Explainer
What Is Photonic Computing? Light for Logic and Links
Photonic computing uses light, made of particles called photons, in place of electric current for part of the work of a computer. The phrase covers two quite different ideas. One uses light to do calculations. The other uses light to move data between chips. They are at very different stages of maturity, and mixing them up leads to confusion.
Two jobs for light
The first job is computing. Here light passes through a circuit that carries out an operation, usually matrix multiplication, the core arithmetic of AI.
The second job is communication. Here light carries data from one chip to another, replacing electrical wires. The computing itself stays electronic.
The second job is the one reaching products today. The first is mostly in start-up hardware, pilots and research.
Computing with light
Light has a useful property. Waves of light can be combined, and the way they add up or cancel is a form of arithmetic. A photonic chip uses this to multiply and add numbers as light travels through it.
One design uses a mesh of interferometers on a silicon chip. An interferometer splits a beam, changes the phase of the parts, and recombines them. A network of them can implement a matrix. Another design sends light through free space and lenses or modulators, so that a whole grid of values is processed together.
The potential benefits are low latency, because light passes through the circuit in a very short time, and low energy for the multiplication step itself. Many wavelengths of light can also travel through the same path at once.
The difficulties are real. Most photonic computing is analog. Analog means the values are continuous quantities, not exact digital numbers, so noise and drift limit precision. Light is hard to make interact with other light, so the nonlinear steps in a neural network are usually done electronically. Every value must be converted from digital to optical and back, and those conversions can consume much of the energy that was saved.
For these reasons, photonic accelerators are best viewed as early-stage devices aimed at specific workloads, usually inference. Companies working here include Lightelligence, Lumai and Q.ANT. Lightmatter began with a photonic computing chip and now also sells optical interconnect.
Moving data with light
Inside an AI data centre, thousands of chips must exchange data constantly. Copper wires lose signal quickly as speeds rise. Beyond a short distance, a copper link needs extra circuitry that uses power and adds delay. Optical fibre loses very little signal over the distances inside a data centre.
An optical link has a laser or other light source, a modulator that encodes data onto the light, a fibre or waveguide, and a photodetector that turns the light back into an electrical signal. Silicon photonics builds modulators, waveguides and detectors on silicon using chip-factory methods. Silicon is a poor light emitter, so the laser usually comes from another material, such as indium phosphide, and is often kept separate from the chip.
Today most optical links use pluggable transceivers, modules that sit at the edge of a switch. The electrical signal must still travel from the chip across the board to the module. Co-packaged optics shortens that path by placing the optical engine in the same package as the switch chip or processor. The aim is lower power per bit and higher bandwidth density.
Optical I/O goes a step further. It puts optical chiplets next to a processor so that the processor itself talks in light. Ayar Labs and Lightmatter have worked on this, and Broadcom and NVIDIA have announced co-packaged optics for network switches.
What to be sceptical about
Claims of large efficiency gains deserve a close look. Check whether the figure covers the whole system, including lasers, converters and control electronics, or only the optical core.
Check the precision. A result at four or eight bits is not comparable with a result at 16 bits.
Check the workload. A photonic chip that excels at one matrix size may not help with a full model.
For interconnect, ask about reliability. A data centre has many thousands of links, and lasers are a known source of failures. Serviceability matters as much as speed.
Why it matters
The pressure behind both ideas is the same. Moving data now costs a large share of the energy and time in AI systems, and electrical links are reaching physical limits. Light offers a way around that.
The nearer-term impact is likely to come from optical interconnect, where the technology is a replacement for an existing component. Photonic computing may find a place in particular workloads, but it has yet to show it can compete broadly with digital chips.
Companies mentioned
- Lightmatter Lightmatter develops photonic interconnect technology for AI data centers, using light rather than electrical signals to move data between chips. Its Passage product line includes the M1000, a photonic interposer, and the L200 and L200X, co-packaged optics engines. The M1000 is built on the GlobalFoundries Fotonix silicon photonics platform, and the L200 was developed with Alphawave Semi. Lightmatter is a privately held company whose products target the bandwidth limits that arise when connecting large numbers of GPUs and other accelerators.
- Celestial AI Celestial AI developed the Photonic Fabric, an optical interconnect that moves data between chips and memory using light rather than copper wires. The aim is to give AI processors high bandwidth and low latency over longer distances, with less power loss. Marvell Technology announced that it had completed its acquisition of the company in early 2026, and Celestial AI's teams and technology now sit within Marvell's Data Center Group. The company was previously an independent startup based in Santa Clara, California.
- Ayar Labs Ayar Labs makes optical input and output for chips. Its TeraPHY chiplet is a small silicon photonics die that sits in the same package as a processor and sends data as light, and its SuperNova is a separate multi-wavelength laser source that feeds light to those chiplets over standard single-mode fibre. This replaces electrical links, which lose energy and range at high speeds. The company was founded in 2015 and describes its chiplets as compatible with the UCIe chiplet standard. Directories place its headquarters in San Jose, California.
- Lightelligence Lightelligence develops photonic products for AI systems, with two segments, optical interconnect and optical computing. Optical interconnect, which links chips using light, produces its current revenue, while optical computing, which does calculations with light, is a longer-term effort. Yichen Shen, who holds a physics PhD from MIT, founded the company in 2017. It listed on the Hong Kong stock exchange on 28 April 2026 under the code 1879, and filings name Shanghai Xizhi Technology as the operating company.
- Lumai Lumai is building an optical AI accelerator that does matrix multiplication, the core arithmetic of neural networks, with beams of light travelling through three-dimensional free space instead of etched waveguides on a chip. The company argues this avoids scaling limits of integrated photonics. It spun out of the University of Oxford in 2022 and is based in Oxford, England. In April 2025 it announced more than 10 million dollars of venture funding in a round led by Constructor Capital. Performance claims, such as large efficiency gains over silicon, come from the company and are unverified.
- Q.ANT Stuttgart photonic computing company spun out of TRUMPF that sells the Native Processing Server, a plug-in photonic co-processor for AI and HPC. It raised a 62 million euro Series A in July 2025 (expanded to $80 million), unveiled its second-generation NPU 2 in November 2025 and agreed a four-year partnership with Jülich Supercomputing Centre.
- Ranovus Ranovus designs optical engines that move data between chips in data centers using light. Its Odin engines combine a quantum dot laser, ring-shaped modulators that encode data onto the light, and photodetectors on a silicon photonics chip. The engines are aimed at pluggable optical transceivers and at co-packaged optics, where the optics sit beside the switch or processor chip to reduce power use. Ranovus was founded in 2012, works from Ottawa, and has named IBM, TE Connectivity and SENKO as packaging partners.
- OpenLight OpenLight, based in Santa Barbara, California, offers a silicon photonics platform with lasers built into the chip. Silicon cannot easily emit light, so most designs attach separate laser parts. OpenLight instead bonds indium phosphide material onto silicon wafers, allowing lasers, modulators, amplifiers and detectors on one chip. Customers design with its process design kit, and it is validated at the foundry Tower Semiconductor. The technology traces to Aurrion, a UC Santa Barbara spinout that Juniper Networks bought in 2016. OpenLight targets 800G and 1.6T links for AI data centers.
- Broadcom Broadcom is a semiconductor and infrastructure software company. Its chip lines include custom AI accelerators it calls XPUs, Ethernet switching and routing silicon, and Ethernet network adapters, which are used in AI and cloud data centers. It also owns infrastructure software, notably VMware Cloud Foundation for private cloud. Broadcom is mostly fabless and names TSMC as a key manufacturing partner. It works with large cloud customers on custom chips and sells networking silicon to system builders. It is headquartered in Palo Alto, California.
- Optalysys Optalysys develops photonic chips that use light to carry out Fourier transforms, a core operation in fully homomorphic encryption, which allows data to be processed while it remains encrypted. The aim is to make encrypted computing fast enough for practical use, where conventional processors are far too slow. The company is based in Leeds, England, and was founded in 2013 by Nick New and Robert Todd. Its backers include Northern Gritstone, and the UK government's National Security Strategic Investment Fund took part in a later round.
Company details come from the Compute Compass directory and may be out of date. This is not investment advice.