ZKSF logo, a neon quantum brainZKSF
← All articles

What It Costs to Rent a Quantum Computer

Last updated · 9 min read · ZKSF team

You can rent time on a quantum computer this afternoon, from a laptop, with no contract and no sales call. The model has a name, Quantum Computing as a Service, abbreviated QCaaS and also called QaaS, Quantum as a Service, and what it covers is set out here. What follows is what it costs. Every figure is a provider list price, taken from the same price table jobs are charged against, and the per shot cost calculator will price a run on any of these machines against your own budget.

The short version

  • You do not rent a quantum computer by the hour. You rent a task, and inside it a number of shots
  • Most of these will run a single shot. We sent one to IQM Garnet and one to IQM Emerald, about thirty cents each
  • Price follows the clock, not the qubit count. A trapped-ion shot is orders of magnitude slower than a superconducting one
  • A simulator is exact below about 30 qubits for a hundredth of a cent, with no queue

Everything here is runnable on your own circuit. Try it in the console

One correction to the premise first, because everything downstream depends on it. With a single exception below, you do not rent a quantum computer by the hour. You rent a task, and inside that task a number of shots.

A shot is one execution of your circuit followed by one measurement. Your task queues, runs, returns its counts, and then the machine belongs to somebody else. Nobody hands you the machine, and at retail nobody sells you an hour of one.

If the question you actually have is what one of these machines costs to buy or build rather than to use for an afternoon, that is a different article and the answer is in quantum computing cost in 2026, which also carries the receipts from real billed runs on this service. This page is only about access: the per-device price list, the shot bounds, and the things no price covers.

What a single run costs

Eight machines from seven manufacturers, all reachable without an enterprise agreement. The last column is a complete 1,000-shot experiment, which is a normal size for a result you would actually publish. Quandela Belenos sits at the top because it charges per job rather than per shot, so a thousand shots costs it almost nothing extra; what that buys is covered in the photonic write-up.

device                      modality          qubits  per task   per shot   1,000 shots
Quandela Belenos            photonic              12  EUR 0.30 EUR 0.000001    EUR 0.301
Rigetti Cepheus-1-108Q      superconducting      108     $0.30    $0.000425       $0.725
IQM Garnet                  superconducting       20     $0.30    $0.00145         $1.75
IQM Emerald                 superconducting       54     $0.30    $0.0016          $1.90
QuEra Aquila                neutral atom         256     $0.30    $0.01           $10.30
AQT IBEX Q1                 trapped ion           12     $0.30    $0.0235         $23.80
IonQ Forte Enterprise 1     trapped ion           36     $0.30    $0.08           $80.30
Pasqal FRESNEL              neutral atom         100     billed by the clock, see below

The task fee is a flat thirty cents on every device here, whatever the machine and whatever the circuit. Everything that varies lives in the shot column.

A note on where these sit relative to us. We resell every one of them at list with zero markup, so the numbers above are what the provider charges and what you pay. Pasqal was market context when this was first written and is now on the same footing as the rest, which is why the clock-based line below matters to a budget rather than being a curiosity.

Why one machine costs 110 times another

The identical 1,000-shot experiment is $0.725 on Rigetti and $80.30 on IonQ. That is not a pricing strategy, it is two different physical technologies with a per-shot ratio of 188 to 1.

A superconducting processor is a circuit on a chip. It is driven with microwave pulses, reset, and driven again, and the cycle is fast enough that a thousand repetitions is a rounding error of machine time.

A trapped-ion processor holds individual atoms in an electromagnetic trap and addresses them with lasers one interaction at a time. Every shot is slower by orders of magnitude, and the price follows the clock.

What you buy for the higher price is gate fidelity and full connectivity, which is a real thing to want. The full comparison of the two is here.

Pasqal bills by the clock instead

Neutral-atom machines are the exception to the per-shot rule. Pasqal's pay-as-you-go tier 1 rate is EUR 500 per hour of machine time, and the contract fixes an effective circuit repetition rate of 0.25 Hz. That is four seconds per shot, so a shot costs roughly EUR 0.56.

Put that in the same terms as the table and a 1,000-shot run is about four thousand seconds of machine time, or approximately EUR 556. Against $0.725 for the same shot count on Rigetti. These are two currencies and not a like-for-like conversion, but no exchange rate closes a gap of that shape.

This is not a criticism of the machine, it is a statement about what it is for. You do not run a thousand shots of a gate circuit on a neutral-atom device, because it does not run gate circuits.

It runs analog dynamics: a register of atoms driven by laser pulses, evolving continuously under a Rydberg Hamiltonian, with the answer read from where the atoms end up. That is a different experiment with different shot economics, and we have measured what it produces.

Why neutral-atom machines are not gate-based covers the architecture behind that split, including why they reach more qubits than anything else on public cloud.

The per shot calculator prices any of these against your own shot count, and live engine status says which of them is actually accepting jobs right now.

A free tier for neutral-atom development

Pasqal's cloud carries a device called EMU_FREE that is explicitly free and consumes no credits. Their paid emulators (EMU_MPS, EMU_SV, EMU_TN) bill against your contract; EMU_FREE does not. If you want to develop a neutral-atom sequence rather than execute a final one, that is the entry point, and it costs nothing.

The same logic applies to us. Our analog engine runs the identical Pulser sequences exactly, up to 14 atoms, on a CPU at $0.69 an hour, and it applies the same validation rules the real hardware applies. A sequence that this engine accepts is one the hardware would accept, which is a much better moment to discover a problem than after paying for the task.

What the shot ceilings actually mean

Every device has a floor and a ceiling on shots per task, and those bounds cap what a single run can cost you.

device                     min shots   max shots   most one task can cost
Quandela Belenos                   1     800,000                 EUR 1.10
QuEra Aquila                       1       1,000                   $10.30
Rigetti Cepheus-1-108Q            10      50,000                   $21.55
IQM Garnet                         1      20,000                   $29.30
IQM Emerald                        1      20,000                   $32.30
AQT IBEX Q1                        1       2,000                   $47.30
Pasqal FRESNEL                     1         100                   $60.00
IonQ Forte Enterprise 1          100       5,000                  $400.30

The floors are the surprising column. Most of these machines will genuinely run a single shot. We submitted one to IQM Garnet and one to IQM Emerald and both came back, at about thirty cents each, which is the cheapest way there is to confirm a circuit compiles and executes on real superconducting hardware.

Rigetti will not go below 10, and IonQ will not go below 100, which puts the cheapest possible interaction with IonQ at $8.30 even when all you want to know is whether your circuit is accepted.

The ceiling column is the one that matters for a budget, because it caps what a single task can cost you no matter what you do. The worst case ranges from EUR 1.10 on Quandela Belenos, which charges per job, to $400.30 on IonQ.

Renting by the shot in the ZKSF console: add credit, price a run before it starts, and see what the last ones cost
Renting by the shot in the ZKSF console: add credit, price a run before it starts, and see what the last ones cost. Try it yourself in the console

You probably do not need the hardware

For most of what people want a quantum computer for, renting one is the wrong purchase, because a simulator returns the same answer and costs a rounding error.

engine                      what it does                        rate
CPU (all engines)           exact, stabilizer, tensor network,  $0.69 / hour
                            Pauli propagation, analog
GPU                         exact statevector to 30 qubits      $3.00 / hour
GPU (31-32 qubits)          exact statevector, large card       $8.00 / hour
TPU                         neural network quantum states,      $1.85 / hour
                            ground states to 40 spins
minimum charge per circuit                                      $0.0001

The one-hundredth-of-a-cent minimum covers a CPU circuit's first second, and only the time past that is metered at $0.69 an hour. The overwhelming majority of circuits finish inside that second, which means the overwhelming majority of simulated circuits cost exactly $0.0001.

What a real comparison actually costs

Those are list prices. This is a bill: satellite observation tasking, the same circuit at 500 shots on five engines, including two superconducting quantum processors.

Run on our engines

Satellite observation tasking at 14 requests, seed 20260902, whose exact optimum is value 32.5128. On 25 September the same instance ran on exact.tpu, a Google TPU, which returned 31.2164 and the smallest gap on the table. Submitted to each kind of compute we offer, on 16 and 25 September 2026 at 500 shots. Every figure below is a real job on the service, priced as any customer would be priced.

DeviceEngineKindQubitsResultCost
CPUmps.quimb.cpuCPU1425.1546, gap 7.36 certificate$0.0001
CPUexact.cpuCPU1420.0569, gap 12.46 certificate$0.0001
NVIDIAexact.gpuGPU1425.1546, gap 7.36 certificate$0.0001
IQMqpu.iqm.garnetQPU1426.8778, gap 5.64 certificate$1.025
Rigettiqpu.rigettiQPU1427.0176, gap 5.49 * certificate$0.5125
IQMqpu.iqm.emeraldQPU1426.9421, gap 5.57 certificate$1.100
Google Cloud TPUexact.tpuTPU1431.2164, gap 1.30 certificatebest outcome$0.0776
Google Cloud TPUneural.tpuTPU—the tasking QUBO is diagonal, which is not the shape a neural ansatz is for—

* The Rigetti row is a separate sample of ours on this same instance, with the QAOA angles re-optimised for it. The steps are in the docs.

A note on the hardware certificates: they state Hellinger fidelity against the exact distribution. For an optimisation circuit that distribution is spread across many outcomes rather than concentrated on one, so the figure is low by construction and is not a measure of whether the device found a good answer. The result column above is.

The same problem is yours to run: every instance here is seeded, so it rebuilds exactly. Open the console and a cost estimate is free before anything executes.

Five engines, two of them real quantum hardware, for $2.13 all in. The three simulator rows cost three hundredths of a cent between them and the hardware carries the rest, which is the per-shot economics above expressed as a single afternoon of work. The benchmark itself is on the space and satellites page.

Renting hardware to obtain a correct answer is the expensive way to get a wrong one.

Two companion articles go further on this. CPU vs GPU vs TPU vs QPU is the decision procedure with the crossover points marked, and quantum emulator vs simulator settles what those two words mean, since vendors use them interchangeably and they do not return the same thing.

What you cannot rent at any price

Exclusive machine time. Dedicated device reservations exist, but they are enterprise contracts negotiated with the manufacturer, not something on the per-task price list. If you have arrived here expecting to book a quantum computer for an afternoon, that product exists and it is not priced in dollars per shot.

A predictable finish time. Tasks are scheduled against a shared queue, and several devices only accept work during published availability windows. A task submitted on a Friday evening may return on Monday. Price and latency are unrelated on these machines.

A correct answer. This is the one that costs people real money. A hardware run returns a distribution shaped by gate errors, readout errors and decoherence, and none of that is marked in the output. The counts look exactly like the counts a perfect machine would give you.

A logical qubit. Every machine on every price list rents physical qubits carrying their raw error rates. An error-corrected logical qubit is not for sale anywhere in 2026 at any price. What can be priced is the exchange rate between the two, and it is measured rather than quoted: run a surface code memory experiment at your device's error rate, count how often the logical value comes back wrong, and read off how many physical qubits one logical qubit would take. That runs on a classical engine for a fraction of a cent and returns a logical error rate with a 95 percent confidence interval. The measurement and the numbers it produces are in logical qubits vs physical qubits.

Knowing whether the run was worth the money

Which is the reason this service exists. Every hardware run here can be certified against an exact reference of the same circuit, reporting a Hellinger fidelity between what the machine returned and what it should have returned. The certificate is a public URL that anybody can fetch without an account, and an open-source checker recomputes the number without calling us.

That converts a $0.725 experiment from a pile of counts into a result with a stated distance from the truth. It is also the only honest way to compare two machines. A cheap shot on a noisy device and an expensive shot on an accurate one are not the same purchase, and the fidelity number is where that difference becomes visible instead of remaining a marketing claim.

How to actually rent one

Install the client, submit a circuit, name a device. There is no procurement step.

pip install qsim-sdk

Ask for a free estimate before you commit, which returns the exact charge for the quantum processor run you are about to make. Then submit with the engine named, and read the counts and the certificate when the task returns. The documentation has the full sequence, and the worked circuits each link to the certificate from the run that produced them, on the machine that produced it.

If you want to spend nothing at all first, the browser sandbox builds and runs a small circuit in the page with no account and nothing submitted anywhere.

Common questions

Can you rent a quantum computer?

Yes, and without an enterprise agreement or an institutional affiliation. What you rent is time on a shared machine rather than the machine itself. You submit a circuit, it is queued against everyone else's work, and you are billed for that task. Eight processors from seven manufacturers are reachable this way, and the cheapest complete run costs well under a dollar.

How much does it cost to rent quantum computer time?

Every task carries a flat $0.30 fee, and the rest depends on the machine. A 1,000-shot experiment is $0.725 on Rigetti Cepheus-1 and $80.30 on IonQ Forte Enterprise 1, a spread of about 110 times for the identical circuit. Pasqal FRESNEL does not bill by the shot at all. It charges EUR 500 per hour of machine time, which works out at roughly EUR 0.56 a shot.

How do you rent a quantum computer, step by step?

Install the client, ask for a free estimate, then submit with a device named. There is no procurement step and no minimum spend. The estimate returns the exact charge for a quantum processor run, which matters because the same circuit can differ by two orders of magnitude in price depending on which machine you send it to.

Run your own 100-qubit circuit, with an error bar.

Share this articleLink copied