Google Shows Magic State Cultivation Working on Willow

What changed

On 15 October 2026, Google Quantum AI said it had shown that magic state cultivation works on the Willow processor. Research scientist Emma Rosenfeld and quantum software engineer Craig Gidney wrote that the protocol cut errors about 40 times compared with uncultivated states, and that the team recorded T states at a logical fidelity of 99.99 percent. The blog says the work was published the same day in Science.

Magic states are a known bottleneck. The lab recalls that, since 2004, researchers have expected a useful quantum computer to spend a large share of its capacity just preparing these states, because a universal set of operations depends on them and a used magic state cannot be reused. Google cites estimates that as much as 95 percent of a machine might have to be reserved for that preparation.

In 2024, Craig Gidney, Noah Shutty, and Cody Jones turned an idea from the literature into the cultivation protocol. A noisy state is injected into a small region of the chip. A series of coordinated measurements then throws away flawed attempts, so the surviving state can be grown into a larger encoding. Rosenfeld wrote that early lab data appeared to break the laws of physics. Months of checks traced the contradiction to the benchmark itself: drift in electrical equipment and temperature variation had invalidated an assumption in the measurement method. The team added control to reduce that noise and built a benchmarking method that uses cultivation to measure itself.

On Willow, that stack produced the 40 times error reduction and the 99.99 percent T state figure. The group also moved a purified state from a distance 3 color code, which suits cultivation, onto a distance 5 surface code, the encoding Google already uses more widely. The blog calls that grafting a first check that the 2024 theory fits this hardware. It also says cultivation still has to be shown at a larger scale before it is realistic for practical computing.

The cover photograph is an Air Force Research Laboratory quantum lab. It is not the Willow processor and not Google's chip.

Why it matters

Error corrected quantum computing is limited by how many good magic states a machine can supply, not only by how many physical qubits it has. If cultivation holds up, a buyer of quantum hardware can treat magic state overhead as an engineering cost that has started to move, rather than as a fixed tax that consumes almost the whole machine. Google is explicit that this is not yet a useful computer. It is a hardware demonstration that a theoretical saving survives real noise.

The result also changes how to read fidelity claims. The lab says an ordinary benchmark was lying because of slow equipment drift. Teams comparing Willow with other processors should ask whether the fidelity number was measured with a method that can see that class of noise.

Who is affected

Groups building superconducting processors, and the software teams writing distillation and cultivation schedules, now have a public Willow data point to beat or reproduce. National and university labs running their own cryogenic systems are in the same race. The photographs in this piece are other quantum laboratories, not Willow and not Google's chip.

Photo: U.S. Army Research Laboratory, via Wikimedia Commons. Quantum teleportation research in Adelphi, Maryland. Not Willow.

Photo: Shishirdasika / Wikimedia Commons. Quantum structure laboratory at CSIR NPL in Delhi, not Willow.

What to do next

Ask Google for the Science citation details before putting the 99.99 percent figure into a board note, and keep the comparison label: it is versus uncultivated states on Willow, not versus a named rival machine. If you are scoping a quantum program, separate this overhead result from earlier Willow claims about verifiable advantage. They are different experiments.

What to watch

Whether a larger code distance still shows the same error reduction, and whether other hardware groups reproduce cultivation rather than only cite it. Also watch whether the Science paper's methods section matches the blog's account of the drift problem. This desk does not yet have a DOI from the blog post.

Sources

• Primary — Google Quantum AI, Bringing the magic: Google breaks open a bottleneck in quantum computing (15 October 2026). Establishes the Willow result, the 40 times error cut, the 99.99 percent T state fidelity, the 2024 cultivation paper by Gidney, Shutty, and Jones, the benchmark drift finding, and the color code to surface code grafting. The authors say the work was published in Science the same day.