1. The starting point: AI's electricity demand and the hopes placed on quantum computers

As AI grows, so does the electricity needed to run it. According to the International Energy Agency (IEA), data centres used around 415 terawatt-hours (TWh) in 2024, about 1.5% of the world's electricity consumption, and that figure is set to more than double to around 945 TWh by 2030. The IEA names AI as the most important driver of that growth.

Against this backdrop, the claim that "quantum computers will break AI's limits" comes up often. Yet physicist John Preskill sounded a note of caution. In his 2018 paper on Noisy Intermediate-Scale Quantum (NISQ) technology, he wrote that noise in quantum gates will limit the size of circuits that can be executed reliably. He also wrote that a 100-qubit quantum computer "will not change the world right away."

This article follows what the papers and official announcements actually show, rather than the expectations.

Golden multi-stage dilution refrigerator chandelier assembly used to cool superconducting quantum processors to millikelvin temperatures
A cryogenic dilution refrigerator system used to cool superconducting quantum processors near absolute zero for quantum computing experiments · PhOtOnQuAnTiQuE from Earth France · CC BY-SA 2.0 · Wikimedia Commons

2. What the quantum advantage experiments actually showed

In 2019 Google researchers reported a result in Nature (Arute et al.). Their 53-qubit superconducting processor, called Sycamore, took about 200 seconds to sample one instance of a random quantum circuit a million times. Their benchmarks at the time indicated that a state-of-the-art classical supercomputer would need approximately 10,000 years for the equivalent task. They described the result as an experimental realization of quantum supremacy "for this specific computational task."

The same paper also added a caveat. It noted that simulation methods had improved steadily. The authors expected lower simulation costs than reported to eventually be achieved. Still, they expected hardware improvements on larger processors to keep outpacing them.

The challenge came quickly. IBM researchers argued in arXiv:1910.09534 that such circuits could be simulated with high fidelity in a matter of days. Their approach used secondary storage on Summit, the supercomputer at Oak Ridge National Laboratory.

Later, Pan, Chen and Zhang reported in arXiv:2111.03011 that they generated one million uncorrelated samples for the 53-qubit, 20-cycle Sycamore circuit. It took about 15 hours on a cluster of 512 GPUs, at a target fidelity of about 0.0037.

Researchers at the University of Science and Technology of China (USTC) pursued similar experiments. Their superconducting processor Zuchongzhi 2.1 has 66 qubits and performed random circuit sampling at a scale of up to 60 qubits and 24 cycles in about 4.2 hours; the team estimated that simulating the same task classically would take about 4.8×10⁴ years (Zhu et al., arXiv:2109.03494). That paper's abstract opens by saying quantum hardware must be upgraded "to withstand the competition of continuously improved classical algorithms and hardwares." The group's photonic quantum computer Jiuzhang observed up to 76 output photon clicks in a Gaussian boson sampling experiment (Zhong et al., arXiv:2012.01625).

In short, these experiments show a contest between quantum devices and classical computation on specific sampling tasks. They are not results in which an AI model was trained faster.

A square quantum processor chip mounted on a metal substrate against a dark background
Google's Sycamore quantum processor chip, from a Google video (2019) · Google ( https://www.youtube.com/channel/UCK8sQmJBp8GCxrOtXWBpyEA ) · CC BY 3.0 · Wikimedia Commons

3. What quantum machine learning is trying to do

A widely cited starting point for quantum machine learning is the 2019 Nature paper by Havlíček and colleagues (arXiv:1804.11326).

The authors begin with kernel methods, such as support vector machines (SVMs). They note that these methods are ubiquitous. But kernel methods run into limits when the feature space becomes large and kernel functions become expensive to estimate.

The authors then explain where quantum computers could help. A core element of quantum speed-ups, they say, is exploiting an exponentially large quantum state space. This is done through controllable entanglement and interference.

The paper proposed two methods and implemented both experimentally on a superconducting processor. The first is the "quantum variational classifier." It uses a variational quantum circuit to classify training data. The second is the "quantum kernel estimator." It uses the quantum computer to estimate the kernel function, then optimizes the classifier with those values. Both methods use the quantum state space as the feature space of the classification problem.

The wording matters. The paper says that a quantum-enhanced feature space that is only efficiently accessible on a quantum computer provides "a possible path" to quantum advantage. What it offers is a direction still to be verified, not an established advantage.

This makes clear that a large quantum state space and better classification results are two separate things. That is why the two should be kept apart whenever the phrase "quantum machine learning" comes up.

4. Today's leading strategy: variational quantum algorithms (VQAs)

A review by Cerezo and colleagues in Nature Reviews Physics (2021, arXiv:2012.09265) sums up the situation. Simulating complicated quantum systems is very hard for classical computers. So is solving large-scale linear algebra problems. Quantum computers promise a solution. But fault-tolerant quantum computers will likely not be available in the near future. Current devices have limited numbers of qubits, and noise limits circuit depth.

The leading response to these constraints is the variational quantum algorithm (VQA). It trains a parameterized quantum circuit with a classical optimizer. The quantum device estimates a cost function. A classical computer then uses that value to update the circuit parameters.

Gradients can also be estimated on the quantum hardware itself, using the "parameter-shift rule." This rule runs the circuit again with shifted parameters. This is a different structure from the popular picture, in which "the quantum computer calculates and a GPU runs backpropagation."

Two algorithms fall within this framework. The Variational Quantum Eigensolver (VQE) targets ground-state energies of molecules. The Quantum Approximate Optimization Algorithm (QAOA) seeks approximate solutions to combinatorial problems.

The Cerezo review says VQAs appear to be "the best hope" for obtaining quantum advantage. But it also states plainly that challenges remain in their trainability, accuracy and efficiency.

In VQAs, the quantum device and the classical computer do not replace each other; they mesh within a single computation. That the same review names VQAs the "best hope" while listing open challenges can be read as a sign that the field is still at an early stage.

5. The wall that blocks training: barren plateaus

The best-known trainability problem is the "barren plateau." It is described in a paper by McClean and colleagues in Nature Communications (2018, arXiv:1803.11173). The paper notes that random circuits are often proposed as initial guesses for parameterized quantum circuits. But it shows that this choice is unsuitable for hybrid algorithms run on more than a few qubits. The reasons are the exponential dimension of Hilbert space and the complexity of estimating gradients.

Specifically, consider a wide class of reasonable parameterized quantum circuits. For these, the probability that the gradient along any reasonable direction is non-zero to some fixed precision is exponentially small in the number of qubits. The authors relate this to the 2-design characteristic of random circuits and conclude that solutions must be studied.

In plain terms, adding qubits does not by itself improve training. Without careful initialization and circuit design, training can stall altogether.

This shows that the "trainability" challenge flagged by the review in the previous section is not an abstract worry. It also suggests that the hardware figures that follow cannot be judged by qubit count alone.

6. Hardware: three approaches and the published numbers

Superconducting qubits are used by Google's Sycamore and by IBM. IBM's hardware and roadmap page lists several processors: Heron r1 with 133 qubits, Heron r2 and r3 with 156 qubits, and Nighthawk with 120 qubits. The Heron family uses tunable couplers. IBM describes Quantum System Two as its flagship system. It calls it the cornerstone of "quantum-centric supercomputing."

Trapped-ion qubits hold ions in place with electromagnetic forces. IonQ's technology overview page explains that the ions are isolated from the environment to minimize noise and decoherence. Because the qubits are not connected by physical wires, every qubit can interact with every other directly.

Quantinuum announced on April 16, 2024 a two-qubit gate fidelity of 99.914(3)% across all qubit pairs. The result came from its commercially available H1-1 system. Both of these are company statements and should be read as such.

Neutral-atom qubits use arrays of uncharged atoms. A Nature paper by Bluvstein and colleagues (Harvard and others) was published online in December 2023. It reported a programmable processor based on logical qubits. The processor operated with up to 280 physical qubits in reconfigurable neutral-atom arrays.

The team reported three results. They improved a two-qubit logic gate by scaling surface code distance from 3 to 7. They created logical GHZ states fault-tolerantly. And they realized sampling circuits with up to 48 logical qubits.

The figures for the three approaches measure different things: qubit counts, gate fidelity, and logical qubit counts. The gap between 280 physical qubits and 48 logical qubits likewise shows that the two kinds of numbers must be read separately.

A black cylindrical quantum computer housed inside a glass cube in a dark room
IBM Quantum System One in Ehningen, Germany (2021) · IBM Research · CC BY 2.0 · Wikimedia Commons

7. Error correction, and the hopes for chemistry simulation

Bluvstein and colleagues open their paper with a clear statement. Suppressing errors, they write, is the central challenge for useful quantum computing, and large-scale processing requires quantum error correction. They add that "logical" qubits come with an overhead. A logical qubit's information is encoded across many physical qubits for redundancy. This overhead poses significant challenges to large-scale logical quantum computing.

This fits the Cerezo review's view that fault-tolerant machines are not a near-term prospect.

Molecular and materials chemistry is often named as a potential "killer application" for quantum computers. However, an analysis by Lee and colleagues (arXiv:2208.02199) gathered the evidence for ground-state energy estimation, the most common task in quantum chemistry. It concluded that evidence for an exponential quantum advantage across chemical space has yet to be found.

The authors say quantum computers may still prove useful for quantum chemistry. But they add that it may be prudent to assume exponential speed-ups are not generically available for this problem.

Error correction is a challenge for the devices, whereas the analysis above concerns the problem itself: ground-state energy estimation. Only by separating the two can the "killer application" hopes be read with a cool head.

8. What governments and institutions are doing

Quantum-resistant cryptography. On August 13, 2024, the U.S. National Institute of Standards and Technology (NIST) released three finalized encryption standards. They are designed to withstand attack by a quantum computer. One is FIPS 203 (ML-KEM). The other two, FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA), are for digital signatures. They are the first completed standards from NIST's post-quantum cryptography standardization project.

United States. The National Quantum Initiative (NQI) was established by the National Quantum Initiative Act in 2018. It was later amended by the National Defense Authorization Act for Fiscal Year 2022 and the CHIPS and Science Act of 2022. Federal agencies carry it out. It includes 14 National Quantum Information Science (QIS) Centers and a quantum industry consortium.

European Union. The Quantum Technologies Flagship was launched in 2018, following the Quantum Manifesto of 2016, with an expected EU budget of €1 billion over 10 years. It funds quantum computing, quantum simulation, quantum communication, and quantum sensing and metrology. Two prominent projects from its initial phase were OpenSuperQ, based on superconducting circuits, and AQTION, a trapped-ion system.

Japan. RIKEN, a Japanese research institute, worked with Fujitsu and other Japanese institutes on a domestically built 64-qubit superconducting quantum computer. At the end of March 2023, they put it online via the cloud for researchers (RIKEN).

From October 5 of that year, RIKEN and Fujitsu began offering a hybrid platform to joint-research companies and institutions. It combines a newly developed 64-qubit quantum computer with Fujitsu's 40-qubit quantum computer simulator (RIKEN).

South Korea. On June 27, 2023, the Ministry of Science and ICT announced the government's first Quantum Science and Technology Strategy. It set goals to invest about 3 trillion won by 2035 and reach 85% of leading countries' technology level. It also aims to train 2,500 core researchers, secure a 10% share of the global quantum industry and foster 1,200 quantum companies (Korea.kr, in Korean).

The Korea Research Institute of Standards and Science (KRISS, in Korean) developed a 20-qubit quantum computing system in 2024. It plans to offer that system to industry through a cloud service. It has also announced plans to develop a 50-qubit system.

Side by side, the national approaches diverge: laws and research centers, long-term budgets, cloud access, and investment targets. That cryptography standards have already been finalized shows that preparation is moving ahead of the machines themselves.

9. Takeaway: a time for verification, not replacement

Checked against the sources, the claim that "quantum computers will transform AI" breaks into three parts. First, quantum advantage experiments are real but limited to specific sampling tasks, and the contest with classical computation continues. Second, quantum machine learning has proposed interesting methods, but practical advantage remains a "possible path," and training problems such as barren plateaus have been identified. Third, the error correction needed for large-scale computation is at the stage of logical-qubit experiments.

So the useful question today is not "when will quantum computers replace GPUs." It is "who has verified that a quantum method beats classical methods on which problem, and under what conditions."

When you read a new announcement, check three things. What was the task? Was the classical method it was compared with state of the art? Do the figures come from a paper or from a company announcement? These checks let you tell hype from real progress.