Electronics moves electrons. Photonics moves photons — and light is fast, hot-free in the channel, and can carry many colours at once. Here's how it works, from waveguides to optical matrix multiply.
Photonics is the science and engineering of generating, guiding, modulating and detecting light to carry and process information — the optical counterpart to electronics. Where electronics pushes electrons through wires and transistors, photonics sends photons through fibres, waveguides and optical devices.
It already runs the modern world quietly: every byte that crosses the internet's backbone travels as pulses of light in optical fibre. But photonics is now pushing into the chip itself — for moving data between processors, and even for doing computation directly with light.
The flip side: light is wonderful at moving and linearly transforming information, but it has no easy equivalent of the transistor — photons don't naturally switch each other the way one electronic signal gates another. That single fact shapes everything about where photonics wins (communication, linear math) and where electronics still rules (logic, memory, control).
Just as electronics has resistors, transistors and wires, photonics has a small set of building blocks:
Generates the photons. On-chip systems usually use a laser for a clean, single-wavelength beam to carry data.
The "wire" for light — a tiny channel (often silicon) that guides the beam by total internal reflection. The fundamental on-chip routing element.
Encodes electrical data onto the light by changing its intensity or phase — turning bits into optical pulses. The bridge from the electronic to the optical world.
Converts light back into an electrical signal at the receiving end — the reverse of the modulator.
Split, combine and filter light. Ring resonators pick out specific wavelengths — essential for WDM and compact modulators.
String these together and you can build a complete optical link or even a computing fabric on a chip.
You can't just send a laser across a chip in open air. A waveguide traps light in a high-refractive-index core (e.g. silicon) surrounded by lower-index cladding (e.g. silicon dioxide). Light hitting the boundary at a shallow angle is totally internally reflected, so it zig-zags along the core, staying confined — exactly how an optical fibre works, shrunk onto a chip.
How do you put data on light, or build a tunable optical operation? With interference. The workhorse is the Mach-Zehnder Interferometer (MZI):
So a voltage on the phase shifter controls the output brightness — that's a modulator (encoding bits as light). The same device, configured carefully, also applies a precise multiply/weight to a beam — which is what makes MZIs the building block of optical computing.
Here's photonics' superpower for communication: different wavelengths (colours) of light don't interfere. So you can send many independent data streams down one fibre at once, each on its own colour, and separate them at the far end. This is Wavelength-Division Multiplexing (WDM), and it's why a single hair-thin fibre can carry terabits per second.
For decades, photonic devices were discrete and expensive. The breakthrough is silicon photonics: building waveguides, modulators, detectors and couplers in silicon, using the same CMOS fab processes as computer chips. That means optical functions can be integrated, mass-produced and placed right next to electronics at low cost.
The one stubborn piece is the laser — silicon is a poor light emitter, so lasers are often made from other materials (like indium phosphide) and bonded on, or attached externally. Everything else — guiding, modulating, detecting — silicon does well. Silicon photonics is what turned optics from lab apparatus into a manufacturable chip technology.
Photonics' biggest real-world impact right now isn't computing — it's communication between chips. Inside data centres, copper links are hitting a wall: they lose signal over distance, burn energy, and can't carry enough bandwidth for AI clusters of thousands of accelerators.
Optical links solve all three: low loss, high energy-efficiency per bit over distance, and huge bandwidth via WDM. The frontier is co-packaged optics — putting the optical engine right beside the processor so data converts to light the instant it leaves the chip, instead of travelling far as electricity first. For large AI systems, where chip-to-chip bandwidth is a critical bottleneck (see the memory wall), optical interconnect is rapidly becoming essential.
The boldest goal: do the math itself with light. Neural networks are dominated by matrix multiply (see What Is an AI Chip?), and light is naturally good at linear operations:
Arrange a mesh of MZIs and you can implement an entire matrix: encode the input vector as the amplitude/phase of input beams, send them through the mesh, and the output light intensities are the matrix-vector product — computed at the speed of light, in the analog domain, with very little energy spent in the optical path. It's the optical cousin of the systolic array and the memristor crossbar: do the multiply-accumulate in the physics, not in shuttled data.
Optical compute is analog — limited precision, noise and calibration drift — and you still pay energy converting between electronic and optical at the edges (modulators, detectors, ADC/DAC). There's no good optical memory, and no true optical transistor, so light handles the linear math while electronics still does the nonlinear activations, control and storage. It's powerful for the matrix part, not a whole computer.
AI made two things suddenly precious: moving oceans of data between chips, and multiplying giant matrices. Photonics targets both — optical interconnect for the data movement, optical meshes for the matrix math. That's why it's one of the most-watched long-term bets in AI hardware.
Realistically: optical interconnect is here and growing fast (co-packaged optics in next-gen AI systems), while optical computing is promising but still maturing — strongest as an accelerator for specific linear workloads rather than a general replacement for digital chips.
Beyond chips, photonics quietly powers a huge range of technology:
| Field | Photonics inside |
|---|---|
| Communications | Fibre-optic internet backbone, undersea cables, data-centre links |
| Sensing & LiDAR | Self-driving cars, industrial & environmental sensors |
| Medical | Endoscopy, optical coherence tomography, laser surgery |
| Displays & imaging | Cameras, lasers, AR/VR optics |
| Manufacturing | Laser cutting, welding, EUV lithography light sources |
| Quantum | Photonic qubits & quantum communication |
Photonics carries and transforms information with light using waveguides, modulators (often MZIs), detectors and many wavelengths at once (WDM). It already dominates communication, and silicon photonics is bringing optical interconnect onto the chip to beat the copper bottleneck. Computing with light is real for linear math (matrix multiply at light speed) but limited by analog precision, the lack of an optical transistor/memory, and conversion costs — so it augments electronics rather than replacing it.
The science of generating, guiding, modulating and detecting light to carry and process information — the optical counterpart to electronics.
Building photonic devices (waveguides, modulators, detectors) in silicon with CMOS fab processes, so optics can be integrated cheaply beside electronics.
A mesh of Mach-Zehnder interferometers scales beams (multiply) and interferes them (add); the output intensities are the matrix-vector product, at the speed of light.
It attacks AI's two costs — moving data between chips (optical interconnect) and multiplying matrices (optical meshes). Interconnect is shipping; compute is maturing.
Related: What Is an AI Chip? · Beyond HBM: Memory Alternatives · Systolic Array Lab · Transistor Evolution