We keep a quantum research lane open, we evaluate it under protocols fixed before the run, and we publish the result either way. Most of what we have published about quantum is negative. That is the point.
Quantum computing is real, and the marketing around it is mostly not. A firm that sells you a quantum claim today is selling you a hope. We would rather sell you the engine that works and tell you the truth about the rest — including when the truth is that our own quantum results lost.
The most useful thing we can publish is not an opinion — it is subtraction. Below are qubit counts and two-qubit gate fidelities as published by the vendors themselves. The practical-size column is arithmetic on those figures under the standard one-qubit-per-name encoding: finite gate fidelity bounds how much circuit you can run before the signal is gone.
| System | Kind | Qubits | 2-qubit fidelity | Practical universe |
|---|---|---|---|---|
| IonQ Forte | gate-model | 32 | 99.5% | ≤ 20 names |
| Google Willow | gate-model | 105 | 99.7% | ≤ 25 names |
| IBM Eagle r3 | gate-model | 127 | 99.5% | ≤ 30 names |
| IBM Heron r2 | gate-model | 133 | 99.9% | ≤ 35 names |
| D-Wave Advantage2 | annealer | n/a | n/a | n/a |
Qubit counts and fidelities are the vendors' own published figures, recorded April 5, 2026, and will drift as hardware improves. The practical-universe column is derived arithmetic on those recorded figures, not a measurement of anyone's machine. D-Wave Advantage2 is an annealer: it uses a direct one-qubit-per-name embedding with no two-qubit gate depth, so the gate-model arithmetic does not apply to it and we do not print a specification we cannot source.
The subtraction. A selection problem over a universe of 64 names needs 64 qubits under the standard encoding. The strongest gate-model system in the table above is practical to roughly 35. Every gate-model system listed is therefore short by 29 to 44 names — before anyone argues about algorithms. A real portfolio is not 64 names; it is thousands. That gap is the whole story, and it is not close.
Our quantum programme has run a series of controlled evaluations against classical controls, with the success criterion written down before the run. Every arm shares the same problem, the same decoding, and the same scoring, so a win cannot come from the plumbing. Here is where that landed.
A single quantum arm produced a genuine, audited quality edge over a strong classical control at small universe sizes — measured, paired, and survivable under re-audit.
Carried one layer further — to the place it would have to earn its keep for a customer — that same edge ties or loses to the classical control, and costs more per solve to obtain.
Independent lines of attack were each taken to a pre-registered decision point. They tie classical methods at matched cost, add nothing measurable, or do not fit a production latency budget at all.
We report these the same way we would report a win. A negative result that was pre-registered is worth more than a positive one that was not: it is the only kind of evidence that cannot have been fitted after the fact.
Everything we sell today runs on classical hardware and is measured as such. The quantum programme exists so that the day a device is genuinely ready for a problem shaped like yours, we will know it on evidence rather than on a vendor's slide — and so will you, because we will have published the losses on the way there.