Is quantum computing useful for machine learning today?

Status as of: August 2026.

Quantum machine learning (QML) is the class where the honest answer is not just "not yet" — it is "the clock ran backwards." Starting in 2018, the field's flagship exponential speedups fell to classical algorithms, and they have not recovered. This is the feasibility verdict for a decision-maker: what QML promised, the three dated blows that shrank the promise, what published benchmarks measure today, and the one narrow front where a learning advantage has actually been measured on hardware.

What does QML promise, and why does it sound so good?

Two pitches carried a decade of enthusiasm. First, amplitude encoding: N qubits hold 2^N amplitudes, so in principle a large dataset fits into exponentially few qubits and gets processed by quantum linear algebra (what a qubit actually stores). Second, quantum kernels: a circuit maps data into a huge Hilbert space where a simple separator might find structure classical feature maps miss. On paper both looked exponential. The fine print — how data gets into amplitudes, and what the classical rival is allowed to do — is where the decade actually happened (Aaronson, Nature Physics 11, 291, 2015).

Blow 1 (2018–2020): dequantization levelled the table

The flagship exponential claims — recommendation systems, PCA, clustering, support-vector classification — all assumed the quantum machine receives its data in an efficient quantum-access format. In 2018 Ewin Tang gave the classical side the analogous access (ℓ²-norm sampling) and matched the recommendation speedup classically (arXiv:1807.04271, 2018; STOC 2019). By 2020 a single framework had dequantized essentially every low-rank QML speedup: recommendations, PCA, clustering, qSVM, low-rank regression (Chia et al., STOC 2020). The exponential gap collapsed to polynomial. The mechanism and the full map of what fell versus what stands are in our dequantization post — this post only needs the verdict: the exponential pitch on classical data did not survive a fair data-access model.

Blow 2 (2018–2025): barren plateaus, and a sharper 2025 twist

Variational QML models — the trainable-circuit kind you can actually run on today's machines — hit a structural training wall: for wide classes of circuits, gradients vanish exponentially in qubit count (McClean et al., Nat. Commun. 9, 4812, 2018), and the 2025 review classes barren plateaus among the field's main barriers (Larocca et al., Nat. Rev. Phys. 7, 174–189, 2025). The 2025 twist cuts deeper: for many circuit families, provable absence of barren plateaus tends to imply the model is classically simulable (Cerezo et al., Nat. Commun. 2025). The corridor between "trainable" and "still quantum" is narrow, and nobody has shown it is inhabited — the same wall we documented for VQE in chemistry.

Blow 3 (standing since 2015): the data-loading wall

QRAM at scale does not exist. Loading N classical numbers into amplitudes generally costs on the order of N operations — which by itself erases an exponential end-to-end gain on classical data (Aaronson, 2015). This is not an engineering backlog item anyone has cleared: as of August 2026 there is no scalable QRAM demonstration. Every end-to-end QML pipeline on classical data still pays this toll at the door.

What do published benchmarks actually measure?

When quantum and classical models are run on the same tasks with tuned baselines, the pattern is consistent — and it is the same lesson as a weak baseline ruining a benchmark, read from the other side:

Read the split in that table carefully: on classical data the scoreboard shows zero wins; the measured advantages live on quantum data.

What about the HSBC–IBM ~34% result?

The one recent headline that reads like a commercial QML win: HSBC and IBM reported up to ~34% improvement in bond-RFQ fill prediction using quantum-generated features (HSBC media release, 25 Sep 2025). We examined it in the financial-risk verdict and won't re-litigate it here: it is feature generation inside a classical pipeline, the team's own paper attributes part of the gain to hardware noise, the noiseless simulation does not show the advantage, and there is no independent replication as of August 2026. It does not move this verdict. None of this is an accusation — it is a real experiment read with its own caveats.

What is still standing?

Three things, in decreasing order of evidence:

1. Learning from quantum data — measured. When the input is itself quantum — states coming from a sensor or an experiment, not numbers from a database — a quantum memory can provably and measurably beat every classical strategy. The exponential reduction in required experiments was demonstrated on hardware in 2022 (Huang et al., Science 376, 1182), and in September 2025 a photonic platform learned a 100-mode bosonic process with ~11.8 orders of magnitude fewer samples than the conventional scheme, using imperfect (~5 dB) EPR entanglement (Science 389, 1332, 2025). Real advantage, real hardware, still a physics proof-of-concept — and the input is quantum states, not your CSV.

2. Conditional kernel separations — on paper. There is a rigorous, robust speedup for a classification problem built on the discrete logarithm (Liu, Arunachalam & Temme, Nat. Phys. 17, 1013, 2021). It is a cryptographic construction, not a natural dataset — it proves the separation can exist, not that your data has it.

3. Polynomial speedups — surviving, unmeasured, economically questionable. What dequantization left standing on classical data is polynomial. No polynomial QML speedup has been measured end-to-end, and quadratic-class speedups are not expected to pay on early fault-tolerant machines once error-correction overhead is priced in (Babbush et al., PRX Quantum 2, 010103, 2021). The decade's own summary: quantum advantage "has narrowed to a smaller and more clearly defined set" of algorithms than ten years ago (Khan, SSRN, 28 Apr 2026) — and QML's exponential entries were the ones that narrowed most. The honest list of what is proven, class by class, is in which problems have proven quantum advantage.

What would change this verdict?

Three observable events, none published as of August 2026: (1) an end-to-end QML win on classical data against a strong, tuned classical baseline on the same instance, published with reproducible artifacts; (2) a scalable QRAM-class data-loading demonstration; (3) a trainable circuit family proven both barren-plateau-free and classically hard to simulate. Watch for those, not for headline percentages.

What we know / what we don't know

We know: the flagship exponential QML claims on classical data fell to dequantization between 2018 and 2020 (Tang; Chia et al.). We know published benchmarks keep finding tuned classical models ahead — 12 models over 160 datasets (Bowles et al., 2024), 7 model pairs including RL (Yu et al., Jul 2026). We know a genuine, measured learning advantage exists on quantum data (Science 2022; Science, Sep 2025). And we know Rosetta has no sealed runs in this class: our sealed series are optimization and quantum-walk experiments — we have no QML measurements of our own and claim none.

We don't know: whether kernels over genuinely quantum data — quantum sensors feeding quantum memories — become the one structural door into commercially relevant QML; the measured wins are physics proofs-of-concept, not products. Whether the surviving polynomial advantages ever pay end-to-end on fault-tolerant hardware — Babbush et al. argue quadratic does not, on the first generations. Whether an inhabited corridor of trainable-but-classically-hard circuits exists — absence of proof is not proof of absence. And the zero on the scoreboard is an absence of published demonstrations, not a theorem of impossibility.

Sources

• Tang — A quantum-inspired classical algorithm for recommendation systems (arXiv:1807.04271, 2018; STOC 2019)
• Chia, Gilyén, Li, Lin, Tang & Wang — Sampling-based sublinear low-rank framework for dequantizing QML (STOC 2020; arXiv:1910.06151)
• Aaronson — Read the fine print (Nature Physics 11, 291–293, 2015)
• McClean et al. — Barren plateaus in quantum neural network training landscapes (Nature Communications 9, 4812, 2018)
• Larocca et al. — Barren plateaus in variational quantum computing, review (Nature Reviews Physics 7, 174–189, 2025)
• Cerezo et al. — Does provable absence of barren plateaus imply classical simulability? (Nature Communications, 2025; arXiv:2312.09121)
• Bowles, Ahmed & Schuld — Better than classical? The subtle art of benchmarking QML models (arXiv:2403.07059, 2024)
• Yu et al. — Quantum vs. classical machine learning: a unified empirical comparison (arXiv:2607.01197, Jul 2026)
• Huang et al. — Quantum advantage in learning from experiments (Science 376, 1182–1186, 2022)
• Quantum learning advantage on a scalable photonic platform (Science 389, 1332–1335, 25 Sep 2025)
• Liu, Arunachalam & Temme — A rigorous and robust quantum speed-up in supervised machine learning (Nature Physics 17, 1013–1017, 2021)
• Babbush et al. — Focus beyond quadratic speedups for error-corrected quantum advantage (PRX Quantum 2, 010103, 2021)
• HSBC — quantum-enabled algorithmic trading, media release (25 Sep 2025)
• Khan — Quantum Algorithms: A Decade in Review (SSRN, 28 Apr 2026)
• PostQuantum — Quantum Machine Learning in 2026: State of the Field (accessed Aug 2026)

Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.