Google Willow Magic State Cultivation Cuts Quantum Errors 40-Fold

What happened

Google Quantum AI has taken a technique that until now existed only on paper and run it on real hardware. The technique, called "magic state cultivation," ran on the company's Willow processor and reduced errors by about 40 times compared with states prepared without it. The team reported the result in Science on October 15.1 Google also said it produced T-states, a particular kind of magic state, with a record logical fidelity of 99.99%.1

The headline number needs context, and some of the early coverage left it out. The 40-fold gain is not an improvement in the whole Willow chip. It applies to one task: preparing magic states.20 One analysis put it directly: the 99.99% figure describes the fidelity of a single prepared state, not overall computing accuracy, and Willow itself is not a new chip.18 Even Sundar Pichai's post was loose on this point. He described the gain as "vs. previous methods," while Google's own research blog makes the comparison against uncultivated states.101 Some outlets also mislabeled the technique as magic state "distillation."7 That is actually the name of the older method cultivation is meant to replace.14

Read correctly, the result is narrower than the "40-fold" headlines suggest, and more important.

Why magic states are the bottleneck

Error-corrected quantum computers can perform some operations cheaply. Universal computation, though, also needs a resource called magic states. They are hard to make with low enough error, and each one is destroyed when it is used, so a machine has to keep producing new ones.13 The coverage widely repeats one estimate: producing magic states could take up as much as 95% of a quantum computer's capacity, leaving little for actual computation.1113

The standard answer has been distillation, proposed by Sergei Bravyi and Alexei Kitaev in 2004. It refines many noisy magic states into a few clean ones, at the cost of large dedicated "factories" of qubits.14 Cultivation takes a different approach. Google researchers Craig Gidney, Noah Shutty and Cody Jones proposed it in 2024.13 A noisy state is placed in a small patch of the chip and checked with a series of coordinated measurements. Attempts that fail the checks are thrown away, and the ones that pass are expanded into a larger, usable encoding.13 The checks ask a yes-or-no question about whether the state is the one wanted, instead of collapsing it outright. That is what lets the state survive being inspected.14

The engineering behind the number

Every account of the experiment agrees that it was not a simple translation from theory to hardware. Early on, first author Emma Rosenfeld saw data that appeared to break the laws of physics. After months of work, the team traced the problem to lab noise from electrical equipment drift and temperature changes, which was undermining their usual benchmarking method.1 Google responded with extra system controls and a new benchmark that uses cultivation to measure its own performance.1 One report adds that the team used reinforcement learning to improve the processor's calibration. It also notes that the new method measures a lower bound on fidelity, a cautious choice meant to avoid overstating quality.20

The details show how much the result depends on discarding work. The experiment used 54 of Willow's 105 qubits, and only about 8% of attempts passed every check. The other 92% were thrown out.14 That rate is the price of the filter. Whether it holds up at scale will decide how much of the 95% overhead cultivation actually saves.

Google also showed what it calls "grafting." A purified state was moved from a distance-3 color code, which suits cultivation, into a distance-5 surface code, the more mature scheme Google's architecture is built on.1 This step matters more than it might seem. It suggests cultivation can plug into the error-correction approach Google already uses, rather than requiring a separate architecture.

Not every result was positive. When the researchers kept the grafted states in storage for longer, the logical error rate came out about seven times higher than simulations predicted. They pointed to leakage as the likely main cause: qubits drifting out of the two states used to encode information. Neither the simulations nor the experimental circuits accounted for it.20 This detail barely appears in the headline coverage, and it is a fair warning that holding magic states until they are needed is its own unsolved problem.

The cryptography angle: real, but distant

Crypto-focused outlets quickly tied the result to Bitcoin.19 The connection is not invented. In a May 2025 paper, Gidney cut his estimate of the qubits needed to factor a 2048-bit RSA key from 20 million noisy qubits, his 2019 figure, to under 1 million. Magic state cultivation was one of three main reasons for the lower number.14 This experiment is the first hardware test of an assumption that had already been built into those threat estimates.14

The distance remaining is large. The paper says factoring RSA-2048 would need magic states with errors around one in 10 billion, roughly a million times cleaner than what Willow just achieved.17 Google expects that halving Willow's underlying hardware noise would bring magic-state errors to about two in 100,000, because errors in the process shrink with the cube of the chip's noise level.17 That cubic relationship is the most important technical point in the paper. It means modest hardware improvements could produce outsized gains in magic-state quality. It also means today's chips are still far from attack-grade.

The experiment did not break any encryption. Willow's 105 qubits are far below the fewer than 500,000 physical qubits Google estimated in March would be needed to break the elliptic-curve cryptography Bitcoin uses.19 Google's team considers molecular simulation, which needs error rates of roughly one in a million to one in 100 million, a more realistic near-term target.14

The reason this still matters for post-quantum cryptography is that it confirms a trend Google is already acting on. In March, the company moved its post-quantum migration target up to 2029, citing progress in hardware, error correction and resource estimates.22 Google Cloud has since published a roadmap. It aims to address "store now, decrypt later" risk for customer workloads by the end of 2027 and to have signature and identity protections in place by the end of 2028.27 Critics say Google is blurring an internal migration deadline with the arrival of a code-breaking machine, which most experts expect somewhere between the 2030s and the 2050s.30 That criticism misses the point. Migration takes years, and the cultivation result shows that the assumptions behind falling resource estimates are being confirmed in the lab one by one.

How this fits the industry race

The announcement came during a busy stretch for Google's error-correction work. On the same day, Google Quantum AI, Google DeepMind and MIT described modular hyperbolic codes said to cut error-correction overhead by more than 30 times.9 Two days earlier, Google researchers posted a preprint on "virtual error cancellation," a technique that suppressed leftover error bias more than 1,000-fold in simulations but has not yet been tried on hardware.5 Together these suggest Google is working on the cost of fault tolerance at several levels: codes, magic states and decoding.

Competitors are going after the same bottleneck. In July, Quantinuum and academic partners used a topologically ordered state on its H2 system to perform universal gates, describing it as a possible way around costly distillation. Quantinuum's Helios is reported to run 50 error-corrected logical qubits, and Infleqtion has shown 30 entangled logical qubits on neutral-atom hardware.31 IBM's published roadmap includes demonstrating magic state distillation across multiple modules in 2028, ahead of its Starling system in 2029, which is planned to have 200 logical qubits running 100 million gates.32 Google has published much less of a dated roadmap than IBM.36 Results like this one serve as its public record of progress instead.

The bottom line

The fair verdict is that a bottleneck has been opened, not removed. Google itself says cultivation still has to be shown at larger scale and that many technical challenges remain.1 Even so, the result changes the conversation in a real way. The magic-state factory used to be the most expensive line item in fault-tolerance planning. Engineers can now point to a measured hardware result for it, where before they had only a theoretical saving. With an 8% acceptance rate, a leakage problem and a million-fold gap to cryptographic relevance, the road ahead is long. But estimates of how many qubits a code-breaking machine needs have been falling for years, and this experiment adds hardware evidence that the trend is real.