Six ways to build a qubit: quantum computing modalities compared

Every quantum computer needs qubits that hold a quantum state long enough to compute, respond to precise control, interact with each other on demand, and can be read out. Six physical approaches lead today. Each makes different trade-offs between speed, quality and scale, and none has yet delivered a large fault-tolerant machine. This guide explains how each one works and how they compare, with figures as of October 2026. Each modality also has its own in-depth explainer, linked below.

The six modalities at a glance

Superconducting circuits. Electrical circuits on a chip, cooled to about 10 millikelvin (mK), that behave like artificial atoms. They have fast gates and a leading record in error correction. Google, IBM, AWS, Rigetti and Alice & Bob build them.

Trapped ions. Single charged atoms held by electric fields above a chip and controlled with lasers or microwaves. They have the highest published gate fidelities and connect every qubit to every other, but they are slow. Quantinuum, IonQ and AQT build them.

Neutral atoms. Uncharged atoms held in focused laser beams and entangled through highly excited Rydberg states. They form the largest arrays, with thousands of atoms. QuEra, Atom Computing, Pasqal and Infleqtion build them, with Harvard, MIT and Caltech leading the research.

Silicon spin qubits. The spin of single electrons or holes in quantum dots tens of nanometers across, built like transistors. They are tiny and suit chip fabs, but systems are still small. Intel, Diraq, HRL Laboratories, Quantum Motion and QuTech lead the work.

Photonic qubits. Single photons in waveguides on a chip. Photons barely lose their quantum state and travel well through fiber, but they do not interact, so entangling them is probabilistic. PsiQuantum, Xanadu and Quandela build them.

Topological qubits. A proposed qubit stored non-locally in pairs of Majorana zero modes in superconducting nanowires, which could be protected from local noise by design. Microsoft leads the approach; no qubit operation has been published, and the evidence is disputed.

How they compare

Two-qubit gate fidelity. The error of each entangling gate largely decides how much error correction costs. Trapped ions lead: Quantinuum reports 99.921% averaged across all pairs of its 98-qubit Helios, and IonQ 99.99% on a single pair; both are company figures. Neutral atoms typically reach 99.5 to 99.6%, with a record of 99.854% in an April 2026 preprint from Harvard and MIT. Across a whole superconducting chip, Google's Willow averages 99.67% for controlled-Z gates and Rigetti reports a 99.1% median; the best peer-reviewed single pairs reach 99.90% for transmons (RIKEN and Toshiba, 2024) and 99.92% for fluxonium (MIT, 2023). Silicon spin devices mostly publish 99 to 99.8%, with HRL's best pair at 99.91%. PsiQuantum reports 99.22% for photonic fusions, counting only events in which the photons were detected. No Majorana qubit operation has been published.

Gate speed. Compare the whole operation, including any movement a gate needs. Superconducting and spin qubits complete gates in tens to hundreds of nanoseconds. A neutral-atom entangling pulse takes 0.2 to 0.5 microseconds (µs), but bringing the atoms together takes about 0.1 to 1 millisecond (ms) per gate. A trapped-ion gate takes tens to hundreds of microseconds, and about 1 to 10 ms once shuttling and re-cooling are counted, by our estimate from measured layer times. One round of error correction takes 1.1 µs on Google's Willow but an estimated 1 to 10 ms on neutral atoms, where imaging dominates. Photonic machines could in principle run at gigahertz clock rates, but each entangling step succeeds only part of the time.

Coherence. Trapped ions hold superpositions for seconds, and neutral atoms typically for around a second, with up to 12.6 seconds in Caltech's cesium array. Superconducting transmons keep their energy for about 0.07 to 0.3 ms in processors (68 µs on average on Willow), and the best laboratory transmon reaches 1.68 ms (Princeton, 2025). The best foundry-made silicon spin device keeps its phase for 30 µs without echo pulses and 0.8 ms with them, while a 12-qubit Intel array averages 1.3 µs without echo. These are different measures (energy lifetime for transmons, phase coherence for spins), so they are not a strict ranking. Photons barely decohere; loss is their main error. There is no coherence data yet for a Majorana qubit.

Connectivity. Trapped ions connect every qubit to every other within a zone, and across the chip by moving ions. Neutral atoms reach any partner by being moved. Photonic qubits route through switches and fiber. Superconducting qubits are wired to three or four fixed neighbors on a square lattice, although IBM has shown long-range couplers. Spin qubits couple only to adjacent dots, mostly in 1D chains or small 2D grids.

Qubits today. Neutral atoms lead in raw count: Caltech held 6,100 atoms as qubits in 2025, though not all were linked by gates. IBM's largest current superconducting chips run 156 qubits (a 2023 demonstration chip had 1,121), and Google's Willow 105. Quantinuum's Helios runs 98 ions, and IonQ has announced 256-qubit chips. Spin processors run about 6 to 18 qubits, with HRL's largest at 18 qubits made from 54 quantum dots, and photonic machines about 12. Microsoft describes its Majorana 1 chip as eight qubits, a claim critics dispute.

Manufacturing. Superconducting, spin and photonic chips are made with semiconductor lithography, and spin and photonic chips already come from 300 mm CMOS lines at Intel, imec and GlobalFoundries; IBM moved its superconducting wafer production to a 300 mm fab in 2025. Trapped ions use microfabricated chip traps but need benches of lasers and optics. Neutral atoms need no chip at all: the processor is a pattern of light in a vacuum cell, and lasers dominate the cost. Majorana devices depend on specialized semiconductor growth.

Cooling. Superconducting qubits need dilution refrigerators at about 10 mK, as do Majorana devices, which also need a magnetic field of about 1 to 2 tesla. Spin qubits usually run at 10 to 150 mK, and above 1 kelvin has been shown. PsiQuantum runs its photonic chips at about 2 kelvin, where its superconducting detectors work. Ion traps run in vacuum at room temperature or cryogenically, and neutral-atom machines are room-temperature apparatus around atoms laser-cooled to microkelvin.

Error correction. Google's Willow showed logical errors falling by a factor of 2.14 each time its surface code grew, with a 1.1 µs cycle, and bosonic GKP codes have gone beyond break-even (Yale in 2023, and Yale with Google in 2025). Harvard showed neutral-atom error suppression below threshold in 2025, and Harvard and MIT kept more than 3,000 atoms running for over two hours by reloading them. Quantinuum reports 48 error-corrected logical qubits on Helios, a company figure. Spin qubits have run small repetition codes. PsiQuantum has not publicly shown a logical qubit, and there is none yet from Majorana devices.

What each approach is betting on

Superconducting qubits bet on speed; trapped ions on near-perfect gates; neutral atoms on sheer numbers; silicon spin qubits on the fabs that make transistors; photonics on chips linked by fiber; and Majorana qubits on protection built into the hardware, which is still unproven.

No approach has yet run a fault-tolerant algorithm at useful scale. Company roadmaps aim for fault-tolerant machines from about 2028 to the early 2030s: QuEra's Libra (256 logical qubits, 2028), IBM's Starling (200 logical qubits, 2029), Quantinuum's Apollo (2029), Microsoft's scalable Majorana machine (2029) and Alice & Bob's Graphene (100 logical qubits, 2030). DARPA's Quantum Benchmarking Initiative asks whether utility-scale quantum computing is achievable by 2033. These are plans, not results.

Quick answers

Which type of qubit is best? None leads on every measure. Trapped ions lead on gate quality and connectivity, neutral atoms on qubit count, superconducting circuits on speed and error-correction cycles, and silicon spins and photonics on compatibility with chip manufacturing. The answer depends on whether a machine is limited by error rates, by size or by run time.

What is a quantum computing modality? The physical system used to make qubits: superconducting circuits, trapped ions, neutral atoms, electron spins, photons or topological states. It sets how qubits are built, controlled, connected and cooled.

Which qubits are the fastest? Superconducting and silicon spin qubits, with gates of tens to hundreds of nanoseconds. Google's Willow runs a full error-correction cycle in 1.1 µs, against about 1 to 10 ms for neutral atoms.

Which modality has the most qubits? Neutral atoms: Caltech held 6,100 atoms as qubits in 2025, and Harvard and MIT ran more than 3,000 continuously for over two hours. The number that can all be entangled at high fidelity in one computation is smaller.

Do all quantum computers need extreme cold? No. Superconducting and Majorana qubits need millikelvin refrigerators, and spin qubits usually do, though they have run above 1 kelvin. PsiQuantum's photonic chips run at about 2 kelvin, although some photonic detectors need cooling below 1 kelvin. Trapped-ion and neutral-atom machines work in vacuum chambers around atoms cooled by lasers.

Sources

Each in-depth explainer lists its own sources. The main ones behind the comparison:

Google Quantum AI, Willow spec sheet (2024)

Quantinuum Helios launch (November 2025)

Lukin group, 99.854% neutral-atom entangling gates (arXiv 2604.25987, preprint)

Caltech, a tweezer array with 6,100 highly coherent atomic qubits, Nature (2025)

HRL Laboratories, a digitally controlled silicon quantum processing unit (arXiv 2604.16216)

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

Microsoft Azure Quantum et al., interferometric single-shot parity measurement, Nature 638, 651 (2025)

The Quantum Insider, IBM's fault-tolerance roadmap from Loon to Starling (June 2025)

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.