A Smaller Quantum Computer Could One Day Do Useful Work, a Google Test Suggests

For about 20 years, quantum computer designers have known where the bill would come due. To run the programs that make these machines worth building, a computer needs a steady supply of special prepared states, nicknamed magic states, and making clean ones is expensive. Google’s own team writes that some estimates have shown up to 95% of a quantum computer’s capacity might go to making them, leaving little room for the actual computing. Say that out loud and it sounds absurd.

On October 15, 2026, Google Quantum AI reported in Science that a cheaper recipe, called magic state cultivation, works on a real chip. On its Willow processor, cultivation cut the error in these states by a factor of 40, to a fidelity of at least 99.99%. If that holds up at scale, the team says, it could substantially reduce the size of the quantum computer needed for practical tasks.

What is a magic state, and why does it run out?

Error-corrected quantum computers handle some operations easily. A few essential ones are much harder, and the standard way to run them is to feed in a magic state, which gets used up in the process. So a working machine must keep making fresh, very clean ones. The older method for cleaning them up, called distillation, is the costly part.

How does cultivation work?

Craig Gidney, Noah Shutty and Cody Jones of Google developed cultivation in 2024, refining an idea from earlier research. The team’s blog post compares it to a filter: inject a noisy state into a small patch of the chip, run a careful series of checks, and throw away any attempt that shows an error. What survives is much purer.

That filter is strict. Only about 8% of attempts passed every check in the headline result. The appeal is that each attempt is small and quick, so tossing most of them can still cost less than the old approach.

Where did machine learning come in?

According to the supplementary material of the preprint version, the team let a reinforcement learning system tune the chip’s control settings for about four hours just before collecting the main data, aiming to lower the rate of detected errors. It says this improved both the logical error and the share of attempts kept by about an order of magnitude, and was crucial for the reported results. Google described the method in an earlier paper; here it was the last step of calibration in a much larger engineering effort.

Was it smooth sailing?

Not at all, and this is the most human part of the story. Emma Rosenfeld, the paper’s first author, who ran the experiments, kept seeing data that seemed to defy the laws of physics. After months of checking, she traced it to a flawed assumption in the team’s usual benchmark, which slow drift in lab electronics and temperature had quietly undermined, according to Google’s blog post. The fix was a new way to measure the result that uses cultivation to check itself.

The team also moved a cultivated state from a small color code into a larger surface code, a more established kind of error-correcting code, showing the cleaned state can survive the handoff. In a storage test afterward, though, the preprint’s supplement reports an error rate about seven times higher than simulations predicted, which the authors say is likely due mainly to qubits leaking out of their working states.

Who could this help?

Nobody outside a lab will run a program on this tomorrow. But the paper notes that very clean magic states are what some quantum simulations and optimization problems will require, roughly 100 to 10,000 times cleaner than the ones shown here. A cheaper supply means the machine those scientists are waiting for could be smaller, and arrive sooner, than one built around distillation.

The result was first posted as a preprint in December 2025 and has now passed peer review. All but one of its nearly 300 authors work at Google, which builds the chip. Google says the next job is showing cultivation at a larger scale. For a field that has budgeted around this cost for two decades, that is welcome news.