D-Wave quantum supremacy claim

The D-Wave quantum supremacy dispute

What this is about, in one paragraph

In March 2025, D-Wave Systems (a publicly traded company that sells quantum annealing computers) published a peer-reviewed paper in Science stating that its hardware had simulated the behavior of a class of magnetic materials faster than any classical computer could, for a problem with genuine physical meaning rather than a contrived benchmark. Within days, two independent academic groups posted preprints showing that classical software — run on ordinary computing hardware, in one case starting on a laptop — could reproduce large parts of what D-Wave's quantum processor had done. D-Wave disputed that the classical work covered the same ground. A third academic group published a further, narrower objection in July 2026. The exchange is still open as of this writing (16 August 2026).

Some background before the specifics: a quantum annealer is a special-purpose quantum computer built to settle into low-energy configurations of a physical system, rather than to run general-purpose algorithms. A spin glass is a model magnetic material made of many small magnets ("spins") whose interactions are randomly frustrated, so the system has no simple, orderly resting state — predicting how it evolves over time is a genuinely hard computational problem that also describes real disordered magnetic materials. The dispute below is about whether classical (ordinary, non-quantum) software can predict that evolution as accurately as D-Wave's hardware did, and if not, by how much and in which cases.

The claim

D-Wave's paper is "Beyond-classical computation in quantum simulation," by Andrew D. King and 62 coauthors (D-Wave Quantum Inc., with academic collaborators including Anders W. Sandvik, Roger G. Melko, Juan Carrasquilla and Marcel Franz), published in Science 388, 199–204 (2025). Preprint: arXiv:2403.00910 (submitted 1 March 2024; revised 14 October 2025).

The specific problem: simulating the real-time quantum dynamics of programmable Ising spin glasses — random magnetic interactions laid out on square, cylindrical, dimerized-cubic, diamond, and dimerized-biclique lattice geometries — as the system is driven from a simple starting state toward a disordered target ("quenched" dynamics), on systems up to 576 qubits.

The specific comparison: D-Wave benchmarked its quantum annealer's output against three families of classical simulation method — matrix product states (MPS), projected entangled-pair states (PEPS), and neural quantum states (NQS) — and against an estimate of what the largest of those methods would cost on the Frontier supercomputer at Oak Ridge National Laboratory. The paper's own words:

"We show that several leading approximate methods based on tensor networks and neural networks cannot achieve the same accuracy as the quantum annealer within a reasonable time frame. Thus, quantum annealers can answer questions of practical importance that may remain out of reach for classical computation." — abstract, arXiv:2403.00910

The headline number, on the largest instances:

"MPS would take millions of years on the Frontier supercomputer per input to match QPU quality. Memory requirements would exceed its 700 PB storage, and electricity requirements would exceed annual global consumption." — arXiv:2403.00910

Important nuance: that figure is an extrapolation. D-Wave ran MPS simulations at smaller, tractable sizes, measured how their cost scaled as system size grew, and projected that scaling out to the full problem size — it is not the record of an actual multi-million-year computation. The paper is explicit that the extrapolation, not a completed run, is what licenses the "beyond classical" claim; the entire dispute that follows is about whether that extrapolation, and the general conclusion drawn from it, holds up.

Challenge 1: EPFL — a narrower classical method, a narrower claim

Who: Linda Mauron and Giuseppe Carleo, École Polytechnique Fédérale de Lausanne (EPFL).

Paper: "Challenging the Quantum Advantage Frontier with Large-Scale Classical Simulations of Annealing Dynamics," arXiv:2503.08247 (submitted 11 March 2025 — two days before D-Wave's press conference; still an arXiv preprint, not published in a peer-reviewed journal as of this writing).

What they showed: using a classical algorithm called time-dependent variational Monte Carlo (t-VMC) with a "Jastrow-Feenberg" trial wavefunction — a different classical method from the ones D-Wave tested — they simulated spin-glass annealing dynamics on lattices up to 128 spins (three-dimensional diamond lattice), with correlation errors below 7%, using computational resources that scale polynomially rather than exponentially with system size.

"Our approach achieves accuracy comparable to that of quantum processing units while requiring only polynomially scaling computational resources." — arXiv:2503.08247, abstract

Their framing of what this means:

"Our findings call for a significant shift in the computational 'advantage' frontier identified in Ref. [King et al. 2024]." — arXiv:2503.08247

What they did not claim: their largest simulated system (128 spins) is smaller than D-Wave's largest (576 spins); they tested one lattice geometry (3D diamond), not the full set D-Wave used; and they did not claim that MPS, PEPS, or NQS — the specific methods D-Wave benchmarked against — were mischaracterized. Their claim is that a different classical method, not tried by D-Wave, closes the gap in the regime they tested. It is a claim about where the frontier sits, not a claim that D-Wave's specific comparisons were wrong.

Funding: the paper's acknowledgments state it "was supported by SEFRI [Switzerland's State Secretariat for Education, Research and Innovation] through Grant No. MB22.00051 (NEQS – Neural Quantum Simulation)" — a grant explicitly aimed at developing classical neural/variational simulation methods as alternatives to specialized hardware.

Challenge 2: Flatiron Institute — the one that made it back into Science

Who: Joseph Tindall, Antonio Mello, Matt Fishman, E. Miles Stoudenmire, and Dries Sels (Center for Computational Quantum Physics, Flatiron Institute; Sels is also affiliated with New York University).

Paper: "Dynamics of disordered quantum systems with two- and three-dimensional tensor networks," arXiv:2503.05693 (v1 submitted 7 March 2025, five days before D-Wave's announcement — meaning this work was independently underway before the Science paper's public unveiling). This paper was itself later published in Science 392, 868–872 (21 May 2026), DOI 10.1126/science.adx2728 — meaning a classical rebuttal to a Science quantum-supremacy claim was, fourteen months later, itself peer-reviewed and printed in the same journal.

What they showed: a "belief propagation" (BP) algorithm from 1982, combined with lattice-specific tensor networks, tracking entanglement growth during time evolution. Their own summary:

"We find that state-of-the-art accuracies can be achieved with modest computational resources." — arXiv:2503.05693, abstract

Their result is uneven by design, not uniformly better everywhere — this matters for reading the claim precisely:

"In the case of cylindrical and diamond lattices our simulations reach accuracies which are well beyond those of the quantum annealer... in the case of the dimerized cubic lattice the accuracies we reach are comparable to the annealer." — arXiv:2503.05693

They also state their own method's limitation directly, rather than leaving it implicit:

"This is likely because on the 36 and 54 qubit dimerized cubic lattices a periodic boundary is present... which creates a loop of size 3 in the tensor network. This small loop increases the error in our methods." — arXiv:2503.05693

What they did not claim: their paper covers square/cylindrical, diamond, and dimerized-cubic lattices. It does not address the dimerized-biclique lattice that also appears in D-Wave's original problem set (King et al. tested "dimerized biclique graphs" among their lattice types) — a geometry Tindall et al. simply do not attempt. They also do not claim to reproduce D-Wave's highest-order correlator measurements. Their paper does not contain a sentence declaring D-Wave's overall conclusion false; it demonstrates that one specific classical method, in specific regimes, matches or beats the quantum result, and is candid about where it does not.

Funding: "ongoing support through the Flatiron Institute, a division of the Simons Foundation" (a private non-profit research foundation, not a D-Wave competitor); Sels separately acknowledges AFOSR (U.S. Air Force Office of Scientific Research) Grant FA9550-21-1-0236.

A media detail worth being precise about: press coverage widely described early calculations as having been run "on a laptop." That detail traces to reporting about how Tindall initially prototyped the method using the Flatiron-built ITensor software library, not a claim inside the paper itself that the published, full-scale results were laptop-generated. I was unable to independently verify the laptop detail against a primary source (see Notes below) and flag it here rather than assert it as established.

D-Wave's response

D-Wave engaged the Flatiron paper with a full technical counter-paper, not just a statement:

Paper: Alberto Nocera (University of British Columbia — also a coauthor of D-Wave's original Science paper), Jack Raymond, William Bernoudy, Mohammad H. Amin, and Andrew D. King (all D-Wave Quantum Inc., Amin also at Simon Fraser University), "Evaluating classical simulations with a quantum processor," arXiv:2508.15759 (submitted 21 August 2025; still an arXiv preprint, not journal-published as of this writing).

This paper's method is itself notable: rather than arguing from theory, the authors used D-Wave's own quantum processor as a source of "ground truth" data to test whether Tindall et al.'s predicted error-scaling behavior for their BP-based tensor network method actually holds up in regimes too large for classical methods to cross-check independently.

"Our observations run contrary to previous scaling predictions, demonstrating the need for caution when extrapolating the accuracy of classical simulations of quantum dynamics." — arXiv:2508.15759, abstract

"Our findings contradict a recent prediction about error scaling in a novel classical algorithm based on belief-propagation-gauged tensor-network states (BP-TNS)." — arXiv:2508.15759

Does this engage the specific objection, or reframe it? In substance, it engages: it targets Tindall et al.'s specific error-scaling predictions with new data, rather than rebutting the general "quantum supremacy is disputed" narrative. Its own conclusion is notably cooperative in tone rather than declaring victory:

"Our results demonstrate that the virtuous cycle of competition between classical and quantum simulations can lend insight in both directions." — arXiv:2508.15759

Separately, D-Wave issued public statements — reported by multiple technology and trade outlets — under headlines including "D-Wave's Quantum Supremacy Result Stands," reportedly arguing that the Flatiron work "does not reproduce the full scope" of the original result and does not reach "the hardest problem instances." I was not able to open D-Wave's own press release directly (see Notes below); I can confirm from independent secondary reporting that D-Wave made public statements to this effect, but I cannot verify the exact wording against the primary source, so I am not quoting it as a direct citation. This public-statement register is broader and more categorical than the arXiv technical paper's narrower, data-specific rebuttal — worth noting as a real difference in register between D-Wave's peer-facing science and its public-facing communication, even though I could not verify the exact public language.

One relevant fact I could verify: none of Mauron, Carleo, Tindall, Mello, Fishman, Stoudenmire, or Sels appear anywhere in the 63-author list of D-Wave's original Science paper. The classical challengers are independent of the original claim's authorship. Nocera, by contrast, is a repeat collaborator — a coauthor of both the original claim and D-Wave's rebuttal to Flatiron.

Challenge 3: a narrower, more recent objection — Neural Quantum States

Who: Wladislaw Krinitsin, Nikita Alert, Matteo Rizzi, and Markus Schmitt (Forschungszentrum Jülich, with joint appointments at the Universities of Regensburg and Cologne).

Paper: "Comment on 'Beyond-classical computation in quantum simulation,'" arXiv:2607.08811 (submitted 9 July 2026 — the most recent substantive contribution to this dispute that I could find).

This is the narrowest of the three challenges — it does not touch the tensor-network comparisons at all, only D-Wave's specific claim about neural quantum states (NQS):

"One of the findings indicates that Neural Quantum States... fail to reach the same accuracy as the quantum processor. In this comment we revisit these concerns, demonstrating that NQS can provide competitive results in some of the cases when accounting for the Monte-Carlo noise and large autocorrelation times between samples obtained from the final state." — arXiv:2607.08811, abstract

Their technical point is about how error was measured, not about the physics: D-Wave's NQS comparison, they argue, mixed genuine method error with statistical sampling noise, making NQS look worse than it is.

"The estimated values of the comparison metric do not fully correspond to method-specific limitations, but include significant contributions from statistical noise." — arXiv:2607.08811

What they did not claim: their own conclusion is scoped to "some of the cases" and to "the considered instance of the disorder realization" — singular, not a general claim that NQS matches the quantum annealer across the board. Their acknowledgments note "insightful comments on the results by A. King and R. Wiersema" — i.e., a collegial exchange with D-Wave authors during preparation, not an adversarial one. No funding source is listed in the paper. I found no published D-Wave response to this specific comment as of 16 August 2026.

Stakes

Stated factually; a stake is context, not an accusation.

• D-Wave Quantum Inc. is a publicly traded company (Nasdaq/NYSE: QBTS). Quantum supremacy/advantage claims are central to its investor narrative; independent reporting found via search (not independently opened — see Notes) describes a sharp single-day stock decline around the time Tindall et al.'s paper appeared in Science in May 2026, which I mention as context for the claim's commercial weight but do not cite as a verified figure. D-Wave also announced a $550 million acquisition of Quantum Circuits Inc. in January 2026 to expand into gate-model quantum computing, indicating the company's broader competitive positioning in the field beyond annealing.
• Mauron and Carleo (EPFL) are funded, per their own paper, by a Swiss government grant specifically for developing classical neural/variational quantum simulation methods (SEFRI Grant MB22.00051, "Neural Quantum Simulation") — an institutional and research interest in demonstrating that classical variational methods can match specialized quantum hardware.
• Tindall, Mello, Fishman, Stoudenmire, and Sels (Flatiron Institute / NYU) are supported by the Flatiron Institute (a division of the Simons Foundation, a private non-profit) and, for Sels, a U.S. Air Force Office of Scientific Research grant. Flatiron does not sell competing hardware and has no commercial stake in D-Wave's stock performance; its Center for Computational Quantum Physics has an institutional interest in advancing tensor-network methods generally.
• Krinitsin, Alert, Rizzi, and Schmitt (Forschungszentrum Jülich) work in variational/neural quantum simulation, the same research area as the method (NQS) they are defending — a direct research-reputation stake in showing NQS was undersold, comparable in kind to Carleo's stake in t-VMC. No funding section is disclosed in their paper.
• Nocera is affiliated with the University of British Columbia (not a D-Wave employee) but has coauthored both D-Wave's original claim and its rebuttal to Flatiron — an ongoing research collaboration with D-Wave.
• None of the classical-side authors (Mauron, Carleo, Tindall, Mello, Fishman, Stoudenmire, Sels, Krinitsin, Alert, Rizzi, Schmitt) appear as coauthors on D-Wave's original paper; I checked the full 63-name author list directly.

Where it stands now

Has anything been settled? No single point has been conceded by either side in a way that ends the dispute, but the scope of disagreement has narrowed considerably since March 2025. Nobody in this exchange now disputes that classical tensor-network and variational methods can match D-Wave's quantum annealer on some lattice geometries and system sizes — that much is empirically demonstrated and, in Tindall et al.'s case, itself peer-reviewed in Science. What remains contested is narrower and more specific: whether any classical method reaches the hardest instances D-Wave tested (particularly the dimerized-biclique geometry and highest-order correlators), and whether Tindall et al.'s predicted error-scaling trend, once extrapolated further, would eventually catch up — a question Nocera et al.'s August 2025 paper argues the current evidence does not support. The exchange is live, not concluded: the most recent substantive contribution is the July 2026 Krinitsin comment, and I found no sign either side considers the matter closed.

What would settle it

A pre-registered, adversarial joint benchmark would go furthest: before running anything, D-Wave and a classical team (or several, since different methods have different strengths) would publicly agree on a fixed set of problem instances — including the dimerized-biclique lattice and the highest-order correlators that classical work has not yet addressed — a fixed error tolerance, and a fixed compute/time budget, all specified in advance so neither side can adjust the target after seeing results. D-Wave would need to publish full raw correlator data for those instances in an open, versioned repository (not just summary statistics), so classical methods could be scored without requiring access to D-Wave's proprietary hardware. If a classical method met the pre-registered target within the pre-registered budget, that would resolve the specific empirical question for that instance class. It would not, however, settle the underlying theoretical question — whether classical simulation of these dynamics is fundamentally exponential in system size, or merely difficult with today's best algorithms — because that is a computational-complexity claim, and no one has a mathematical hardness proof for this problem. Empirical benchmarks, however carefully designed, can narrow that question but cannot close it the way a proof would.

What an outsider can and cannot check

Can check, without special access:

• Both classical challenges are built on open-source software: the Flatiron team's tensor-network work uses ITensor (a Flatiron-developed, publicly available Julia library); the EPFL work uses NetKet (an open-source, JAX-based library). Anyone with a standard workstation or modest cluster can, in principle, rerun these classical simulations and check the reported accuracies against the methods' own descriptions.
• The papers' stated error metrics, lattice sizes, and scaling arguments are laid out in the preprints/publication text and can be checked mathematically or reproduced computationally without needing D-Wave's cooperation.
• D-Wave's original "millions of years on Frontier" figure is explicitly an extrapolation from smaller, actually-run MPS simulations — the extrapolation's assumptions are described in the paper and can be scrutinized, though (see below) the underlying quantum comparison data cannot be independently regenerated by most readers.

Cannot check without D-Wave's hardware or cooperation:

• The actual quantum annealer output — the primary experimental data this entire dispute is about — comes from D-Wave's proprietary Advantage2 processor. Cloud access to D-Wave hardware exists commercially, but an independent party cannot freely rerun D-Wave's exact experiment, calibration, and gauge choices at will the way they can rerun open-source classical code.
• D-Wave's own rebuttal to Flatiron (Nocera et al., arXiv:2508.15759) uses the quantum processor's output as the ground truth against which Tindall et al.'s classical predictions are tested — meaning the specific evidence D-Wave offers against the classical challenge is itself only regenerable by someone with access to D-Wave's hardware. This is a real asymmetry: the classical side's work is fully open and independently reproducible; part of D-Wave's counter-evidence is not.
• Any claims in D-Wave's press statements that I could not open directly (see Notes) cannot be checked by me or, easily, by a general reader — press releases are not data.
• The stock-price and financial context mentioned under Stakes comes from search-engine summaries of pages I could not open; I flag it as unverified rather than presenting it as confirmed fact.