Google Reports a Major Advance in Quantum Computing Error Reduction

Preface

Quantum computers could eventually help solve problems in areas such as drug discovery and materials science. Their potential, however, depends on overcoming a central engineering challenge: quantum information is highly vulnerable to noise, so useful calculations require extensive error correction. Google Quantum AI reports progress on a particularly costly part of that challenge. In a paper published in Science on Thursday, the team demonstrated a technique called magic state cultivation on its Willow processor, reducing errors by about 40-fold and producing T-states with 99.99% logical fidelity. The result brings a proposal previously explored in theory and simulations onto real hardware. It does not mean that quantum computers can now break modern encryption. Instead, it offers evidence that one important resource for future quantum programs may be produced more efficiently, while the gap between this experiment and cryptographic applications remains substantial.

Lazy bag

Magic states are single-use resources needed for certain operations in useful quantum programs, and estimates cited by Google suggest producing them can consume up to 95% of a quantum computer’s capacity. Traditional magic state distillation reduces errors by processing many noisy states, but requires large, dedicated qubit factories. Cultivation uses a different approach: it tests a rough state, discards attempts that fail checks, and progressively develops the survivors. On Willow, Google reports an error reduction of about 40 times and 99.99% fidelity, although only about 8% of attempts passed every check. The experiment used at most 54 of the processor’s 105 qubits. Breaking RSA-2048 would require magic-state errors near one in 10 billion, so the result is a technical advance, not evidence that encryption is currently at risk.

Main Body

Google Quantum AI’s reported demonstration addresses a bottleneck that sits between basic quantum error correction and the execution of larger, more capable programs. Quantum computers use qubits, which can represent a blend of 0 and 1. This property allows quantum algorithms to manipulate information in ways that differ from ordinary computing. But qubits are fragile: heat, electrical noise, and other disturbances can alter their states and compromise a calculation. As a result, practical quantum computing depends on identifying and correcting errors without destroying the information being processed.

One common strategy is to spread information across multiple physical qubits so that a system can detect and correct certain faults. This approach can protect many operations, but a useful quantum program also needs harder operations that cannot be protected in the same way. Magic states provide a route to performing those operations. They are prepared in advance and consumed when used, making them comparable to a limited, single-use resource rather than a reusable component.

The cost of preparing sufficiently clean magic states has therefore become a major concern. Google cites estimates that production could use as much as 95% of a quantum computer’s capacity. If a large fraction of a machine must be dedicated to making the resources required for computation, relatively little capacity remains for the program itself. Lowering this overhead could make a future fault-tolerant machine more efficient, although it would not by itself solve the many other challenges involved in building one.

The established approach to improving magic-state quality is distillation. Physicists Sergei Bravyi and Alexei Kitaev described the standard method in 2004. It starts with many imperfect, noisy magic states and combines them to produce a smaller number of cleaner states. The method can be effective, but it requires substantial hardware and operations. In large-scale designs, the qubits and supporting processes devoted to distillation can take the form of extensive “factories” whose role is to produce magic states rather than execute the main computation.

Google researchers proposed magic state cultivation in 2024 as a potentially less expensive alternative. Rather than relying on a conventional factory that distills batches of states, cultivation begins with a rough magic state in a small region of the processor. The system then performs a sequence of checks. If a check detects a flaw, that attempt is discarded. If it passes, the state can be extended through further steps. In effect, the method filters out unsuccessful attempts while allowing viable states to develop. The researchers’ earlier work suggested that, in some circumstances, further distillation “may never be needed in practice.” That was a proposal, however, not a hardware demonstration.

The new result tests the approach on Google’s Willow processor. In its account of the work, Google said the protocol achieved an “extraordinary 40-fold reduction in errors” relative to un-cultivated states and produced T-states with a record logical fidelity of 99.99%. Logical fidelity describes how reliably the encoded quantum information represents the intended state after the relevant error-protection procedures. The result is significant because it moves cultivation beyond theory and computer simulations to an experiment on a physical quantum processor.

That improvement comes with a clear efficiency trade-off. Google reports that only about 8% of attempts passed every check. In other words, the processor rejected more than nine out of every 10 tries to keep one clean state. This yield matters when assessing the method’s practical value: higher-quality outputs are useful, but a system must also account for the qubits, time, and discarded attempts needed to produce them. The demonstration shows that cultivation can work on hardware; it does not yet establish that it will be cheaper or faster than every alternative at larger scale.

The experiment is relevant to discussion of encryption because magic-state cultivation features in estimates of how much quantum hardware might be required for certain cryptographic attacks. In 2025, Google researcher Craig Gidney estimated that RSA-2048 could potentially be broken using fewer than a million noisy qubits. That was a 20-fold reduction from his 2019 estimate of 20 million. The cultivation technique is one of the ideas contributing to the newer estimate. The Willow demonstration lends experimental support to a component of the broader technical picture, but it does not demonstrate the full computation needed to factor an RSA-2048 key.

The distinction between a resource estimate and an actual attack is essential. The Willow experiment used at most 54 of the processor’s 105 qubits. According to the paper, factoring an RSA-2048 key would require magic states with errors around one in 10 billion. Google’s reported 99.99% fidelity result is not close to that requirement: the target states for the proposed code-breaking calculation would need to be roughly a million times cleaner than those achieved in this experiment. Moreover, estimates of the qubit resources required for a future attack do not mean that the necessary machine currently exists or that protected accounts are presently exposed to this demonstration.

Google has also discussed quantum computing’s implications for Bitcoin’s elliptic curve cryptography, the mathematical system that helps prevent someone from deriving a private key from a public address. In March, a Quantum AI whitepaper co-authored with Ethereum Foundation researcher Justin Drake and Stanford cryptographer Dan Boneh put the estimated cost of breaking that cryptography at fewer than 500,000 physical qubits. The company has also pointed to its Quantum Echoes experiment last October, which it described as the first verified quantum speedup on real hardware. These developments form part of a wider research effort, but they should be distinguished from the magic-state experiment: none of them means that Bitcoin wallets are being drained or that the present Willow demonstration can derive private keys.

The next step identified by Google is to run cultivation at a larger scale and improve the processor’s underlying hardware. The team projects that halving Willow’s hardware noise would reduce magic-state errors to about two in 100,000. Its projection is based on the claim that errors in the cultivation process shrink with the cube of the chip’s noise level. This is a forecast for what could follow from improved hardware, not a result already achieved in the reported experiment. Testing the method across larger systems will be important for determining whether the projected gains persist and whether the process can deliver enough states at a useful rate.

Overall, the finding is best understood as a meaningful but bounded engineering advance. It provides evidence that cultivation can reduce errors on a real chip and may help address the cost of a resource required by advanced quantum programs. The low pass rate, the limited scale of the experiment, and the much higher quality needed for cryptographic applications remain important constraints. Future progress will depend on combining reliable error correction, lower hardware noise, and resource-efficient production. The study narrows one part of the gap between quantum-computing proposals and working hardware, while leaving clear technical milestones ahead.

Key Insights Table