, ~50% less power │ CPO ............. optics beside the chip — the endgame (~2027+) │ Micro LED ....... wide-and-slow, ultra-low cost (<¥50) │ OCS ............. switches in the light domain (Google TPU) │ ├─ BOTTLENECKS ..... optical chips (EML / CW / pump) — booked to 2028, >30% gap │ optical engine · FAU (hand-built) · isolator (Coherent) │ └─ WHERE IT'S USED in the DC (800G→1.6T) · DCI (long-haul) · in-rack (CPO/NPO)
The bottom line
In the grand AI story, the GPU is the brain — it supplies the "light of intelligence," the raw compute. The optical module is the nerve that connects those brains: the "light of communication" that lets compute work together and intelligence flow. Without it, the strongest GPU in the world is just an island. The competing routes — EML, silicon photonics, LPO, CPO, Micro LED — are all racing toward the same horizon: pulling the light closer and closer to the chip until the chip simply emits its own. And as ever, the edge isn't the cleverest design — it's owning the part of the light nobody else can make.
An explainer built from optical-module supply-chain research; figures are stated as given in the source, not independently re-derived.