Google Grows Cleaner Quantum Parts and Throws Out 92 Percent

Google researchers have cut the error in a key ingredient of fault-tolerant quantum computing 40-fold on real hardware, using a method called magic-state cultivation on the company's Willow chip ^[2][3][4].

The result, published in Science on October 15, 2026, is the first experimental demonstration of cultivation on a superconducting quantum processor, according to the paper ^[1][3]. The authors report a logical-state fidelity of 0.9999(1), measured against state injection without cultivation ^[3][4]. It addresses one of the most expensive line items on the road to a quantum computer that does useful work.

It is not a 40-fold drop in every error a quantum processor makes. Nor does it make a whole quantum computer 40 times more capable. The number describes the quality of one kind of prepared state ^[3][4].

Why the machines need a special ingredient

Every operation on a quantum bit carries a small chance of error, and a long calculation strings together billions of operations. The standard answer is error correction: spread one reliable "logical" qubit across many physical ones, and keep checking them for faults.

That protection comes easily for some operations and not for others. The harder ones, known as non-Clifford operations, are needed alongside the easy ones for a computer that can run any quantum algorithm ^[2][4]. To perform them, the machine consumes magic states, special quantum states that act like a one-use part. Each time the computer uses one, it is gone, so it must make more ^[2][4].

The states have to be very clean. The paper's introduction gives example targets. Some quantum simulations and optimization problems would need errors of about 10^-6 to 10^-8, or one in a million to one in 100 million. Factoring a 2048-bit RSA key, the kind of encryption protecting much of the internet, would need about 10^-10, or one in 10 billion ^[4].

The old way and the new one

The established method is magic-state distillation. It takes many imperfect states and runs them through repeated error-detection steps to produce fewer, cleaner ones. It works, but the paper calls it resource-intensive ^[4].

Google says some estimates put the share of a quantum computer's capacity devoted to making magic states as high as 95 percent ^[2]. That is an estimate of what a full machine might spend, not something this experiment measured.

Cultivation takes another route. Researchers prepare a noisy state inside an error-protected logical qubit, then run coordinated fault-tolerant measurements that check it and refine it. Any attempt that fails the checks is thrown away ^[2][4]. It is a strict quality inspection built into the making of the part, rather than a separate refinery afterwards.

The inspection is strict. The team kept only about eight percent of attempts, roughly one in 12 ^[3][4]. That cost is the price of the cleaner product, and the open question is whether the bargain holds at scale.

The engineering test

Making a clean state is half the job. It also has to be handed to the code that will use it. The team began cultivation in a distance-3 color code and transferred the result into a distance-5 surface code ^[2][4]. Both are families of error-correcting code, and the distance is a rough measure of how large and protective each one is. The surface code is the more widely used architecture, so the handover tests a route between the two.

The measured error after cultivation is around 10^-4, or one in 10,000 ^[4]. That is far better than the starting point, yet still short of the targets above. Cultivation alone does not reach them, and the paper does not claim it does ^[4].

Fighting drift

The lab work was not tidy. Google researchers Emma Rosenfeld and Craig Gidney wrote that drift in electrical equipment and variations in temperature disrupted the team's earlier measurement benchmark ^[2]. The team added control measures and developed a new benchmarking method ^[2].

It is a reminder of how delicate these machines are. The thing being measured was an error of a few parts in 10,000, and the room itself could swamp it.

The study first appeared on arXiv, the open preprint server, on December 15, 2025, ten months before Science published it ^[3]. Rosenfeld is the first author and the corresponding author on that record ^[3].

Why a 40-fold gain on a chip matters

My view is that this is more than a footnote. Cultivation had a theoretical case before it had a hardware one, and a theory that survives contact with a real chip, with its drifting electronics and stray heat, is worth far more. The paper shows the idea works in a physical device, that its output can be handed to a surface code, and that the gain is large.

The wider stake is the size of the future machine. If making magic states eats most of a computer's capacity, as some estimates suggest, then any method that makes them cheaper brings forward the day when a useful machine is buildable. This result does not prove that saving. It makes the saving credible.

Google's own researchers are cautious. They say larger-scale demonstrations and further technical work are needed before cultivation is realistic for practical quantum computing ^[2].

What to watch

Four questions will decide how much this matters. Does the error reduction hold as codes grow and operations multiply? How much hardware and time does cultivation really save in a full architecture, once the 92 percent of discarded attempts are counted? Can it be fitted into the larger error-correction systems that will consume magic states? And can other teams, on other hardware, reproduce it?

None has an answer yet. The first test is the one Google has set itself: bigger codes, more rounds, and the same clean states at the end.