Photonic's Peer-Reviewed QLDPC Paper Reports 2x and 3.5x Qubit Savings, Not the Preprint's 10x — and Its Aug. 25 Release Publishes No Number at All

by Margaret Ashworth · August 26, 2026

Senior Analyst, Frontier Tech Wire

Photonic Inc., the British Columbia-based company building optically linked silicon spin qubits, announced on Tuesday that its work on a family of quantum error-correction codes had been published in Nature Communications. The release went out at 11:00 ET on Aug. 25 via GlobeNewswire and framed the result as a step toward useful quantum computing with far less overhead than the surface code, the error-correction scheme most of the industry has organized around.

The claim in the release is qualitative. It says the codes require meaningfully fewer physical qubits than the surface code at the sizes tested. It does not say how many fewer. It does not give a code distance, a code rate, a logical error rate or a threshold. For a company announcing a peer-reviewed error-correction result, that is a conspicuous set of omissions — particularly because the paper it is announcing contains all of them.

The paper is titled "Computing efficiently in QLDPC codes." It carries thirteen authors led by Alexander J. Malcolm, with Stephanie Simmons — Photonic's chief quantum officer — as the final author. Most are at Photonic; Joschka Roffe is at the University of Edinburgh and Armanda O. Quintavalle at Freie Universität Berlin. The journal lists it as Nature Communications volume 17, article number 7286 (2026), DOI 10.1038/s41467-026-73061-9.

Here are the numbers the release left out, taken from the published paper itself. For the larger instance they simulated, the authors report that the [225, 16, 8] SHYPS code is comparable in performance to the [[784, 16, 7]] scaled distance-7 surface code while reducing qubit count by a factor of 3.5. They also state that the [225, 16, 8] code has a lower pseudo-threshold, approximately 0.35 percent, than the [[784, 16, 7]] surface code's roughly 0.8 percent, while its logical error rate decreases more aggressively in the sub-threshold regime. Of the smallest code in the family they are blunter still: "The smallest member of the SHYPS family, the [49, 9, 4] code, confers no pseudo-threshold advantage, but significantly outperforms the [[81, 9, 3]] scaled distance-3 surface code in both physical qubit overhead (2x reduction) and logical error rates below pseudo-threshold." The notation is the paper's own: single brackets for the SHYPS subsystem codes, double brackets for the stabilizer surface codes they are measured against.

Those are checkable statements, and they are considerably more useful than the release's phrasing. Each one names the code, gives its parameters, names the specific surface-code baseline it is measured against — the paper says it compares SHYPS codes of even distance to k copies of rotated surface codes — and states both the qubit ratio and the error-rate regime in which the comparison holds. Read against them, "meaningfully fewer physical qubits" is a weaker claim than the company is entitled to make. It is also, as it happens, the only formulation that covers both of the paper's two headline instances, which do not save the same amount.

The codes belong to the QLDPC family, short for quantum low-density parity check. The general SHYPS construction in the paper scales as [n(r), k(r), d(r)] equal to [(2^r - 1)^2, r^2, 2^(r-1)], which is how a 49-physical-qubit block ends up encoding nine logical qubits at distance four.

The genuinely novel part is not storage but logic. QLDPC codes have been known for years to store quantum information more efficiently than the surface code; the open problem has been computing inside them without giving that advantage back. The paper's central result is that arbitrary logical Clifford operations can be implemented transversally, with any m-qubit Clifford operation executed in at most order-m rounds of syndrome extraction. The authors quantify the payoff against lattice surgery, the standard surface-code approach: "For a moderately sized code with distance d = 20, a SHYPS-compiled CNOT gate would achieve an order-of-magnitude reduction in depth relative to the equivalent compiled using lattice surgery (4 vs. roughly 40)."

Now the caveat that neither the release nor most coverage of it states plainly: none of this happened on hardware. Every result in the paper comes from circuit-level noise simulation. The authors ran two kinds — quantum memory benchmarking of the [49, 9, 4] and [225, 16, 8] codes against scaled surface codes under a circuit-level noise model, and logical operation simulations on randomly sampled Clifford circuits across 18 logical qubits, held in two blocks of the [49, 9, 4] code, at depth 126. The abstract says as much in its own words: the authors run circuit-level simulations of depth-126 logical circuits. There is no experiment on a physical device reported anywhere in the paper.

That matters because the release uses the word "demonstrated." Simmons is quoted saying the paper introduced the first demonstrated QLDPC code family capable of performing logic efficiently — "not just storing information, but computing with it, using a fraction of the qubits error correction has always demanded," as carried in the Aug. 25 release and in The Quantum Insider's account of it. In error correction, "demonstrated" is a word that readers reasonably hear as "run on qubits." Here it means demonstrated in simulation. The distinction is the whole distinction between a code design result and a hardware result, and a release announcing a peer-reviewed paper is the right place to make it.

There is also a timing question worth putting on the record, because this desk has flagged the same pattern before. The paper is not new. It went up on arXiv as preprint 2502.07150, whose page carries a date of Feb. 12, 2025 — more than eighteen months ago. The Nature Communications article page itself carries a publication date of May 20, 2026, roughly three months before Tuesday's announcement. The news on Aug. 25 is therefore the announcement of a peer-reviewed publication that had already been peer-reviewed and published, of results that had been publicly readable since early 2025.

The gap between those two versions is not cosmetic, and it is the reason the numbers in circulation should be handled carefully. The arXiv preprint claims in its abstract that the codes are "much less resource intensive, requiring up to 10x fewer physical qubits per logical qubit than practical surface code implementations," and reports that "[t]he pseudo-threshold of the [49, 9, 4] SHYPS code matches that of the [[225, 9, 4]] scaled distance-4 surface code, approximately at p = 0.08%", using one-fifth of the physical qubits. None of that survives into the published paper. The peer-reviewed version reports a 2x reduction for the [49, 9, 4] code against a different and smaller baseline, the [[81, 9, 3]] surface code, and says that code confers no pseudo-threshold advantage at all; the 10x figure, the 0.08 percent pseudo-threshold and the [[225, 9, 4]] comparison do not appear in it. Anyone still citing the 10x number — and it is the one that travelled — is citing a superseded preprint, not the peer-reviewed result Photonic announced on Tuesday.

None of that makes the underlying work less interesting. Efficient transversal Clifford logic inside a high-rate QLDPC code is a real contribution, it survived review at a serious journal, and it has co-authors outside the company. The distance-20 CNOT depth comparison is the kind of concrete number that lets other groups argue with it.

The codes also carry a structural consideration that shapes who can use them. In its discussion the paper reduces the remaining reasons to prefer the surface code over QLDPC codes to two factors, connectivity and simplicity — an acknowledgment that the surface code's appeal is its tolerance of purely local, planar wiring, which is what most superconducting layouts natively provide and what high-rate QLDPC constructions generally are not built for. Photonic claims long-range connectivity as a property of its optically linked spin-qubit architecture. The paper does not put a required connectivity degree on the SHYPS family, so how much of this advantage is portable to other hardware is a question the paper leaves open rather than one it answers.

The practical test for readers is the one this desk applies to every technical milestone claim: what number did the company actually publish, against what baseline, and was it measured or modeled. In this case the numbers exist and are real ones — a 3.5x qubit reduction for the [225, 16, 8] code against a matched [[784, 16, 7]] surface code, a 2x reduction for the [49, 9, 4] code with no pseudo-threshold advantage, and a four-cycle CNOT at distance 20 against roughly forty for lattice surgery. They are in the paper, and all of them are modeled rather than measured. They are not in the press release, and the modeling caveat is not there either.