Google's Willow chip cuts quantum magic-state errors 40x

Google has shown that a technique called "magic state cultivation" works on real quantum hardware, cutting errors in quantum computing's costliest ingredients by about forty-fold. The experiment ran on the company's Willow chip, and the results were described in the journal Science on October 15, 2026. According to Google Quantum AI, it is the technique's first run on physical hardware rather than in theory or simulation.

Magic states are special quantum resources that universal quantum computers need to run programs no ordinary machine can handle. Producing them with few errors has been a stubborn bottleneck for the field for two decades. Estimates cited by Google suggest that as much as 95 percent of a quantum computer's total capacity could be consumed simply making them, and each magic state is destroyed the moment it is used, so the cost never stops.

Willow is the superconducting processor at the center of Google's quantum program, and it has become the team's testbed for error-correction research. That work drew wide attention when Google reported logical error rates that fell as the team added more qubits to the code, the long-sought condition known as operating below the error-correction threshold. The new cultivation result is the next step in that same campaign.

A filter, not a factory

Traditional approaches, known as distillation, build these states through repeated, resource-hungry purification rounds. Cultivation takes a different route. A noisy candidate state is injected into a small patch of the chip, and rounds of coordinated measurements probe it for errors. Failed attempts are discarded; survivors are purified as they grow into a larger, usable encoding. The approach was outlined in 2024 by Google researchers Craig Gidney, Noah Shutty and Cody Jones, and the new work is its debut on superconducting hardware.

The team also showed that a purified state can be grafted from a distance-three color code into a distance-five surface code, a different and more established error-correction scheme. Google presents the transfer as evidence that cultivation can slot into existing quantum architectures rather than requiring machines to be rebuilt around it.

The headline number, with an asterisk

On Willow, cultivated T-states, the variety of magic state needed for universal computation, reached a logical fidelity of 99.99 percent, a record for the technique. The experiment used at most 54 of the processor's 105 qubits, and about eight percent of preparation attempts survived the filtering process. Google disclosed the result in a blog post from its Quantum AI team, authored by researchers Emma Rosenfeld and Craig Gidney, alongside the paper and a preprint on arXiv. The paper appeared as a preprint in December 2025, according to its arXiv listing, before passing peer review.

That yield means the headline figure comes with an asterisk: the fidelity figure applies only to the attempts that passed inspection. In a statement accompanying the announcement, Google said the result still has to be demonstrated at a larger scale before it can be called practical, and that many technical challenges remain.

Why anyone outside a lab should care

A universal, fault-tolerant quantum computer could eventually crack the encryption that protects banking and digital assets, which is why each incremental advance draws attention from the security world. With an error rate of roughly 1 in ten thousand among the attempts that passed inspection, the new result is still far too noisy for that kind of work, according to reporting on the paper. But it attacks what has been the most expensive line item in every blueprint for such a machine.

Scientific American, covering the study, noted that the demonstration turns a long-theorized idea into something that runs on an actual chip. If the method scales, quantum engineers would spend far less of every machine's capacity on making magic states, leaving more room for the chemistry, materials and optimization problems that motivate the field. The idea of cheap magic has been on the drawing board since 2004. Now it has a hardware result attached to it, joining a run of physics stories on the site, from optical neural networks reading hidden light patterns to the broader science desk.