Quantum computing

Quantum machine learning.

A real research field, with real results, and no demonstrated advantage over classical methods on the kind of data businesses hold. That last clause is the whole story, and it is the clause most often left out.

WHAT THE FIELD ACTUALLY BUILDS

Variational quantum circuits

A parameterised quantum circuit whose parameters are tuned by a classical optimiser — structurally analogous to training a neural network, with the circuit as the model. The dominant approach on near-term hardware because the circuits are shallow enough to run.

Quantum kernel methods

Using a quantum circuit to compute a similarity measure between data points that is hard to compute classically, then feeding that kernel to an ordinary classical algorithm. Elegant, and the case where a provable separation is easiest to construct — on carefully chosen data.

Quantum-inspired classical methods

Algorithms derived from quantum reasoning that run on classical hardware. Repeatedly, a proposed quantum speedup has prompted someone to find a classical algorithm nearly as fast — a pattern known as dequantisation, and a genuinely useful output of the field.

Learning on quantum data

Models trained on measurements from physical quantum systems rather than on classical datasets. This is where the strongest evidence of advantage sits, and it is a different problem from the one most businesses have.

THE HYBRID LOOP

VARIATIONAL CIRCUIT TRAINING

Classical data

your records, images, text
Encode

the expensive step
Quantum circuit

parameterised gates
Measure

many shots, statistical
Classical optimiser

updates the parameters
Repeat

until convergence, or budget

Solid: one training iteration Dashed: the loop, run thousands of times

Almost every near-term quantum machine learning method is this loop. Note where the classical work sits: at both ends.

THE DATA-LOADING PROBLEM

This is the objection that separates a serious proposal from a pitch, and any claim of quantum advantage in machine learning has to answer it.

To process classical data on a quantum computer, the data must first be encoded into a quantum state. For a dataset of any realistic size, that encoding is itself expensive — in many proposed schemes, expensive enough to consume the entire theoretical speedup before any computation happens. It is not a matter of faster interconnects; it is a structural feature of how quantum states are prepared.

Which is why the promising results involve data that is already quantum. There is no loading step for a measurement taken from a physical quantum system, so the bottleneck does not arise. That is a genuine and interesting research direction. It is also, for a company whose data is invoices and support tickets, not applicable.

When a vendor proposes quantum machine learning on your business data, the question that settles it is: how is our data encoded into the quantum state, and what does that encoding cost relative to the claimed speedup? A good answer is technical and specific. In the proposals we have seen, the absence of an answer is the most common outcome, and it is dispositive.

WHERE THE EVIDENCE STANDS

There is recent published work demonstrating performance separations at small scales — on the order of tens of noisy qubits — under carefully constructed conditions. These are real results and worth reading. What they are not is evidence that a business classification or forecasting task will run better on quantum hardware, and the papers themselves are generally careful about that distinction in a way that the coverage of them is not.

On classical real-world data — customer records, images, sensor streams, text — there remains no demonstrated quantum advantage for machine learning that survives a fair comparison against a well-tuned classical baseline. After more than a decade of concerted effort and considerable funding, that absence is itself the most important finding in the field, and treating it as a temporary state of affairs awaiting better hardware is an assumption rather than a conclusion.

Our position, stated plainly so it can be held against us: if your machine learning problem is hard, the highest-expected-value interventions are better data, better features, a better classical model and a better evaluation set — in that order. Quantum is not on the list for a production system in 2026. We will update this page when that changes, and the change will be visible in the literature before it is visible in a vendor deck.

QUESTIONS

QUESTIONS — 3

There is no proposed mechanism by which they would, and several reasons to think they would not — the data-loading bottleneck, the scale of the parameter counts involved, and the fact that transformer training is dominated by dense linear algebra, which is exactly what classical accelerators are optimised for. Treat any such claim as requiring extraordinary evidence.

Genuinely useful and frequently overlooked. Several run on ordinary hardware today and have produced real improvements in sampling and optimisation. If a vendor offers quantum-inspired methods, that is a more credible conversation than a quantum one, and the honest framing is that it is a classical algorithm with an interesting intellectual origin.

If your organisation has a research function and a long horizon, a small ongoing investment in understanding is defensible — mostly so you can evaluate claims. If you are trying to solve a business problem this year, the same budget spent on data quality will return more, and we would say so even though it means a smaller engagement for us.

Keep reading

Resource

How to Choose a Model

→

Technology

Evaluating AI Systems

30-minute technical call · no deck