Does quantum computing speed up machine learning?
No, not for machine learning on ordinary data. As of August 2026, no quantum speedup has been shown for any practical machine-learning task. The big "exponential" claims of the 2010s were mostly knocked down by ordinary "dequantized" algorithms, starting with Ewin Tang in 2018. The one proven advantage shown on hardware is for learning from quantum data, and almost no machine-learning work uses that.
No exponential speedup — the flagship claims were 'dequantized' (matched by ordinary algorithms) in 2018 The quantum recommendation-system algorithm was the star example of exponential QML speedups. Ewin Tang showed that an ordinary algorithm, given the same data access the quantum one needs, solves the problem with only a polynomial slowdown. No exponential advantage ever existed. Later work dequantized most of the low-rank QML toolkit. Lesson for developers: check what data-access assumptions a QML speedup claim sneaks in.
QML circuits run on real hardware — with zero demonstrated advantage The quantum-kernel idea is reasonable. A quantum circuit maps data into a 'feature space' that is hard to compute classically. Then a standard SVM (a common classifier) sorts it. This has run on hardware since 2019. What has never been shown is any dataset where it beats ordinary ML. The known cases where quantum kernels provably win are built from cryptographic structure (discrete logarithms), not from data anyone actually has.
Established theory works against QML: training breaks down at scale Variational QML trains a quantum circuit with an ordinary optimizer, in the style of gradient descent (always stepping downhill). McClean and colleagues proved that for common circuit families the training landscape is exponentially flat. At useful qubit counts, every gradient estimate looks like noise. Later work made the dilemma sharper. Many designs known to avoid barren plateaus turn out to be simulable on ordinary computers, which would make the quantum hardware unnecessary. This is a problem in the method itself, not an engineering one.
Yes — but only when the data itself is quantum This is the one rigorous learning advantage shown on hardware. The input is quantum states from an experiment, not a spreadsheet file. A quantum memory stores two copies and measures them together. It then provably needs exponentially fewer samples than any classical strategy. It is real, and it is not your ML workload. It applies to studying quantum devices and quantum experiments, not to images, text or tables of data.
What happened to the exponential QML speedups?
Why care? If quantum computers made AI much faster, a lot of software would change. So the claims deserve a careful look. QML means quantum machine learning: using a quantum computer to learn patterns from data.
First, what does "exponential" mean? A speedup is exponential when the gap grows like 2, 4, 8, 16 as the problem grows. After 10 steps that is 2^10 = 1,024 times. After 20 steps it is about a million times. That is the kind of gain the 2010s papers promised.
That wave of QML rested on quantum linear algebra, the math of big grids of numbers called matrices. Algorithms like HHL promised to solve systems of equations exponentially faster. But the fine print always held three traps:
- Loading the data. The data must be loaded into quantum amplitudes, the numbers that say how strongly qubits lean toward each result. That is costly. Papers usually assumed it away with "QRAM", a quantum memory that still does not exist.
- A well-behaved matrix. The matrix must be well-conditioned. That means small changes in the input don't cause wild swings in the answer.
- Reading the answer. The output is a quantum state, and you cannot read it out cheaply.
In 2018, Ewin Tang, then an undergraduate, made the key move. She gave an ordinary algorithm the same kind of data access the quantum one assumed: the ability to sample from the data. With that, the ordinary algorithm matched the quantum recommendation algorithm, up to polynomial factors. A polynomial factor is a slowdown like n² or n³, not 2^n.
So the exponential gap came from an unfair comparison. The quantum side got generous assumptions, and the ordinary side got none. Think of a race where one runner gets a bike, then claims to be the faster runner. Unlike that race, nobody cheated on purpose. The unfair start was hidden in the assumptions. Most low-rank QML fell to the same technique. Low-rank means the data can be summed up by a few main patterns.
What survives? Speedups for sparse, high-rank, well-conditioned problems. So far, that niche maps poorly onto real machine learning.
Is there any real quantum advantage in machine learning?
Yes, one, and it is narrow. Sometimes the data is itself quantum: states coming out of a physics experiment or a quantum device. Then a proven exponential advantage exists in sample complexity, the number of experiments you need. It has been shown on Google hardware. If you study how quantum devices behave, this matters to you. If you train models on ordinary data, it does not.
For ordinary data, the honest scoreboard reads:
- No proven start-to-finish speedup.
- No real-world win on any benchmark people use.
- A built-in training obstacle, called barren plateaus, for the variational methods that fit today's hardware.
A variational method uses a quantum circuit with adjustable dials. An ordinary computer turns the dials to improve the results. A barren plateau is a flat landscape. Imagine hiking blindfolded to find the lowest valley, but the ground is almost perfectly flat for miles. You can't tell which way is down. With more qubits, the ground gets exponentially flatter. Unlike a real hike, you can't just walk further to find a slope. The flatness covers almost the whole map.
Some made-up problems built from discrete logarithms show quantum classifiers can win in principle. A classifier is a program that sorts data into groups. But that is a statement about complexity theory, not about your data.
You can run a small variational classifier yourself in the lab. It shows what two-qubit QML actually looks like.
What would have to change for the answer to change?
Watch for these, roughly from most to least important:
- A start-to-finish demo on ordinary data that beats the best ordinary method. It must include loading the data and reading out the answer. Every claim so far has failed this test.
- QRAM that exists. Most exponential-speedup claims quietly assume a quantum memory that can reach N items in log(N) time with almost no noise. For a million items, log₂(N) is only about 20 steps. Nobody has built one at any useful size.
- Circuits that train but can't be copied cheaply. Researchers need designs that avoid barren plateaus without becoming something an ordinary computer can simulate.
Until then, treat "quantum machine learning" in a sales pitch as a warning sign. Related: can quantum computers solve optimization problems faster?
Classifications follow the QPU137 editorial policy: every applied label carries a date, source, scale, and hardware, or it does not render. Found an error? Report it.