Qubit count is an inventory number, not a capability measure
A quantum processor can contain many qubits and still be unable to execute a useful circuit. Physical qubits decohere, gates introduce errors, measurements are imperfect, and control operations can create correlated faults. Capability depends on the usable circuit that survives this noise: how many qubits participate, how deep the circuit runs, which operations are available, and whether the output can be trusted. Reporting only the largest qubit count is comparable to describing a telescope by the number of components while omitting resolution and calibration.
NIST notes that contemporary quantum computers have hundreds of interconnected qubits and roughly one error per thousand operations, while many advanced algorithms require errors to be suppressed far more strongly. The relevant unit is increasingly the logical qubit: quantum information encoded across physical qubits with errors detected or corrected. Even then, one must ask what logical operations were performed, at what error rate, for how many cycles, and with what overhead. A logical memory is not yet a logical processor.
Read demonstrations as claims with boundaries
IBM's 2023 Nature study ran circuits on 127 superconducting qubits with up to 60 layers of two-qubit gates and 2,880 CNOT operations. Error-mitigated expectation values agreed with exactly verifiable cases, and the authors explored regimes where selected classical tensor-network approximations became difficult. The careful interpretation is the paper's own: evidence for utility before fault tolerance and a foundation for testing candidate applications. It did not establish a generally useful quantum advantage or a production workload that beats the best classical method on cost, time and accuracy.
That distinction matters because classical algorithms improve in response to quantum claims. A credible study publishes the circuit, error model, mitigation method, sampling cost and verification route. It compares against relevant classical baselines rather than brute force alone, and it states whether the advantage is asymptotic, empirical or projected. Benchmark experiments are valuable metrology. They become misleading only when a narrow result is translated into broad economic capability without carrying its assumptions forward.
Below-threshold error correction changes the scaling question
Google's Willow surface-code result provides a more structural metric. Increasing code distance by two suppressed the logical error rate by a measured factor of 2.14 ± 0.02, culminating in a distance-7 memory using 101 qubits with 0.143% ± 0.003% logical error per correction cycle. The logical lifetime exceeded that of the best constituent physical qubit by a factor of 2.4 ± 0.3. This is important because additional redundancy improved the encoded information rather than making it worse, the signature of operation below threshold.
The result remains a memory experiment, not a fault-tolerant application. Useful computation requires many logical qubits, a universal set of logical gates, fast decoding, state preparation, routing and repeated operations at a total failure probability compatible with the algorithm. Progress should therefore be tracked as a vector: error-suppression factor, logical error per cycle, logical gate fidelity, decoder latency, physical-qubit overhead and duration. No single scalar can substitute for this system view.
Different platforms illuminate different bottlenecks
A neutral-atom team reported programmable experiments with up to 48 logical qubits, including a four-dimensional hypercube circuit on 48 logical qubits encoded in 128 physical atoms. Superconducting systems emphasize fast gates and integrated control; trapped ions offer high-fidelity operations and flexible connectivity; neutral atoms provide reconfigurable arrays. These demonstrations should not be reduced to a race over one headline number. Each platform trades speed, fidelity, connectivity, loss, cooling, control complexity and manufacturability differently.
The research question is whether an architecture can move from a controlled demonstration to repeatable fault-tolerant operations with acceptable resource overhead. Cross-platform comparison should normalize the task, error definition, post-selection, runtime and classical assistance. A result obtained after discarding failed runs is scientifically informative, but it is not equivalent to deterministic computation. Likewise, encoded qubits that store information are not directly comparable with encoded qubits that execute deep logical circuits.
Application value requires end-to-end resource accounting
An application claim begins with a problem whose quantum algorithm has a defensible advantage, then includes data loading, error correction, logical gates, measurement, classical pre- and post-processing, and the probability of obtaining an acceptable answer. Chemistry and materials simulation are plausible long-term candidates because quantum systems naturally represent quantum states, but practical advantage depends on required precision and resources. Optimization and machine-learning proposals often rely on heuristics and must be compared against rapidly improving classical solvers.
Cryptanalytic resource estimates illustrate why assumptions matter. A 2025 preprint estimates that RSA-2048 could be factored in less than a week with fewer than one million noisy qubits, assuming a nearest-neighbour grid, 0.1% gate error, one-microsecond surface-code cycles and ten-microsecond control reaction. This is a major reduction from earlier estimates, but it is neither evidence that such hardware exists nor a date prediction. It is a conditional engineering calculation that informs security planning.
A practical scorecard for quantum claims
For hardware, record physical error distributions, coherence, connectivity, calibration stability and yield. For error correction, record code distance, logical error, suppression with scale, decoder latency and the fraction of runs retained. For computation, record logical gate set, circuit depth, wall-clock time and total success probability. For applications, publish the full resource estimate, best classical baseline, verification method and sensitivity to assumptions. A claim should state whether each number is measured, simulated, estimated or targeted.
The National Academies concluded that significant scientific and engineering challenges remain on the path to scalable quantum computers. That remains a productive stance: neither dismissal nor inevitability. Below-threshold memories, encoded processors and controlled many-body experiments are real progress. Commercially decisive fault-tolerant applications remain an open engineering programme. The most credible organizations will make that gap visible, use milestones that retire specific uncertainties, and avoid turning a laboratory benchmark into a promise the evidence does not support.
Scope and limitations
The cited experiments use different devices, codes, tasks and definitions, so their metrics are not a league table. Some application-resource figures are simulations or scenario estimates rather than measurements. Hardware and classical algorithms continue to change; any procurement or security decision should revisit the underlying assumptions and latest peer-reviewed evidence.
References
Source review: 20 August 2026. Quantitative values retain their original definitions, periods, and boundaries.
- 01Evidence for the utility of quantum computing before fault tolerance
Nature · 2023
www.nature.com ↗ - 02Quantum error correction below the surface code threshold
Nature · 2025
www.nature.com ↗ - 03Logical quantum processor based on reconfigurable atom arrays
Nature · 2023
www.nature.com ↗ - 04How to factor 2048 bit RSA integers with less than a million noisy qubits
arXiv · 2025
arxiv.org ↗ - 05Quantum Computing Explained
National Institute of Standards and Technology · 2025
www.nist.gov ↗ - 06Quantum Computing: Progress and Prospects
National Academies Press · 2019
nap.nationalacademies.org ↗