For most commercial workloads, a quantum computer offers no advantage today. A short list of activities is still worth an organisation's time in 2026, and one of them, preparing cryptography for future quantum attacks, has a deadline that does not depend on when large machines arrive. This article goes through the list and gives a way to test any quantum claim you are shown.
The state of the hardware
Current processors hold from a few dozen to more than a thousand physical qubits, depending on the technology: superconducting circuits, trapped ions, neutral atoms or photonics. The qubit count matters less than the error rate. Two-qubit gate errors on good devices are in the range of one in a hundred to one in a thousand. A circuit with a few thousand such gates produces output that is mostly noise, and that sets the size of problem that can be run directly.
Error correction is the way past this limit. It combines many physical qubits into one logical qubit with a lower error rate. Experiments published in 2024 showed logical error falling as the code was made larger, which is the behaviour the theory requires. Useful fault-tolerant computing needs thousands of logical qubits, and each may need hundreds or thousands of physical qubits. The estimate for factoring a 2048-bit RSA key shows the gap: a 2019 analysis by Gidney and Ekerå put the requirement at about 20 million noisy qubits, and a 2025 paper by Gidney reduced it to under one million. Both figures are far beyond any machine that exists.
The rest of this article follows from that gap. Algorithms with proven large speed-ups, such as Shor's algorithm for factoring, need fault tolerance. Algorithms that fit on today's devices have no proven advantage for commercial problems.
1. Post-quantum cryptography: practical and urgent
This is not quantum computing, but quantum computing is the reason for it. RSA and elliptic-curve cryptography will be breakable by a sufficiently large fault-tolerant machine. Encrypted traffic recorded today can be stored and decrypted when that machine exists. Any data that must remain confidential for ten or twenty years is therefore exposed by the key exchange used to send it now.
NIST published three post-quantum standards in August 2024: FIPS 203 (ML-KEM) for key establishment, and FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA) for signatures. Mainstream libraries and browsers support hybrid key exchange, which pairs a classical algorithm with a post-quantum one. The work for an organisation is an inventory of where public-key cryptography is used, a ranking by data lifetime and upgrade difficulty, and a phased migration. Embedded devices and long-lived certificates take the longest. Our quantum computing solutions and cybersecurity practices handle this together.
2. Building skills on real hardware
Cloud access to quantum processors is billed by use and needs no capital purchase. A small team can learn circuit design, noise behaviour and error mitigation on real devices for modest cost. The return is judgement: engineers who have run experiments can evaluate vendor claims and recognise when the technology becomes relevant to their own problems. For universities, the same access supports teaching laboratories and publishable experiments.
3. Simulating quantum systems
Chemistry and materials science are the most natural applications, because the systems being modelled are themselves quantum mechanical. On current hardware the molecules that can be simulated are small enough for classical methods to handle well. The value today lies in method development: better ansatz design, measurement reduction and error mitigation that will carry over to larger machines. This is research with a long horizon, suited to organisations that already fund computational chemistry.
4. Optimisation: benchmark before believing
Scheduling, routing and portfolio problems are often presented as quantum applications, through QAOA on gate-based machines or through quantum annealers. No one has demonstrated an advantage over the best classical solvers on industrially relevant instances. Mixed-integer programming, local search and simulated annealing are mature, and they improve every year.
The exercise can still pay off. Writing a business problem in the mathematical form that a quantum optimiser requires clarifies the objective and the constraints, and the same formulation often runs well on a classical or quantum-inspired solver. A feasibility study should report all of them side by side and recommend whichever wins. Today that is usually the classical one.
5. Quantum machine learning: a research topic
Quantum kernels and variational classifiers can be trained on small datasets, and they sometimes match classical models on those datasets. Several obstacles stand between that and practical use. Loading classical data into a quantum state is expensive. Training landscapes flatten as circuits grow, a problem known as barren plateaus. For some proposed quantum models, researchers have found classical algorithms that perform equally well. The honest position is that quantum machine learning is an open research question. It is reasonable to study it, and it should not be on a product roadmap. Where the business need is a working model, our AI and machine learning practice will use classical methods.
6. Quantum technologies that are not computers
Some quantum products are available now. Quantum random number generators supply entropy from a physical quantum process, are sold as chips and cards, and integrate easily into key generation. Quantum key distribution is deployed on a limited number of fibre and free-space links. It needs dedicated hardware, works over limited distances without trusted relay nodes, and for most organisations post-quantum algorithms are the more practical protection. Quantum sensors, including atomic clocks, magnetometers and gravimeters, are the most mature part of the field, though they address measurement and not computation.
Five questions to ask about any quantum claim
- What classical method was it compared with, and was that method properly tuned and given the same time budget?
- Was the result obtained on real hardware with noise, or on an ideal simulator?
- How large was the problem instance, and how does the result change as the instance grows?
- How many runs were made, and what is the statistical uncertainty?
- Does the speed-up include the time to load data and read out results?
A claim that cannot answer these is a demonstration. It is not evidence of advantage.
What a technology leader can do this year
Start the cryptographic inventory, because it is needed whatever the pace of hardware progress. Choose one candidate workload and commission a feasibility assessment with a classical baseline, so that the organisation has a documented position. Give two or three engineers time to learn on cloud hardware. Do not buy quantum hardware. Review the position annually, since error-correction progress is the signal that will change it.