Quantum optimisation.

Combinatorial optimisation has had seventy years of intense classical attention, driven by logistics, chip design, airline scheduling and finance — industries with enormous incentives and enormous budgets. The result is a body of solvers that are extremely good: commercial mixed-integer programming solvers, specialised heuristics, simulated annealing, and metaheuristics tuned over decades against real instances.

Comparisons against quantum methods often skip that baseline, or use a weak version of it. A demonstration that a quantum annealer beats naive random search is not evidence of anything. The comparison that matters is against a well-configured commercial solver given the same time budget on the same instance — and in the published comparisons we have reviewed that survive scrutiny, the classical solver generally wins or ties.

There is a second, quieter finding worth knowing about. Repeatedly, a proposed quantum advantage has prompted researchers to look harder at the classical side and find a better classical algorithm. The field calls this dequantisation, and it has happened often enough to be a reasonable prior. Sometimes the most valuable output of a quantum proposal has been the classical algorithm it provoked.

Four questions, in order. They are not hostile, and a serious vendor will have answers ready.

1. What was the classical baseline? Named solver, version, configuration, and how much time it was given. If the baseline is unnamed or the time budget is unequal, the comparison carries no information.

2. Is the problem instance representative? Quantum methods can be tuned to instances with structure that suits them. Ask whether the demonstration used your kind of instance at your kind of size, and whether the result held across a distribution or on one example.

3. Does the comparison include preparation time? Encoding a problem into a form the hardware accepts is real work with real latency, and it is frequently excluded from the quoted runtime.

4. What happens as the problem grows? A method that ties at small scale and degrades faster than the classical one as instances grow is not a path to anything. Ask for the scaling curve, not the single data point.

Our advice for a business with a real optimisation problem is unambiguous: solve it classically first, properly, with a good solver and someone who knows how to configure it. In our experience the achievable improvement from that alone is typically larger than anything currently on offer from quantum hardware, and it is available now.