What quantum machine learning is trying to do
Quantum machine learning (QML) commonly encodes classical data into quantum states, transforms features with parameterized circuits or quantum kernels, and uses a classical optimizer to update parameters. Entanglement and large Hilbert spaces may represent complex correlations with relatively few trainable parameters.
An open 2026 review in Discover Computing compared work in energy forecasting, finance, health, materials, generative modeling and clustering. Some small datasets showed competitive accuracy with fewer parameters. Parameter efficiency, however, is not the same as faster execution.
Why it is not yet faster than classical AI
- Data loading: Turning large classical datasets into quantum states can erase theoretical speedups. Devices assumed by some algorithms, such as practical QRAM, do not yet exist.
- Noise and shallow circuits: NISQ hardware cannot reliably execute deep circuits, while barren plateaus can make gradients vanish.
- Repeated measurement: A quantum state does not yield a complete answer in one measurement. Error mitigation and gradient estimation can require many circuit shots.
- Quantum–classical communication: Hybrid training alternates QPU execution with CPU or GPU optimization. The review found that communication and circuit runtime prevent parameter efficiency from becoming wall-clock speedup.
- Improving classical baselines: Quantum-inspired classical algorithms have closed some early theoretical gaps. QML must beat a well-tuned modern classical model, not a weak baseline.
Three development scenarios
Near term—AI helps quantum. Machine learning becomes useful first in decoding, calibration, circuit search and experiment automation rather than QML applications.
Medium term—small, structured hybrid problems. QPUs are tested as coprocessors for chemistry and materials data that is natively quantum, compact feature maps, and specialized optimization kernels. Evaluation must include the full pipeline from loading to final output.
Long term—fault-tolerant quantum-native learning. With enough logical qubits, deep reliable circuits and efficient state preparation, quantum linear algebra, sampling or correlation routines could accelerate parts of AI pipelines. This would still not mean replacing all Transformer training.
A credible QML claim must beat a strong classical baseline at equal quality in end-to-end time, cost and energy, including encoding, shots, error correction, classical pre- and post-processing, and communication. Broad practical evidence meeting that standard does not yet exist.