Pillar A · State as of 2026-10-01

Can quantum computing accelerate generative AI (LLMs)?

As of October 1, 2026, quantum computing does not accelerate generative AI. The number of demonstrated quantum speedups on any LLM workload — training or inference — in the public record is zero. The scale gap is the story. Meta's Llama 3.1 flagship was pre-trained with 3.8×10²⁵ FLOPs on 15.6T tokens using up to 16,000 H100 GPUs (arXiv:2407.21783), while the largest LLM-adjacent run on real quantum hardware used 128 qubits to compute one cosine similarity — and ran slower than the classical baseline (arXiv:2512.02619). Rosetta Quantum keeps this verdict at 0 until a measured, reproducible run at matched budget says otherwise.

→ Leer en español

State as of: 2026-10-01

Status as of: October 1, 2026.

What is being claimed?

Three families of claims are in circulation, and none of them, as published, includes an end-to-end speedup on a real generative-AI workload. The first is the headline version: quantum computers will speed up the training or inference of large language models. It appears in vendor keynotes and investor notes, almost always in future tense and almost never with a benchmark attached. The second is subtler: quantum-generated data will improve AI models. That is the framing of Quantinuum's Gen QAI, announced on February 4, 2025 with the promise of "significantly enhancing the fidelity of AI models, allowing them to tackle challenges previously deemed unsolvable" (Quantinuum press release, 4-Feb-2025) — a launch, with no public benchmark attached to date. The third is the only one with measurements behind it: quantum generative models can learn and sample distributions that classical models cannot. That is the claim of Google Quantum AI's "generative quantum advantage" paper (Huang et al., arXiv:2509.09033, 12-Sep-2025) — measured, but on synthetic distributions, and carrying the authors' own caveats, which we quote below.

What has actually been measured?

Parity or slower, at scales nine or more orders of magnitude below production AI. The public record contains three recent hardware data points, each with its own fine print. First: the largest LLM-adjacent run on a real QPU that we could find used 128 qubits on IBM's ibm_fez to compute a cosine similarity between sentence-transformer embeddings (Laine, arXiv:2512.02619, Dec-2025). The quantum result was 0.8635 against a classical 0.8682, and the author states the fine print himself: the implementation "is not currently practical, as classical methods are faster." Second: IonQ ran a quantum generative model (QCBM) for change detection on real satellite SAR imagery from Capella Space, and the published comparison reports that the quantum and classical methods "achieved similar peak filtered F1 scores" (Quantum Zeitgeist, 24-Sep-2026) — parity, with a claimed edge only on strongly skewed pixel distributions. Third: Google's generative-advantage experiments ran on a 68-qubit superconducting processor mapped to 816 effective shallow qubits, and the authors write that 816 shallow qubits "are not yet expected to be in the beyond-classical regime" and that the "precise remaining steps towards useful generative quantum advantage remain an open question" (arXiv:2509.09033). The caveats sat in the papers; the headlines dropped them — the pattern our guide to reading quantum arXiv papers documents case by case.

Why don't the numbers line up?

Because generative AI is the exact workload shape that quantum hardware is worst at: enormous data in, enormous data out. The honest quantum regime is the opposite — big compute on small data, as our catalog of problems quantum will not accelerate lays out. The table below puts the two sides of the 2026 record next to each other, source by source.

Quantity	Frontier LLM (measured)	Quantum record (measured)	Source
Model parameters	405B (Llama 3.1, dense)	QML demos run at toy scale; no LLM has been trained on a QPU	arXiv:2407.21783 · arXiv:2504.03192
Training data	15.6T tokens	one embedding pair per run in the 128-qubit experiment	arXiv:2407.21783 · arXiv:2512.02619
Training compute	3.8×10²⁵ FLOPs	no equivalent exists on quantum hardware	arXiv:2407.21783
Hardware per run	up to 16,384 H100 GPUs at 700 W TDP each	68 and 128 physical qubits in the runs cited above	arXiv:2407.21783 · arXiv:2509.09033
I/O bandwidth	classical interconnects	roughly four orders of magnitude below classical	Hoefler et al., CACM 66(5), 2023

Two structural walls stand behind that table, and both have their own pages in this series. The data-loading wall: feeding classical data into a quantum state at scale requires QRAM, a 2008 proposal that has never been built at scale — our QRAM explainer covers why that asterisk sits under most QML speedups. The I/O wall: Hoefler, Häner and Troyer computed that quantum input/output bandwidth runs roughly four orders of magnitude below classical interconnects (CACM 66(5), 2023); a 15.6T-token corpus is the least quantum-shaped dataset imaginable. Energy does not rescue the claim either. Meta's paper states each H100 ran at 700 W TDP; at the stated 16,384-GPU peak that is ≈11.5 MW in GPUs alone — our arithmetic, declared, on the paper's own figures — against the ≈26 kW of wall power measured for a current QPU in our energy ledger. But watts are not verdicts: a machine that cannot run the workload at all is not an efficient way to run it.

What happened to the quantum-ML speedups on paper?

Most of them fell to dequantization, and the strongest survivor carries fine print its own authors wrote. Tang's 2018 result (arXiv:1807.04271) removed the exponential speedup of quantum recommendation systems, and the HHL fine print did the rest for the linear-algebra family — our dequantization explainer and the HHL fine-print post track that history claim by claim. The strongest surviving theoretical claim for large-model training is Liu et al. in Nature Communications (10-Jan-2024): a quantum algorithm scaling as O(T²·polylog(n)) in model size — but only while the network stays sparse, and the authors state plainly that "there is no guarantee that our hybrid quantum-classical algorithm will necessarily outperform all other conceivable classical algorithms" and that in real training "the sparsity ratio will generically decay" (Nature Communications, s41467-023-43957-x). It is theory plus classical simulation; no quantum hardware was involved. Quantum transformers exist as a research program — a survey counts 28 papers since 2022, at small scale, naming barren plateaus and data loading as the open bottlenecks, with no systematic comparison against production classical transformers (arXiv:2504.03192). The broader scoreboard for quantum machine learning has not moved since our QML verdict.

What is actually shipping in the LLM economy, then?

The only "quantum" technology in production in the LLM economy of 2026 is quantum-inspired — meaning classical software. Multiverse Computing raised $215M (Series B, 12-Jun-2025, led by Bullhound Capital) for CompactifAI, an LLM-compression product built on tensor networks, described by the company as "a quantum-inspired approach to simplifying neural networks": it compresses Llama, DeepSeek and Mistral models and runs on CPUs, GPUs and edge devices — no quantum computer involved (insideHPC, 12-Jun-2025). Tensor networks being the classical workhorse is an old theme of this ledger: they are the classical rival that keeps winning. Meanwhile, the direction that is demonstrably real runs the other way: AI for quantum computing. A Nature Communications review (Alexeev et al., 2-Dec-2025) catalogs machine learning across the quantum stack — decoding, calibration, circuit design — and argues that "many of QC's biggest scaling challenges may ultimately rest on developments in AI" (s41467-025-65836-3). The traffic between the two fields is real; today it flows almost entirely from AI into quantum, not the reverse.

The offers currently on the table, with their status:

Offer	What it actually is	Status (2026-10-01)	Source
Gen QAI (Quantinuum)	quantum-generated training data for AI models	launched Feb-2025; no public benchmark to date	PR 4-Feb-2025
Generative quantum advantage (Google)	sampling synthetic distributions on 68 qubits	measured; authors: not yet beyond-classical at demonstrated scale	arXiv:2509.09033
Sparse-training speedup (Liu et al.)	theory, O(T²·polylog(n)), sparsity required	paper + classical simulation only	Nat. Commun. 10-Jan-2024
CompactifAI (Multiverse)	quantum-INSPIRED compression — classical tensor networks	shipping; vendor claims 4×–12× faster inference	insideHPC 12-Jun-2025
QCBM on SAR imagery (IonQ)	quantum generative model for change detection	F1 similar to classical	Quantum Zeitgeist 24-Sep-2026
Quantum transformers	research program, 28 papers since 2022	small-scale; no LLM-scale comparison	arXiv:2504.03192

What would have to happen for this verdict to change?

The verdict will change the day someone demonstrates an advantaged solve on a generative-AI workload: same task, same budget, quantum side faster or cheaper, artifacts published for re-running. Getting there will require machines and infrastructure that do not exist today — fault-tolerant hardware orders of magnitude beyond current logical-qubit counts, and a data-loading path (QRAM at scale) that has never been built — and any candidate speedup will then have to survive the dequantization test that killed the last generation of QML claims. Google's own authors frame the remaining steps toward useful generative advantage as an open question. Rosetta Quantum will re-run this question quarterly and will update this page under the same URL with a new date; if a speedup is demonstrated and verified, the 0 on this page will change the same day.

What we don't know

We don't know what sits in private labs: this verdict covers the public record only, and a frontier-lab result could exist unpublished. We don't know whether quantum-generated training data (the Gen QAI lane) will improve a real model's fidelity — no public benchmark exists yet in either direction. We don't know where the beyond-classical threshold for Google's generative models actually sits: the authors project experiments toward 34,304 shallow qubits, and state the threshold for useful tasks is open. And we don't know whether IonQ's narrow edge on strongly skewed pixel distributions would survive a classical baseline tuned for that specific regime — the comparison published so far reports parity on the headline metric.