Google Cracked Quantum Computing’s 20-Year Bottleneck

Google Quantum AI says its magic state cultivation protocol reduced errors about 40-fold on the Willow processor, and the work was published in Science on October 15. The underlying preprint, posted to arXiv in December 2025 by Emma Rosenfeld, Craig Gidney and a large Google team, supplies the numbers. In the team's test circuit, a state prepared by injection alone showed a logical error of 4.4 × 10⁻³. With cultivation added, the error fell to between 1 × 10⁻⁴ and 2 × 10⁻⁴. The authors report a state fidelity of 0.9999(1) and kept 8 percent of attempts to get there.
Why T gates need prepared states
A quantum computer built on the surface code can run most of its operations cheaply, but T gates and other non-Clifford operations have no native implementation. The standard workaround is to prepare a special resource called a magic state, consume it to apply the gate, and prepare another for the next one. The preprint states that factoring a 2048-bit RSA key would call for magic state infidelities near 10⁻¹⁰, and that some quantum simulation and optimization tasks need roughly 10⁻⁶ to 10⁻⁸. Google's announcement adds that some estimates put as much as 95 percent of a machine's capacity into producing these states, and that each state can be used only once.
Distillation, the established method, combines many noisy states through logical operations to obtain a few clean ones. In 2024, Gidney, Noah Shutty and Cody Jones proposed cultivation as a cheaper route. Their paper estimated that cultivation fits inside a single surface code patch and uses about as many physical gates as a lattice surgery CNOT gate of similar reliability, and it suggested that further distillation might never be needed in practice. Those estimates came from simulation. The Willow experiment tests the idea on hardware.
How the protocol runs
The team began with a distance-3 color code on Willow, a processor with 105 transmon qubits. A noisy T state is injected into the code, and a measurement of the logical Hadamard operator, carried out by phase kickback through helper qubits, checks whether the state sits on the magic state axis. Any detector that fires causes the attempt to be discarded and restarted. When every check passes, the circuit runs in reverse to restore the code's stabilizers and to turn the helper qubits into additional error detectors. The improvement comes from this post-selection, and the cost appears as discarded attempts. In the measured circuit, which included two extra rounds of cultivation and three error
correction cycles, 8 percent of the data was kept. State injection alone kept 76 percent.
The color code contributes a great deal to the result. When the authors removed the error correction cycles, the logical error grew about 100-fold. The full circuit contains 216 two-qubit gates and 54 measurements, and its logical error of 1 to 2 × 10⁻⁴ is below the error of each of those individual operations. Each of the four error correction cycles added to the circuit uses 30 two-qubit gates and 6 measurements, and together they improved logical performance by roughly a factor of 100.
The team also tuned the hardware with reinforcement learning. Throughout the experiments, control parameters were adjusted to minimize the number of error detection events in the error correction circuit, a technique the group describes in a separate November 2025 preprint. Simulations of the cultivation circuit matched the measured error and retention rates when the authors assumed a depolarizing noise model with a strength of 2.3 × 10⁻³, which gives a rough picture of how the device behaved during the runs.
Measuring an error of one in 10,000
Characterizing a state that clean became a research problem of its own. Standard logical state tomography would need about 10⁸ shots to resolve errors near 10⁻⁴, according to the preprint. It can also overestimate purity when coherent noise is present, to the point that some pure states produce a Bloch vector longer than one even with perfect measurements. Google's post describes Rosenfeld spending months with data that appeared to defy physics before tracing it to an incorrect assumption in the team's usual measurement technique. The post attributes the interference to noise from electrical equipment drift and temperature variation in the lab.
The team's answer, called kickback tomography, uses cultivation as the measuring device. A first round prepares the state, and later rounds measure it along the T axis inside one fault-tolerant circuit. That cuts the number of measurements needed to roughly 10⁴. The method yields an upper bound on infidelity, and the headline value is written 1(1) × 10⁻⁴, with an uncertainty as large as the value itself. The authors also report that device drift moves the logical error by about 1 × 10⁻⁴ over an hour. The preprint is at arxiv.org/abs/2512.13908 and its data is deposited at doi.org/10.5281/zenodo.17872003.
Grafting into a surface code
A cultivated state has to leave the small color code, because the error accumulated per cycle in a distance-3 code would erase the gain. The team grafted the state into a distance-5 code that is compatible with the surface code. That grafted code mixes color code and surface code stabilizers, so standard matching decoders cannot process it, and the data was decoded with Tesseract, a search-based most-likely-error decoder.
The experiment used 54 qubits in its largest layout. The color code first shifts to a geometry compatible with the surface code stabilizers in a single extension cycle, and the extra data qubits are initialized in the basis that produces the most error detectors. The remaining cycles run in the grafted code, and the authors tomographed the logical state after two and three cycles in total.
The grafting results are more modest than the headline. After grafting, the distance of the state from the surface of the Bloch sphere matched the injection-only case, because both were limited by idle error in the distance-5 code. The authors call the experiment a proof-of-principle demonstration of state escape, while Google's announcement describes the transfer as smooth. Retention rates differed by about 10 percent between runs of two and three cycles, which the authors suggest may come from flux transients or calibration drift. Four cycles would be needed before the grafted code could switch to a matchable one, and that step was left for later work.
Comparison and open items
Google's announcement calls the 99.99 percent fidelity a record. The preprint describes it as similar to recent trapped-ion demonstrations, with the added benefit of faster superconducting hardware. Against the targets quoted above, the measured upper bound sits 100 times above the 10⁻⁶ end of the simulation range and one million times above the figure for RSA-2048.
The authors expect infidelity to scale as the cube of the processor's noise strength. Halving the noise should bring the upper bound to about 2 × 10⁻⁵ using only a distance-3 code for cultivation, according to the preprint's supplementary material. The 2024 theory paper reported simulated logical error rates of 2 × 10⁻⁹ at a circuit noise of 10⁻³ and 4 × 10⁻¹¹ at 5 × 10⁻⁴, including growth to a distance-15 code. The fitted noise strengths for the Willow runs were 2.3 × 10⁻³ and 3.0 × 10⁻³. Other groups have since proposed cultivation variants on the surface code itself that would skip the color code stage, and the preprint cites several.
Google's post says cultivation still has to be demonstrated at larger scale. The preprint lists real-time implementation of T gates, grafting into larger distances, further decoder work for the grafted code, and a matchable code after grafting as the next steps.

