Google grew 'magic states' on real hardware — the costliest part of a fault-tolerant quantum computer just got cheaper

Google Quantum AI reported in Science on 15 October that it has demonstrated magic state cultivation on its Willow processor, the first time the technique has run on physical hardware. The team prepared magic states with a fidelity of at least 0.9999 and errors roughly 40 times lower than uncultivated states, at the cost of discarding most attempts — about 8% were kept. Cultivation was proposed by Craig Gidney and colleagues in 2024. Google says the next step is to run it at larger scale, and that halving the chip's hardware noise would push magic-state errors much lower. The tracker records the result.

Why it matters

Error correction protects quantum information, but it only makes a certain set of operations — the Clifford gates — cheap. Every algorithm worth running also needs non-Clifford gates, and the standard way to get them is to feed in a special resource state, a magic state, of very high quality. Producing those states has been the most expensive part of most fault-tolerant designs: the older method, distillation, uses large "factories" of qubits to boil many noisy states down into a few clean ones, and estimates have put that overhead at the majority of a machine's hardware.

Cultivation grows a good magic state in place, checking it as it goes and throwing away the attempts that fail. It trades a low acceptance rate — the 8% here — for a large saving in qubits. If it scales, the number of physical qubits needed for a useful fault-tolerant computer falls substantially, which changes the arithmetic behind every company's roadmap, not only Google's.

Two things keep this in proportion. It is a component result on one chip, not a fault-tolerant computer, and Google says as much. And it is peer reviewed in Science, which in a field where most "records" are company announcements makes it one of the more trustworthy quantum results of the year. It also lands a week after DARPA moved six rival approaches to hands-on government testing: independent checking is arriving in this field from two directions at once.

What to watch

Cultivation at larger code distances, whether other hardware platforms reproduce it, and whether resource estimates for fault-tolerant algorithms are revised downward as a result.