How much energy does a quantum computer use?

Status as of September 25, 2026. A working quantum computer today draws tens of kilowatts. Google ran its Sycamore processor at roughly 26 kW of total wall power (The Conversation, Oct 2022; breakdown in RAND, 2023), and D-Wave rates its Advantage system at 25 kW maximum (datasheet, 2022). An exascale supercomputer draws about a thousand times more: Frontier is listed at 24,607 kW on the June 2026 TOP500 (TOP500); the ×1,000 ratio is our own arithmetic on those two published figures. That gap is real and measured. What does not follow from it is the popular conclusion — that quantum computing is already the green alternative. Energy per machine is not energy per solved problem, and on the one task where a quantum machine claimed a win, the best classical rerun now spends less energy than the quantum run did (Fu et al., 2024). Rosetta Quantum reads the energy question the way it reads every claim: measured first, promised second.

How many kilowatts does a real quantum machine draw today?

Tens of kilowatts — a large house, not a power plant. The two machine-level figures on the public record that this run could verify: ≈26 kW for Google’s Sycamore, which RAND’s 2023 review breaks into ≈10 kW for the dilution-refrigerator compressor, 10–13 kW for chilled-water cooling and ≈3 kW for classical control electronics (Parker & Vermeer, RAND, 2023), and 25 kW rated maximum for D-Wave’s Advantage annealer, per the vendor’s own datasheet (D-Wave, 2022). Two properties of that number matter more than its size. First, the draw is roughly constant: the refrigerator runs around the clock whether the QPU is computing or idle, so the energy bill is dominated by uptime, not workload. Second, at today’s scales the draw barely tracks qubit count — Sycamore’s 53 qubits and Advantage’s 5,000+ qubits sit within the same ≈25 kW envelope, because what is being cooled is a cryostat, not a transistor per qubit.

Where does the electricity actually go?

Into refrigeration, not computation. A superconducting QPU operates at 15–25 millikelvin, and holding it there is a fight against thermodynamics. IBM’s Goldeneye — the largest quantum-ready cryostat yet built, with 1.7 m³ of experimental volume, cooled to ≈25 mK — delivers about 10 milliwatts of cooling power at the 100 mK stage and ≈24 W at 4 K (IBM, Sep 2022). Set those milliwatts against the ≈10 kW a dilution-refrigerator compressor draws from the wall (RAND, 2023) and the shape of the machine appears: roughly a million watts in per watt of heat removed at the coldest stage — our arithmetic on the two published figures. The energy of a quantum computer is not the cost of flipping qubits; it is the cost of keeping a fraction of a cubic meter colder than deep space, continuously. Trapped-ion and photonic machines skip the dilution refrigerator (vacuum chambers, lasers, and ≈4 K stages instead), but we could not find published system-level wall-power specs for those platforms in this run’s sources — an absence worth naming, not a verdict.

How does that compare with a supercomputer?

The comparison is ≈1,000× in quantum’s favor on wall power — and it means less than it sounds. Here is the ledger, every row from a published source:

Watts, however, are not verdicts. The honest unit is energy per solved problem, and that unit only exists where both machines actually solve the same problem. On problem classes with no demonstrated quantum advantage — today, that is every commercial class on Rosetta’s proven-advantage list — the supercomputer’s megawatts buy answers and the cryostat’s kilowatts buy physics experiments. A machine that draws 1,000× less power but does not solve your problem is not greener; it is idle. (What a quantum run costs in dollars rather than watts is its own ledger.)

Does the «quantum is green» argument survive the data?

As a conditional, yes; as a present-tense claim, no. The serious version was put on the record in 2022: Sycamore consumes far less than a supercomputer on its sampling task — but that task «has no identified ‘useful’ application to date», and the projected savings assume qubits «of a quality that is still out of reach in practice today» (The Conversation, Oct 2022). Jaschke & Montangero turned the condition into a threshold: a green quantum advantage exists where the QPU spends less energy than the best classical emulation at equal fidelity, and their estimate places the crossover near an entanglement power-law exponent α ≲ 1 — favorable for random circuits, unfavorable for structured algorithms like the quantum Fourier transform (Quantum Sci. Technol., 2023). Then the flagship example inverted. In July 2024, Fu et al. ran Sycamore’s sampling task on 2,304 A100 GPUs in 17.18 seconds at matching XEB fidelity (0.002), spending 0.29 kWh against the 4.3 kWh a Sycamore run of the same 3-million-sample size consumes at ≈26 kW — «outperforming Google’s quantum processor Sycamore in both speed and energy efficiency», in the paper’s own words (arXiv:2407.00769). That is the paper our Sycamore retrospective flagged as unverified yesterday; this run verified it at its arXiv source. Zhao et al. then cut classical simulation energy by another two orders of magnitude against prior GPU runs while re-running the task faster than the quantum machine (Natl. Sci. Rev. 12, nwae317). The scoreboard as of today: no task on the public record shows quantum hardware demonstrably spending less energy than the best classical run of the same task.

What will a fault-tolerant machine consume?

Nobody has measured one, because none exists — and the published estimates disagree by ≈18× (our arithmetic on the two figures). RAND extrapolates ≈125 MW for a cryptanalytically relevant quantum computer breaking RSA-2048 — 6.25 W per physical qubit across 20 million qubits, 7.1 hours per key — and flags its own extrapolation as extremely rough (RAND, 2023). Fellous-Asiani et al., optimizing the full stack with their Metric-Noise-Resource methodology, estimate the same job will run at ≈7 MW in 1.5 hours (PRX Quantum, 2023). Both numbers will remain estimates until someone meters a real machine. Either way, the kilowatt era is a property of today’s lab-scale devices, not a law of the technology: at useful scale, quantum computers will be megawatt facilities too — and the grid that would host them is a ledger Rosetta has already scored.

What we don’t know

• Wall power for most commercial QPUs is unpublished. We could not find system-level power specs for IonQ, Quantinuum or IBM production systems in this run’s public sources; the ≈26 kW and 25 kW figures cover two machines, not the industry.
• How cryogenic draw scales. Whether the roughly constant kilowatt envelope holds through the next 10× in qubit count — RAND flags economies of scale as an unknown in both directions.
• Where the green-advantage threshold really lands. The α ≲ 1 crossover is a model estimate, not an end-to-end measurement on commercial workloads.
• Our own scope. Rosetta Quantum has not metered any quantum computer’s wall power in-house; every figure above is from the linked public record.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim.