Tom McCarthy

State of Qubit Modalities

December 3rd, 2025. Neutral atoms are improving quickly. Trapped ions and superconducting have state of the art fidelities and attracted the most investment to date but are improving quite slowly. Photonics and semiconductor platforms are less proven.

There are at least 5 major types of qubits, all racing to achieve a large scale quantum computer. They are: trapped ions, neutral atoms, superconducting qubits, photonics, and semiconductor qubits. Two approaches in the same category can be quite different - particularly in the photonics and semiconductor categories.

3 modalities have moved beyond demonstrations with 1-10 qubits. Those modalities are neutral atoms, trapped ions and superconductors. I think the biggest surprise of the past 3 years in quantum computing has been the emergence of neutral atoms as a real contender.

When evaluating modalities, there are a couple of key metrics. SOTA machines today have pretty good fidelity, maybe to the point where qubit count is much more important to improve than fidelity. The number of qubits that can be simultaneously used for computation in a device is usually less than the number of qubits in a system. I call this ‘total good qubits’ below. Fidelities matter too, as does the ratio of qubit lifetime (decoherence time) to gate time; this limits how many operations you can perform on a qubit during its lifetime. I don’t include this last metric below.

Notes on the below specs:

  • ‘Total qubits’ is physical qubits in a single machine.
  • ‘Total good qubits’ is the number of qubits available at the machine’s quoted fidelity. This is a bit stricter than total qubits. Suppose a machine has 10 qubits at 99% fidelity and 90 at 98% fidelity. This is either a 10-qubit 99% fidelity machine, or a 100 qubit 98% fidelity machine, but it is not a 100-qubit, 99% fidelity machine!
  • Single-qubit and two-qubit gate fidelities can be compared like-for-like, but the overall fidelity of some systems depends on modality-specific issues. For example, the atoms in neutral atom machines sometimes leave the computing zone and don’t return, so two-qubit fidelity might miss that. Similarly, most of the two-qubit fidelities quoted were only achieved by a single pair of qubits in a system, rather than most connected qubits in the device.

I generally don’t count machines that have been announced by companies without an accompanying paper describing the device in detail and showing how they benchmarked it. Too easy to get away with tricks!

Edison Scientific’s literature agent was very helpful in writing this piece. I’ve linked my searches below. ChatGPT was also quite helpful. The literature searches on Edison cost only 1 credit, vs 200 credits for Kosmos, their heavyweight agent.

Please email me with any comments or corrections - tommccarthyprojects@gmail.com.

Summary

The figures in each cell are, generally, from different machines. The single-qubit and two-qubit gate fidelities are typically for smaller systems - sometimes just one or two qubits.

Modality Total Qubits Total Good Qubits 1Q Gate Fidelity 2Q Gate Fidelity
Neutral Atoms 6,139 256 @ 99.6% 2Q 99.993% 99.71%
Superconducting 156 105 @ 99.67% 2Q 99.9926% 99.94%
Trapped Ions 98 98 @ 99.92% 2Q 99.999985% 99.97%
Photonics 20 6–12 (variable) 99.6%* 99.69%*
Semiconductors 11 11 @ 99.64% 2Q 99.99% 99.9%

*Photonics fidelities are conditional on successful photon detection.

Neutral Atoms

SOTA here is the work from the Lukin and Endres groups, at Harvard and Caltech, respectively. They have created arrays of thousands of coherent atoms / qubits, though haven’t yet shown full control and programmability of the largest of these areas. In June, the Lukin group shared a paper demonstrating a 3,000-qubit coherent machine that could run continuously, and a separate paper running error-correction schemes and some computations on arrays of 448-qubits. In March 2024, the Endres group at Caltech showed off a 6,100 array of coherent qubits with single-qubit gates, but did not perform two-qubit gates or any sophisticated computations. More computation requires better control of atoms. Perhaps that will come soon. The Lukin group uses Rubidium-87 atoms, and the Endres group uses Cesium-133 atoms. Atom Computing claims to have a machine - the AC1000 - with over 1,200 qubits and competitive fidelities, but they have not shared much about it.

Metric Value Source Date Notes
Total qubits 6,139 qubits Manetsch et al Mar 2024 Cesium-133 atoms
Total good qubits 256 qubits @ 99.6% 2Q fidelity Bluvstein et al Jun 2025 Rubidium-87; 99.9% local 1Q gates
1Q fidelity 99.993(2)% PRL 131.030602 Jul 2023 Cesium-133 atoms
2Q fidelity 99.71(5)% Tsai et al Aug 2024 Strontium-88 atoms

Note on total good qubits: SOTA here is Bluvstein et al and similar work from the Lukin group over the past 2–3 years. They mention 99.9% fidelity for local single-qubit gates and 99.6% for CZ gates, as well as readout (0.3%) and movement (1%) errors. They can arrange up to 448 qubits, but I don’t think they benchmark the whole 448-qubit array. They do mention entangling 16 blocks of 16 qubits. They use Rubidium-87 atoms.

Gate speed is the most commonly mentioned issue with neutral atom QCs, and I have not seen any published plans to improve the gate speed by 10x or more. I think there is a clear route to tens of thousands of qubits in a single vacuum chamber in the next few years. The Methods section in Bluvstein et al outlines some potential improvements in error rates and getting to 10x below-threshold performance. This includes 99.99% single-qubit gates, 99.85% two-qubit gates, and much reduced movement and readout loss.

Edison: SOTA, SOTA again, 1Q fidelity, 2Q fidelity.

Superconducting

Google is one of the leaders in superconducting qubits. Recently, they have focused on error-correction demonstrations and improving fidelity in their chips rather than increasing qubit count, resulting in a 105-qubit chip, Google Willow. The Zuchongzhi-3, by a Chinese collaboration, has similar specs. Google's chip is particularly interesting because it was the first quantum computer used to verify that the surface code reduces errors exponentially well.

Metric Value Source Date Notes
Total qubits 156 qubits IBM Heron r3 Jul 2025 99.75% median 2Q (CZ), 98.99% mean 2Q fidelity
Total good qubits 105 qubits @ 99.67% 2Q fidelity Google Willow Dec 2024 First surface code exponential error demo
1Q fidelity 99.9926% Li et al Feb 2023 Transmons
2Q fidelity 99.94% Lin et al Nov 2024 CNOT gate on fluxonium qubits

Note on total qubits: IBM Heron r3 launched July 2025 with 99.75% two-qubit fidelity (CZ gates, median). However, you can see 6-7 qubits with >1% error rate on Qiskit. Calculating the mean two-qubit fidelity for the IBM chip gives a 98.99% two-qubit fidelity. I respect IBM's transparency in publicly displaying their chips and specs.

Note on total good qubits: The Zuchonghzi-3 is a 105-qubit chip and reports 99.62% two-qubit fidelity, but their reported experiment uses 83 qubits out of the 105 available qubits.

IBM has announced larger machines (Condor, Osprey) previously, but they do not appear to be pursuing them, and I can't find any publications detailing their specs. I think manufacturing precision and consistency is one of the big obstacles to better superconducting systems. Each superconducting circuit is sensitive to microwave signal timing and frequency. Ideally, they all require the exact same control signals, but small differences and imperfections will result in each qubit needing a slightly different local environment - or else accept that control signals will work for some qubits and disrupt others, and try to mitigate / correct that somehow.

This is a biased source, but a professor mentioned to me that he decided that superconducting qubits were plateauing in 2015. This is actually when a lot of investment was taking place, so his view and the commercial / investor view were different. His rationale was that experimental work and publications in superconducting qubits had slowed down quite a bit by then. I find this gap - and others like it - between researcher and investor / founder viewpoints interesting.

Edison: SOTA, 1Q fidelity, 2Q fidelity.

Trapped Ions

Quantinuum leads the field in trapped ions. They recently launched Helios, a 98-qubit machine with solid gate fidelities. Oxford Ionics has led the field in gate fidelities, but these have been for much smaller systems, and have not been extended to computation on 10s of qubits. Oxford Ionics was acquired by IONQ for $1B in June 2025.

Metric Value Source Date Notes
Total qubits 98 qubits Quantinuum Helios Nov 2025 Barium ions
Total good qubits 98 qubits @ 99.92% 2Q fidelity Quantinuum Helios Nov 2025 All-to-all connectivity
1Q fidelity 99.999985% Smith et al Dec 2024 Oxford/Osaka; Calcium ions; <10-7 error
2Q fidelity 99.97% Löschnauer et al Jul 2024 Oxford Ionics; Calcium ions

Edison: SOTA, 1Q fidelity, 2Q fidelity.

Photonics

Photonics-based quantum computing is mostly based around approaches that are quite different to the other modalities, and the usual metrics don't always apply. Photons get used up, or are transitory, so you can't treat them as static qubits, like Josephson junctions in superconducting machines. They are generated, acted on, and discarded. A photonics-specific metric of interest is overall transmission quality, or conversely, loss; what percentage of photons get transmitted fully from source to detector? A lot of the experiments and published results include fidelities that are conditional on successful detection of photons. Reporting this as a headline figure avoids the elephant in the room, i.e. how many of the photons got to the detector?

Metric Value Source Date Notes
Largest machine 20-qubit cluster O'Sullivan et al Sep 2024 Device seems capable of universal computation
Total good qubits 12 qubits @ 56% and 95% losses Xanadu Aurora Jan 2025 ~1% loss required for fault-tolerance
6 qubits @ 93.8% 2Q fidelity, 91.6% loss Quandela Ascella Jun 2023 -
1Q fidelity 99.6% Quandela Ascella Jun 2023 Conditional on detection, 91.6% loss
99.98% SPAM PsiQuantum Feb 2025 -
2Q fidelity 99.69% entangling Shi et al Aug 2022 Cold atom system, different to photonic chips
99.22% fusion PsiQuantum Feb 2025 Conditional on detection

Overall, I am not very confident that my figures for photonics are well chosen, or all that comparable to each other. Transmission loss appears to be the biggest issue.

Edison: Loss, SOTA, SOTA again, 2Q fidelity.

Semiconductors

Because we are so good at making transistors, some people argue that qubits based on silicon - or Germanium or Gallium - chips and photolithography are a better approach than the other modalities. As with photonics, there are a few different architectures within silicon-based quantum computing, and no clear winner yet.

Metric Value Source Date Notes
Total qubits 11 qubits UNSW/SQC Jun 2025 Phosphorous in silicon; universal gates
Total good qubits 11 qubits @ 99.64% CROT UNSW/SQC Jun 2025
1Q fidelity 99.99% Wu et al Jul 2025 5-qubit system; 99.9% simultaneous
2Q fidelity 99.9% CZ UNSW/SQC Jun 2025 Specific nuclear spin pair
99.8% CPHASE Mills et al Jun 2022
99.8% CZ Tanttu et al Mar 2024

Note on total qubits: Intel reported a 12-qubit device in October 2024 but they only report (de)coherence times, and don't demonstrate universal gates, whereas the 11-qubit device appears to be capable of universal computation.

Edison: SOTA, 1Q fidelity, 2Q fidelity.

Again, please email me with any corrections or clarifications. Thanks in advance and thanks for reading! tommccarthyprojects@gmail.com