Quantum computing vs AI and supercomputers: 5 myths

Quantum computers are not a faster kind of AI chip, and they will not replace GPUs or supercomputers. The research and the vendors’ own plans point the other way: quantum processors are being attached to classical supercomputers as accelerators for a narrow class of problems, they need large amounts of classical computing to work at all, and AI is one of the tools that makes them more reliable. For a company deciding where to put its technology budget in 2026, the practical answer is AI and classical computing for today’s problems, and quantum for learning, a few targeted tests and security.

Five beliefs come up again and again in boardrooms and vendor pitches. Here is what the evidence says about each.

Myth 1: Quantum computers will replace classical computers and GPUs

They will not, and nobody building them says they will. A quantum processing unit (QPU) works as an accelerator. The classical side prepares the circuits, reads out and processes the results, and corrects errors while the computation runs.

The industry is building exactly that. In June 2025 IBM and RIKEN unveiled an IBM Quantum System Two with a 156-qubit Heron processor next to RIKEN’s Fugaku supercomputer, so that “each paradigm” can “seamlessly perform the parts of an algorithm for which it is best suited.” In August 2025 IBM and AMD announced a partnership around a future where “quantum computers work in tandem with powerful high-performance computing and AI infrastructure.” On October 28, 2025 NVIDIA announced NVQLink, an architecture to couple GPUs with quantum processors, supported by 17 QPU builders, five control system builders and nine US national laboratories. By November 17, 2025 NVIDIA said more than a dozen supercomputing centers in Asia, Europe and the Middle East had adopted it.

Error correction makes the dependence concrete. IBM reported in November 2025 that it could decode errors in real time “(less than 480 nanoseconds)” using classical hardware. NVIDIA’s CUDA-Q QEC release of April 2026 decodes on GPUs at about 2.1 microseconds per round over NVQLink. A large fault-tolerant quantum computer will be surrounded by more classical computing, not less.

NVIDIA’s own CEO shows how the framing changed. In January 2025 Jensen Huang said that “15 years for very useful quantum computers” would “probably be on the early side,” and quantum stocks fell sharply, according to Fortune. Two months later, at the Quantum Day of NVIDIA’s GTC conference, he joked that it was “the first event in history where a company CEO invites all of the guests to explain why he was wrong.” By October he was announcing NVQLink and saying that “every NVIDIA GPU scientific supercomputer will be hybrid, tightly coupled with quantum processors.” That is a chip company planning for quantum as a partner to its GPUs, not a replacement.

Myth 2: Quantum computers will make AI training much faster

Not with any known algorithm on any foreseeable machine. The main obstacle is data. Training a modern model means moving enormous datasets, and quantum computers are slow at loading classical data.

The clearest analysis comes from Torsten Hoefler, Thomas Häner and Matthias Troyer, published in Communications of the ACM in May 2023. Their comparison puts a GPU’s data bandwidth at about 10,000 Gbit/s against roughly 1 Gbit/s for an optimistic future quantum computer. They conclude that “quantum computers will be practical for ‘big compute’ problems on small data, not big data problems,” and that “many current machine learning training approaches” will not “achieve quantum advantage with current quantum algorithms in the foreseeable future.”

The problem has been known for a decade. In a 2015 Nature Physics commentary titled “Read the fine print,” Scott Aaronson showed that several celebrated quantum machine learning speedups assume the data is already loaded; if preparing the input takes polynomial time, “the exponential speedup of HHL vanishes in the very first step.” In 2018 Ewin Tang found a classical algorithm showing that a famous quantum recommendation algorithm “does not in fact give an exponential speedup over classical algorithms.”

Benchmarks point the same way. When Bowles, Ahmed and Schuld at Xanadu, a quantum computing company, compared quantum classifiers with standard methods in 2024, they found that “out-of-the-box classical machine learning models outperform the quantum classifiers.” Google posted a preprint in September 2025 claiming that a 68-qubit processor “can learn to generate distributions beyond the reach of classical computers,” which is interesting research but is not yet peer reviewed and is far from training a language model.

Myth 3: Any quantum speedup is a business win

A speedup only matters if it survives the overhead of the hardware. Quantum operations are much slower and more expensive than classical ones, so a modest algorithmic advantage can be wiped out in practice.

Hoefler, Häner and Troyer address this directly. After comparing the speed of quantum and classical hardware, they “come to the clear conclusion that quadratic speedups are insufficient for practical quantum advantage.” Quadratic speedups are what Grover-style search offers, and they underpin many optimization and machine learning proposals. The authors say “small data problems and quantum algorithms with super-quadratic speedups are essential,” and name chemistry and materials science as “the most promising application.” That matches where real progress has appeared: Google’s Quantum Echoes experiment of October 2025 measured a physics quantity about 13,000 times faster than a supercomputer, and Google said quantum-enhanced NMR “could become a powerful tool in drug discovery,” not that it already is.

Myth 4: Once a quantum computer wins, classical computing is out of the race

Classical algorithms keep improving, often in response to quantum claims. IBM’s 2023 “utility” experiment on a 127-qubit processor was soon matched by classical methods. Tindall and colleagues reported tensor network results “significantly more accurate and precise than the results obtained from the quantum processor,” and Begušić, Gray and Chan showed simulations “orders of magnitude faster than the quantum experiment.” In March 2025 Tindall’s group answered D-Wave’s claim of beyond-classical simulation by showing that “state-of-the-art accuracies can be achieved with modest computational resources.”

Classical machines also simulate quantum computers, which lets teams learn and test quantum software without quantum hardware. The Latin America Quantum Computing Center, opened in May 2021 by SENAI CIMATEC and Atos in Salvador, Brazil, is built around a classical machine that can “simulate quantum systems and algorithms of up to 35 qubits.” Our article is quantum computing useful yet? reviews how other advantage claims held up.

The energy question shows the same nuance. A 2020 study by Google, NASA and Oak Ridge researchers found that simulating Google’s random circuit sampling task took 21.1 MWh on the Summit supercomputer, while a quantum computer draws about 15 kW. The same paper warns that the task is “particularly favorable to quantum computers and an advantage is presently not achievable in practice for most other computational problems.”

Myth 5: AI and quantum compete for the same future

They are increasingly the same project. AI is helping quantum computers work, and GPUs are part of every serious quantum roadmap.

The best-known example is AlphaQubit, a neural network decoder from Google DeepMind and Google Quantum AI published in Nature on November 20, 2024. On data from Google’s Sycamore processor it made “6% fewer errors than tensor network methods” and “30% fewer errors than correlated matching.” Google was candid about the limit: “it’s still too slow to correct errors in a superconducting processor in real time.” On the GPU side, platforms such as CUDA-Q run circuits on GPU-accelerated simulators, so teams can test quantum code before paying for hardware time.

The reverse direction, quantum helping AI, is the speculative one. If it comes, the research above suggests it will be on small, structured problems such as generating training data from chemistry simulations, not on training large models.

Where does Latin America fit?

On the hardware side, the region is starting from a small base. None of the supercomputing centers NVIDIA named as NVQLink adopters in November 2025 is in Latin America. Brazil is ahead: besides SENAI CIMATEC’s simulator, the Eldorado institute in Campinas announced in September 2026 the purchase of a 5-qubit IQM Spark for installation in the first quarter of 2027, described as the first IQM system in South America. Our overview of Latin American quantum strategies covers Brazil’s other machines and Chile’s plan, and quantum computing in Colombia covers the Colombian groups.

For companies, the more useful point is that access does not depend on local hardware. Cloud platforms and GPU simulators are available from anywhere, so the constraint is people who can pick the right problem and judge the results.

What should a company do with this?

Put AI and classical computing first for current business problems; that is where returns are today. Then use quantum budget for three things:

1. Learning. A small team that understands hybrid quantum-classical workflows will be ready when error-corrected machines arrive. Quantum training for teams is built for that.
2. Targeted tests. Pick small-data, compute-heavy problems, such as molecular simulation in healthcare and pharma or materials in mining and materials, and test them against a strong classical baseline. Our guide to choosing a first quantum use case explains the scoring, and quantum use case discovery does it with you.
3. Security. The same hardware progress shortens the timeline to machines that break today’s public-key encryption, which is a deadline rather than an option. Start with a quantum readiness assessment.

When a vendor offers “quantum AI,” ask three questions: what is the classical baseline, how is the data loaded, and has the result been peer reviewed? The glossary defines QPU, hybrid quantum-classical computing, CUDA-Q and the other terms used here.