Key Takeaways: Quantum Cloud Platforms
The chapter in one line
A 4,096-shot Bell job occupies the device for 6.92 milliseconds, and the bill for identical work differs by 3,718× depending only on how the provider counts.
What the chapter measured
| Claim | Measured |
|---|---|
| Bell, 4,096 shots | 6.93 ms of device time (1.69 µs/shot) |
| QFT-8, 4,096 shots | 43.20 ms (10.55 µs/shot) |
| Local simulation, same jobs | 22–74 ms — same order of magnitude |
| Utilization at a 5-min queue | 2.31 × 10⁻⁵ — 43,340× wall clock |
| 100 circuits batched vs separate | ~99× faster |
| 120 VQE iterations, session vs jobs | 10 hours waiting → 5 minutes |
| Book's experiments | 614,400 to 18,456,984 shots |
| 18.5M shots of device time | 31.2 seconds |
| Same run, per-minute | $50 |
| Same run, per-shot superconducting | $7,432 (149×) |
| Same run, per-shot trapped ion | $185,542 (3,718×) |
| 2-qubit error spread, one device | 1.79e-03 to 1.00 (dead links) |
| T1 spread | 15.2 – 483.0 µs |
| 14-qubit layout, 24 seeds | fidelity 0.5755 – 0.7911, 2.03× error |
| Same test, 4 qubits | exactly zero variation |
The six things to carry forward
1. The device is not the bottleneck — the queue is the product. At a five-minute queue you wait 43,000× longer than you compute, and the processor spends 99.998% of your job's lifetime serving somebody else. Local simulation of the same circuits took 22–74 ms, the same order of magnitude as hardware execution.
2. Therefore batching beats transpiling, and batching is not a quantum technique. One hundred circuits in one job instead of one hundred jobs is worth ~99× — more than any transpiler setting. The first question about a slow quantum workflow is "how many jobs?", not "how deep is the circuit?"
3. This reframes Part V without contradicting it. Chapters 28, 29, and 31 optimized fidelity, and all of them improve the 7 milliseconds while touching none of the five minutes. Different problems, different levers. A team that conflates them optimizes the wrong one.
4. ★★★ You are not paying for device time. Chapter 36's VQE run is 18.5 million shots — which is 31.2 seconds of hardware. That is $50 under a per-minute model, $7,432 per-shot, or $185,542 on trapped ions. Per device-second: $2, $238, $5,948. The bill tracks access, and the unit of access differs by three orders of magnitude for identical work.
The trapped-ion premium buys something real — all-to-all connectivity, longer coherence, better two-qubit fidelity. But it means a shot count chosen on superconducting hardware is a budgeting decision, not a physics decision, when moved.
5. A per-shot and a per-minute price are not comparable without the circuit duration, because the duration is the conversion factor. A per-shot model charges the same for a 1.69 µs Bell circuit and a 10.55 µs QFT; a per-minute model charges 6× more for the second. Neither is wrong — they price different things.
6. ★★ A hardware result without its execution metadata is not reproducible, and a result that is not reproducible is not verifiable. Four of the fields that determine your result — job ID, execution timestamp, calibration snapshot, physical qubit assignment — come from the provider and are not in your code. If you do not capture them, they are gone. The same code, on the same backend, on a different day, is a different experiment.
The toy-circuit trap
The chapter's sharpest methodological finding, and it is the book's sixth instance of the same error.
Transpiling a 14-qubit circuit with 24 seeds gave fidelities from 0.5755 to 0.7911 — 2.03× the error — driven by a 49-to-112 spread in two-qubit gate count. That is routing overhead: the transpiler inserts swaps to satisfy connectivity, and how many depends on where it started.
The identical experiment on a 4-qubit circuit gave exactly zero variation across all 24 seeds.
Both measurements are correct. The 4-qubit result is a true measurement of a false general claim — a 4-qubit chain fits the coupling map without routing, so of course the seed does not matter. A team that tests seed sensitivity on a toy circuit concludes the transpiler is deterministic, and is wrong for every circuit they actually care about.
| Chapter | Sample that was too small | What it hid |
|---|---|---|
| 27 | 200 runs | a false-failure rate off by 6.7× |
| 28 | 2 circuits | levels 2 and 3 differ in 14 of 40 |
| 33 | 1 train/test split | an accuracy that was a draw |
| 34 | 1 prediction | the effect ran the other way |
| 37 | 1 seed per depth | a regression that did not exist |
| 38 | 32 test bits | a detection claim needing 37 |
| 39 | a 4-qubit test circuit | 2.03× layout variance at real widths |
Note what is different here: the earlier six were too few samples. This one had 24 samples and the wrong system. Sample size is not the only dimension a measurement can be too small in.
What went into vqelab
vqelab.platform, 20 tests passing:
job_cost()raisesAmbiguousPricingErrorwithout a circuit duration — the conversion factor between pricing models.PriceBook.as_ofis required. An undated price rots without telling you.utilization()andbatching_speedup()— the ratios that decide workflow design, neither of them a quantum quantity.layout_variance_is_measurable()returns False below 8 qubits, so the toy-circuit conclusion cannot be drawn accidentally.ExecutionRecord.is_reproducibleis False until all four provider-side fields are captured — not "the code is committed."
Next: Chapter 40 closes the book. What this field actually employs people to do, what forty chapters of measurement add up to, and how to tell a real opportunity from a well-funded one.