Google Willow Chip Grows ‘Magic States,’ Cutting Errors 40x

Google Quantum AI said on Oct. 15, 2026, that magic state cultivation, a technique its researchers first proposed in 2024, now works on real hardware. In a paper published in the journal Science, the team reported that running the protocol on its Willow superconducting processor cut errors in a key type of quantum state by about 40 times, reaching a measured fidelity of 99.99%.

Image: Google’s earlier Sycamore quantum processor. Photo: Google, via Wikimedia Commons (CC BY 3.0).

What a magic state is, in plain terms

Ordinary computers build every calculation from a handful of simple logic gates. Quantum computers have their own basic set, known as Clifford gates, but those alone can be imitated by a normal computer. To do something a classical machine cannot, a quantum computer also needs at least one “non-Clifford” operation, most often called a T gate.

That extra gate is produced by consuming a carefully prepared qubit called a magic state. Emma Rosenfeld, the study’s lead author and a research scientist at Google Quantum AI, told Scientific American that the state “has to be made very, very precisely,” because any flaw in it becomes a flaw in the gate. Once a magic state is used, it is gone, so a working machine has to keep producing fresh ones.

The traditional recipe, called distillation, turns many imperfect states into fewer accurate ones. It is expensive. Rosenfeld said up to 95% of a quantum computer could end up devoted just to making high-fidelity magic states, and Google’s own announcement repeated that estimate as the bottleneck researchers have chipped away at since 2004.

How cultivation works on Willow

Cultivation, proposed in 2024 by Google researchers Craig Gidney, Noah Shutty and Cody Jones, takes a different route. Instead of combining many states, it works on a single one. A rough magic state is injected into a small patch of qubits, then checked with a series of coordinated measurements. Rather than asking the qubit whether it is a 0 or a 1, which would collapse it, the checks effectively ask whether it is the right magic state. Attempts that fail are thrown away.

States that pass are spread across more qubits for stronger protection, which is where the gardening name comes from: the state starts small, gets checked, then grows. In the experiment, the team used 54 of Willow’s 105 physical qubits. Without cultivation, about four attempts in 1,000 came out faulty. With it, errors fell roughly 40-fold, to an upper bound of about one error in 10,000.

The researchers also moved, or “grafted,” a cultivated state from the small color code it was grown in into a larger distance-5 surface code, the error-correction layout Google uses for its logical qubits. The paper says that step tests the protocol beyond what earlier simulations could check and shows cultivated states can feed into leading error-correction designs.

Getting there was not straightforward. Google said Rosenfeld spent months chasing data that “seemed to defy the laws of physics” before tracing it to lab noise from equipment drift and temperature changes that undermined the team’s usual benchmark. The fix included extra control hardware and a new benchmarking method that uses cultivation to measure itself.

The team also tuned its control settings with reinforcement learning, minimizing the number of error-detection events in the circuit, according to the paper posted on arXiv.

The catch: most attempts are discarded

The filter is strict. Only about 8% of attempts survive the checks, meaning 92% are rejected. Kenneth Brown, director of the Duke Quantum Center, who was not involved, told Scientific American that “the Google experiment shows that cultivation works in practice,” while pointing to that heavy discard rate.

The result is also far from what large applications need. Breaking RSA-2048 encryption would require magic states that fail only once in roughly 10 billion tries, according to the Scientific American report. Simulating molecules, which Rosenfeld described as more attainable, would need error rates between one in a million and one in 100 million. Google’s team now sits near one in 10,000.

Why it matters for the quantum race

The fidelity Google reported is similar to recent trapped-ion demonstrations, the paper notes, but superconducting chips like Willow run faster. If cultivation scales, it could shrink the size of the machine needed for useful work, because far less of the chip would be tied up making magic states. Google called the result “a crucial step towards scaling up quantum computing to practical use.”

The news also rippled through crypto circles online, where some users raised long-term questions about encryption, though the gap between one error in 10,000 and the one-in-10-billion level cited for RSA shows how far away that remains.

The work adds to a busy month for science and AI research. The Boston Report recently covered CERN and Mistral AI’s plan for a Large Physics Model, and earlier looked at Google’s SynthID Detector opening to the public. For another recent discovery, see our report on the first known green star.

Sources

• Google: Bringing the magic: Google breaks open a bottleneck in quantum computing
• Science: Magic state cultivation on a superconducting quantum processor
• arXiv: Magic state cultivation on a superconducting quantum processor (preprint)
• Scientific American: Google’s quantum Willow chip may be able to grow ‘magic states’