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 that they had used a quantum computer to improve predictions of whether corporate bond trades would go through at a quoted price. The reported improvement was up to 34%. They used real trading data and IBM’s Heron processors, making this an obvious candidate for the long-awaited business application of quantum computing.

There was a complication, though. The improvement disappeared when the same quantum circuit ran on a noiseless classical simulator. The authors acknowledged this in the preprint’s abstract and suggested that noise in the quantum hardware might have contributed to the result. Scott Aaronson called the claim a “qombie”, his term for a quantum-advantage claim that keeps circulating despite problems with the evidence.

I’ll come back to what the experiment actually showed. But it illustrates why I find the phrase “quantum AI” increasingly unhelpful. It can describe a neural network that helps correct errors in a quantum computer, a proposed quantum algorithm for training neural networks, or ordinary software with quantum branding. Those are very different things, with very different evidence behind them.

I’ve written about quantum machine learning and the intersection of PQC and quantum AI before. Here, I want to separate the main uses of the term and look at what each has actually delivered.

What people mean by “quantum AI”

I find seven distinctions useful. They overlap in places, particularly where machine learning meets quantum hardware control, but they help avoid treating progress in one area as evidence for a completely different claim.

1. Using AI to build and run quantum computers

This is where some of the clearest progress is happening. Quantum computers are difficult to calibrate, control, and keep free of errors. Machine learning is proving useful for several of those jobs.

Google DeepMind’s AlphaQubit, published in Nature in November 2024, is a good example. It uses a recurrent, transformer-based neural network to decode the surface code: interpreting error-detection measurements to work out what has gone wrong with the encoded quantum information. On data from Google’s Sycamore processor, it reported 6% fewer errors than tensor-network methods and 30% fewer than correlated matching. The hardware tests used distance-3 and distance-5 codes, with further validation on simulated data up to distance 11.

Accuracy was only part of the problem. Superconducting processors can generate roughly a million rounds of error-detection measurements per second. A decoder has to process that stream fast enough to avoid an ever-growing backlog, while also meeting the system’s feedback deadlines.

AlphaQubit 2, first posted in December 2025 and updated in March 2026, addresses that bottleneck. The preprint reports decoding faster than one microsecond per cycle on commercial accelerators, for surface codes up to distance 11 and color codes up to distance 9. Color codes are interesting because they can offer resource savings, but fast, accurate decoding has been a persistent difficulty. A separate group at the University of Science and Technology of China has explored a parallel-decoding architecture, benchmarked using the Zuchongzhi 3.2 processor.

There is useful work beyond decoding, too. Researchers have used reinforcement learning to design error-correction codes and encoders, allowing the search to account for the hardware’s noise. A RIKEN-led group used neural networks to design approximate GKP codes with lower resource requirements than conventional constructions. On IBM hardware, reinforcement learning achieved a readout assignment error of about 0.46% while making readout and resonator reset nearly three times faster than the default procedure.

Machine learning also helps with error mitigation. This is different from error correction: mitigation improves estimates obtained from noisy computations, usually through extra measurements and processing, while correction protects information encoded in logical qubits. Liao and colleagues’ ML-QEM work in Nature Machine Intelligence used classical models, including random forests and graph neural networks, to match or outperform zero-noise extrapolation in experiments covering up to 100 qubits, while reducing sampling overhead by more than half.

These results have different levels of maturity, and a promising decoder benchmark still leaves substantial integration work. Even so, there is plenty here to justify taking AI’s contribution to quantum computing seriously.

2. Connecting quantum processors to conventional supercomputers

Another set of announcements concerns the infrastructure around quantum computing. Researchers need to coordinate CPUs, GPUs, and quantum processing units, or QPUs, within the same system. That includes running classical calculations quickly enough to support quantum error correction and control, as well as scheduling larger hybrid applications.

NVIDIA’s NVQLink, announced in October 2025, is one example. It provides an architecture for connecting quantum control systems to GPU computing over high-speed Ethernet, with microsecond-scale communication. CUDA-Q provides a programming environment for combining classical and quantum operations.

The scheduling side matters as well. Pasqal’s integration of CUDA-Q with the Quantum Resource Management Interface, or QRMI, allows quantum resources to fit into established HPC workflows. A researcher can submit a hybrid job through a scheduler such as Slurm instead of managing the quantum processor as a separate service.

This is necessary engineering. Much of it supports the work in the previous section, and some enables experiments with hybrid applications. The presence of a GPU and a QPU in the same workflow, however, tells us very little about whether that application gains anything from quantum computation. That still has to be measured.

3. Learning from quantum data

Learning directly from quantum systems is a particularly interesting case because there are already rigorous results showing an advantage for specific tasks.

Huang and colleagues’ 2022 paper in Science proved that quantum methods can require exponentially fewer experiments to learn certain properties of quantum systems. They also demonstrated the effect on Sycamore, under the hardware’s noise conditions.

The qualification matters: this is an advantage in the number of experiments needed for particular learning problems involving quantum data. It does not establish a general speedup for machine learning, and it does not transfer automatically to spreadsheets, images, or language-model training.

It does suggest where useful applications might emerge. Characterizing quantum devices, learning about many-body systems, and processing information from quantum sensors are natural places to investigate. I would watch those areas closely, while keeping the demonstrated learning tasks separate from the commercial applications they might eventually support.

4. Quantum-inspired software running on classical hardware

Quantum mechanics has also inspired algorithms that run entirely on conventional computers. Tensor networks, Ising-model formulations, and quantum-inspired sampling methods all fall into this broad group.

Toshiba’s Simulated Bifurcation Machine and SQBM+, Fujitsu’s Digital Annealer, and Hitachi’s CMOS annealer are examples of classical approaches to optimization associated with this space. They can be useful products on their own merits. Their performance needs to be assessed against other solvers on the particular problem being solved.

Tensor networks are especially relevant to quantum machine learning. Developed in quantum many-body physics, they are used both in classical ML and to simulate quantum circuits. They can produce useful classical algorithms while also making it harder for a proposed quantum algorithm to establish an advantage. A comparison that overlooks strong tensor-network methods may set the classical baseline too low.

“Quantum-inspired” is a reasonable description when the vendor explains the connection and makes clear what hardware runs the software. The problem is selling the same product as though a quantum computer were responsible for its performance. That is the kind of quantum washing buyers should challenge.

5. Quantum machine learning as a research field

Quantum machine learning covers a wide range of approaches, including quantum kernels and trainable quantum circuits. There is substantial research here, but the commercial question remains unresolved: can a quantum method solve a useful learning problem on classical data better than the best practical classical alternative, after all costs are included?

I have yet to see a broadly accepted result that settles that question at a commercially meaningful scale. Three recurring problems explain much of the difficulty.

The first is training. In many variational quantum circuits, gradients become exponentially small as the system grows. These “barren plateaus” can make it impractical to find useful parameters. Larocca and colleagues’ review, first posted in 2024 and published in Nature Reviews Physics in 2025, explains how circuit structure, entanglement, noise, and the choice of loss function affect the problem. Barren plateaus are not inevitable, but avoiding them is a central design constraint.

The second is getting data into the quantum computer. A fast quantum calculation is of limited use if preparing its input takes as much work as the classical calculation it replaces. Many proposals rely on quantum random-access memory, or qRAM, to provide efficient access to classical data. We do not have that capability at the scales those proposals would need.

The third is dequantization: finding classical algorithms that recover much of an apparent quantum advantage under comparable assumptions. Ewin Tang’s 2018 recommendation-systems result was an important example, showing a classical counterpart with only a polynomial slowdown. Related work has since addressed principal component analysis, low-rank regression, and other linear-algebra tasks.

More recently, Sweke and colleagues examined when random Fourier features can dequantize variational QML models for regression, and where that approach reaches its limits.

This work helps researchers identify which proposals still have a plausible case for quantum advantage. There are constructed learning problems with rigorous speedups under stated assumptions. Turning those results into useful applications remains a separate challenge, particularly when efficient data access is part of the assumption rather than part of the demonstrated system.

6. Using quantum computers for mainstream AI

This is the prospect behind many of the bigger commercial claims: faster training of large neural networks, cheaper inference, or some other reduction in the GPU resources that AI consumes. It overlaps with QML research, but the specific promise of accelerating today’s AI workloads deserves its own examination.

The practical obstacles are substantial. Hoefler, Häner, and Troyer’s analysis in Communications of the ACM explains why the amount of data and the size of the speedup matter so much. Quantum hardware carries significant overhead, particularly once error correction is included. A quadratic improvement in an algorithm’s scaling may require an enormous problem before it outweighs that overhead. Loading classical data can make the comparison worse.

Their analysis points toward problems with relatively small inputs and very large computational demands, ideally with more than a quadratic quantum speedup. AI workloads need to be examined against those requirements individually, including the cost of accessing datasets and model weights.

There are nevertheless concrete proposals worth reading.

Zlokapa, Neven, and Lloyd’s algorithm for training wide and deep classical neural networks works in the neural tangent kernel regime, where sufficiently wide networks behave like kernel methods. It uses quantum sparse matrix inversion to obtain a large theoretical speedup. The end-to-end exponential advantage depends on efficient preparation of the input state and readout of the result. The authors provide numerical evidence using MNIST, but that does not show that the same conditions hold for production training workloads or that the full algorithm runs competitively on quantum hardware.

Quantum annealing offers another approach. A 2025 paper in Communications Physics reported training classical neural networks with D-Wave hardware, finding better performance scaling than backpropagation on the tested problems. Annealers make this line of work experimentally accessible without waiting for a fault-tolerant gate-based computer. The next question is how well the result holds up across workloads and against well-tuned classical alternatives, including quantum-inspired solvers.

For inference, Guo, Nakaji, Aspuru-Guzik, and colleagues’ Quantum Transformer paper develops quantum versions of familiar transformer components: self-attention, residual connections, layer normalization, and feed-forward networks. The work includes methods for handling operations such as softmax and nonlinear activations. The authors examine input assumptions using open-source language models and identify conditions under which a speedup could be possible. Data access and the possibility of competitive classical algorithms remain central to the assessment.

Rattew and colleagues’ work, published at ICLR 2026, develops a coherent quantum implementation of multilayer networks with nonlinear activations, including constructions resembling ResNets. Its results make the importance of data access particularly clear. The authors establish a quadratic speedup for shallow bilinear-style networks without the stronger access assumptions, and a quartic speedup when efficient quantum access to the weights is available.

Those are meaningful algorithmic results. To judge their practical value, we need to know how the required access would be implemented, what it costs, and how the complete system compares with optimized classical inference.

Some researchers are also questioning whether accelerating existing neural-network designs is the most productive starting point. In a March 2026 article, Xanadu’s QML team argued for building around operations that quantum computers perform naturally, including the Quantum Fourier Transform and associated spectral methods. That is a research direction whose usefulness still needs to be demonstrated.

A framework proposed by Masot-Llima and colleagues similarly tries to distinguish models that classical methods can reproduce from those that may retain a quantum advantage. These efforts are useful because they make the assumptions easier to examine. They also show how much of the field is still working out which problems to pursue.

7. “Quantum AI” trading scams

Unfortunately, the phrase also appears in investment schemes with no credible connection to this research.

The UK’s Financial Conduct Authority lists a firm called Quantum AI as unauthorized and warns the public against dealing with it. Hong Kong’s Securities and Futures Commission has warned about suspected virtual-asset fraud using the same name, including suspected use of AI-generated Elon Musk videos and images to imply his involvement.

Neither quantum computing nor AI provides a basis for the guaranteed returns these schemes advertise. The technical language and celebrity endorsements make the pitch look credible to people who have little opportunity to assess it.

There is a security-awareness issue here as well. Employees who encounter these promotions may submit personal details or work email addresses to fraudulent sites. It is worth covering them in awareness training, alongside the more familiar investment and impersonation scams.

What the HSBC and IBM result actually tells us

The HSBC experiment sits in the research on using quantum computation to improve learning from classical data.

The team used production-scale corporate-bond trading data to generate features on IBM’s Heron processors. Those features became inputs to classical machine-learning models. The quantum processing happened offline, as a separate component that the trading models could query. It was outside the live order-routing path.

The reported gain was up to 34% in out-of-sample test scores relative to the comparisons used in the study. That describes predictive performance; it does not measure how much faster a quantum computer completed the task, and it does not establish an improvement in live trading returns.

The noiseless simulation result leaves an important question open. If hardware noise helped, could a classical random transformation or another feature-engineering method provide a comparable benefit? Aaronson’s criticism focuses on the lack of evidence that the improvement requires quantum computation. The noise explanation remains a hypothesis to investigate, rather than an established account of what happened.

My reading is that the study demonstrates a way to incorporate quantum-generated features into a trading-model workflow at realistic data scale. That is useful integration work. Establishing a quantum advantage would require stronger evidence about the source of the benefit, reproducibility, competitive classical baselines, and the cost of the complete process.

Noise contributing to a result does not automatically invalidate it. But if noise is central to the claimed benefit, understanding and benchmarking that contribution becomes part of the work needed to support the claim.

What I would ask about a quantum AI claim

I would start by asking what the quantum computer actually does. Is machine learning helping operate the hardware? Is the hardware processing quantum data? Or is a quantum algorithm supposed to improve a conventional AI workload?

That answer determines what evidence is relevant. A better decoder can be evaluated through accuracy, throughput, and feedback latency. A proposed improvement to AI training needs a comparison with strong classical training methods, including the cost of data preparation, quantum execution, and readout.

I would also ask whether the reported result comes from hardware, simulation, or complexity analysis. All three are useful in research, but they support different conclusions. Simulating a quantum algorithm is a legitimate way to study it; it does not demonstrate a hardware speedup or automatically turn the algorithm into “quantum-inspired” software.

For a commercial claim, I would want the benchmark, the classical baseline, the assumptions about data access, and enough detail for independent scrutiny. Peer review helps, but the comparison itself still needs to match the claim being sold.

There are several results that would make me more optimistic about quantum computing for mainstream AI. One would be a reproducible advantage on a useful classical-data learning task that survives serious attempts to improve the classical baseline. Another would be a full inference implementation that beats an optimized classical system after including data access, error correction, and readout. A hybrid result with a clearly isolated and reproducible quantum contribution would also be significant, even if the application were narrow.

For now, my assessment is uneven because the evidence is uneven. AI is already helping researchers improve quantum computers. Learning from quantum data has produced important theoretical and experimental results. Using quantum computers to accelerate mainstream AI remains a much less settled proposition.

This is the same difficulty I’ve discussed in writing about quantum hype and denial. Broad claims of a quantum AI breakthrough get ahead of the evidence, while blanket dismissal overlooks useful work already happening in labs. I’m interested in the applications where the quantum contribution can be explained and measured. Naming that contribution would make most discussions of “quantum AI” considerably more useful.