The Quantum AI Taxonomy: Seven Things People Mean When They Say “Quantum AI” – and Why Confusing Them Costs Money

In September 2025, HSBC and IBM announced a 34% improvement in predicting whether corporate bond trades would fill at a quoted price, using IBM Quantum Heron processors on production-scale requests for quote. The headlines wrote themselves: quantum and AI, working together, delivering real results in real markets. Reuters ran it straight. LinkedIn amplified it. Conference slides picked it up within days.

What the headlines omitted appears in the preprint’s own abstract. The improvement was not reproducible when the same quantum computation ran on a noiseless classical simulator, and the authors suggest hardware noise may have contributed to the effect. Scott Aaronson, director of the Quantum Information Center at UT Austin, labelled it a “qombie,” meaning a zombie claim of quantum advantage, and in a follow-up comment called the paper “scientifically risible.”

The bottom line: “quantum AI” is not one thing. It is at least seven distinct categories that vendors, journalists, investors, and executives conflate routinely, and confusing them produces bad procurement, bad investment, and bad analysis. Some of these categories represent peer-reviewed, demonstrated gains. Some are legitimate theory without near-term evidence. Some are classical algorithms marketed under a quantum label. At least one is a vehicle for outright fraud. I have covered QML and the PQC-QAI intersection before; this piece maps the entire territory.

The taxonomy

The seven categories run from strongest evidence to weakest. That ordering is deliberate, because it inverts the way most vendors tell the story.

1. AI for quantum computing

This is where the intersection delivers today, and it is the most undersold category. Machine learning is producing measured, peer-reviewed improvements across the quantum stack. These are shipped results rather than roadmap items.

The flagship result is Google DeepMind’s AlphaQubit, published in Nature in November 2024. AlphaQubit is a recurrent, transformer-based neural network that decodes the surface code, the error-correction scheme most superconducting quantum computers rely on. Tested on Google’s Sycamore processor, it reduced errors by 6% compared to tensor network methods and by 30% compared to correlated matching, measured on distance-3 and distance-5 data and validated on simulated data up to distance 11.

The caveat at the time was throughput. Superconducting processors generate roughly a million syndrome cycles per second. A decoder slower than about a microsecond per cycle falls behind and cannot run in a production fault-tolerant system.

That limit was addressed a year later. AlphaQubit 2, posted as a preprint in December 2025 by the same Google DeepMind and Quantum AI effort, reports near-optimal logical error rates for both the surface code and the color code, plus real-time decoding faster than one microsecond per cycle on commercial TPU accelerators (surface code to distance 11, color code to distance 9). The color-code result matters because color codes need fewer physical qubits than surface codes for a given logical error rate, but have long been considered too hard to decode in real time. An independent group at the University of Science and Technology of China attacked the same throughput problem with a parallel-decoding architecture benchmarked on the Zuchongzhi 3.2 processor.

AI is doing more than decoding. Reinforcement learning now searches the code-design space directly instead of tuning human-designed candidates, producing error-correction codes and encoders from noise-aware agents. A RIKEN-led group used neural networks to design approximate GKP codes that outperform conventional constructions with a fraction of the resources. Model-free RL on IBM hardware reached a qubit-readout assignment error of (4.6 ± 0.4) × 10⁻³ while running readout and resonator reset nearly three times faster than the default. Neural inverse design that maps target Hamiltonian parameters to transmon geometry now produces usable layouts in 97% of cases, with mean errors of 0.7% on qubit frequency and 1.6% on anharmonicity.

Machine learning is also improving error mitigation, which is a different problem from error correction. Mitigation reduces bias in expectation values at a sampling cost on noisy hardware, whereas correction builds logical qubits from physical ones. Liao et al.’s ML-QEM work in Nature Machine Intelligence showed classical ML models, including random forests and graph neural networks, matching or beating zero-noise extrapolation on up to 100 qubits while cutting the sampling overhead by more than half.

When someone says “AI and quantum are working together,” this is the defensible version.

2. HPC-quantum orchestration

This is the infrastructure engineering of heterogeneous computing: scheduling workloads across CPUs, GPUs, and QPUs as co-managed resources in the same environment. It is real, deployed, and growing, and it is one of the most miscategorized items on this list.

The leading example is NVIDIA’s NVQLink, announced in October 2025. It connects quantum control systems to GPU-accelerated classical compute over standard 400 Gb/s Ethernet, with measured round-trip latency under four microseconds. The launch brought together 17 QPU builders, five controller manufacturers, and nine U.S. national laboratories. CUDA-Q supplies the programming model. QRMI, the Quantum Resource Management Interface demonstrated with Pasqal in March 2026, exposes QPUs as Slurm-native schedulable resources so hybrid jobs can be submitted through standard HPC schedulers.

Reference deployments at the Jülich Supercomputing Centre, Israel’s IQCC, and Munich’s LRZ Euro-Q-Exa show what the architecture looks like in practice. A quantum processor sits on dedicated fiber to a GPU cluster, with decoding and control running on the GPUs in real time.

This is mostly Category 1 packaged as infrastructure. The GPUs are helping run the quantum computer, doing decoding, calibration, and mitigation. Marketing calls it “hybrid quantum-AI.” Engineering calls it a classical coprocessor for quantum error correction.

3. Learning from quantum data

This is the one category where a provable, hardware-demonstrated exponential advantage for a learning task exists. Huang et al.’s 2022 paper in Science proved, and demonstrated on Sycamore, an exponential reduction in the number of experiments needed to learn properties of quantum systems, robust enough to survive then-available noise.

The result applies to quantum-native tasks such as many-body characterization, quantum sensor output, and process tomography. Spreadsheets, images, and transformer workloads are classical data, and the advantage does not transfer.

For anyone watching where genuine quantum-enhanced ML might first appear commercially, this is the category to track. It points toward quantum sensing and quantum networking pipelines. Enterprise data platforms are the wrong place to look.

4. Quantum-inspired classical algorithms

These borrow structures from quantum mechanics, such as tensor networks, Ising-model formulations, and amplitude-inspired sampling, and run on classical hardware throughout. Some deliver genuine commercial value. None require a qubit.

Toshiba’s Simulated Bifurcation Machine and its SQBM+ successor use quantum-inspired dynamics to attack combinatorial optimization on GPUs and FPGAs, scaling to very large variable counts. Fujitsu’s Digital Annealer and Hitachi’s CMOS annealer follow the same approach. Published benchmarks often show these matching or beating actual quantum annealers on time-to-solution.

Tensor networks, born in quantum many-body physics, are now a workhorse of classical machine learning and the primary tool for classically simulating quantum circuits. They serve two functions at once. They let classical methods compete with quantum ones, and they set the bar any quantum ML claim has to clear.

These are legitimate classical tools. The word “quantum” describes their intellectual ancestry; the hardware is standard GPUs and FPGAs. When a vendor sells “quantum-powered optimization” that executes on an NVIDIA GPU with no QPU in the loop, this is the category, and the honest label is quantum washing.

5. Quantum machine learning, the academic field

QML is an active research field with serious people doing serious work. It is also in a sober, self-critical phase. The honest assessment in 2026 is that no proposed quantum speedup for classical-data machine learning has held up as a commercially significant, end-to-end result.

Three structural problems have reduced the field’s plausible scope.

The first is barren plateaus. In many variational quantum circuits, the workhorse of NISQ-era QML, gradients vanish exponentially with system size and training becomes intractable. The definitive review is Larocca et al. (2024; journal version Nature Reviews Physics, 2025), which traces the failure through expressibility, entanglement, and noise. Barren plateaus are not universal. Whether they appear depends on the ansatz, the initialization, the observable, the loss function, and the noise model. But they are common enough to have reshaped how the field designs circuits.

The second is data loading. Putting classical data into a quantum state can cost as much as the computation it was meant to accelerate. Many QML proposals assume quantum RAM that does not exist at scale and may never be practical. At a November 2024 QTML panel, Scott Aaronson and Aram Harrow both treated qRAM-dependent results with skepticism and stressed fair benchmarking against optimized classical baselines.

The third, and the most consequential, is dequantization. Since Ewin Tang’s 2018 result showing that the quantum recommendation-systems algorithm has a classical counterpart with only polynomial slowdown, a line of work has dequantized QML algorithms for PCA, low-rank regression, and related linear-algebra primitives. Sweke et al. (Quantum 9, 1640) established necessary and sufficient conditions under which random Fourier features dequantize variational QML for regression, and mapped which circuit structures remain out of reach.

Dequantization did not end QML. It showed where advantage can and cannot live, and the remaining zone is much smaller than the 2019-era enthusiasm suggested. There are rigorous constructed problems where quantum kernels provide an end-to-end speedup under stated assumptions, but they function as existence proofs; none has produced a commercially compelling demonstration.

6. Quantum computing for mainstream AI

This is the dream: quantum-accelerated training of large neural networks, quantum-enhanced inference, quantum optimization for the workloads that already consume billions of dollars in GPU time. It is the weakest category by evidence and the strongest by marketing spend. It also has more serious theoretical work behind it than most skeptical coverage acknowledges, which is why the proposals below get a careful walk-through.

The structural obstacle. Hoefler, Häner, and Troyer’s analysis in Communications of the ACM is the clearest statement of why quantum-for-AI is hard. Their conclusion is that quantum computers will be practical for “big compute” problems on small data, not big data problems. Two mechanisms drive that. First, any problem limited by accessing classical data will be solved faster classically, because reading data into a quantum state is slow. Reasonable future quantum designs read on the order of gigabits per second while modern processors already hit terabits per second. Second, quadratic speedups are not enough. Once you pay the constant-factor and error-correction overhead of a fault-tolerant machine, the problem must reach a scale where the overhead stops dominating. Their guidance is to focus on super-quadratic, ideally exponential speedups on small-data problems. Mainstream AI training is the opposite: modest arithmetic intensity on enormous datasets.

Training proposals. The most-cited concrete result is Zlokapa, Neven, and Lloyd’s quantum algorithm for training wide and deep classical neural networks. It approximately trains a network to O(1/n) error for a training set of size n by performing sparse matrix inversion in O(log n) time, exploiting the neural tangent kernel regime where wide networks behave like kernel machines. The authors are explicit about the condition: to achieve an end-to-end exponential speedup over gradient descent, the data distribution must permit efficient state preparation and readout. They demonstrate numerically that MNIST satisfies those conditions. Whether production-scale training data does is an open question, and the qRAM assumption reappears here as it does everywhere else in the category.

A different angle comes from quantum annealing. A late-2025 paper in Communications Physics reported using D-Wave annealers to train classical neural networks, finding a higher scaling exponent than classical backpropagation on the problems tested. Annealing sidesteps the gate-model fault-tolerance requirement entirely, which makes it near-term testable, but it also inherits the long-running question of whether annealing advantages survive comparison against well-tuned classical solvers, including the quantum-inspired ones in Category 4.

Inference proposals. This is where the last two years produced the most substantive work. Guo, Nakaji, Aspuru-Guzik and colleagues’ Quantum Transformer paper develops quantum subroutines for the actual building blocks of a transformer: self-attention, residual connections with layer normalization, and the feed-forward network. The technical contributions include implementing element-wise functions of block-encoded matrices, handling softmax through Gibbs-like state preparation, and approximating GELU activations via polynomial methods. Their per-token cost is roughly Õ(√N d log²(1/ε)). They validate the input-norm assumptions against real open-source LLMs and identify regimes where a polynomial speedup could arise. They are also explicit about the qRAM and dequantization caveats.

Rattew et al. went further in Accelerating Inference for Multilayer Neural Networks with Quantum Computers, accepted at ICLR 2026. They present what they describe as the first fully-coherent quantum implementation of a multilayer neural network with non-linear activation functions, mirroring ResNet architectures with residual blocks, multi-filter 2D convolutions, sigmoid activations, skip connections, and layer normalization. Their complexity analysis is honest about the dependency on data access assumptions: without any assumptions, they establish a quadratic speedup for shallow bilinear-style networks; with efficient quantum access to the weights, they obtain a quartic speedup.

That gradient is the whole story of the category in one result. The speedup depends on the data-access assumption you are willing to grant. The strongest assumptions are the ones nobody knows how to build.

Why the field’s own labs are redirecting. In a March 2026 position piece, Xanadu’s QML team argued that “the worst starting point” for quantum ML research is “attempting to speed up neural networks with quantum computers.” Their alternative builds around operations natural to quantum hardware, particularly the Quantum Fourier Transform, as the basis for spectral methods. It is more intellectually honest than bolting a quantum circuit onto a neural network, and it is also explicitly prospective. No empirical demonstration accompanies it.

A December 2025 framework paper by Masot-Llima et al. sorted QML models into three classes: classically simulable, classically tractable in function space, and robustly quantum. It maps where opportunity might exist. It does not claim that advantage has been achieved.

How to read a vendor claim in this category. Ask for the peer-reviewed benchmark against a state-of-the-art classical baseline. Then ask whether the quantum computation ran on an actual QPU or a simulator. Then ask what data-access assumption the speedup depends on, because a quartic speedup conditional on qRAM is a different claim from a quadratic speedup with no assumptions. If the answer to the second question is “simulated,” you are in Category 4.

7. “Quantum AI” as fraud

This is not an edge case. “Quantum AI” is the operating name of a multi-jurisdiction investment-fraud cluster that financial regulators have flagged on several continents. The UK’s Financial Conduct Authority lists it as an unauthorised firm. Hong Kong’s SFC added Quantum AI to its alert list, warning that it claims to provide cryptocurrency trading services using underlying AI technology. The scheme has used AI-generated deepfake videos of Elon Musk and Martin Lewis, among other public figures. Ireland’s Central Bank, Germany’s BaFin, and Australia’s ASIC have issued parallel warnings.

No quantum computer can predict financial markets. No AI system of the kind advertised can either. The two buzzwords are stacked because each one supplies enough technical mystique to discourage due diligence.

The practical relevance for anyone running security awareness: employees encounter these ads, click through, and sometimes hand personal data or a corporate email address to the phishing infrastructure behind them. The “Quantum AI” brand is an attack surface.

The HSBC/IBM case, anatomy of a claim

HSBC and IBM produced the most-cited commercial example of “quantum and AI working together.” The gap between the result and the headlines makes it a useful template for reading the next one.

What they did. The team passed real trading data, production-scale RFQs across the European corporate bond market, through a quantum feature extractor running on IBM Heron processors. The circuit generated new input features. A classical ML pipeline still did the evaluation, and the quantum feature generation ran as a decoupled offline component, outside any live order-routing path.

What they found. Up to a 34% relative gain in out-of-sample test scores against common classical techniques.

What the preprint says. The gain was not reproducible under noiseless classical simulation of the same quantum circuit. The authors name hardware noise as a plausible cause and call for further investigation.

The skeptical reading. Aaronson’s argument is that the advantage looks like hardware noise acting as a random feature transform, a stochastic perturbation that happened to help this model on this task. If that is right, a classical random-noise injection should do similar work. Commenters on his blog noted that the classical baselines were standard industry methods; nobody searched exhaustively for the best available feature engineering.

The defensible reading. This is a proof of interoperability and integration. A major bank showed that quantum hardware output can feed a production-class trading pipeline, that its systems can ingest quantum-processed features, and that the architecture works at production data scale. That is an engineering result. It is not evidence of a quantum computational speedup, and it will not be until the noise hypothesis is tested and the classical baselines are exhausted.

The question that matters is never “did the quantum computer do something?” It is whether the result was benchmarked against the best classical method, whether the advantage is reproducible and explained, and whether the explanation requires quantum mechanics.

The one question that resolves most confusion

When someone says “quantum AI” in a meeting, a deck, or a press release, one question resolves most of the confusion:

Is AI improving quantum hardware, or is quantum computing improving AI?

If AI is improving quantum hardware, meaning decoders, calibration, and code design, you are in Category 1. The evidence is strong and the results are peer-reviewed.

If quantum computing is claimed to accelerate AI, meaning faster training, better inference, or quantum-enhanced ML, you are in Category 5 or 6. The evidence is thin, the theoretical foundations depend on data-access assumptions nobody knows how to build, and you should demand a peer-reviewed benchmark before committing budget.

If neither is happening and the computation executes on classical silicon, you are in Category 4. It may be a good product. The “quantum” label is marketing.

If it is a trading platform promising guaranteed returns, you are in Category 7. Walk away.

That question would have prevented most of the bad coverage, bad investment theses, and bad procurement decisions of the past two years.

What would change this assessment

Three results would move the map.

A reproducible quantum advantage on a classical-data ML task. Peer-reviewed, benchmarked against optimized classical methods, still standing after a year of dequantization attempts. None exists. If one appears, Category 6 leaves “mostly theoretical.” The task will probably have algebraic or combinatorial structure. Image classification and language modeling are the wrong shapes.

A quantum inference result that survives its own assumptions. The Quantum Transformer and ResNet papers give quadratic speedups without data-access assumptions and quartic speedups with them. A demonstration that the assumption-free version beats an optimized GPU implementation on a real workload, at any scale, would be the first concrete evidence that the category is more than a blueprint. That requires logical qubit counts in the hundreds with real-time decoding at scale, which AlphaQubit 2’s sub-microsecond decoding enables but does not deliver.

A hybrid deployment where the quantum contribution is isolated, explained, and reproducible. The HSBC/IBM case qualifies on integration and fails on explanation, because the noise-as-feature hypothesis is still untested. A result whose quantum features still help under noiseless simulation, and survive exhaustive classical benchmarking, would be the first clean commercial proof point.

The two-front problem

The quantum-AI intersection is a clean case of a two-front problem.

The hype front is loud. Vendors stamp “quantum AI” on classical software. Coverage treats hybrid scheduling as “quantum and AI working together.” Trading scams borrow the phrase outright. The corrective is evidence: the benchmark, the baseline, the peer review.

The denial front is quieter and also wrong. Dismissing the whole intersection because the loudest claims are weak ignores the direction where the evidence is strongest. AlphaQubit is a Nature paper. Reinforcement learning is measurably tightening calibration and readout. Neural networks are designing better codes and better qubits. The theoretical work on quantum inference is real mathematics, honestly caveated by its authors, even if it is nowhere near deployment.

The two fronts fail in opposite directions on the same evidence. The hype reading takes Category 6’s blueprints and calls them products. The denial reading takes Category 6’s weakness and concludes the whole intersection is empty, which erases Category 1. Both errors come from treating “quantum AI” as one thing.

It is not one thing. It is seven, and they are not equally real.