How do you evaluate a quantum pilot without burning the budget?
Enterprise quantum programs are not small. In McKinsey's Quantum Technology Monitor 2026 (28 Apr 2026), 300+ global companies are adopting quantum computing; among the 162 McKinsey analyzed in detail, one third allocated more than $10M per year to quantum efforts and 7% more than $50M — while the entire quantum computing industry generated just over $1B in revenue worldwide in 2025, with most applications still "experimental or hybrid." That asymmetry is the pilot problem in one sentence: the spend is production-scale, the evidence is not.
This post is a cost map, not a purchase recommendation. What a pilot buys is evidence — and evidence has a price ladder spanning roughly seven orders of magnitude. Budgets burn when teams buy expensive evidence (hardware runs, integration work) before cheap evidence (a classical baseline, a simulator run) has had the chance to kill the idea for free.
What does each unit of evidence cost?
Prices below are public list prices, dated. The cheap end of the ladder is dominated by classical software that costs nothing and quantum simulators that cost almost nothing; the expensive end is hardware time and program headcount.
| Evidence unit | What it tells you | List cost | Source (date) |
|---|---|---|---|
| Classical baseline (OR-Tools / CP-SAT) | the number to beat | $0 — Apache-2.0 | Google OR-Tools (Aug 2026) |
| Laptop statevector simulation, ≤~30 qubits | candidate's best case, no noise | $0 — open SDKs; RAM is the limit | arithmetic: 2^N × 16 bytes |
| Managed simulator (Braket SV1) | same, past laptop RAM | $0.075/min after 1 free h/month | AWS Braket pricing (Feb–Jul 2026) |
| One QPU task, 1,000 shots (IonQ Forte) | one noisy reality check | ≈$80.30 | AWS Braket pricing (Jul 2026) |
| Error-mitigated task (IonQ, ≥2,500-shot floor) | lower-variance estimate | ≥$200/task | Braket calculator (3 Aug 2026) |
| Variational loop, 100 iters × 1,000 shots | one full QAOA/VQE optimization | $72.50 (Rigetti) – $8,030 (IonQ Forte) | Rosetta cost analysis (Jul 2026) |
| Enterprise quantum program | capabilities, talent, integration | >$10M/yr for 1 in 3 analyzed adopters | McKinsey QTM 2026 (28 Apr 2026) |
What order should the evidence be bought in?
Cheapest-kill-first. Each gate exists to kill the candidate before the next, more expensive gate. A candidate that survives all gates earns hardware spend; a candidate that dies at gate 0 or 1 cost you approximately nothing.
Gate 0 — measure the classical incumbent first. Not a strawman: the best available solver, tuned, on your real instances (OR-Tools and CP-SAT are Apache-2.0; the license price of the number-to-beat is $0). A published benchmark of portfolio optimizers (arXiv:2509.17876, 2025) found QAOA and annealing variants failing to reach Gurobi and in several instances statistically indistinguishable from random sampling — against exactly the kind of baseline a weak pilot would have skipped. If nobody on the team can state the classical number to beat, the pilot has no denominator and no result it produces will mean anything.
Gate 1 — run the quantum candidate on a noise-free simulator. This is the step most pilots invert. For the same circuit, a noise-free statevector simulation is the candidate's best case: real NISQ hardware adds decoherence and gate error on top, it does not subtract. If the candidate cannot approach your Gate-0 number in its best case, hardware cannot save it — and that verdict costs $0 on a laptop up to roughly 30 qubits (a statevector is 2^N × 16 bytes: ~17 GB at N=30) or $0.075/minute on a managed simulator after the free hour, list price Feb–Jul 2026.
Gate 2 — small hardware runs, same instance, same budget. Only if Gate 1 shows signal. The unit prices are known and public: ≈$80.30 for a 1,000-shot IonQ Forte task; a ≥2,500-shot floor per task if you switch on IonQ's error mitigation, i.e. a ≥$200 floor per mitigated data point (Braket calculator, 3 Aug 2026). The purpose is error bars against Gate 0 on identical instances — not a demo.
Gate 3 — scaled pilot. Only past Gates 1–2. This is where the $10³–$10⁴-per-experiment variational loops and the integration headcount live. By construction, most budget dies here — which is why most candidates should die earlier.
Two rules make the gates honest. First, write the kill criterion before running the gate ("QAOA must land within 5% of CP-SAT on n=20 instances"), because criteria written after the run always pass. Second, compare on the same instance and the same time budget — a benchmark against a weak or mismatched classical baseline moves the crossover point by construction, not by physics (why a weak baseline ruins a benchmark).
What does a kill look like in practice? (measured)
Rosetta's verdict V-0012 is one worked example of the gates doing their job. Twenty sealed portfolio-optimization runs at n=12/16/20 (QAOA p=2 vs Google's CP-SAT): CP-SAT reached proven optimum 20 of 20 times; QAOA finished 25–48% away from optimum with no improving trend by size. The quantum side ran on noise-free simulators — its best case — and still lost every run. Total QPU spend to reach that verdict: $0. This is one problem class at small n, not a universal verdict — but it is what Gate 1 doing its job looks like: a defensible "not yet" bought for the price of CPU time, before a single hardware dollar.
The same logic prices the counterfactual. Had those 20 runs been purchased as error-mitigated hardware experiments first, the bill lands in four to five figures — the cost-per-answer arithmetic — to learn what the simulator would have said for free.
What we don't know
- The gates close past ~40 qubits. A statevector doubles per qubit; beyond roughly 40 qubits (≈17.6 TB) brute-force noise-free simulation stops being purchasable at any of the prices above. In that regime the cheap upper bound disappears, and the honest evaluation methodology is genuinely harder — tensor-network baselines help but their coverage is disputed case by case.
- No published survey measures how enterprises actually sequence pilot evidence. McKinsey reports what companies spend, not what order they buy evidence in. We have found no dated, public data on what fraction of quantum pilots run a tuned classical baseline before hardware. The absence of that number is itself a finding.
- Annealers and hybrid solvers blur the gate structure. Annealing-based platforms are not gate-circuit simulable the same way, and commercial hybrid solvers interleave classical layers that make attribution of any win ambiguous — as measured in the routing evidence we reviewed (routing and logistics).
- Prices move. Every dollar figure above is a dated list price; hardware pricing has changed materially over 12 months and will again.
Sources
- McKinsey Quantum Technology Monitor 2026 (28 Apr 2026)
- AWS Braket pricing (verified Aug 2026)
- Google OR-Tools (Apache-2.0)
- Portfolio optimization benchmark (arXiv:2509.17876, 2025)
- Braket cost calculator — IonQ error-mitigation shot floor (accessed 3 Aug 2026)
- Rosetta verdict V-0012 — QAOA vs CP-SAT, 20 sealed runs
Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.