How photonic quantum computers work

A photonic quantum computer encodes qubits in single particles of light that travel through waveguides on a chip. Photons barely interact with their surroundings, so they keep their quantum state over long distances, and the chips can be made in commercial semiconductor fabs and linked with ordinary telecom fiber. The hard part is the other side of the same property: photons do not interact with each other either, so entangling them succeeds only part of the time, and every lost photon is an error. PsiQuantum, Xanadu, Quandela, ORCA Computing and QuiX Quantum build them.

This article is part of a series on the main ways to build a qubit. The overview compares all six side by side.

How the qubit forms

Most photonic machines use dual-rail encoding: one photon shared between two waveguides, where the photon in one guide means 0 and in the other means 1. A beam splitter places the photon in both guides at once, and a phase shifter sets the phase between them. If no photon is detected in either guide, the qubit is known to be lost, so loss becomes a flagged error, an erasure, rather than a silent one.

Single photons come from two kinds of source. PsiQuantum's are heralded: a strong pump laser in a silicon or silicon-nitride ring resonator occasionally converts two pump photons into a pair, and detecting one photon of the pair heralds the other. Each source fires in only a few percent of attempts, because pumping harder risks making two pairs at once. Quandela instead uses quantum dots, artificial atoms that emit about one photon per laser pulse but must be cooled to 4 to 10 kelvin. Xanadu takes a different route and encodes qubits in the amplitude and phase of squeezed light.

Photons do not lose their quantum state as they travel: the coherence of a photon's path or polarization survives kilometers of fiber. The main errors are loss and distinguishability, since photons that differ slightly in color or timing do not interfere perfectly.

How the chips are made

PsiQuantum's chips are built by GlobalFoundries on 300 mm wafers, in a custom process that adds silicon-nitride waveguides, barium titanate for fast switches and niobium-nitride superconducting detectors. Its 2025 Nature paper and a companion preprint report multimode waveguide loss of about 0.5 decibel per meter, crossings that lose 1.6 thousandths of a decibel, and electro-optic switches that lose about 0.1 decibel.

Couplers bring two waveguides close enough for light to tunnel between them, so they act as beam splitters. A Mach-Zehnder interferometer combines two couplers and phase shifters into a tunable beam splitter, the unit cell of programmable photonic circuits.

The detectors are superconducting nanowires laid on the waveguides: 88.9% efficient on average on PsiQuantum's chips, and 96.3% in its next-generation detectors, which also count photons. PsiQuantum cools whole chips to about 2 kelvin in cryostats with 10 to 20 watts of cooling power. That is far warmer than the 10 millikelvin of superconducting qubits, and it lets the company use industrial helium plants. Xanadu's chips run at room temperature, but its detectors need cooling below 1 kelvin.

Chips connect through standard fiber: PsiQuantum transferred qubits between two chips over 42 meters of fiber at 99.72% fidelity, counting only detected photons.

How it is controlled

Single-qubit gates are passive interference: beam splitters and phase shifters rotate the photon between its two rails with very high fidelity. PsiQuantum reports 99.98% for state preparation and measurement, again counting only detected photons.

Two-qubit gates are the hard part. Photons do not interact, so entanglement comes from interfering two photons and detecting them. With simple linear optics that succeeds at most half the time, and extra photons raise it to 75% or more.

Fusion-based quantum computing, PsiQuantum's architecture, is built around this. Small entangled resource states of a few photons are made over and over, and fusions, destructive two-photon measurements between neighboring resource states, stitch them into a large error-correcting structure. Failed fusions and lost photons are reported to a classical decoder, and each photon is measured soon after it is made, so it only has to survive a short path. In models, the scheme tolerates on the order of 10% photon loss per fusion.

PsiQuantum's chips reached 99.22% fidelity for two-qubit fusions (Nature, 2025). That figure counts only events in which the photons were detected; it excludes loss, the hardest part. Fast switches must route successful photons and set later measurements within nanoseconds, because photons cannot wait.

What makes it hard to scale

Photon loss. Every coupler, crossing, switch, fiber joint and detector removes some photons. Fusion-based designs need total loss per photon below roughly 10%, and in practice a few percent. A third-party analysis of Xanadu's Aurora estimates 56% loss on one optical path, against a target of 1% or less.

Probabilistic sources. Heralded sources fire only a few percent of the time, so each reliable photon needs many sources and a switch network, and every switch adds loss.

Switches and detectors. Switches need about 0.1 decibel of loss or less, nanosecond speed and little heat at 2 kelvin. Detectors need efficiencies above 99% and the ability to count photons.

Footprint. A fault-tolerant machine is expected to need tens of thousands of chips linked by fiber, in a building with an industrial cryoplant, according to PsiQuantum.

Synchronization. Photons from different sources and chips must arrive within picoseconds of each other and match in color, across thousands of resonators.

Progress to date. PsiQuantum has not publicly shown a logical qubit, and an outside analysis puts Xanadu's on-chip GKP states about 9 decibels of squeezing short of the threshold for fault tolerance.

What could solve it

Multiplexing. Switch networks or fiber delay loops turn many probabilistic sources into near-deterministic photons, if switch loss is low enough.

Loss-tolerant designs. Better resource states, boosted fusions and decoders that know where photons were lost relax the loss budget.

Lower-loss parts. Low-loss silicon-nitride routing, crossings that lose thousandths of a decibel, and better coupling to fiber.

Deterministic sources. Quantum-dot sources could remove the need for multiplexing, if separate dots can be made to emit identical photons.

Industrial cooling. Working at 2 kelvin lets PsiQuantum use industrial helium refrigerators like those at particle accelerators. Linde is building the plant for its site near Brisbane, with delivery planned for the second half of 2027.

Native networking. Modules link through telecom fiber: Xanadu's Aurora connected 35 chips across four racks with about 13 kilometers of fiber.

Strengths and limits

Strengths. Photons barely decohere, and in dual-rail encoding their loss is flagged; chips come from commercial 300 mm fabs; chips that run at about 2 kelvin rather than millikelvin, although some detectors need cooling below 1 kelvin; native networking over fiber; very high single-qubit and measurement fidelity; and the potential for gigahertz clock rates.

Limits. No deterministic two-photon gate, so entanglement needs heavy multiplexing; tight loss budgets, with published fidelities that exclude loss; detectors that still need cryogenics; an enormous footprint at fault-tolerant scale; and no public logical qubit yet. Current machines have about 12 qubits; Quandela plans 24 in 2026.

Key numbers

Two-qubit fusion fidelity: 99.22%, counting only detected photons (PsiQuantum, Nature 2025).

Chip-to-chip link: 99.72% over 42 meters of fiber, counting only detected photons (PsiQuantum).

State preparation and measurement: 99.98%, counting only detected photons (PsiQuantum).

Loss tolerance: on the order of 10% per photon in fusion-based designs (theory, depending on the noise model).

Detector efficiency: 88.9% on average; 96.3% in next-generation photon-counting detectors (PsiQuantum).

Operating temperature: about 2 kelvin for PsiQuantum's chips; 4 to 10 kelvin for quantum-dot sources.

System scale: 12 qubits on 35 chips across four racks (Xanadu Aurora); 12 qubits in Quandela's Belenos.

Quick answers

Why use photons as qubits? Photons barely interact with their surroundings, so they keep their quantum state, travel through standard fiber, and can be generated and measured on chips made in commercial fabs.

Do photonic quantum computers work at room temperature? Partly. The waveguides work at any temperature, but the best single-photon detectors are superconducting and need cooling, to about 2 kelvin in PsiQuantum's design and below 1 kelvin for Xanadu's detectors.

What is fusion-based quantum computing? PsiQuantum's architecture: small entangled states of a few photons are produced repeatedly and joined by destructive two-photon measurements, called fusions, into a large error-correcting structure, with failures handled by a classical decoder.

What is the biggest challenge for photonic quantum computing? Loss. Every component removes some photons, and fault-tolerant designs need total loss per photon below roughly 10%, ideally a few percent.

Sources

PsiQuantum, a manufacturable platform for photonic quantum computing, Nature (2025)

PsiQuantum, component-level results (arXiv 2404.17570)

Bartolucci et al., fusion-based quantum computation, Nature Communications 14, 912 (2023)

Xanadu, Aurora: 12 qubits, 35 chips, 4 racks (January 2025)

Xanadu, on-chip GKP states on 300 mm silicon nitride (June 2025)

PhotonCap, third-party analysis of Xanadu's loss and squeezing (2026)

PsiQuantum breaks ground at Moreton Bay; Linde cryoplant timeline (June 2026)

About Anyon Computing

Anyon Computing Inc. (also known as Anyon Technologies) is a quantum computing company that provides superconducting modular quantum supercomputers, which it tightly couples with NVIDIA GPUs via NVQLink for data center deployment at scale. Anyon pioneered deploying NVIDIA NVQLink inside a commercial production data center and is among the first superconducting QPU backends integrated with NVIDIA CUDA-Q. Founded by pioneers from Caltech and UC Berkeley. https://www.anyoncomputing.com/

NVIDIA, CUDA-Q and NVQLink are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.