● dense/○ explainer
aileena·machina
investing · 2026 · may

The optical module is the nerve of AI.

Turns electricity into light, sends it down a fiber, turns it back. The whole AI build-out is gated by who can make the light.

— tldr —

GPUs are islands without the optical module. Six tech routes race to make the light. The bottleneck is the optical chip — booked to 2028, gap over 30%.

A GPU on its own is an island. To make thousands of them act as one machine, every chip has to talk — faster than copper wire allows. So the signal leaves the electrical world and travels as light.

↓how the light gets made

optical module · cross-sectionFAUfiber arrayPICsilicon photonicsEICelectronic ICEML
1
2
3
4
5

cross-section · where photons meet electrons

④ FAUfibre coupling —
still built by hand
② PICsilicon photonics —
the waveguide highway
③ EICdrives the laser —
Marvell vs Broadcom
① EMLlaser source —
only 2 suppliers
⑤ power floor≥300 mW EML for 1.6T+
the optical engine ↘
photons → electrons

Transmit. An EML chip modulates electricity into a beam of light carrying the same information.

in plain wordsEML = electro-absorption modulated laser. makes the light.

Travel. The light runs down an optical fiber — fast, far, very little loss.

analogya lighthouse blinking morse,
but unimaginably faster.

Receive. A PD (photodetector) chip at the far end turns the light back into electricity.

in plain wordsPD = photodetector. reads the light.
↓ six routes, racing ↓

→the routes

EML. Mature, reliable, mainstream beyond 2 km. The chip supply is the bottleneck.

route 1today’s default.
chip is short.

Silicon photonics. Optics built in silicon — dense, cheap, low-power; needs an external CW laser. 2025 was called the “year of silicon photonics.”

route 2mainstream at
500 m / 2 km.

LPO. Linear-drive pluggable. Drop the DSP, drive the laser directly — cuts power roughly 50%.

route 3silicon photonics + LPO =
short-reach standard.

CPO. Co-packaged optics. Put the optical engine right next to the switch chip — the signal travels mm, not cm.

route 4the endgame.
volume not before ~2027.

Micro LED. Use Micro LEDs as the emitter, wide-and-slow across many channels. Cost could fall below ¥50 a module.

route 5cheap, emerging,
feasible.

OCS. Optical circuit switch — tiny MEMS mirrors switch data in the light domain, skipping the electrical conversion. Core to Google’s TPU clusters.

route 6not a module —
reshapes how many you need.
↓ where the light gets stuck ↓

⚡the bottlenecks

Optical chips (EML / CW / pump). Dominated by Lumentum & Coherent. Orders booked into 2028. Supply gap over 30%.

choke 1the core bottleneck.
two foreign giants.

The optical engine. The assembled light-making heart. T&S Communications (天孚) supplies it into module makers like Fabrinet and Innolight (旭创).

choke 2the value core.
T&S is the supplier.

FAU (fiber array unit). Aligns fiber to optical chip. Precision is so high it’s still built by hand.

choke 3a quiet,
severe choke point.

Faraday rotators / isolators. Stop stray light reflecting back into the laser. Effectively a Coherent monopoly.

choke 4kept tight on
purpose.

⌖where it goes to work

Inside the data center. GPU to switch — iterating fast from 800G toward 1.6T.

todaythe main use.

Between data centers (DCI). Tens to thousands of km. Demand exploding, high prices, high margins.

long-haulthe Nokia/DCI
world.

Inside the rack (CPO / NPO). Pull the optics off the faceplate and in next to the chip.

tomorrowNvidia & Google
are pushing it.

※the bottom line

The GPU is the brain. The optical module is the nerve that lets the brains talk. All routes race toward one horizon: pull the light closer and closer to the chip until the chip simply emits its own. The edge isn’t the cleverest design — it’s owning the part of the light nobody else can make.

← dispatch