Quantum Computing Modalities: Superconducting Cat Qubits

For other quantum computing modalities and architectures, see Taxonomy of Quantum Computing: Modalities & Architectures. For a step-by-step explanation of the encoding, see Cat Qubits 101.

What It Is

A superconducting cat qubit stores quantum information in an electromagnetic oscillator, typically a microwave resonator, whose dynamics are engineered to protect an encoded two-state system. Its two approximately distinguishable computational states resemble coherent oscillations with equal amplitude and opposite phase, written |α⟩ and |−α⟩. Superpositions of those states are the Schrödinger cats that give the encoding its name. A conventional transmon, by comparison, usually encodes information in the two lowest energy levels of an anharmonic superconducting circuit.

The defining attraction is noise bias: the hardware can make one class of error much rarer than another. Throughout this article, a bit flip exchanges the two opposite-phase computational states, while a phase flip changes their relative sign in a superposition. With suitable stabilization, bit-flip rates can decrease exponentially with the cat’s mean photon number. Single-photon loss still produces predominantly phase errors. This asymmetry can make the remaining error-correction problem substantially cheaper.

That protection has a specific scope. A long bit-flip lifetime does not mean that an arbitrary quantum state remains coherent for that long. Nor does the encoding remove every need to correct bit errors: residual faults, imperfect gates and correlations still matter. The opportunity is to design hardware, gates and codes around a favorable error distribution, as developed in repetition-cat architectures and subsequent work.

For a cryptographically relevant quantum computer, lower fault-tolerance overhead could materially change the engineering requirements. But there is no fixed conversion rate from transmons to cats, and no reliable CRQC date follows from a record lifetime. The relevant question is how many components, correction cycles and non-Clifford resources a complete architecture needs to execute a particular algorithm at an acceptable failure probability.

How It Works

The Cat State in Phase Space

A coherent state is a quantum state that closely resembles a classical oscillation. Its complex amplitude α specifies the field’s amplitude and phase; the coherent state’s mean photon number is |α|². Phase space plots two field quadratures, analogous to position and momentum. The states |α⟩ and |−α⟩ appear as two separated distributions on that plane.

Their overlap is ⟨α|−α⟩ = e^−2|α|², so they become nearly orthogonal as their separation increases. Strictly, an exact computational basis uses orthonormal combinations of the normalized even and odd cats; the coherent states are a useful approximation. The even and odd cats are proportional to |α⟩ + |−α⟩ and |α⟩ − |−α⟩. They remain superpositions of both lobes, even when the lobes are far apart.

Small overlap is part of the intuition, but protection also depends on the engineered dynamics and the physical noise. The explanation cannot simply be that every error must tunnel through an energy barrier. That picture is useful for some Kerr-cat behavior; dissipative stabilization works through an engineered reservoir. In either case, the desired protected space includes quantum superpositions between the lobes. Those superpositions are not an unstable region to be avoided. Guillaud, Cohen and Mirrahimi’s lecture notes provide a fuller treatment.

Stabilization: Keeping the Cat Alive

Dissipative cats. Two-photon driving and engineered two-photon dissipation confine the oscillator to the cat manifold. A common implementation couples the storage mode to a deliberately lossy buffer. In an ideal description, the engineered jump operator is proportional to a² − α², which annihilates both coherent states and every superposition of them. Pair loss alone would not sustain a finite-amplitude cat; the drive and dissipation must work together. This approach is central to Mirrahimi et al.’s 2014 proposal.

A single lost photon acts differently on the two coherent components: a|α⟩ = α|α⟩, whereas a|−α⟩ = −α|−α⟩. It therefore changes their relative sign rather than exchanging the lobes. Stabilization can keep the state within the protected manifold while that phase error remains. Correcting the phase information requires additional redundancy.

Kerr cats. A nonlinear oscillator with a two-photon drive can instead provide Hamiltonian confinement. The Kerr nonlinearity and drive create a protected structure in the oscillator’s rotating-frame dynamics. Grimm et al. demonstrated stabilization and single-qubit control in 2020. Hamiltonian confinement and attraction under engineered dissipation are distinct mechanisms, although practical devices can combine Hamiltonian control with dissipative techniques.

Squeezed and deformed cats. Reshaping the oscillator states and their stabilization can improve the tradeoff between the two error channels. The Rousseau et al. experiment, published as an accepted article in Nature Communications on 22 October 2026, reported a 160-fold improvement in bit-flip time: 22 seconds rather than 138 milliseconds, at operating points with the same 1.3-microsecond phase-flip time. The implemented deformation is known as a “moon cat.” These paired measurements show why bit-flip stability and quantum coherence must be reported separately.

New stabilization hardware. An August 2026 experimental preprint by Paradina et al., revised in September, demonstrated multi-photon dissipation using a DC-voltage-biased Josephson circuit. It provides a possible alternative to the parametric microwave pump used for the relevant coupling. This is a component-level advance toward cat stabilization, rather than a demonstration of a complete fault-tolerant processor.

Gates and Operations

Cat-qubit gates must control the encoded state while managing the error bias. In the coherent-state convention used here, a weak resonant single-photon drive can generate a Z rotation by giving the two lobes different phases. An X operation can exchange the lobes by rotating their positions through controlled changes to the stabilization. They can remain separated during this motion: a bit-flip gate does not universally require shrinking the cat to the origin.

Arbitrary rotations and entangling gates impose additional constraints. A bias-preserving operation must avoid turning the frequent phase faults into the rare bit-flip or combined bit-and-phase faults that the chosen code may not correct. Error propagation between qubits, leakage and correlated faults also need to remain manageable. Preserving a favorable ratio of errors is necessary for some architectures, but it is not sufficient to establish fault tolerance.

Theory has established useful bias-preserving gate constructions, including Puri et al.’s stabilized-cat gates. However, a universal set of ordinary single-qubit and two-qubit unitaries cannot all preserve a fixed dephasing bias: a Hadamard, for example, exchanges X and Z. Universal fault-tolerant computation instead depends on a complete encoded protocol using suitable operations, preparation, measurement and, where required, resource states.

Experimental progress now includes Ocelot’s cat–transmon syndrome-extraction gates and Alice & Bob’s October 2026 report of noise-biased cat–cat CNOTs. These are different interactions in different architectures. Meanwhile, a July 2026 AWS/Caltech proposal develops a coherent cat–cat CNOT based on a vacuum-conditional beam-splitter interaction; its projected logical performance remains a theoretical result.

The Repetition Code Architecture

A phase-flip repetition code distributes one encoded qubit across several cats. It measures joint checks such as X_iX_i+1, which compare neighboring cats without measuring the stored logical amplitudes. With ideal checks, an odd-length code of d data cats can correct up to (d − 1)/2 phase flips. Repeated measurements and decoding are needed when the checks themselves are noisy.

One layout uses d data cats and d − 1 syndrome ancillas, giving 2d − 1 qubits in that count. The ancillas can be transmons or cats, depending on the design. Storage resonators, buffer modes, readout circuits, nonlinear couplers and control hardware must also be counted when assessing actual engineering cost.

AWS’s Ocelot experiment combined five cat data qubits, four transmon syndrome ancillas and five buffers. Its paper reported a minimum overall logical error of 1.65 ± 0.03% per 2.8-microsecond correction cycle at distance 5, compared with 1.75 ± 0.02% averaged over its distance-3 sections. The phase-flip-correcting layer operated below threshold; the overall logical performance was approximately comparable across those distances.

The qualification matters. At fixed, finite bias, a repetition code leaves physical bit flips uncorrected, creating a floor on total logical reliability. Its “distance 5” protects the phase-error channel, whereas a distance-5 surface code protects both Pauli error channels. The familiar count of nine data-plus-ancilla qubits versus 49 for a rotated surface-code patch is therefore a layout comparison, not evidence of equal reliability or a measured fivefold computing advantage.

Architectures for demanding computations consequently go beyond the simplest repetition code. AWS’s 2025 hybrid cat–transmon study uses thin rectangular surface codes to address residual bit errors. Elevator Codes, proposed in January 2026 and revised in September, combine inner phase-error repetition codes with outer codes for bit errors. Cat hardware reduces the correction burden; the appropriate code still depends on the complete error model.

Key Academic Papers

The field has several distinct milestones. Early cat encodings, photon-loss-correcting bosonic memories, stabilized biased qubits and multi-cat error correction should not be treated as interchangeable achievements.

• Cochrane, Milburn & Munro (1999): an early proposal to use superpositions of coherent states as a bosonic encoding against amplitude damping.
• Leghtas et al. (2013; preprint 2012): a hardware-efficient circuit-QED proposal for autonomous error correction in a cavity.
• Mirrahimi et al. (2014): driven-dissipative cat encodings, distinguishing two-component and four-component schemes.
• Ofek et al. (2016): landmark break-even QEC with a four-component cat code. The corrected lifetime reached 320 microseconds, modestly exceeding the uncorrected |0⟩/|1⟩ Fock-state encoding in the same oscillator.
• Guillaud & Mirrahimi (2019): fault-tolerant computation using repetition codes built from biased cat qubits.
• Grimm et al. (2020) and Puri et al. (2020): experimental Kerr-cat control and the theory of bias-preserving gates, respectively.
• Hoshi et al. (2025): entangling operations between cats in Kerr parametric oscillators, distinct from a complete logical error-correction experiment.
• Ruiz et al. (2025): LDPC-cat architecture and resource estimates for low-overhead memory and computation.
• Putterman et al. (2025): Ocelot’s concatenated bosonic logical memory. See also my Ocelot analysis.
• Ruiz et al. (2026): unfolded magic-state distillation, a theoretical scheme for exploiting noise bias in non-Clifford resource preparation.
• Rousseau et al. (2026; preprint 2025): experimentally improved protection through squeezing-inspired cat deformation.

The Vendor and Research Landscape (October 2026)

Alice & Bob

Alice & Bob is developing a computing architecture centered on dissipative cat qubits. Its progress needs to be read at three levels: individual-qubit protection, integrated error correction and the proposed full machine.

The widely reported September 2025 “hour-long” result concerns bit stability. In the company’s detailed experimental account, the controlled single-qubit device had an estimated bit-flip lifetime of 44 minutes, with a 95% confidence interval of 33–60 minutes, alongside a phase-coherence time of 420 nanoseconds. Separate Helium 2 measurements showed longer bit-flip times, but the company cautioned that equivalent quantum control had not been established for those measurements. Neither result means hour-long preservation of an arbitrary quantum state.

In June 2026, the company announced its Helium Quantum System, designed to encode a logical qubit using 18 cats and support a future 48-cat processor. That system announcement should be distinguished from a demonstration that a logical qubit has reached the required reliability.

The more recent 20 October 2026 PROQCIMA report describes three experimental demonstrators: two six-qubit chips and one twelve-qubit chip. Alice & Bob reported noise-biased cat–cat CNOT gates, auxiliary-qubit measurement and reset, repeated distance-3 correction and a physical non-Clifford gate. It also explicitly stated that the measured logical error rate was not yet below threshold. These are company-reported integration milestones, with further improvements required.

Classical processing is also advancing. An October 2026 NVQLink demonstration decoded LDPC-cat syndromes on GPUs using synthetic data. The company explicitly said it did not yet constitute real-time decoding. Likewise, the earlier 9.25-fold CUDA-Q speedup concerned a simulated-syndrome decoding workload, rather than a measured acceleration of hardware QEC cycles.

AWS and Kerr-Cat Research

AWS combines dissipative cat data qubits with transmon ancillas. Ocelot’s significance is the integration of stabilization, syndrome extraction and repeated correction on a chip platform. AWS’s projection of up to a 90% reduction in error-correction overhead concerns a future scaled architecture; it is not a demonstrated cost reduction for a useful quantum computation.

University and research groups also pursue Kerr-cat hardware, gate control and benchmarking. A useful 2026 result is Qing et al.’s benchmarking of a detuned Kerr-cat qubit: the study shows why operational noise bias needs direct characterization. Inferring gate performance from the ratio of idle lifetimes alone can overstate the available protection.

Related Bosonic Platforms

Nord Quantique belongs in the broader bosonic-computing landscape. Its demonstrated work includes Gottesman–Kitaev–Preskill (GKP) error correction, and its subsequent work includes multimode grid-state codes. These use a different encoding from the two-component cats discussed here.

Quantum Circuits developed dual-rail bosonic qubits, encoding information in which of two modes contains a photon and using photon loss as a detectable erasure. D-Wave completed its acquisition of the company in January 2026. Dual-rail, GKP, binomial and cat codes all exploit oscillators, but their protection mechanisms and hardware requirements differ.

Comparison to Other Modalities

Cat Qubits vs. Standard Transmon Superconducting Qubits

Both approaches use superconducting microwave hardware and cryogenic infrastructure, and they can coexist in one processor. The difference is the encoding and the resulting noise model. Transmon relaxation and dephasing are not universally equal, and transmons are not restricted to surface codes. Stabilized cats deliberately seek a particularly strong bias that can support asymmetric codes and reduce some overheads.

The engineering tradeoff is that a cat data qubit can require a storage resonator, stabilization circuitry, pumps and nonlinear couplings. A transmon also needs control and readout hardware. Comparing named “physical qubits” is therefore only a starting point. A meaningful comparison includes full logical error rates, gate times, connectivity, calibration demands and the total hardware needed for the same workload.

Cat Qubits vs. Trapped Ions

Trapped-ion platforms exploit long-lived atomic states and flexible interactions; superconducting cats exploit engineered oscillator protection and microwave control. A high raw gate fidelity, a small encoding ratio and a long idle lifetime measure different capabilities. None determines the winner by itself. The appropriate comparison is a complete fault-tolerant workload, including the time needed to perform it and the errors introduced during repeated correction.

Cat Qubits vs. Neutral Atoms

Neutral-atom platforms offer large reconfigurable arrays and avoid the dilution refrigerators used by superconducting circuits. The atoms themselves are laser-cooled to very low temperatures; describing them simply as “room-temperature qubits” is misleading. Their connectivity, atom-loss handling and logical-code experiments pose a different set of tradeoffs from cat stabilization and microwave integration. A reported logical-qubit count is meaningful only with its code, operation set and achieved error suppression.

Cat Qubits vs. Topological Approaches

Topological approaches seek protection through nonlocal encoding and topological properties. Cat qubits engineer a protected oscillator manifold and a favorable noise bias. Neither principle makes a practical device immune to every error: operations, unwanted excitations and environmental coupling remain relevant. Cat experiments already provide measurable evidence about their specific protection mechanism; claims about another architecture must be assessed against the protection and operations actually demonstrated there.

Advantages

A useful error structure. Suppressing one error channel in hardware can make code design and decoding substantially more efficient. The benefit depends on preserving the bias during computation, but it gives designers a degree of freedom beyond simply improving every physical error rate equally.

Potentially lower total QEC overhead. Repetition-cat, LDPC-cat and asymmetric surface-code proposals offer different ways to exploit the bias. They should be assessed at matched target reliability and with a consistent accounting of resources. The potential saving is an architectural result, rather than a universal property of one isolated oscillator.

Compatibility with superconducting engineering. Cat processors can draw on established microwave fabrication, dilution refrigeration, packaging and quantum control infrastructure. The shared ecosystem helps, although cat stabilization introduces its own requirements for nonlinear elements, filtering and drive management.

Experiments at several levels of integration. The evidence now spans single-cat protection, oscillator control, two-qubit operations and repeated multi-cat error correction. This makes it possible to evaluate specific bottlenecks with hardware data, while keeping those achievements distinct from a completed universal fault-tolerant machine.

Disadvantages and Open Engineering Problems

Phase errors remain costly. Increasing an ordinary cat’s photon population usually increases the rate of photon-loss-induced phase faults. Cat size must therefore be optimized with the code and gate schedule. Maximizing an idle bit-flip lifetime need not minimize logical error during computation.

Residual bit errors eventually matter. A large bias is finite. Long computations can require another code layer or an asymmetric code that also corrects bit faults. Leakage, correlated errors and rare device events cannot simply be omitted because the dominant idealized error is a phase flip.

Operations must retain the advantage. Syndrome extraction, measurement, reset and non-Clifford operations can each reduce the useful bias or introduce additional faults. There is no universal two-qubit fidelity, such as 99.5%, that proves success. The test is whether the full logical protocol reaches its required reliability and throughput.

Scaling adds control and fabrication demands. High-quality storage modes, reproducible nonlinear couplings, pump stability, crosstalk control and automated recalibration all matter. A promising result on one device must survive simultaneous operation across an entire register.

Cryogenic and classical infrastructure remain substantial. Superconducting cat processors still require millikelvin hardware, appropriate cryogenic infrastructure, and suitable facilities. Error decoding must also keep up with the machine. Faster offline simulation or batch decoding is useful progress, but does not establish a real-time control loop.

Impact on Cybersecurity

Cat qubits matter to cybersecurity because fault-tolerance overhead is a major part of the resource cost of quantum cryptanalysis. The defensible way to quantify their impact is through an end-to-end resource estimate with an explicit architecture, noise model, algorithm, runtime and success target.

One such study is Ruiz et al.’s 2025 LDPC-cat architecture. It estimated 95,941 cat qubits and approximately seven days to factor an RSA-2048 integer. Its assumptions included a single-photon-loss to engineered two-photon-dissipation ratio of 10⁻⁵ and a mean photon number of 21. The corresponding earlier repetition-cat estimate used approximately 349,000 cats and four days. These are modeled machines, not demonstrated processors.

The often repeated “up to 200×” advantage is tied to particular comparisons. Alice & Bob’s original announcement compares the approximately 100,000-cat estimate with the older estimate of roughly 20 million conventional physical qubits. It is not a multiplier that can be applied to any newer logical-qubit count or surface-code estimate.

For example, Gidney’s May 2025 analysis estimated RSA-2048 factoring in under a week with fewer than one million noisy qubits under its specified surface-code and hardware assumptions. Combining pieces of that analysis with an unrelated “five physical qubits per logical qubit” ratio or a vendor’s 200× comparison does not produce a valid cat-qubit estimate. Routing, error targets, resource-state production and execution time must be recalculated together.

For practitioners, the implication is that several architectures may reduce the resources needed for cryptanalysis. That strengthens the case for tracking capability rather than relying on one hardware roadmap. Post-quantum migration should be planned around the lifetime of sensitive information, the time needed to change deployed systems and the requirements that apply to the organization. A cat-qubit announcement can change the evidence behind a risk assessment; it cannot by itself provide a dependable Q-day forecast.

Future Outlook

Alice & Bob’s roadmap targets a 100-logical-qubit fault-tolerant machine by 2030. That is a development target, and the roadmap deliberately avoids assigning dates to intermediate scientific milestones. The public evidence does not justify inventing a fixed year for high-fidelity cat gates, cloud deployment or cryptanalytic capability.

The next decisive results should connect the pieces already demonstrated: stable multi-cat hardware, repeated syndrome extraction, lower logical error with added protection, and reliable operations between logical qubits. For universal computation, resource-state preparation and non-Clifford operations must also work within the same error budget. The 2026 unfolded-distillation paper, discussed in my analysis of Alice & Bob’s magic-state work, offers an architectural proposal for reducing that overhead; it does not report an operating magic-state factory.

Cat qubits are best understood as an encoding and protection strategy within superconducting computing, with related ideas extending beyond that platform. Their value will be determined by whether the measured noise bias translates into reliable logical operations at lower total resource cost. The 2025–2026 results make that question increasingly concrete: the task now is to improve and scale the complete error-corrected system.

From the author