Google’s Willow Cuts Magic-State Errors 40-Fold

On October 18, Google Quantum AI reported that magic-state cultivation on Willow cut errors roughly 40-fold compared with uncultivated states. A processor can finish a spectacular short task while accumulating too many errors to run a long algorithm.

Speed established a checkpoint, not an endurance test

In 2019, Google said its Sycamore processor had completed a calculation in 3 minutes 20 seconds that would take contemporary supercomputers more than 10,000 years. The Sycamore claim gave the field a powerful demonstration target: show that quantum hardware can execute a task beyond practical classical simulation.

Sycamore did not measure how many dependable operations a machine could sustain. Google continued to pursue application-facing speed comparisons: in 2025, it said Willow ran its Quantum Echoes algorithm 13,000 times faster than supercomputers, positioning the work around drug and materials research. That result measures the speed of a particular algorithm; it does not establish practical value for either application.

Universal computation needs a supply of special states

Error correction spreads quantum information across physical qubits so that a logical qubit can survive errors in its components. A long, universal fault-tolerant algorithm needs more than protected storage. It must also perform operations that consume specially prepared logical resources called magic states, including the T states Google identifies as necessary for universal computation.

In its Willow magic-state-cultivation report, Google says it experimentally reduced errors in those states by about 40 times compared with uncultivated states. A discussion accompanying the announcement cites projections that T-state production could consume roughly 90% of a fault-tolerant computer’s physical qubits. That is a projected resource budget, not a measurement of Willow’s full-machine overhead.

For a sustained algorithm, engineers need the rate of usable T-state production at a stated logical error rate and physical-qubit cost. Google’s 40-fold result measures the quality of prepared states; it does not establish how many a working machine can supply.

Google’s reliability gains have yet to run together

Google’s 2024 Willow introduction paired 105 improved physical qubits with an error-correction milestone and a roadmap. In the same period, Google Quantum AI and DeepMind introduced AlphaQubit, a machine-learning decoder that they said surpassed existing methods for identifying and correcting quantum errors. Willow addresses physical hardware and encoded reliability; AlphaQubit addresses error identification; cultivation addresses a resource consumed by logical operations.

Google’s reports do not say that AlphaQubit decoded the cultivation experiment or show every reported improvement operating within one long algorithm. Those separate results leave open how the decoder and cultivated states perform together in a corrected circuit.

Rivals count different costs

Microsoft and Quantinuum reported more than 14,000 experiments without errors in April 2024. In September, they reported computation with error correction across 12 highly reliable logical qubits. The first result counts bounded experiments; the second tests protected operations. Neither specifies the duration or resource cost of a large universal algorithm.

QuEra made the conversion from physical inventory to intended logical output explicit. Its 2024 roadmap targeted 100 error-corrected logical qubits from 10,000 physical qubits by 2026—a planned ratio of 100 to one, not a measurement of delivered output. QuEra’s ratio counts physical qubits per planned logical qubit. The roughly 90% projection cited alongside Google’s result concerns the share of a fault-tolerant machine devoted to producing T states. The figures have different denominators; combining them would not yield a factory estimate.

IBM’s Starling roadmap targets a fault-tolerant system for 2029. Google also set a 2029 deadline for its post-quantum-cryptography migration, arguing that quantum frontiers may be closer than they appear. Security teams can rationally replace vulnerable cryptography before a machine capable of breaking it exists; migration itself takes preparation. The shared year says nothing about when either company will demonstrate the state-production rate a long corrected algorithm requires.

Willow’s roughly 40-fold error reduction could shrink a costly part of a future machine. The size of that saving remains open until Google measures how many usable states it can supply while the corrected circuit runs.