Google's Magic State Cultivation Keeps 8% of Attempts and Is Still 1,000,000 Times Short of RSA

Google Quantum AI says it has cut the error in a hard-to-make quantum resource by a factor of 40 on its Willow processor, reaching a fidelity of 0.9999(1) while keeping only 8% of its attempts. That discard rate is the figure that decides whether the technique is cheap, and the company's own announcement does not print it.

The result, called magic state cultivation, was described in a Google blog post dated October 15, 2026 by research scientist Emma Rosenfeld and quantum software engineer Craig Gidney. It is published in Science; the post gives the DOI (10.1126/science.aee7961) but not the paper's title. The same experiment appeared on arXiv in December 2025 as "Magic state cultivation on a superconducting quantum processor", submitted 15 December 2025 with 295 authors.

What magic state cultivation does on Willow

Useful quantum programs need gates that error-correcting codes cannot apply cheaply. Magic states supply those gates, and each one is consumed on use. Google's post says some estimates put up to 95% of a quantum computer's capacity into making them.

The usual route is distillation, which consumes many noisy states to yield one clean one. Cultivation, proposed in 2024 by Gidney, Noah Shutty and Cody Jones, works as a filter instead. A noisy state is injected into a small patch of the chip, coordinated measurements flag errors, and flawed attempts are thrown away. The survivor is grown and then moved into a larger code.

On Willow the team switched a state from a distance-3 color code into a distance-5 surface code. Google reports a 40-fold error reduction against uncultivated states and a record logical fidelity of 99.99% for T-states.

The 8% retention rate is the real cost

The arXiv abstract states that cultivation reaches 0.9999(1) "while retaining 8% of attempts." Eight kept out of 100 means about 12.5 attempts per usable state (1 divided by 0.08). The 92% discarded is the filter doing its job, and it is also the overhead that a larger machine would pay in qubits and cycles.

Quantum Zeitgeist's summary of the paper adds three qualifications. The 1×10⁻⁴ infidelity is an upper bound, not a measured value. The 40-fold gain is against state injection, not against distillation. And the fidelity is described as comparable to trapped-ion demonstrations, not better. Google's blog post carries none of these three.

The post also gives no qubit count and no comparison with distillation overheads. Those would settle the "cheaper" claim, and they are absent from both the announcement and the abstract.

How far 0.9999 is from the target

The same summary cites the infidelity levels that real workloads need: roughly 10⁻⁶ to 10⁻⁸ for some applications and around 10⁻¹⁰ for factoring RSA-2048. Against the 10⁻⁴ upper bound reported here, the gap is arithmetic.

Target	Infidelity needed	Gap from 10⁻⁴
Willow result (upper bound)	10⁻⁴	none
Some applications, upper end	10⁻⁶	100x
Some applications, lower end	10⁻⁸	10,000x
RSA-2048 factoring	10⁻¹⁰	1,000,000x

The 2024 theory paper, per a search summary of its abstract, modelled logical error rates as low as 2×10⁻⁹ at 10⁻³ circuit noise. Those are simulation figures. The Willow experiment is the first hardware number against them, and it is five orders of magnitude from the modelled one.

That is not a failure of the experiment. Cultivation is designed to deliver a moderately clean state cheaply and let a second stage, a short distillation or a larger cultivation, finish the job. The hardware result shows the first stage works; it does not show what the second costs.

What the authors say remains

Google's post says the demonstration must be scaled up before it is practical and that technical challenges remain. It stops there: no timeline, no target qubit count, no date by which a machine would need fewer than 12.5 attempts per state. It also describes the team's debugging: anomalous data traced to drift in electrical equipment and temperature variation, which led to added control to reduce that noise and a benchmarking method in which cultivation measures itself.

For context on Google's compute spending this week, see our coverage of Gemini 4 Argon's API launch, and for the classical-hardware side of the scale argument, Microsoft's MAI-Base-2 training run on GB200 systems.

Why the discard rate matters to chip designers

A discarded attempt still occupies the qubits that held it for the full cycle count. The arXiv abstract's figures include two rounds of cultivation and three error-correction cycles, per the Quantum Zeitgeist summary, so every one of the roughly 12 failed tries per success spends that time on hardware that could be doing something else. At 8% retention, a factory producing one magic state per cycle needs about a dozen parallel cultivation patches running to keep pace, before routing and the second stage are counted. That is my arithmetic from the published rate, not a figure Google gives, and it assumes independent attempts.

The design question for a fault-tolerant machine is therefore how much silicon area goes to patches that mostly fail. Google's 95% estimate for the old approach is the baseline to beat, and neither the post nor the abstract says what share cultivation would take.

What would change this read

A published retention rate at a lower infidelity would move the picture quickly; if Google can reach 10⁻⁶ while keeping more than 8%, the overhead argument weakens. A second group reproducing the 40-fold figure on different hardware would matter more than any further Google result. Until one of those appears, the supportable claim is narrow: cultivation works on one chip, at one code distance, at a price of 12.5 tries per state.