Photonic circuit schematic
SCROLL — LIGHT THE FABRIC ↓
01 / THE WALL

Electrical I/O stopped scaling. Models didn't.

Every hop in today's AI cluster pays for electrical signaling: serialize, drive copper, deserialize, repeat. Above 200G per lane, the energy and area cost of electrical signaling grows faster than the bandwidth it delivers. Co-packaged optics removes the conversion entirely — data leaves the compute die as light.

03 / THE SILICON

Three parts. One optical fabric.

A training processor, a photonic switch, and an inference engine — designed together so light never has to become electricity between them.

AI20P-TAI20SAI20P-I
AI20P-T photonic training processor layout
AI20S active photonic switch fabric layout
AI20P-I photonic inference engine layout
AI20P-T

Train through the bandwidth wall.

PHOTONIC TRAINING PROCESSOR

A systolic array fed directly by optics — activations and gradients leave the die as light, and training domains span racks without an electrical hop.

AI20S

Route in light.

ACTIVE PHOTONIC SWITCH FABRIC

A photonic crossbar between pods. Wavelengths are switched in the optical domain — under a microsecond end to end, no top-of-rack layer at all.

AI20P-I

Inference at optical scale.

PHOTONIC INFERENCE ENGINE

4–5× lower power per token. Co-packaged HBM holds the KV cache next to the compute; light carries everything that has to travel.

FOLLOW THE MODEL →
04 / THE SYSTEM

From package to pod without leaving the optical domain.

Inside the package, HBM and CMEM SRAM keep data next to the systolic array. Between packages, fibers carry it across the node. Between nodes, the AI20S switch routes it — a flat optical fabric with no top-of-rack layer.

AI20P-T VS COMPETITION

PEAK CHIP POWER4–5× LOWER
Competition · ~1000–1200 W
AI20P-T · 180–250 W
PEAK COMPUTE (FP16)1.5× HIGHER
Competition · ~180 TFLOPS
AI20P-T · 275 TFLOPS
RACK-BOUNDARY O-E-O CONVERSIONBYPASSED
NVLink / PCIe · multi-ns
CPO fabric · sub-µs, end to end
Optical fabric circuit schematic

The fabric is light.

Silicon photonics for exascale AI.
01 / ONE OPTICAL DOMAIN

Three parts, designed as one.

AI20P-T trains, AI20S routes, AI20P-I serves — and light never has to become electricity between them.

OPEN THE NODE →
Optical AI server node schematic, top viewOptical AI server node hardware, top view

Open the node.

OPTICAL AI SERVER / TOP VIEW
01 / THE SERVER

No electrical switch plane.

Fibers run from every package straight to the optical fabric — the schematic is the wiring diagram.

02 / THE HARDWARE

From schematic to steel.

The same node, built: packages, fiber bundles and fans in a standard rack unit.

SCALE OUT →
SCALE OUT

One node. Then ten thousand.

EXASCALE ARCHITECTURE
01 / TOPOLOGY

A 3D torus of light.

Every rack connects to the optical fabric directly — no top-of-rack switches, no oversubscription, no cable clutter.

02 / DISTANCE

The fabric doesn't care about distance.

In glass, ten metres costs the same as ten millimetres. The training domain is the whole floor — sub-microsecond, end to end.

03 / ORBITAL

Where light goes next.

Photons carry no charge — radiation doesn't flip them. Low power means less radiator mass per FLOP. The same fabric that scales a datacenter floor is built for orbit.

TALK TO US →
05 / COMPANY

Light becomes infrastructure.

We work with foundries, hyperscalers and systems partners on co-packaged optics, photonic switching and optical compute. If that's your problem space, talk to us.

LOCATION
4, Bellary Rd, 12, K M, Yashoda Nagar, Jakkur, Bengaluru, Karnataka 560064
Nagendra
Dr. Nagendra
PROMOTER
Manjunath
Dr. Manjunath
DIRECTOR
Vamsi Krishnan
Vamsi Krishnan
DIRECTOR