How Do Photonic Quantum Computers Work?

How It WorksHow Do Photonic Quantum Computers Work? No frozen chips, no trapped atoms, no artificial molecules. One family of quantum machines computes with light itself, and it works in ways the rest of the industry finds slightly alien.Key TakeawaysThe qubit is a single particle of light. Information rides on a photon’s path, its timing or its polarisation, and, unlike the platforms it competes with, there is nothing to cool to millikelvin and nothing to trap.The program is written in glass. A programmable maze of waveguides and phase shifters performs the operations, so changing the computation means changing the settings on the chip rather than building a new chip.Photons do not interact, so measurement does the work. Two-photon gates are made to happen by interference and detection rather than by force, which is why measurement sits at the centre of every photonic design.Loss is the enemy, not noise. A photon that vanishes takes its information with it, and no amount of shielding brings it back, so every component is judged on how little light it wastes.Squeezed light is the other road. Continuous-variable machines encode in the wave properties of a light field, its amplitude and its phase, rather than in single photons, and both approaches are being built commercially.The detectors are as hard as the chips. Superconducting nanowire detectors that resolve single photons make the whole architecture possible, and they account for a large part of both the cost and the cryogenics. The question of how photonic quantum computers work has a pleasingly literal answer, because almost every part of the machine does exactly what its name says it does. A source fires single particles of light into channels of glass. A programmable maze of splitters and phase shifters interferes them, superconducting counters at the far end record which channels the light came out of, and the pattern of those clicks is the output. Many runs of the machine build that pattern up. The remarkable part is everything it refuses to need. The photons never require the deep freeze that superconducting processors live in, and they barely notice the electrical noise that plagues every other platform. The chips that guide them come off the same 300 millimetre production lines that make ordinary networking equipment, and photonic firms are betting that the scaling problem is really a manufacturing problem. Light lets them borrow the semiconductor industry wholesale, since a photonic circuit is patterned by the same lithography that makes ordinary optical chips. The bet comes with strange costs. Photons cannot be made to interact directly, so the logic gates other platforms take for granted become games of chance here, and they cannot be stored either. The whole computer must be choreographed in flight. Everything else in the machine follows from those two facts, from the qubit itself right through to the companies racing to build one worth paying for. In this guideOne photon in two channels is the qubitSingle-photon sources still fire at randomThe phase shifter settings are the softwareTwo-photon gates work one time in nineMeasurement does the computingSqueezed light is the second roadClassical algorithms caught the first advantage claimsLight buys manufacturing and pays in lossSeven companies carry the approachNo photonic logical qubit exists yet Loss is the enemy, not noise Superconducting nanowire detectors make it possible One photon in two channels is the qubit The workhorse encoding is called dual-rail, and it is exactly as simple as it sounds, because it needs one photon and two parallel waveguides, the light-carrying channels etched into a chip. Declare that the photon in the top channel means 0. The photon in the bottom channel means 1. A beam splitter can then place that single photon into a genuine superposition of both channels at once, which is everything a qubit needs to be. The same idea wears other clothes when convenient. Polarisation encoding uses horizontal and vertical light in a single channel, which is really the same dual-rail trick played out in a different pair of modes. Time-bin encoding puts the photon into an early or late arrival slot, a form that survives a long run down a stretch of fibre far better. Vendors pick the encoding that suits their hardware, so PsiQuantum and Quandela use path encoding on chips while ORCA Computing runs time bins through loops of fibre. One photon shared between two channels is a complete qubit. A phase shifter between two splitters makes a one-qubit gate. Diagram by Quantum Zeitgeist. Single-photon sources still fire at random Everything downstream depends on sources that emit exactly one photon, on cue. That is much harder than it sounds. The traditional method is spontaneous parametric down-conversion, which shines a laser through a special crystal that occasionally splits one pump photon into a pair. Detecting one member of the pair announces that its twin is on its way, and that heralding is the good news, but the bad news is the word occasionally. The process is random, and turning up the laser to make pairs more often also makes unwanted double pairs that corrupt the computation. The two answers are brute redundancy and better atoms. Multiplexing runs many probabilistic sources in parallel and switches whichever one fired into the circuit, trading hardware for reliability. Deterministic sources replace the crystal with a single quantum dot, an artificial atom grown in a semiconductor pillar that emits one photon nearly every time it is asked. The French firm Quandela builds its machines around dots, and publishes an indistinguishability above 95 per cent, a measure of how nearly identical the emitted photons are to each other. Either way, the source sets the ceiling downstream. The phase shifter settings are the software Once photons exist, the computer itself is a mesh of interference. A result from 1994, refined in 2016 into the layout the industry now uses, showed that any transformation of light across N channels can be built from a grid of two simple parts. Those parts are beam splitters and phase shifters. The splitters mix neighbouring channels, and the shifters delay one channel relative to another. Program the phase shifter settings and you have programmed the machine, since on a photonic chip the settings of the maze are the software and nothing else is. The parts themselves are humble. A programmable splitter is a Mach-Zehnder interferometer, which is two fixed couplers wrapped around a phase shifter. The workaday phase shifter is a microscopic heater that expands the glass beneath it, a method that is cheap to build and painfully slow to switch. The next generation switches with electric fields instead, in materials such as thin-film lithium niobate, which respond far faster than heat can. QCi operates a foundry in Tempe, Arizona dedicated to that material, and PsiQuantum has demonstrated barium titanate phase shifters that it makes on 300 millimetre wafers in California. Two-photon gates work one time in nine Single-qubit tricks are not enough, and here photonics meets its deepest quirk. Photons ignore each other, which is precisely why they carry your internet traffic so cleanly through crowded fibres. A computer, though, needs its bits to interact. The escape hatch was found in 1987 at the University of Rochester by Hong, Ou and Mandel. They sent two identical photons into opposite sides of a 50:50 beam splitter and found that the pair always leaves together through the same port, never one through each. The effect is pure interference with no classical counterpart, and the depth of that never is now the standard measure of whether two photons are identical enough to compute with. This bunching is the raw ingredient of photonic logic, but it comes at a price nature refuses to waive. Entangling gates built from linear optics succeed only some of the time, and the classic postselected CNOT, where the failed runs are simply thrown away afterwards, works one attempt in nine. The first experimental version was demonstrated in 2003.That is nevertheless enough. Knill, Laflamme and Milburn proved in Nature in January 2001 that single photons, linear optics, detectors and enough cleverness about spare helper photons and teleportation add up to universal quantum computing. Every photonic architecture since descends from that paper, and the cleverness it assumed turned out to be the whole industry. Light barely decoheres but barely interacts, which is why an entangling gate succeeds only one attempt in nine. Later architectures absorb those failures rather than retrying them. Diagram by Quantum Zeitgeist. Measurement does the computing Because gates are unreliable, modern photonic architectures turned the circuit model inside out. In measurement-based quantum computing, proposed in 2001, you first weave a large entangled web of qubits, called a cluster state. You then compute purely by measuring qubits one at a time, steering later measurements with earlier results, and the style suits light because measuring photons is what photonics does superbly. All the probabilistic misbehaviour is pushed into preparing the web, before the computation proper begins. PsiQuantum’s version, published in 2023 as fusion-based quantum computing, breaks the web into small identical resource states. Two-photon measurements called fusions stitch those pieces together. Single fusions fail routinely even in theory, and the design simply plans for it, tolerating roughly ten per cent photon loss per fusion in its published form. Photonics enjoys a quiet structural gift here. When a photon goes missing the detectors know exactly where and when, and such flagged erasures are far easier for photonic error correction to repair than the silent errors matter qubits suffer. The awkward corollary is that photons cannot wait around, so architectures park them in kilometre-scale loops of fibre, buying microseconds of patience at the price of yet more loss. The photons never sit still and they never warm up, because the computation happens to them in flight. Squeezed light is the second road That is the first of the two roads photonic quantum computers can take. Everything so far treats light as particles. Xanadu, the Toronto company that reached the public markets in March 2026, computes with light’s wave nature instead, and describes that listing as the first by a pure-play photonic firm. Its machines run on squeezed light, meaning laser fields whose quantum noise has been compressed below the vacuum floor in one direction at the cost of the other. Information lives in continuous wave properties rather than in a single photon’s position. A mathematical construction from 2001 called the GKP code folds those continuous variables back into a proper error-protected qubit, encoded as a comb of peaks that small disturbances visibly shift. The choice looks eccentric until you see what it buys. Logic on GKP qubits needs no more than the splitters, squeezers and standard detectors that photonic chips already do well. The encoding also detects the small drifts that photon loss causes, so error correction is built into the qubit itself rather than bolted on around it. Xanadu put a GKP qubit on an integrated chip in June 2025, in what the company describes as the first error-resistant example anyone has built. Its Aurora system, unveiled in January 2025 and written up in Nature that month, networked 35 photonic chips into one machine that Xanadu says spanned four server racks and 13 kilometres of fibre. Aurora ran 12 physical qubit modes per clock cycle rather than logical qubits. The company is open that driving down optical loss is the remaining battle. Classical algorithms caught the first advantage claims The headlines photonic quantum computers are famous for come from a narrower contest called boson sampling, proposed in 2011, in which the machine merely draws samples from the output of a large interferometer. For a classical computer that is catastrophically expensive. China’s Jiuzhang device claimed exactly such a quantum advantage in December 2020, with 76 detected photons and a claimed hundred-trillion-fold speedup. Xanadu’s Borealis followed in June 2022, and was the first claimed-advantage machine anyone could use over the cloud. The scorekeeping since then is a lesson in why this field rewards patience. Classical algorithms improved relentlessly, cutting Jiuzhang’s billions of years to days by 2022. By 2024 a tensor-network method, which squeezes the quantum state into a small stack of linked tables of numbers, could mimic most of the early experiments outright. That took the shine off the first-generation claims. The Chinese team answered with Jiuzhang 4.0, reported in Nature on 13 May 2026, and that machine injected 1,024 squeezed states into an 8,176-mode processor. This time the team benchmarked against the spoofing algorithms that had undone its predecessors. Both speedups are the teams’ own claims. Samplers are single-purpose demonstrations, not universal computers. Light buys manufacturing and pays in loss The advantages here are real and structural. Photons do not feel stray electric or magnetic fields, so they barely decohere in flight, and they travel through ordinary fibre at telecom wavelengths, the same glass that carries internet traffic. That lets photonic machines grow the way data centres do, room by room and rack by rack, instead of squeezing ever more qubits into one refrigerator. The chips are made on standard 300 millimetre semiconductor lines, with PsiQuantum’s Omega modules coming out of GlobalFoundries’ Fab 8 in New York. Even the cooling counts in light’s favour. The cryogenics are not optional, since single-photon detectors want a few kelvin, but that is a cheap industrial temperature compared with the near-absolute-zero dilution refrigerators that superconducting chips demand. The difference is why one photonic cooling cabinet can host hundreds of chips instead of one or two. The bill arrives elsewhere. Sources and entangling operations are probabilistic, so a fault-tolerant photonic machine needs staggering component counts and fast switching to harvest its lucky events. There is no good photonic memory either, so timing is engineered in glass and fibre rather than granted by a stable atom. PsiQuantum has put a figure on that bill in its own Nature paper. It holds that fault-tolerant operation implies on the order of millions of physical qubits, and that the components behind them will have to be manufactured by the millions. Loss is the ever-present tax, because an amplifier that could top up a quantum signal is forbidden outright by the laws the machine itself runs on. As of mid-2026 no photonic platform has demonstrated a peer-reviewed logical qubit, while superconducting and trapped-ion rivals populate the logical qubit leaderboard. Seven companies carry the approach A concentrated field of specialists is carrying the approach, and our guide to the top photonic quantum computing companies profiles the full roster. Seven firms lead it, each with a different answer to the source problem and a different bet on scaling. PsiQuantumFusion-based architecture manufactured at GlobalFoundries’ Fab 8 in New York; raised $1B in September 2025 at a $7B valuation, with utility-scale sites underway in Chicago and Brisbane. PsiQuantum named Victor Peng, AMD’s president from February 2023 until his retirement in August 2024, interim chief executive on 10 February 2026, and our report on its deployment plans has the context. XanaduThe squeezed-light and GKP camp, listed on Nasdaq as XNDU in March 2026 after its SPAC merger, with a dual listing in Toronto. Aurora demonstrated modular, networked photonic computing, and the company says it is targeting up to 1,000 logical qubits by 2029. QuandelaFrance’s photonic champion, built on deterministic quantum-dot single-photon sources. Its 12-qubit Belenos launched in May 2025, its Lucy machine runs at CEA’s TGCC centre for EuroHPC, and the company says its roadmap promises fault-tolerant machines by 2028 without naming a logical-qubit count. ORCA ComputingLondon-based and defiantly fibre-native, running time-bin qubits through fibre loops with a rubidium vapour quantum memory. A PT-2 serves the UK NQCC and two earlier PT-1 systems went to Poland’s PSNC, and ORCA says June 2026 brought its first deployment inside a private enterprise, with Toyota Tsusho in Japan. QuiX QuantumDutch maker of the silicon-nitride processors several rivals build on. It is now under a German Aerospace Center contract to deliver fully universal photonic machines, and QuiX says that means an 8-qubit system first and a 64-qubit successor after it. AegiqSheffield startup pairing quantum-dot sources with silicon-nitride circuits; its Artemis testbed is one of seven systems hosted by the UK’s National Quantum Computing Centre.

Quantum Computing IncOperates a thin-film lithium niobate photonic foundry in Tempe, Arizona alongside its Dirac entropy-computing line; physicist Yuping Huang was appointed chief executive in December 2025. Our QUBT profile covers the company and its stock in detail. One neighbour is routinely miscounted. Photonic Inc of British Columbia computes with silicon spin qubits, and uses telecom photons only as the links between modules. That is a networking strategy rather than a photonic processor, so the company belongs to a different family despite the name. No photonic logical qubit exists yet The milestone the whole field is chasing is the one it conspicuously lacks. Matter-qubit platforms crossed into error-corrected territory between 2024 and 2026, while photonics has yet to publish a logical qubit. The roadmaps now converge on closing that gap. Quandela aims for dozens of logical qubits by 2028, Xanadu targets its fault-tolerant data centre around 2029, and PsiQuantum is betting that its utility-scale sites will leapfrog intermediate demonstrations entirely. Those dates deserve scepticism. Quantum roadmaps have a long history of slipping. Yet the argument for light will not go away. Photonic quantum computers are unusual in the quantum family, because their errors mostly announce themselves. Their chips already come off commercial semiconductor lines, and their modules network naturally over the fibre the world has already buried. Whether that is enough to overcome sources that fire on a coin flip and gates that work one time in nine is still an open question. The machines that answer it will have been built not by taming light’s strangeness but simply by scheduling it, one photon at a time, in flight. Loss is the enemy, not noise Every hardware platform has one dominant failure mode, and for photonics it is not the one people expect. Matter-based qubits fail by drifting out of their state. A photon does not drift. It either arrives or it does not, and the whole engineering of photonic quantum computing is built around that difference. Loss accumulates at every step. A photon can be absorbed in a waveguide, scattered at an imperfection, lost at a coupling between chip and fibre, or missed by a detector that fails to fire. Each stage keeps some fraction, and the fractions multiply, so a circuit with many components can lose most of its photons before anything is measured. That multiplication is why photonic architectures look different. Many designs are heralded, meaning the machine detects that a step succeeded and only proceeds when it did, which turns an unreliable operation into a slower but dependable one. Others build in enough redundancy that losing some photons is survivable by construction. The compensation is real. A photon that arrives is usually in excellent condition. It has not been sitting in a warm environment accumulating phase errors, and it was travelling at the speed of light and barely interacting with anything. A platform whose errors are dominated by loss rather than decoherence has a different set of problems to solve, and loss is at least conceptually simpler. Superconducting nanowire detectors make it possible This field rests on detecting a single photon reliably. It rests just as much on how recently that became possible. The workhorse is the superconducting nanowire detector, a wire cooled until it superconducts and biased just below the current at which it would stop. A single photon arriving carries enough energy to break the superconductivity in a tiny region. The wire briefly develops resistance, and that produces a measurable voltage pulse. The mechanism is simple to describe. The engineering is not. The wire must be thin enough for one photon to matter, and long enough to have a reasonable chance of catching one. Modern versions detect the overwhelming majority of photons that reach them, recover in nanoseconds and produce very few false counts. That combination arrived over the past fifteen years, and photonic quantum computing at any scale would be impossible without it. The detectors need cryogenic cooling. A platform often described as room-temperature has a refrigerator at the end of it. The remaining difficulty is telling one photon from two. Many protocols need to know exactly how many arrived, and a detector that fires the same way for one or several loses that information. Number-resolving detection exists and is harder. It is one of the specific engineering constraints shaping which photonic architectures are considered practical. Frequently asked questions What is a photonic quantum computer in simple terms?It is a quantum computer whose qubits are single particles of light travelling through channels of glass on a chip. Sources emit the photons, a programmable mesh of beam splitters and phase shifters interferes them, and superconducting detectors count which channels they exit. The pattern of detections over many runs is the result of the computation. How do photonic quantum computers encode a qubit?The standard method is dual-rail encoding, where one photon is shared between two waveguides and the channel it occupies represents 0 or 1, with superpositions created by beam splitters. Polarisation and time-bin encodings are variations on the same idea, and Xanadu instead uses continuous properties of squeezed light with the GKP code. Each vendor picks the encoding that suits its hardware. Do photonic quantum computers run at room temperature?Only partly, and the popular claim overstates it. The photons and the waveguide chips do operate at room temperature, but every leading design relies on superconducting single-photon detectors that must be cooled to a few kelvin. That is still far cheaper cooling than the millikelvin dilution refrigerators superconducting qubits need, which remains a genuine advantage. What is the Hong-Ou-Mandel effect?It is the discovery, made in 1987, that two identical photons entering opposite sides of a balanced beam splitter always leave together through the same port. The effect is pure quantum interference with no classical explanation, and its strength is the standard test of whether two photons are identical enough for computing. It is also the working mechanism inside the fusion gates that stitch photonic architectures together. Why are photonic logic gates probabilistic?Photons do not interact with each other directly, so entangling gates must be engineered from interference and measurement, which only succeed some of the time; the classic linear-optical CNOT works one attempt in nine. The KLM theorem of 2001 proved this is still enough for universal computing, at the cost of heavy redundancy. Modern fusion-based designs plan for the failures rather than trying to eliminate them. What is boson sampling, and is it a universal computer?Boson sampling is a task in which a photonic machine draws samples from the output distribution of a large interferometer, something classical computers find astronomically hard to imitate. Machines like Jiuzhang and Borealis used it to claim quantum advantage. They are single-purpose demonstrations, not universal computers, and several early claims were later weakened by improved classical algorithms. What is fusion-based quantum computing?It is PsiQuantum’s architecture, published in 2023, in which small identical entangled resource states are stitched into a fault-tolerant fabric by two-photon measurements called fusions. Fusions fail often even in principle, and the design absorbs those failures with redundancy, tolerating around ten per cent photon loss per fusion. It replaces long fragile circuits with mass-produced parts and measurements. What is a GKP qubit?A GKP qubit, proposed in 2001, encodes a qubit into the continuous wave properties of light as a grid of peaks in phase space. Small disturbances, which is what photon loss looks like in this picture, shift the grid in measurable ways, so the encoding carries its own error correction. Xanadu says it produced the first error-resistant GKP qubit on an integrated chip in 2025. Which companies are building photonic quantum computers?The companies in the field are PsiQuantum, Xanadu, Quandela, ORCA Computing, QuiX Quantum, Aegiq and QCi, each with a distinct recipe for sources, encoding and scaling. China’s USTC group builds the Jiuzhang research machines. Photonic Inc, despite the name, builds spin qubits linked by photons rather than a photonic processor. When will photonic quantum computers be useful?No photonic platform has yet published a logical qubit, the error-corrected unit that useful machines will be built from, so the honest answer is not before the late 2020s. Vendor roadmaps cluster around 2028 to 2030 for fault-tolerant systems, and quantum roadmaps have a history of slipping. The approach’s manufacturing and networking advantages are the reason the money keeps betting on it anyway. This article is for informational purposes only and does not constitute investment, financial or professional advice. The quantum technology industry evolves rapidly and information may become outdated. Always conduct your own research and consult qualified advisers before making investment decisions. Investing in quantum computing companies involves significant risk, including the potential loss of your entire investment. Past performance is not indicative of future results. More like thisQuantum FeaturesThe Weird Science of Quantum ComputingQuantum FeaturesQuantum Computing Myths and RealityQuantum PeopleChad Rigetti, The Mind Behind One Of The Early Quantum InnovationsQuantum HardwareTop Quantum Hardware Companies 2026 By ModalityStay currentSee today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals. Tags: Dr. Donovan Dr. Donovan is a futurist and technology writer covering the quantum revolution. Where classical computers manipulate bits that are either on or off, quantum machines exploit superposition and entanglement to process information in ways that classical physics cannot. Dr. Donovan tracks the full quantum landscape: fault-tolerant computing, photonic and superconducting architectures, post-quantum cryptography, and the geopolitical race between nations and corporations to achieve quantum advantage. The decisions being made now, in research labs and government offices around the world, will determine who controls the most powerful computers ever built. Latest Posts by Dr. Donovan: New codes aim to link quantum modules despite unreliable connections September 4, 2026 University of Liverpool leads 20-year CERN beam tech project September 4, 2026 Metalens scans entire eye in 3D with 1.