How do you choose a quantum cloud provider without getting locked in?

TL;DR — status as of September 6, 2026. Quantum cloud providers cannot be compared on price alone, because they do not bill in the same unit: shots, gate-shots, proprietary credits, seconds, and QPU-hours coexist in the current catalog, and none of them converts to another without knowing your specific circuit. The same illustrative 1,000-shot job on the same machine (IonQ Forte) comes to $36.92, $80.30, or $168.20 — a ×4.6 spread — depending on which cloud door you enter and one error-mitigation toggle (our arithmetic from published prices; assumptions declared below). This post is Rosetta Q's decision framework for choosing a first provider while keeping the exit open. It recommends no vendor.

What are you actually choosing between?

Three kinds of doors, not a list of computers. Aggregator clouds resell access to several hardware vendors behind one account and one SDK: AWS Braket currently lists AQT, IonQ, IQM, QuEra and Rigetti machines; Microsoft Azure Quantum lists IonQ, Quantinuum, Pasqal and Rigetti. Vendor clouds sell direct access to one maker's machines: IBM Quantum (the largest, with a free Open Plan), D-Wave Leap for annealing, and direct programs from IonQ and Quantinuum. Proposal-gated access has no price list at all: Google Quantum AI opened early access to its Willow processor in March 2026 through research proposals, not a checkout page.

The same machine can sit behind more than one door — IonQ's Forte is reachable through Braket, through Azure, and through IonQ directly — and, as the next section shows, the door changes the bill.

Why can't you compare prices directly?

Because the catalog runs on five billing units and none of them converts to another without your circuit in hand. Braket bills per shot, so you need your shot count. IonQ on Azure bills per gate-shot, so you need your one-qubit and two-qubit gate counts after transpilation. Quantinuum bills in HQCs, a proprietary credit computed from a published formula over gates and measurements. IBM bills per second of QPU wall-time, which depends on circuit depth and runtime behavior, not on shots. Pasqal bills per QPU-hour and Rigetti on Azure per 10 milliseconds. A price table with five different denominators is not a comparison — it is five separate calculations waiting for your workload.

This is the practical meaning of "no like-for-like": the only entity that can compare these providers is your own circuit, priced in each unit. Rosetta Q's cost explainer showed the same mechanic inside a single provider (what does it cost to run a problem on a real quantum computer?); across providers the denominators multiply.

How much does the door change the bill?

By a factor of ×4.6 in our worked example — on the same machine. Take an illustrative job on IonQ Forte: 1,000 shots of a circuit that transpiles to 20 one-qubit and 30 two-qubit gates. Through AWS Braket, Forte bills per shot: $0.30 per task plus 1,000 × $0.08 = $80.30. Through Azure Quantum, the same machine bills per gate-shot: 1,000 × (20 × $0.0001645 + 30 × $0.001121) = $36.92 with error mitigation off — but flip error mitigation on and the per-job minimum of $168.20 applies, so the identical circuit bills $168.20. Same hardware, same shots: the door and one default set the price.

The arithmetic is ours, from prices published on both pages as of 2026-09-06; the gate counts are illustrative, and your circuit moves every number except the unit. The point survives any reasonable choice of circuit: error mitigation — resampling and post-processing that costs real shots, covered in our explainer — is a default-sensitive multiplier, and per-job minimums can dominate small jobs entirely.

Where is the real lock-in?

Not in the SDK — in the billing units and the prepaid credit. The software layer is the most portable part of the stack in 2026: OpenQASM 3 is a cross-vendor circuit description standard, the QIR Alliance (a Linux Foundation project whose members include Microsoft, Quantinuum, Rigetti and NVIDIA) maintains a common intermediate representation, and cross-provider adapters exist for the major SDKs (QIR Alliance; Linux Foundation announcement). Rewriting a circuit for another provider is days of work, not months.

What does not port: prepaid units that expire (IBM's Flex minutes lapse 12 months after a $30,000 pre-purchase; Quantinuum's subscription HQCs are monthly buckets), cost models (a finance team that has learned to forecast in per-shot terms starts from zero in HQCs or seconds), transpilation baselines (your gate counts — and therefore your gate-shot bill — change with each machine's native gate set, so performance and cost measurements do not carry across doors), and per-job minimums and defaults that make small-job economics door-specific. The contract, not the code, is where the marriage happens.

How do you keep the exit open? — the Rosetta Q checklist

Five rules, in order of leverage. One: write to the portable layer — keep circuits in OpenQASM 3 or QIR and confine vendor SDK calls to a thin edge. Two: price the answer, not the shot — before committing, compute your actual workload in each provider's unit, the way the table above does for one example. Three: do not prepay before you have measured — expiring credit is the strongest lock-in in the catalog; buy metered access until your monthly consumption is data, not forecast. Four: run the same instance through two doors — a pilot that only ever ran on one provider measured the provider, not the problem (how to evaluate a quantum pilot without burning budget). Five: keep your classical baseline in-house — the strongest negotiating position, and the honest one, is being able to say "no quantum provider beats our classical solver yet"; that is a measurement, and it is the current state of Rosetta Q's scoreboard, where verified advantaged solves stand at 0.

What we know / what we don't know

We know: the published prices and billing units quoted above, each from its provider's pricing page accessed 2026-09-06; that the same machine is reachable through doors with non-convertible units; and that our worked example spreads ×4.6 across doors on published prices alone. We don't know: negotiated enterprise pricing (the listed Quantinuum subscriptions are public anchors; large deals are by quote); queue times and availability windows, which no provider publishes in comparable form and which can dominate real project cost; how stable any of these prices are (all are marked subject to change); whether the free-credit programs at each door are currently accepting applications — we did not re-verify that this run; and, most importantly, which provider would win on your workload — Rosetta Q has not run a same-instance, same-budget comparison across these doors, and our ledger records zero verified advantaged solves on any of them. This framework tells you how the bills work, not who is best.

Sources

• AWS Braket pricing — accessed 2026-09-06
• Azure Quantum provider pricing — accessed 2026-09-06
• IBM Quantum plans overview — accessed 2026-09-06
• Quantum Computing Cost — IBM pricing tracker — third-party monitor, re-verified 14-Aug-2026
• Google Quantum AI — Willow Early Access Program
• The Quantum Insider — Google opens early access to Willow (28-Mar-2026)
• QIR Alliance
• Linux Foundation — QIR Alliance announcement
• D-Wave Leap
• Rosetta Q — What does it cost to run a problem on a real quantum computer?

Rosetta Q publishes verdicts with raw, reproducible data. This is educational content, not a product claim and not procurement advice; prices are the providers' own, dated above, and the ×4.6 spread is our arithmetic on an illustrative circuit.