Many channels, one path
Different colors of light share a single waveguide without interfering. Each wavelength is its own channel, so one path can carry dozens of data streams at once.
Long-term research · ARK PhotonicsComputing with
Computing with
light.
Today's chips push electrons through copper. We're researching chips that compute with photons instead, guiding light through silicon to make AI faster and far more energy-efficient.
Why light
The bottleneck is moving data.
As AI models grow, moving data across a chip and between chips often costs more energy than the math itself. Light carries information differently.
Less energy per bit
A copper wire spends energy charging and discharging on every bit, and longer wires cost more. Light in a waveguide doesn't, which is why AI data centers are moving their links from copper to optics.
Math in flight
A mesh of interferometers multiplies a vector by a matrix as light passes through it. The multiply finishes in the time light takes to cross the chip, a fraction of a nanosecond.
How it works
A neural-network layer, built from light.
Most of the work in AI is matrix multiplication. A photonic chip does it with interference instead of transistors.
- 01
Encode
A laser is split into channels. Modulators write each input number onto the light in its channel.
- 02
Compute
A mesh of Mach–Zehnder interferometers mixes the channels. Tiny heaters set each one's phase, and those phases are the network's weights. The light that comes out is the matrix product.
- 03
Read out
Photodetectors turn the light back into electrical signals. Electronics apply the nonlinear steps and memory, then feed the next layer.
One platform, two frontiers
Where data centers and quantum meet.
Photonic Chips brings our data center and quantum computing work under one roof, because light is where both are heading.
Light already moves the data.
AI clusters are limited as much by moving data as by computing it, so the industry is moving optics right up against the chips. Nvidia says its co-packaged optics make network switches about 3.5× more power-efficient. Photonic computing asks the next question: what if the math stayed in light too?
Photons make natural qubits.
Photons barely interact with their surroundings, which makes them one of the most robust carriers of quantum information. The same parts (waveguides, beam splitters, phase shifters, and detectors) are the building blocks of photonic quantum computers, one of the leading approaches in the field.
Open problems
Why it isn't solved yet.
Photonic computing has been demonstrated in labs. Beating electronics at scale means solving these.
Heat crosstalk
Most phase shifters are tiny heaters. Heat from one leaks into its neighbors and nudges their settings, so dense meshes drift.
Fabrication variation
Nanometer-scale differences in waveguide width shift every interferometer slightly, so each chip has to be measured and calibrated.
Loss and noise
Every component loses a little light. Losses compound through a deep mesh and limit how precise the analog answer can be.
The electronic boundary
Light is great at linear math, but activation functions and memory still live in electronics, and converting between the two costs energy and time.
Roadmap
The research path.
A long-term program, built in stages that each produce something we can measure.
- Phase 01
Physics and simulation
Model waveguides, interferometers, and full meshes from first principles, then simulate photonic neural-network layers against electronic baselines.
- Phase 02
First silicon
Design test chips, fabricate them on shared multi-project wafer runs at silicon-photonics foundries, and measure loss, crosstalk, and calibration.
- Phase 03
Photonic accelerator
Pair a photonic matrix engine with electronic control to run real AI inference, and measure speed and energy per operation against today's chips.
Collaborate
Build it with us.
ARK Photonics is a long-term research program. If you work in integrated photonics, optical computing, or quantum optics, or want to, we'd like to hear from you.