Comparison

Superconducting vs trapped-ion vs neutral-atom qubits: how the types of quantum computer compare

Superconducting qubits have the fastest gates and run the most shots per hour, but each qubit is wired only to a few neighbours. Trapped ions have the lowest average two-qubit error of any machine on sale and connect every qubit to every other, but their gates are far slower. Neutral atoms make the largest arrays and can be moved to connect freely, but each run takes far longer still. Photonic qubits and analog machines, quantum annealers and analog atom simulators, are further from running general circuits. No type is best for every job: the right one depends on the circuit, its width, its depth, how far apart its qubits interact and how many shots it needs.

Updated

The short answer

Superconducting
The fastest gates: a two-qubit gate in 42 ns[5] on Google's Willow. Each qubit couples only to its neighbours on the chip, so distant qubits are joined with extra SWAP gates. IBM, Google, Rigetti and IQM build them
Trapped ion
The lowest average gate error in service: a two-qubit error of 7.9×10−47.9 \times 10^{-4}[10] on Quantinuum's Helios, with every qubit connected to every other. Gates are slow, about 900 microseconds[15] for one two-qubit gate on IonQ's Forte
Neutral atom
The largest arrays: over 6,100 atoms[24] held in one lab array, and 260 qubits[17] in QuEra's gate-based Gemini. Atoms are moved to connect any pair, and a machine runs about 1 shot per second[17]
Photonic
Qubits made of light, mostly at room temperature. The largest cloud machine we found, Quandela's Belenos, has 12 qubits[25]; optical loss is the obstacle to scaling
Analog machines
D-Wave's annealer has 4,400+ qubits[28] and QuEra's Aquila 256 atoms[18], but neither runs gate circuits: each evolves one physical model
Which is best
None for every job. Wide circuits with long-range interactions favour all-to-all ions or movable atoms; deep circuits that need many shots favour superconducting chips; the error per gate after routing decides the rest

The types of qubit side by side

Each value names the device and the source it comes from, and the date where the source gives one. Vendors report different statistics (mean, median, typical, a layered benchmark), so read across a row with care; the caveat below the device table says more.

AspectSuperconductingTrapped ionNeutral atomPhotonic
What the qubit isA circuit of aluminium and Josephson junctions on a chip, cooled near absolute zeroThe internal states of one ion, held by electric fields in a vacuum and driven by lasers or microwavesThe internal states of one neutral atom, held by a laser tweezer in a vacuumA photon in one of two paths (dual rail) or a mode of squeezed light
Two-qubit gate accuracyWillow: mean CZ error 0.33% ± 0.18%[4]. IBM Boston: layered error 2.37×10−32.37 \times 10^{-3}[1] (4 November 2026)Helios: error 7.9×10−47.9 \times 10^{-4}[10], averaged over all gate zonesGemini: global gate fidelity 99.2%[17]. Harvard and QuEra in 2023: 99.5% on up to 60 atoms in parallel[21]Belenos: fidelity (99.0 ± 0.8)%[25]. PsiQuantum: fusion fidelity 99.22% ± 0.12%[26]
Two-qubit gate timeCZ in 42 ns[5] on Willow; ~60 ns[7] on Rigetti's CepheusZZ gate about 900 microseconds[15] on average on IonQ ForteThe Rydberg pulse is short, but moving atoms leaves about 200 microseconds[22] between gate layersSet by the clock of the photon source and detectors; the limit is loss, not time
ConnectivityNeighbours only: IBM Heron is a heavy-hex lattice[3], Willow has 3.47 on average, 4-way typical[4]All-to-all: one ion chain on IonQ Forte; on Helios ions move between 8 gate zones[11] so any pair can meetAny-to-any by moving atoms; Gemini lists 20[17] parallel gate pairsAny pair of modes, through programmable interferometers, paid for in extra optical modes
CoherenceWillow mean T1 68 ± 13 microseconds[4] on its error-correction chipQuantinuum reports a memory error instead: 6×10−46 \times 10^{-4}[11] per qubit per average circuit layer on HeliosUp to 12.6 seconds[24] in a lab tweezer arrayPhotons barely decohere; they are lost
Shots per secondUp to 109.7 kHz[1] maximum circuits per second on IBM Phoenix; 63,000[4] on WillowNot published by the vendors we checked; gate times put it far below superconducting chipsGemini: 1 shot per second[17]Not published for Belenos
Large device in serviceIBM Heron, 156 qubits[1] on the cloudQuantinuum Helios, 98 qubits[11]; IonQ lists Tempo at 100 qubits[14]QuEra Gemini, 260 qubits[17], installed in JapanQuandela Belenos, 12 qubits[25] on the cloud
Runs inA dilution refrigeratorUltra-high vacuum, with lasers or microwave electronics to drive the ionsA vacuum cell at room temperature[17]Mostly room temperature; Xanadu's photon-number detectors sit in a cryostat[27]
Leading buildersIBM, Google, Rigetti, IQM, USTCQuantinuum, IonQQuEra, Pasqal, Atom Computing, InfleqtionPsiQuantum, Xanadu, Quandela

Analog machines are left out of this table because they do not run gates; they have their own section below.

Superconducting qubits

A superconducting qubit is an electrical circuit: a small loop of superconducting metal interrupted by Josephson junctions, which makes it behave like an artificial atom with two lowest energy levels. Microwave pulses rotate a single qubit, and a two-qubit gate is made by tuning a coupler between two neighbouring qubits. Because everything is a chip and a pulse, gates are fast: Google's Willow performs a CZ gate in 42 ns[5] and repeats whole circuits 63,000[4] times a second. Google describes the strength plainly: superconducting circuits have run millions of gate and measurement cycles, each taking about a microsecond[6].

The costs come with the chip. Each qubit is wired only to its neighbours, at most three on IBM's heavy-hex Heron lattice[3] and four on square grids such as IBM's Nighthawk and Google's Willow, so two distant qubits interact only after SWAP gates walk them together. Coherence is short, a mean T1 of 68 microseconds[4] on Willow's error-correction chip, and the chip must sit in a dilution refrigerator. Quality also varies across one chip: IBM's live page for its Heron r3 machine Boston listed a best pair at 5.3×10−45.3 \times 10^{-4}[1] and an error per layered gate across a 100-qubit chain[2] of 2.37×10−32.37 \times 10^{-3}[1] on 4 November 2026. The path to scale is lithography: more qubits per chip, chips linked by couplers, and error correction, where superconducting chips showed the first memory below threshold.

Trapped-ion qubits

A trapped-ion qubit is two internal energy levels of one charged atom, barium in Quantinuum's Helios, held in a vacuum by electric fields. Lasers or microwaves drive the gates, and two-qubit gates use the shared motion of the ions as a bus. Every ion of an element is identical, so there is no fabrication spread, and averaged over a whole machine the gates are the most accurate on sale: Helios reports a two-qubit error of 7.9×10−47.9 \times 10^{-4}[13] averaged over all its gate zones. In a two-ion lab test, Oxford Ionics, now part of IonQ, measured an error of 8.4×10−58.4 \times 10^{-5}[16], in a preprint.

Connectivity is all-to-all. IonQ's Forte keeps its ions in one chain where any pair can be entangled; Quantinuum's machines shuttle ions between zones, 4 of the 8[11] on Helios able to do two-qubit gates, so any pair can meet. The price is time. A two-qubit gate on Forte took about 900 microseconds[15] on average in a 2024 benchmark, and moving ions adds more, so a deep circuit takes far longer per shot than on a superconducting chip. Scaling means longer chains or more zones and junctions on one trap, and the largest machine in service, Helios, has 98 qubits[11]. IonQ lists its Tempo system at 100 qubits[14] and a 99.9%[14] two-qubit fidelity.

Neutral-atom qubits

A neutral-atom qubit is an uncharged atom, rubidium in QuEra's machines, held in a focused laser beam called an optical tweezer. Thousands of tweezers can be made from one laser, which is why neutral atoms hold the size record: a Caltech array trapped over 6,100 atoms[24] with a coherence time of 12.6 seconds[24]. A two-qubit gate excites atoms to a Rydberg state, a huge orbit in which a nearby atom blocks its neighbour's excitation, and the same laser pulse can drive many pairs at once: Harvard and QuEra reached 99.5%[21] fidelity on up to 60 atoms[21] in parallel.

Connectivity comes from motion. Atoms are carried in moving tweezers to the partner they need, so any pair can interact, and whole rows move together. Moving is slow, though, with about 200 microseconds[22] between gate layers in Harvard's logical processor; Google, which added neutral atoms to its programme in 2026, puts their cycle times in milliseconds[6]. Loading and imaging a fresh array is slower still: QuEra's gate-based Gemini, with 260 qubits[17] and a global two-qubit fidelity of 99.2%[17], lists a repetition rate of 1 shot per second[17]. Atoms can also be lost from a tweezer mid-circuit, an error the other types do not have. The apparatus runs at room temperature, and the scaling path is more tweezers and continuous reloading.

Photonic qubits

A photonic qubit is light. In the dual-rail encoding a single photon travels down one of two waveguides, and beam splitters and phase shifters on a chip rotate it. Photons barely interact, which keeps them coherent but makes a two-qubit gate hard: it is done with extra photons and a measurement that heralds success, so it works only part of the time. Quandela's Belenos, the one photonic gate machine on a public cloud we found, has 12 qubits[25] in 24 modes[25] and a two-qubit fidelity of (99.0 ± 0.8)%[25].

The large photonic programmes skip the circuit model and build fault-tolerant machines from measured entangled resource states. PsiQuantum showed chip-made components with a fusion fidelity of 99.22% ± 0.12%[26], and Xanadu's Aurora ran 35 photonic chips[27] as one networked machine. Loss is what stands in the way: Aurora's authors put the loss budget for fault tolerance at about 1%[27] and measured about 56%[27] on the heralding paths. The appeal is manufacturing in existing chip fabs and networking over optical fibre.

Analog machines: quantum annealers and analog atom simulators

Some quantum computers do not run gates at all. A quantum annealer, D-Wave's Advantage2 with 4,400+ qubits[28] and 20-way[28] connectivity, uses superconducting flux qubits whose couplings are set to a problem's Ising energy and slowly turns off a transverse field, so it finds low-energy states of that one model. It solves QUBO and Ising problems and nothing else; how to write one is in the QUBO and quantum annealing guide, and how it compares with the gate-based QAOA on the same problem is in quantum annealing vs QAOA.

Analog neutral-atom machines place atoms in a chosen pattern and drive them all with the same laser, so the array evolves under one Rydberg Hamiltonian. QuEra's Aquila has 256 atoms[18] and is on Amazon Braket. Pasqal's spec sheet for Orion lists 140 qubits[19] nominal and an effective repetition rate of 0.25 Hz[19], the figures of its newest generation, Orion Gamma. Fresnel, the Pasqal machine open through Azure Quantum, is the earlier Orion Beta with 100 qubits[20]. Both companies' machines are simulators of a physical model rather than programmable circuits, which is why a qubit count and a gate count do not describe a workload for them. What analog and digital machines have simulated so far, and which results classical methods later matched, is in the quantum simulation guide. D-Wave's claim of a beyond-classical simulation on an annealer, and the classical replies to it, are on the quantum advantage page.

The QPUs on EigenOrbit Benchmarks, with their specs today

EigenOrbit Benchmarks lists QPUs from five hardware families and routes circuits onto each one. This is a representative subset, with each number re-checked against the vendor's own page or calibration data on 4 November 2026 unless the row gives another date.

DeviceTypeQubitsConnectivityTwo-qubit gate, as publishedSource and date
IBM Boston (Heron r3)Superconducting156[1]Heavy-hex latticeLayered error 2.37×10−32.37 \times 10^{-3}[1]IBM live calibration, 4 Nov 2026
IBM Kingston (Heron r2)Superconducting156[1]Heavy-hex latticeLayered error 3.63×10−33.63 \times 10^{-3}[1]IBM live calibration, 4 Nov 2026
IBM Phoenix (Nighthawk r2)Superconducting120[1]Square grid[3]Layered error 3.65×10−33.65 \times 10^{-3}[1]IBM live calibration, 4 Nov 2026
Rigetti Cepheus-1-108QSuperconducting107 active[8] of 108[7]Square lattice of chiplets, 193 active couplers[8]Median CZ fidelity 98.7%[8] (our median over all pairs); Rigetti's GA figure 99.1%[7]Rigetti live calibration, 4 Nov 2026; press release, 7 Apr 2026
IQM EmeraldSuperconducting54[9]Square latticeMedian fidelity 99.5%[9]IQM, at launch, 5 Aug 2025
Quantinuum HeliosTrapped ion98[11]All-to-allError 7.9×10−47.9 \times 10^{-4}[10]Quantinuum performance page, read 4 Nov 2026
Quantinuum H2-1Trapped ion56[12]All-to-allError 1.2×10−31.2 \times 10^{-3}[10]Quantinuum performance page, read 4 Nov 2026
Quantinuum H2-2Trapped ion56[12]All-to-allError 8.3×10−48.3 \times 10^{-4}[10]Quantinuum performance page, read 4 Nov 2026
IonQ ForteTrapped ion36[14]All-to-allFidelity 99.6%[14]IonQ systems page, read 4 Nov 2026
Quandela BelenosPhotonic12[25]Programmable interferometerFidelity (99.0 ± 0.8)%[25]Quandela product page, read 4 Nov 2026
QuEra AquilaNeutral atom, analog256[18]Atoms placed in a chosen 2D patternNo gates: analog evolution onlyQuEra product page, read 4 Nov 2026
Pasqal Fresnel (Orion Beta)Neutral atom, analog100[20]Atoms placed in a chosen 2D patternNo gates: analog evolution onlyAzure Quantum provider page, read 4 Nov 2026

Benchmarks also lists IonQ Forte Enterprise, the on-premises Forte with the same published figures, and IQM Garnet. It reads IonQ Forte from the device's own calibration export on Amazon Braket, which reports a worse two-qubit error than IonQ's page. The neutral-atom gate machines, QuEra Gemini among them, are not on Benchmarks because none takes circuits from outside users through a public cloud.

Why the best quantum computer depends on the circuit

A circuit asks for two-qubit gates between particular pairs. On an all-to-all machine every gate runs as written. On a lattice chip a gate between distant qubits first needs SWAPs, and each SWAP costs three two-qubit gates, so routing multiplies the two-qubit gate count by a factor r≥1r \geq 1 that depends on the circuit and the chip. A random MaxCut graph pays a lot; a nearest-neighbour circuit laid out along the lattice pays nothing. With independent errors, the probability that a run has no error is roughly

p≈(1−e2)r g2 (1−e1)g1 (1−eSPAM)np \approx (1-e_2)^{r\,g_2}\,(1-e_1)^{g_1}\,(1-e_{\text{SPAM}})^{n}

for g2g_2 two-qubit and g1g_1 single-qubit gates on nn qubits. For small errors, an all-to-all machine with two-qubit error eae_{\text{a}} beats a lattice chip with error ele_{\text{l}} on that circuit when ea<r ele_{\text{a}} < r\,e_{\text{l}}. That is why an ion trap can win a long-range circuit against a chip with a better headline error, and lose a nearest-neighbour one. The model ignores crosstalk, leakage and idle errors, so it flatters every device.

Depth brings in time. A shot takes about the circuit depth times the time per layer, plus readout and reset. Superconducting layers take tens of nanoseconds and their qubits decohere within tens of microseconds[4], so their budget is many fast layers; ions and atoms have far longer coherence but layers that take hundreds of microseconds or more. For a fixed number of shots that difference becomes hours. IBM's Phoenix can run up to 109.7 kHz[1] of simple circuits; QuEra's Gemini runs 1 shot per second[17]. A variational algorithm that needs millions of shots is practical on the first and not on the second.

Shots also set the bill. Vendors charge per shot, per task, per circuit size or per minute on the machine, and the same circuit can cost very different amounts on two types of hardware; what a run costs on each, and why, is in how much quantum computing costs. Width is the last limit: a circuit wider than a machine does not run at all, which today rules out photonic hardware for most workloads.

Where each type stands on error correction

Error correction is what will make any of these machines run long algorithms. The details, and what below threshold means, are on the quantum error correction page.

TypeBest result we foundWhat is still missing
SuperconductingGoogle's Willow: a surface-code memory below threshold, with logical errors falling by Λ=2.14±0.02\Lambda = 2.14 \pm 0.02[5] per step in code distance and a decoder that kept up in real time at distance 5[5]A larger Λ\Lambda, and many logical qubits computing together; fast cycles make real-time decoding hard
Trapped ionQuantinuum's Helios: a colour code with Λ=12\Lambda = 12[29] and 5×10−65 \times 10^{-6}[29] logical error per cycle, in a preprintDecoding was done after the run; slow cycles make every logical operation slow
Neutral atomHarvard, MIT and QuEra: a surface code 2.14 times[23] below threshold and logical circuits on up to 448 atoms[23]Deep circuits with many rounds, and handling atom loss at scale
PhotonicXanadu's Aurora: a distance-2 repetition code[27] with real-time decodingOptical loss far above the fault-tolerance budget
Analog machinesNo error correction in the gate sense; annealers and analog simulators run uncorrectedA way to correct an analog evolution at scale

How to choose a machine for your circuit

Start from the circuit, not the spec sheet.

  1. Write down width, depth and shots

    The number of qubits rules machines in or out. The two-qubit gate count after compilation and the number of shots decide error and time. An optimisation loop that needs many shots per step points to fast hardware.

  2. Look at which qubits interact

    If the gates follow a line or a grid, a lattice chip runs them with few SWAPs. If they reach across the register, as in QAOA on an arbitrary graph or a quantum kernel with all pairs, an all-to-all ion trap or a neutral-atom machine avoids the routing overhead.

  3. Route it and estimate the success probability

    Transpile the circuit onto each candidate's real coupling graph, count the two-qubit gates, and multiply out the published errors as in the formula above. EigenOrbit Benchmarks does this for every device it lists.

  4. Check the analog option only for its own problem

    An annealer runs QUBO and Ising problems; an analog atom simulator runs Rydberg dynamics. If your problem is one of those, compare it with tuned classical solvers first; if it is a general circuit, they do not apply.

  5. Run the same circuit on two or three types

    Specs predict, shots decide. Run one circuit on a superconducting chip and an ion trap, with the same number of shots, and compare the results with a simulator's exact answer.

    Run it on a real QPU with EigenOrbit RouterIn eigenkit: pip install "eigenkit[router]", then provider=RouterConfig(device=...) (docs).

Run it on a real QPU with EigenOrbit RouterThe same circuit on QPUs from several vendors and hardware types, with one account instead of one per vendor, to see which one your circuit suits.In eigenkit: pip install "eigenkit[router]", then provider=RouterConfig(device=...) (docs).

Questions

What is the difference between superconducting and trapped-ion qubits?

A superconducting qubit is an electrical circuit on a chip, cooled in a dilution refrigerator; a trapped-ion qubit is one charged atom held in a vacuum by electric fields. Superconducting gates are far faster and run many more shots per hour, but each qubit only reaches its neighbours. Trapped-ion gates are slower but, averaged over a machine, more accurate, and any qubit can interact with any other.

What are the main types of qubits?

The gate-based types in use today are superconducting circuits, trapped ions, neutral atoms and photons. Spin qubits in silicon and topological qubits are earlier in development. Quantum annealers and analog neutral-atom simulators also use qubits, but they evolve one physical model instead of running gate circuits.

Which quantum computer is best?

None is best for every job. Trapped ions have the lowest average gate errors and all-to-all connectivity, superconducting chips are the fastest, and neutral atoms reach the largest qubit counts. Which wins depends on the circuit: how wide and deep it is, how far apart its interacting qubits are and how many shots it needs. Routing the same circuit onto each machine and estimating its error is the fair comparison.

How does a neutral-atom quantum computer work?

Single atoms are held in optical tweezers, focused laser beams, in a vacuum cell at room temperature. Lasers drive single-qubit gates, and two-qubit gates excite atoms to Rydberg states, where one atom blocks its neighbour's excitation. Atoms are moved by tweezers so any pair can interact. Some neutral-atom machines run gate circuits; others, such as QuEra's Aquila, run analog simulations only.

What is a photonic quantum computer?

A quantum computer whose qubits are photons, steered through chips of waveguides, beam splitters and phase shifters, mostly at room temperature. Photons keep their quantum state well but are easily lost, and two-qubit gates succeed only part of the time, so today's photonic machines are small. PsiQuantum and Xanadu aim at fault-tolerant machines built from entangled photonic resource states.

Can I run the same circuit on different types of quantum computer?

Yes, on the gate-based ones. A circuit written in Qiskit or OpenQASM can be compiled for superconducting, trapped-ion and some photonic and neutral-atom machines, and each compiler maps it onto the hardware's own gates and connectivity. Quantum annealers and analog simulators do not accept circuits.

Sources

  1. IBM. IBM Quantum Platform: compute resources (live calibration data, read on 4 November 2026). IBM Quantum Platform, 2026.
  2. IBM. QPU information: how the calibration metrics are defined. IBM Quantum documentation, 2026.
  3. IBM. Processor types. IBM Quantum documentation, 2026.
  4. Google Quantum AI. Willow spec sheet. Google Quantum AI, 2024.
  5. R. Acharya et al. (Google Quantum AI and collaborators). Quantum error correction below the surface code threshold. Nature 638, 920–926, 2025. arXiv:2408.13687.
  6. H. Neven (Google Quantum AI). Building superconducting and neutral atom quantum computers. Google blog, 2026.
  7. Rigetti Computing. Rigetti announces general availability of 108-qubit system (press release, filed with the SEC as exhibit 99.1). Rigetti press release, 7 April 2026, 2026.
  8. Rigetti Computing. Cepheus-1-108Q instruction set architecture, with per-qubit and per-pair calibration (read on 4 November 2026). Rigetti Quantum Cloud Services API, 2026.
  9. E. Stuart (IQM). New on IQM Resonance: IQM Emerald, almost 3x qubits for the same reliable quality. IQM blog, 2025.
  10. Quantinuum. Performance validation: machine specifications of Helios-1, H2-1 and H2-2. Quantinuum systems documentation, 2026.
  11. Quantinuum. Quantinuum Helios product data sheet, version 1.2 (9 June 2026). Quantinuum, 2026.
  12. Quantinuum. Quantinuum System Model H2 product data sheet, version 4.00 (17 September 2025). Quantinuum, 2025.
  13. A. Ransford, M. S. Allman, J. Arkinstall, J. P. Campora et al., J. G. Bohnet. A 98-qubit trapped-ion quantum computer with all-to-all connectivity. Nature 655, 81–86, 2026. arXiv:2511.05465.
  14. IonQ. Compare IonQ quantum systems. IonQ, 2026.
  15. J.-S. Chen, E. Nielsen, M. Ebert, V. Inlek, K. Wright, V. Chaplin, A. Maksymov, E. Páez, A. Poudel, P. Maunz, J. Gamble. Benchmarking a trapped-ion quantum computer with 30 qubits. Quantum 8, 1516, 2024. arXiv:2308.05071.
  16. A. C. Hughes, R. Srinivas, C. M. Löschnauer, H. M. Knaack, R. Matt, C. J. Ballance, M. Malinowski, T. P. Harty, R. T. Sutherland. Trapped-ion two-qubit gates with >99.99% fidelity without ground-state cooling. arXiv preprint, 2025. arXiv:2510.17286.
  17. QuEra Computing. Gemini-class gate-model quantum computer. QuEra Computing (product page, read on 4 November 2026), 2026.
  18. QuEra Computing. Aquila: 256-qubit analog neutral-atom quantum computer. QuEra Computing (product page, read on 4 November 2026), 2026.
  19. Pasqal. Quantum, integrated into your HPC center: Pasqal Orion QPU specs. Pasqal (product page, read on 4 November 2026), 2026.
  20. Microsoft. Pasqal provider. Azure Quantum documentation (read on 4 November 2026), 2026.
  21. S. J. Evered, D. Bluvstein, M. Kalinowski, S. Ebadi, T. Manovitz et al., M. D. Lukin. High-fidelity parallel entangling gates on a neutral-atom quantum computer. Nature 622, 268–272, 2023. arXiv:2304.05420.
  22. D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz et al., M. D. Lukin. Logical quantum processor based on reconfigurable atom arrays. Nature 626, 58–65, 2024. arXiv:2312.03982.
  23. D. Bluvstein, A. A. Geim, S. H. Li, S. J. Evered, J. P. Bonilla Ataides et al., M. D. Lukin. A fault-tolerant neutral-atom architecture for universal quantum computation. Nature 649, 39–46, 2026. arXiv:2506.20661.
  24. H. J. Manetsch, G. Nomura, E. Bataille, X. Lv, K. H. Leung, M. Endres. A tweezer array with 6,100 highly coherent atomic qubits. Nature 647, 60–67, 2025. arXiv:2403.12021.
  25. Quandela. Belenos. Quandela (product page, read on 4 November 2026), 2026.
  26. K. Alexander, A. Benyamini, D. Black, D. Bonneau et al. (PsiQuantum). A manufacturable platform for photonic quantum computing. Nature 641, 876–883, 2025. arXiv:2404.17570.
  27. H. Aghaee Rad, T. Ainsworth, R. N. Alexander, B. Altieri et al. (Xanadu). Scaling and networking a modular photonic quantum computer. Nature 638, 912–919, 2025.
  28. D-Wave Quantum Inc. D-Wave announces general availability of Advantage2 quantum computer, its most advanced and performant system. D-Wave press release, 20 May 2025, 2025.
  29. A. Lavasani, A. P. Reed, J. M. Dreiling, A. J. Park et al., A. C. Potter. Quantum error correction well below the color code threshold. arXiv preprint, 2026. arXiv:2610.24077.

Run it yourself with eigenkit

eigenkit is EigenOrbit's open-source Python library: write a circuit once, run it on any simulator or QPU, and test it against a classical baseline on your own data before you invest in it or rule it out.

$ pip install eigenkit