Quantum cloud test shows why qubit counts are not enough

A quantum processor’s qubit count describes its physical register, not how reliably those qubits work together. A preprint posted on November 10, 2026, examines that gap by testing nine cloud-accessible processors from IBM, IQM, Rigetti, AQT and Oxford Quantum Circuits. Eight use superconducting circuits; one uses trapped ions.

The study measures a particular capability: preparing a Greenberger–Horne–Zeilinger, or GHZ, state. It is not a general ranking of computing power, and its largest reported sizes require a statistical qualification.

What the shared state tests

An ideal GHZ state is an equal quantum superposition of every participating qubit being zero and every participating qubit being one. Crucially, it preserves coherence between those possibilities. Merely obtaining all-zero and all-one measurement results would not establish that coherence: a classical mixture could produce the same outcomes.

Preparing and checking a GHZ state therefore probes more than individual components. Errors in entangling operations, loss of coherence while qubits wait, unwanted interactions and measurement mistakes can all reduce the observed fidelity—the overlap with the ideal target state.

For this target, fidelity above one-half is sufficient to demonstrate genuine multipartite entanglement: entanglement involving the complete group rather than a mixture of states separable across different divisions. Fidelity at or below one-half does not prove that such entanglement is absent; this particular test may simply fail to detect it.

A cloud-service benchmark, not a hardware-only comparison

The researchers used direct fidelity estimation, randomly sampling 2,952 measurement settings for each tested state size. This exploits the mathematical structure of GHZ states to estimate fidelity without reconstructing the entire quantum state. At fixed statistical precision, the protocol’s sample requirement does not grow with qubit count, although preparing and measuring larger circuits still involves more hardware operations.

The main experiments used an uncertainty allowance of 0.1 and a stated confidence level of 90% for each fidelity estimate. No additional error mitigation was applied.

The software workflows were not identical. IBM and Qmio circuits used Qiskit’s highest standard optimization level, while Amazon Braket devices used provider-managed compilation. These workflows can select physical qubits differently. The comparison consequently measures the processor together with its execution software, not an isolated hardware limit.

Passing is not the same as certifying

The search accepted a state size whenever its estimated fidelity plus the uncertainty allowance reached one-half. With the chosen allowance, an estimate of just 0.4 could pass.

That means the data did not exclude fidelity at the entanglement threshold. It does not mean the experiment established fidelity above it. The stricter certification rule instead requires the estimate minus its uncertainty allowance to reach the threshold, corresponding here to an estimate of at least 0.6.

Under the permissive rule, IBM Phoenix reached 27 qubits, Basquecountry 24 and Pittsburgh 21. IQM Emerald reached 20, IBM Miami 18 and IQM Garnet 13. AQT IBEX Q1 passed with all 12 of its qubits; Rigetti Cepheus-1 reached four and Qmio three.

Phoenix’s 27-qubit fidelity estimate was about 0.438. Calling that certified 27-qubit entanglement would misrepresent the result.

Only four processors met the stricter certification rule at any sampled size: Phoenix and Basquecountry at 16 qubits, Garnet at eight and IBEX Q1 at 12. The other processors were not exhaustively tested at smaller sizes, so their absence from this list does not establish that they cannot certify entanglement.

Why there is no universal winner

IBEX Q1’s full-register result had an estimated fidelity of about 0.711. Its available register, rather than a measured failure at larger sizes, stopped the search. Expressing performance as a percentage of installed qubits therefore favours this smaller machine; the larger IBM systems supported larger passing states in absolute terms.

Device-wide median calibration figures also failed to predict the ordering reliably. Pittsburgh had better median gate and readout errors than Basquecountry, yet its passing boundary was smaller. The researchers suggest that measurement errors, waiting time, unwanted interactions and the specific qubits selected all matter. Their measurements do not isolate those contributions. A supplementary calibration-aware qubit-selection experiment on Rigetti did not improve its passing boundary.

Each size in the main comparison was measured once, and the search assumed fidelity would decrease with size. Some results violated that assumption, so the reported boundary is not an exhaustive maximum. Limited repeat experiments on three IBM processors found mostly stable boundaries despite fluctuating fidelities, but do not establish stability across every platform.

The preprint’s useful lesson is narrower than a vendor league table: qubit counts, software choices and statistical decision rules must be read together before a benchmark result becomes a claim about usable quantum capability.