Marin Ivezic
Two results that look like reports from different fields
In 2022, a team led by Hsin-Yuan Huang showed in Science that for specific families of tasks, absorbing quantum data coherently reduces the number of experiments required exponentially, compared with measuring first and analyzing the results classically. They demonstrated it on circuits of up to 40 superconducting qubits (Science 376, 6598).
On ordinary tabular business data, meanwhile, logistic regression and gradient-boosted trees remain the models to beat, and no quantum learning method has displaced them in a way the field accepts.
Both statements are honest, and read side by side they look like dispatches from two different fields. In a useful sense they are. They belong in different columns of the same ledger, and keeping those columns apart is most of what quantum machine learning diligence consists of in 2026.
A definition first, because the term gets stretched. By quantum machine learning we mean the headline version, the one the funding rounds are selling: using a quantum computer to learn from data, to train models, to classify and generate, faster or better than a classical machine can. There is a quieter reverse direction, in which machine learning runs quantum hardware better through error-correction decoders and calibration routines. That half already earns its keep, and it is a separate subject.
Three columns
We keep a three-column scorecard, and we ask learners to place a claim in one of them before arguing about it.
Learning from quantum systems
Quantum data means measurements of something that is itself quantum: a molecule, a material, a quantum sensor, or the output of another quantum device. The Huang result belongs here. It is proved in theory and demonstrated on hardware, and it is also narrow, because the advantage holds for particular families of questions about quantum processes rather than for data that started life in a database.
Constructed and highly structured tasks
In 2021, Yunchao Liu, Srinivasan Arunachalam, and Kristan Temme built a classification problem out of the discrete logarithm (Nature Physics 17, 1013–1017). An efficient quantum kernel method – a similarity score between two data points computed by a quantum circuit – solves it, while no efficient classical learner does meaningfully better than guessing, provided discrete logarithms stay classically hard. Two conditions travel with the result. The dataset was engineered from a cryptographic problem precisely so that quantum structure would pay, and the classifier requires a fault-tolerant machine that nobody has built.
This column shows that the door can open under a standard hardness assumption. It says nothing about whether your data walks through it.
Ordinary business data
This column is empty. On customer records, images, sensor streams, and text, no quantum learning result commands broad acceptance as an end-to-end advantage once a properly tuned classical baseline and the full hardware cost are counted. More than a decade of work and a great deal of capital have not filled it. A pitch that assumes otherwise is asking you to price an outcome the literature has not produced.
Why the third column stays empty
Three mechanisms explain the gap, and each one tells you where not to expect a miracle.
Data loading. Running a quantum algorithm on classical data means encoding that data into a quantum state first. For generic data, the encoding can cost as much as the computation it was supposed to accelerate, which quietly erases the speed-up. Scott Aaronson named the trap in 2015, in an essay called “Read the Fine Print”, and it hasn’t gone away.
The size of the prize. Some quantum approaches to learning and optimization inherit only a quadratic speed-up. A quadratic gain is real and fragile at the same time. Once you add the constant factors and error-correction overhead of a fault-tolerant machine, the problem has to be enormous before the quantum route pulls ahead. This applies to search-based proposals, not to kernel or spectral methods, which have different bottlenecks and deserve separate treatment.
Dequantization. In 2018, Ewin Tang showed that granting a classical algorithm sampling access comparable to what the quantum recommendation-system proposal quietly assumed makes the exponential separation disappear (arXiv:1807.04271). The pattern repeated through a 2021 reassessment by Jordan Cotler, Hsin-Yuan Huang, and Jarrod McClean (arXiv:2112.00811). Dequantization drew a boundary rather than ending the field, and the boundary is the useful part: wherever a classical sampler can imitate the access model, look elsewhere.
What survives is structure a classical method can’t cheaply fake, meaning quantum structure or deep algebraic structure. Everything on the watchlist below follows from that.
Five questions, and one real claim run through them
You don’t need a physics degree to read these announcements well. You need five questions, and most hype fails the first.
- Where is the primary evidence? A paper, a dataset, a reproducible experiment, or a technical report. A press release without one is a press release.
- Real hardware, or a simulator? A simulator can validate a circuit construction. It does not show that physical hardware wins after noise, sampling, and runtime are counted.
- What is the classical baseline, and was it tuned? Many reported quantum wins evaporate against a well-built classical model on the same task.
- Does the speed-up survive data loading and error correction? An advantage that assumes free data encoding, or a flawless fault-tolerant machine, is an advantage on paper.
- Is the dataset natural or constructed? A result on a cryptographically engineered dataset is a proof of principle about the world, not about your data.
Questions two and three catch more claims than the other three combined, and dequantization is question three with a decade of practice behind it. Tang did not build new hardware. She built a better classical baseline for a comparison that had already been declared won, and the separation went with it.
Run the discrete-logarithm classifier through the same five and the pattern comes out clearly. Question one clears, because there is a peer-reviewed paper with a proof in it. Question two fails, because the classifier needs a fault-tolerant machine nobody has built. Question five fails as well, because the dataset was engineered from a cryptographic problem so that quantum structure would pay. None of that is a criticism of the paper, which claims exactly what it proves and no more. It is a warning about what happens when a result like that reaches a slide deck with the two failures dropped. We use that example with learners for exactly that reason: it is real work, correctly reported, and it still is not a win in column three.
Where to point attention
Three areas are worth tracking, and they are not the ones in the headlines.
Learning from quantum data comes first. As quantum sensors, quantum networks, and quantum simulators produce more intrinsically quantum data, the class of problems in column one widens on its own.
Quantum simulation feeding classical scientific machine learning comes second. A quantum computer that computes a molecular property classical methods cannot reach becomes a data source, and the model trained downstream can stay entirely classical. The advantage sits in producing the data, not in the learning.
Structured and symmetric problems come third. The more promising direction inside the field points toward exploiting spectral and group structure with tools such as the quantum Fourier transform, rather than bolting quantum circuits onto neural networks. This is early work, mostly feasibility analysis for fault-tolerant machines rather than benchmarks on real data. It is still the right question, and some of the people asking it are the field’s more reliable skeptics.
What we would not wait for is a quantum large language model, or a quantum classifier that beats gradient-boosted trees on a table of customer records. Those are the headlines, and they are the least likely outcomes.
What this changes for your roadmap
Security and technology leaders can leave quantum machine learning off the near-term risk register without much anxiety. The quantum development with an actual clock attached is the cryptographic one. NIST published ML-KEM, ML-DSA, and SLH-DSA as final standards in August 2024, FN-DSA is still to come, and federal and sector migration timelines are already written (NIST post-quantum cryptography project). Crypto-agility work has deadlines; quantum AI has a research program.
Capital allocators and research leaders should read the same evidence more optimistically. Experimental advantage in learning quantum processes is no longer theoretical, and the boundary between quantum and classical learning keeps getting sharper. Portfolio companies that generate quantum data, or that work directly in sensing, hardware, and scientific simulation, are the ones for whom this frontier is near-term. When a quantum-AI pitch crosses the desk, run it through the five questions before it runs through the budget. The filter costs nothing.
Building the judgment in-house
The scorecard and the five questions are the output of a specific kind of literacy: enough of the underlying mathematics to know what a kernel method actually claims, and enough practice to read a benchmark table against its own abstract. That combination is what our programs are built to produce. Program names, prerequisites, and current pricing in USD are listed at quantumacademy.com/.
For the cryptographic work that does carry deadlines, the migration methodology is set out at pqcframework.org. For a longer technical treatment of the same evidence base, see PostQuantum.com’s analysis of quantum machine learning.