Updated Material. July 2026

Abstract

Quantum processors use superposition, interference, and entanglement to perform certain structured computations differently from classical computers. Artificial intelligence relies heavily on optimisation, sampling, linear algebra, probabilistic inference, and representation learning, creating several possible points of contact between the two fields. Quantum systems may eventually accelerate selected components of AI pipelines, particularly when the data are quantum-native or the problem has mathematical structure that a quantum algorithm can exploit. However, no general practical quantum advantage has yet been established for real-world AI workloads. The principal obstacles are hardware noise, error-correction overhead, classical-data loading, measurement costs, strong classical competitors, and inconsistent benchmarking. This explanatory survey distinguishes demonstrated results, research prototypes, theoretical possibilities, and corporate road maps. Evidence and road maps are current as of July 2026.

What Quantum Computing Can and Cannot Do

A classical bit has a definite value of zero or one. A qubit can exist in a superposition described by amplitudes associated with both values. A register of qubits therefore has a state space whose mathematical size grows exponentially with the number of qubits.

This does not mean that a quantum computer simply tests every possible answer and then reveals the correct one. Measurement returns a limited classical result. A useful quantum algorithm must organise interference so that desirable outcomes become more likely and undesirable outcomes cancel. The computational advantage, when it exists, comes from the structure of the algorithm rather than from unlimited access to the full quantum state.

Entanglement allows correlations that cannot be represented efficiently by many classical descriptions. These properties make quantum processors promising for selected problems in simulation, optimisation, sampling, cryptography, and linear algebra. They do not make quantum computers universally faster than conventional processors.

Qubit Quality and Scalability

Quantum states are extremely sensitive to noise. Useful large-scale computation will therefore require logical qubits, in which information is distributed across multiple physical qubits so that errors can be detected and corrected.

There is no universal physical-to-logical qubit ratio. The required overhead depends on the physical error rate, the correction code, the connectivity of the processor, the target logical error rate, the depth of the algorithm, and the operations that must be performed. A logical memory is also not equivalent to a complete fault-tolerant computer. The latter must support reliable state preparation, logical gates, measurement, real-time decoding, and control across many logical qubits.

Google's Willow processor demonstrated below-threshold surface-code error correction in a 101-physical-qubit, distance-7 logical memory. The reported logical error rate was (0.143% ± 0.003%) per error-correction cycle. Increasing the code distance suppressed the logical error rate, and the logical memory outlived its best constituent physical qubit. This was a major experimental result, but it did not constitute a general-purpose fault-tolerant computer. Nature

Neutral-atom systems have demonstrated another route to logical processing. A programmable processor operating with up to 280 physical atoms implemented several forms of encoding, including computationally complex sampling circuits with up to 48 logical qubits. The experiment demonstrated important error-correction and logical-control components, but not unrestricted fault-tolerant computation at application scale. Nature

Quantum low-density parity-check codes may substantially reduce physical-qubit overhead relative to conventional surface-code designs. IBM's bivariate bicycle code, for example, encodes 12 logical qubits using 144 data qubits and 144 additional check qubits in its proposed implementation. Such codes require demanding long-range connectivity, decoding, logical-gate, and manufacturing capabilities that have not yet been integrated at scale. Their potential reduction in overhead should therefore be treated as an architectural projection rather than an established system-level saving.

Bosonic codes provide another approach by encoding information in the states of oscillators. AWS's Ocelot chip is a research prototype designed to combine bosonic cat qubits with additional error correction. Estimates that this architecture could greatly reduce future correction costs are projections, not measurements from a complete fault-tolerant computer. Nature

Microsoft announced Majorana 1 in 2025 and an upgraded Majorana 2 chip in 2026 as steps toward topologically protected quantum computation. The supporting peer-reviewed experiment demonstrated single-shot fermion-parity measurement, a necessary ingredient for the proposed architecture. The paper explicitly stated that the measurements did not by themselves establish that the detected states were topological. Independent researchers remained sceptical of the stronger topological-qubit claims in 2026. The million-qubit chip remains a design objective, not a demonstrated processor. Nature research paper, Nature assessment

Current Evidence for Quantum Advantage in AI

Quantum machine learning has produced important theoretical results and small experimental demonstrations, but no broadly accepted end-to-end advantage has been established for practical AI workloads. Many reported improvements occur in ideal simulations, on small datasets, or against classical baselines that do not represent the strongest available methods.

Some quantum algorithms offer large asymptotic speedups under restrictive assumptions. The HHL algorithm for linear systems is an instructive example. Its exponential speedup requires a sparse and well-conditioned matrix, efficient preparation of the input quantum state, efficient access to matrix elements, and a task that needs only a limited property of the solution. Reading the complete solution vector would remove much of the advantage. These conditions are not generally guaranteed together in practical machine-learning workloads. Physical Review Letters

A 2025 systematic review of quantum machine learning in digital health identified 169 eligible studies. The reviewers excluded 123 from the final synthesis for insufficient methodological rigour, and only 16 examined realistic conditions through hardware or noisy simulation. The remaining evidence did not establish a clear performance benefit over classical methods. This result concerns digital health rather than all quantum machine learning, but it illustrates the field's benchmarking problem. npj Digital Medicine

The most credible research targets currently include quantum-native data, specialised sampling problems, quantum chemistry, selected optimisation structures, quantum kernels, generative modelling, and probabilistic inference. Their promise remains problem-specific. A successful result in one narrow task cannot establish that quantum processing will generally accelerate AI.

Data Input and Measurement

A quantum computer cannot absorb an arbitrarily large classical dataset without computational cost. Encoding millions of classical values into quantum amplitudes may require enough operations to eliminate the expected speedup. Quantum random-access memory has been proposed as one solution, but scalable fault-tolerant QRAM is not currently available.

The data-loading problem is less severe when the input is generated directly by a quantum sensor, quantum simulation, or another quantum process. Quantum learning may therefore become useful earlier for quantum-native data than for conventional text, image, or business datasets.

Measurement is not an absolute barrier. For classification and sampling tasks, samples may be exactly the required output. The practical cost appears when reliable probabilities, expectation values, gradients, or small differences between candidate outputs require many repeated circuit executions, known as shots. The necessary number of shots depends on the required precision, confidence level, hardware noise, and separation between possible results.

Amplitude amplification and amplitude estimation can reduce repetition costs for particular structured problems. They do not provide a universal method for extracting the full contents of an arbitrary quantum state.

Engineering and Economic Constraints

Current quantum processors require specialised control equipment, calibration, error mitigation, and repeated measurement. Superconducting platforms generally require dilution refrigeration. Neutral-atom and trapped-ion systems require precise lasers and vacuum systems. Photonic architectures can reduce some cooling requirements but still face demanding source, detector, loss, and interconnection problems.

For small variational circuits that remain easy to simulate classically, current cloud quantum execution adds task charges, shot charges, queueing, and repeated-measurement costs without a demonstrated computational advantage over classical execution. Direct economic comparisons are meaningful only when the quantum and classical systems solve the same task to the same accuracy and include data preparation, control, error correction, measurement, and post-processing.

Room-temperature qubits, photonic processors, and improved compilers may eventually reduce some infrastructure and execution costs. Their economic effect cannot yet be expressed through a reliable universal percentage.

Theoretical Gaps and the Moving Classical Frontier

Complexity theory does not currently establish an unconditional separation between BQP, the class of problems efficiently solvable by a quantum computer with bounded error, and BPP, its probabilistic classical counterpart. Theoretical quantum speedups remain valuable, but their practical relevance depends on whether their assumptions can be satisfied by real data and hardware.

Classical simulation is also a moving target. Tensor-network methods such as matrix product states and projected entangled-pair states, together with variational Monte Carlo and other approximate techniques, continue to extend the range of quantum systems that can be simulated classically. Some early quantum-performance claims have been reproduced or challenged by improved classical algorithms. Quantum advantage must therefore be tested against the best available classical method at the time of the experiment.

A valid benchmark must account for:

• data preparation and encoding;
• circuit compilation and execution;
• noise and error correction;
• measurement and statistical uncertainty;
• classical pre-processing and post-processing;
• the strongest relevant classical baseline;
• equal accuracy and confidence requirements;
• time, monetary cost, energy use, and hardware utilisation;
• open code, data, parameters, and hardware configuration;
• replication by researchers independent of the original hardware provider.

Performance on an ideal simulator is evidence about an algorithm, not proof that current quantum hardware can provide the same result. A larger qubit count is also not sufficient. Fidelity, connectivity, circuit depth, logical operations, measurement quality, and classical control may matter more than the nominal number of physical qubits.

Integration with Classical AI

Practical quantum systems are expected to operate as specialised accelerators inside classical computing environments. Frameworks such as PennyLane, Qiskit Machine Learning, TensorFlow Quantum, and CUDA-Q support the construction of hybrid workflows in which classical processors optimise parameters, prepare data, call quantum circuits, and analyse measurement results.

This model introduces communication latency and orchestration costs. For iterative algorithms, repeated movement between a cloud quantum processor and a classical optimiser can dominate execution time. Co-located control electronics and high-performance classical computing may reduce this problem, but the full hybrid pipeline must be benchmarked rather than the quantum circuit in isolation.

The more immediate interaction may run in the opposite direction: AI for quantum computing. Machine-learning methods are already used in qubit calibration, device control, noise modelling, circuit compilation, experiment design, readout, error decoding, and the search for improved correction codes. These applications do not require quantum processors to outperform classical AI. Instead, classical AI helps make quantum hardware more stable, efficient, and programmable. Nature Communications

Principal Research Directions

Quantum-Native Learning

Quantum processors may analyse data produced by quantum experiments or sensors without first converting the complete state into a large classical representation.

Quantum Chemistry and Materials

Molecular and many-body systems are natural targets because they are themselves quantum mechanical. Hybrid quantum-classical methods have already produced scientifically meaningful small-scale simulations, although broad practical advantage over the best classical techniques remains unproven.

Optimisation

QAOA, quantum annealing, and related methods investigate combinatorial problems in scheduling, routing, resource allocation, and scientific design. Their value depends on outperforming strong classical heuristics under complete cost accounting.

Sampling and Generative Modelling

Quantum systems naturally generate samples from probability distributions. The central question is whether the relevant distribution is useful, difficult to reproduce classically, and accessible with an acceptable number of measurements.

Quantum Kernels and Classification

Quantum feature maps may represent certain data structures in ways that are difficult to reproduce classically. Demonstrating benefit requires a realistic dataset, a strong classical kernel baseline, and evidence that data-loading costs do not eliminate the gain.

Quantum Reinforcement Learning

Quantum subroutines may assist with exploration, policy evaluation, or sampling. Current evidence is primarily theoretical or based on small demonstrations, and no general training advantage has been established.

Road Maps and Research Targets for 2026 to 2030

The horizon descriptions below distinguish present results, near-term targets through 2028, mid-term targets from 2029 through 2031, and long-term or undated objectives. Corporate targets should be understood as announced plans, not independently validated forecasts.

Google Willow

Willow represents a demonstrated present result. Google achieved below-threshold error correction in a logical memory with an error rate of (0.143% ± 0.003%) per correction cycle. The result establishes improved logical-memory performance as code distance increases, but it does not constitute a complete fault-tolerant computer.

Neutral-Atom Logical Processing

Neutral-atom logical processing is a demonstrated research prototype with present results and continued near-term development. Researchers have implemented logical circuits with up to 48 encoded qubits and several important error-correction and logical-control components. Application-scale universal fault-tolerant computation has not yet been demonstrated.

AWS Ocelot

Ocelot is a research prototype with a long-term or unspecified deployment horizon. It demonstrates components of a bosonic error-correction architecture intended to reduce the physical resources required for future fault-tolerant systems. Estimates of large reductions in hardware overhead remain architectural projections.

Microsoft Majorana Programme

The Majorana programme combines a research prototype with disputed platform claims and has a long-term or unspecified horizon. Fermion-parity measurement has been demonstrated, but the claimed topological nature of the underlying states and the scalability of the proposed qubit architecture remain contested. The planned million-qubit processor has not been demonstrated.

IBM Starling

Starling is a corporate road map with a mid-term target of 2029. IBM describes it as a planned system containing 200 logical qubits and capable of running quantum circuits comprising 100 million quantum gates on 200 logical qubits. These specifications are development objectives rather than the capabilities of an existing machine.

Photonic and Room-Temperature Architectures

Photonic and room-temperature platforms remain research programmes with long-term or uncertain deployment horizons. They may reduce some cooling, packaging, and integration costs, but unresolved problems include photon loss, source quality, detector performance, fabrication, interconnection, control, and large-scale manufacturing.

Conclusion

Quantum computing should not be understood as a faster replacement for conventional AI hardware. Its plausible role is narrower and potentially more important: accelerating selected subproblems whose mathematical or physical structure can be exploited by a quantum algorithm.

Progress in error correction, neutral atoms, superconducting systems, bosonic codes, photonics, and topological research is substantial. Quantum processors are already useful scientific research instruments. Hybrid systems have performed selected chemistry and materials calculations, including active-space problems beyond the scale of straightforward exact diagonalisation. These results depend heavily on classical co-processing and do not yet establish a broadly accepted practical advantage over the best classical methods. Communications Physics

The decisive questions are whether useful data can be prepared efficiently, whether logical operations can be performed at sufficiently low error rates, and whether the complete quantum-classical pipeline can outperform the strongest classical alternative in time, cost, accuracy, and energy use. Quantum processors may eventually become specialised accelerators for optimisation, sampling, chemistry, materials research, and quantum-native learning. The timing and economic scale of that transition cannot yet be predicted reliably.

* * *

References

Google Quantum AI and Collaborators. “Quantum Error Correction Below the Surface Code Threshold.” Nature 638, 920-926, 2025.

Bluvstein, D., Evered, S. J., Geim, A. A., et al. “Logical Quantum Processor Based on Reconfigurable Atom Arrays.” Nature 626, 58-65, 2024.

Bravyi, S., Cross, A. W., Gambetta, J. M., Maslov, D., Rall, P., and Yoder, T. J. “High-Threshold and Low-Overhead Fault-Tolerant Quantum Memory.” Nature 627, 778-782, 2024.

Putterman, H., Noh, K., Hann, C. T., et al. “Hardware-Efficient Quantum Error Correction via Concatenated Bosonic Qubits.” Nature 638, 927-934, 2025.

Microsoft Azure Quantum, Aghaee, M., Alcaraz Ramirez, A., et al. “Interferometric Single-Shot Parity Measurement in InAs-Al Hybrid Devices.” Nature 638, 651-655, 2025.

Castelvecchi, D. “Microsoft Upgrades Controversial Quantum Chip: Researchers Are Still Sceptical.” Nature, 2026.

Harrow, A. W., Hassidim, A., and Lloyd, S. “Quantum Algorithm for Linear Systems of Equations.” Physical Review Letters 103, 150502, 2009.

Bernstein, E., and Vazirani, U. “Quantum Complexity Theory.” SIAM Journal on Computing 26, 1411-1473, 1997.

Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., and Lloyd, S. “Quantum Machine Learning.” Nature 549, 195-202, 2017.

Cerezo, M., Verdon, G., Huang, H. Y., Cincio, L., and Coles, P. J. “Challenges and Opportunities in Quantum Machine Learning.” Nature Computational Science 2, 567-576, 2022.

Gupta, R. S., Wood, C. E., Engstrom, T., Pole, J. D., and Shrapnel, S. “A Systematic Review of Quantum Machine Learning for Digital Health.” npj Digital Medicine 8, 237, 2025.

Xu, S., Hann, C., Foxman, B., Girvin, S. M., and Ding, Y. “Systems Architecture for Quantum Random Access Memory.” IEEE/ACM International Symposium on Microarchitecture, 2023.

Amazon Web Services. “Amazon Braket Pricing.” Accessed July 22, 2026.

Tindall, J., Fishman, M., Stoudenmire, E. M., and Sels, D. “Efficient Tensor Network Simulation of IBM’s Eagle Kicked Ising Experiment.” PRX Quantum 5, 010308, 2024.

Begusic, T., Gray, J., and Chan, G. K. L. “Fast and Converged Classical Simulations of Evidence for the Utility of Quantum Computing Before Fault Tolerance.” Science Advances 10, eadk4321, 2024.

Alexeev, Y., Farag, M. H., Patti, T. L., et al. “Artificial Intelligence for Quantum Computing.” Nature Communications 16, 10829, 2025.

PennyLane. “PennyLane Documentation.” Accessed July 22, 2026.

Qiskit Machine Learning Development Team. “Qiskit Machine Learning.” Accessed July 22, 2026.

Google. “TensorFlow Quantum.” Accessed July 22, 2026.

NVIDIA. “CUDA-Q Documentation.” Accessed July 22, 2026.

Mandelbaum, R. “IBM Lays Out Clear Path to Fault-Tolerant Quantum Computing.” IBM, 2025.

Kaliakin, D., Shajan, A., Liang, F., et al. “Accurate Quantum-Centric Simulations of Intermolecular Interactions.” Communications Physics 8, 396, 2025.