Hybrid quantum proof of concept

A quantum proof of concept is worth doing when you already have a specific problem, a classical baseline, and a question that only an experiment can answer. It is not worth doing to find out whether “quantum works for us.” That question belongs in use-case discovery, which is cheaper and faster.

When the conditions are right, a PoC gives you something slides can’t: your data, run through a real hybrid quantum-classical workflow, on real hardware, measured against the best classical method you have. The result is usually “not yet, and here is when to look again.” Sometimes it is “this formulation scales better than we expected, keep going.” Either answer is useful if it is honest.

What can a quantum proof of concept prove?

A quantum proof of concept can prove that a formulation of your problem is correct and runs end to end, show how solution quality changes as the problem grows, and estimate what hardware the full-size problem would need. On 2026 hardware it cannot prove business-level quantum advantage.

The reason is size and noise. Current machines have somewhere between tens and a few hundred physical qubits in commercial access, with error rates that limit how deep a circuit can go before the output turns to noise. Error mitigation helps, at the cost of many more runs. So the problem you test is a reduced version of the one you care about, and at that reduced size a good classical solver will almost always win.

What you learn instead is the trend. If the quantum approach degrades gracefully as you add variables, and the resource estimate for the full problem lands within reach of published hardware roadmaps, the problem is worth watching closely. If it degrades fast, or the estimate needs millions of physical qubits, you know that too.

Picking the algorithm

The problem decides the method, and most PoCs use one of three.

QAOA (the quantum approximate optimization algorithm) runs on gate-based machines. It encodes an optimization problem, typically as a QUBO, into a parameterized circuit, and a classical optimizer tunes the parameters over many runs. It suits routing, scheduling and assignment problems with binary decisions.

VQE (the variational quantum eigensolver) uses the same hybrid loop to estimate the lowest energy of a molecule or material. It is the usual starting point for chemistry and materials questions, for example comparing candidate catalysts or battery electrolytes at small scale.

Quantum annealing is a different kind of hardware, built specifically to find low-energy states of optimization problems. Annealers have many more qubits than gate-based machines but less flexibility, and whether they beat classical heuristics is still debated problem by problem. When D-Wave claimed quantum supremacy on a materials simulation in March 2025, EPFL researchers Mauron and Carleo argued in a March 2025 preprint that large-scale classical simulations could handle those problems. We test them when the problem maps cleanly onto their connectivity.

All three are hybrid. A classical computer does most of the work (preparing the problem, running the optimizer, post-processing results), and the quantum processor handles one step inside the loop.

Simulators first, hardware second

Every PoC starts on a simulator, because simulators are free of hardware noise, cheap to run, and let you check that the formulation is right before paying for QPU time. Qiskit Aer, the simulators in Amazon Braket and NVIDIA’s CUDA-Q are the common choices.

Exact state-vector simulation needs memory that doubles with every qubit, so it tops out somewhere in the 30s to low 40s of qubits even on large GPU clusters. Tensor-network simulators go further for circuits with limited entanglement. Past that point, and to learn how noise affects your specific circuits, you need real hardware.

Which cloud quantum computers can you use?

A quantum proof of concept can run on three cloud platforms that cover most of the market: IBM Quantum for IBM’s superconducting processors, Amazon Braket and Azure Quantum for machines from several vendors under one account, plus D-Wave’s Leap service for annealing. None of them requires buying hardware.

IBM Quantum gives access to IBM’s own superconducting processors, including the Heron family, through Qiskit; in November 2025 IBM also announced Nighthawk, a processor with more than 20 percent more couplers than Heron. Amazon Braket offers several vendors under one API, among them AQT and IonQ trapped-ion systems, Rigetti and IQM superconducting processors, and QuEra’s neutral-atom machines. Azure Quantum offers IonQ, Quantinuum, Rigetti and Pasqal, and Microsoft publishes which providers are available in each region, which matters for teams in Latin America. For annealing, D-Wave sells access to its Advantage and Advantage2 systems through its own Leap service.

We choose based on the problem, the qubit connectivity it needs, queue times, and your cloud contracts. Running the same circuits on two different hardware types is often part of the PoC, because results can differ more than people expect.

Benchmarking against classical, done properly

Each run is compared with a classical baseline on the same instance, measured on solution quality, wall-clock time (including queue time) and cost. Without that, a PoC result means nothing.

The baseline is a real one: a commercial or open-source solver (Gurobi, CPLEX, OR-Tools, or a tuned heuristic like simulated annealing) configured by someone who knows it, or for chemistry, established classical methods run at comparable accuracy. If your team already has a production solver, that is the bar.

We report the unflattering numbers too. If the classical solver finds a better answer in two seconds and the quantum run took four hours of queue time, the report says so on the first page.

What are the phases of a quantum PoC?

A quantum PoC for one problem usually runs eight to twelve weeks in five overlapping phases: scoping the hypothesis, formulating the problem alongside a classical baseline, validating on simulators, running on cloud hardware, and reporting results with a resource estimate and a go, wait or stop recommendation.

The code and notebooks stay with your team. If you want them to keep the work going, pair the PoC with training for your engineers, or later with quantum software development if the results justify it.

What drives the cost of a quantum PoC?

The cost of a quantum PoC is driven mostly by people time, not hardware. Formulating the problem, building the classical baseline and analyzing results take more hours than the quantum runs, and QPU charges, which depend on the provider and the number of shots and tasks, are usually the smaller line item.

We estimate both once the problem is scoped.

Horizon matters here. We list this service as “near” because the hardware that would make a PoC decisive for most businesses doesn’t exist yet. For companies in logistics and transport or mining and materials with a well-defined problem and a strong baseline, running a PoC now builds the formulation, the data pipeline and the team knowledge that let you move quickly when it does.

AndesQubit is opening engagements in stages. If you have a candidate problem and a baseline, send us a short description.