Quantum computers versus supercomputers

Every machine in this dataset is a classical computer: transistors, bits, deterministic logic, however many million cores are wired together. A quantum computer is not a faster version of that. It is built out of a different physical resource, it fails in a different way, and for the overwhelming majority of what the systems in this dataset actually do (PDEs on a grid, molecular dynamics, dense linear algebra, neural network training) it currently offers nothing at all. Where it is useful, and increasingly deployed, is next to classical HPC systems, not instead of them.

A different computational model

A classical bit is 0 or 1. A qubit is a two-level quantum system (a superconducting circuit, a trapped ion, a neutral atom, a photon) that can be placed in a superposition of both states at once, described by two complex amplitudes rather than a single value. A register of n qubits is described by 2^n amplitudes, which is why quantum state grows exponentially with qubit count and why classical simulation of a quantum computer becomes intractable past roughly 50 qubits.

Entanglement is the second resource: qubits can be correlated in a way that has no classical analogue, so that the state of the whole register is not reducible to the states of its parts. Computation proceeds by applying quantum gates, unitary operations that rotate the state of one or more qubits, building up a circuit, and then measuring, which collapses the superposition to a classical bit string with a probability given by the amplitudes.

That last step is the catch that separates a quantum computer from a magic exponential parallel machine. Measurement gives you one sample, not the whole superposition. A quantum algorithm has to be designed so the amplitudes of wrong answers cancel out through interference and the amplitude of the right answer is reinforced, so a small number of measurements is enough. Very few known problems admit that kind of algorithm, which is why quantum advantage is narrow rather than general.

Where the hardware actually is

The field is in what physicist John Preskill named the NISQ era: Noisy Intermediate-Scale Quantum. Devices today have on the order of tens to a few hundred physical qubits with gate error rates high enough that circuits decohere, meaning lose their quantum information to the environment, after a limited number of operations. There is no error correction running by default; a physical qubit is not yet a reliable logical qubit.

The clearest recent evidence of progress toward fixing this is Google’s December 2024 result on its Willow processor, published in Nature: a distance-7 surface code built from 101 physical qubits ran below the error-correction threshold, meaning the logical error rate fell as more physical qubits were added rather than rising, at a measured 0.143% error per cycle, with the logical qubit’s effective lifetime exceeding its best constituent physical qubit’s lifetime by a factor of 2.4. That is the first experimental demonstration of this specific, long-sought crossover: proof the error-correction approach can in principle scale, not evidence that a large fault-tolerant machine exists.

IBM’s own public roadmap illustrates the distance still to cover. Its 2026 processor, Nighthawk, is a 120-qubit device aimed at deeper circuits rather than more qubits. Its fault-tolerant target, IBM Quantum Starling, is scheduled for 2029 and is specified as roughly 200 logical qubits running on the order of 100 million quantum gates, reached via intermediate systems (Kookaburra in 2026, Cockatoo in 2027) that demonstrate quantum low-density parity-check codes and multi-module entanglement. Two hundred logical qubits by 2029, if achieved on schedule, is a meaningful fault-tolerant machine and still a small one next to what most proposed cryptographic or simulation applications require.

What each machine is actually good at

Three quantum algorithms carry almost the entire theoretical case for quantum advantage, and it is worth being precise about how proven each one is:

• Shor’s algorithm factors integers exponentially faster than the best known classical method, which is why it threatens RSA-style public-key cryptography. It also has the widest gap between theory and demonstrated hardware: the largest integer factored by a genuine implementation of Shor’s algorithm on real quantum hardware is 21, a 2012 result that relied on prior knowledge of the answer to simplify the circuit. Larger factoring claims (56,153; a reported 48-bit hybrid result) used different, non-Shor methods that do not scale the same way. The exponential speedup is real mathematics; a cryptographically relevant demonstration of it does not exist yet.
• Grover’s algorithm gives a quadratic speedup for unstructured search, proven and well understood, but a quadratic speedup is a far weaker practical case than an exponential one: it takes an enormous problem before the crossover with classical search is worth the overhead of running it on noisy hardware.
• Quantum chemistry and materials simulation is the strongest near-term case, for a structural reason rather than a benchmark result: a quantum system simulating another quantum system is a natural fit rather than an indirect encoding, so resource requirements grow far more gently than for classical simulation of the same molecule. This is theorized and partially demonstrated on small systems, not yet proven at a scale classical HPC cannot already reach; essentially every EuroHPC quantum deployment below lists chemistry and materials science as a first target application.
• QAOA and other variational optimization algorithms run on current NISQ hardware for combinatorial optimization (logistics, scheduling, portfolios), but there is no proven advantage over the best classical heuristics at any instance size that matters commercially. This is the most oversold category in the field.

Against that, the actual daily work of the systems in this dataset (partial differential equations on a grid, climate and weather models, computational fluid dynamics, nuclear stockpile stewardship calculations, dense linear algebra, and neural network training, covered in what makes a supercomputer different and AI clusters versus HPC systems) has no known quantum algorithm that helps at all. A quantum computer is not a general accelerator; it is a specialised co-processor for a short list of problem shapes, most of which are not what a national HPC facility spends its cycles on.

Hybrid quantum-classical computing, which is what is actually being built

Given that narrow and mostly theorized advantage, almost nobody is building a standalone quantum computer meant to replace an HPC system. What is actually being deployed, especially across EuroHPC, is a QPU physically co-located with a classical machine, connected through the same job scheduler, running variational hybrid algorithms in which the classical system executes the optimizer loop and the QPU evaluates the quantum circuit on each iteration.

LUMI, hosted by CSC in Kajaani, Finland, is receiving LUMI-IQ, an IQM Halocene H4 superconducting quantum computer. It will be delivered in 2027 with 150 physical qubits and early quantum error correction, upgraded from 2028 toward real-time error correction, and is being built into the LUMI AI Factory environment alongside LUMI’s classical and AI capacity. Funding comes from EuroHPC JU plus Finland, Czechia, Norway and Poland.

Karolina, at IT4Innovations in Ostrava, Czechia, already has its quantum computer running: VLQ, an IQM Star 24 system with 24 superconducting qubits in a star topology, inaugurated September 2025 and integrated into Karolina. The same LUMI-Q consortium (a separate project from LUMI-IQ, despite the near-identical name) is meant to let that quantum resource also be reached from LUMI and from Poland’s EHPCPL.

Leonardo, hosted by CINECA in Bologna, now has two on-premises quantum computers integrated with it. SOL is a Pasqal Orion neutral-atom system with 140 qubits, engineered for tight integration with Leonardo through Pasqal’s QRMI resource-management stack and compatible with CUDA-Q and Qiskit. NOX is an IQM Radiance 54 superconducting system, CINECA’s first on-premises superconducting machine, providing what IQM describes as a production environment for hybrid classical-quantum workflows.

MareNostrum 5, at BSC in Barcelona, sits alongside MareNostrum-Ona: by 2026, three quantum computers in the Torre Girona chapel that housed earlier MareNostrum generations, two digital superconducting transmon-qubit systems from the Quantum Spain project plus an analog annealer built from fluxonium qubits by Qilimanjaro. BSC describes this as making MareNostrum 5 one of the first supercomputers anywhere combining classical, digital-quantum and analog-quantum computing in one facility.

At Jülich, JUNIQ connects several quantum systems to specific classical machines rather than one QPU to one computer. A D-Wave Advantage annealer with up to 5,760 superconducting qubits is slated to connect to JUPITER, Germany’s EuroHPC exascale system. An IQM Spark 5-qubit gate-based system and a Pasqal JADE neutral-atom simulator (100-plus qubits, part of the HPCQS project) both connect to JURECA-DC. Separately, JUWELS Booster’s GPUs are used within JUNIQ to classically simulate small quantum circuits, a second and distinct kind of overlap: using an HPC system to model a quantum computer rather than pairing it with one.

MeluXina, LuxProvide’s EuroHPC system in Luxembourg, is earlier in the process: EuroHPC JU opened procurement in July 2026 for MeluXina-Q, specified at a minimum of 10 semiconductor spin qubits rising toward more than 80 over its operational life, to be integrated into MeluXina. As of this writing it is a procurement in progress, not a deployed system, and is included on that basis.

Outside EuroHPC the pattern is similar but earlier-stage. At Oak Ridge, an IQM Radiance system named Pathfinder (20 superconducting qubits) was deployed in mid-2026 in a dedicated lab on the ORNL campus, the same site that hosts Frontier. The source describing it is explicit that Pathfinder connects to a separate NCCS test bed in a different building and does not claim a working link to Frontier itself; the stated goal is an open hybrid-computing software stack, earlier-stage work than Europe’s on-scheduler integrations. At NERSC the overlap runs the other direction: Perlmutter has been used to classically simulate the physical design of a superconducting quantum chip built by Berkeley Lab’s Advanced Quantum Testbed, using nearly all of its 7,168 GPUs for a 24-hour run, which is HPC helping build a quantum computer rather than HPC paired with one.

Where this goes

Nobody credibly knows when, or whether on any fixed date, fault-tolerant quantum computing arrives at a scale that matters, and the honest position is to say so rather than pick a number. The US Department of Energy has issued a grand challenge for a scientifically useful fault-tolerant quantum computer by 2028, which the physicist Steven Girvin, among others, has called “a very optimistic but worthy goal” given that the field remains, in his words, far from true fault tolerance. IBM’s own roadmap targets 2029 for a 200-logical-qubit machine, a real milestone and still well short of what most proposed cryptographic or large-scale simulation applications would need. A 2026 analysis modeling the arrival of a cryptanalytically relevant quantum computer found expert surveys clustering around 2038 to 2040, while bottom-up physical hardware-scaling models clustered closer to 2052, a full decade of disagreement between two reasonable methods, with a resulting 80% probability range spanning roughly 2032 to 2060.

What is not uncertain is the near-term shape of the field: narrow, proven advantages on a short list of problems, a much longer list of theorized advantages that have not been demonstrated at meaningful scale, and a growing number of classical HPC systems, several of them in this dataset, with a quantum processor now sitting in the same room, on the same scheduler, running the classical optimizer loop around a quantum circuit. This is a genuinely new and fast-moving area, the EuroHPC deployments above have mostly landed within the last two years, and this site is likely to expand this chapter as more of them come online.

The machines themselves, with their qubit counts, operators and published benchmark figures, are in the quantum section; the benchmarks page explains why no ranked quantum list exists.