On October 15, 2026, Google Quantum AI said its paper in Science shows magic state cultivation working on physical hardware. Sundar Pichai wrote that his team "has demonstrated 'magic state cultivation' on physical hardware for the first time," with a 40x error reduction on the Willow processor and a state-of-the-art result.
Google's own post says the result is 99.99% fidelity. This is a real step for fault-tolerant quantum computing. It is not a code-breaking moment, and some viral summaries are running ahead of the evidence.
TL;DR: the numbers and the caveats
| Question | Answer |
|---|---|
| What was demonstrated? | Magic state cultivation, first time on physical hardware |
| Chip | Willow, 105 qubits (per press summaries of the announcement) |
| Headline result | 40x error reduction versus not using cultivation |
| Fidelity | 99.99% |
| Catch | The preprint reports keeping about 8% of attempts, so the method filters aggressively |
| Venue | Science, per Google Quantum AI's post on X |
| Does it break Bitcoin? | No. That needs hundreds of thousands of qubits or more |
| Who proposed the method? | Craig Gidney, Noah Shutty and Cody Jones of Google Quantum AI, in a 2024 paper |
What is a magic state, in plain language?
Quantum computers make mistakes constantly, so they use error-correcting codes that spread one logical qubit across many physical qubits. Inside those codes, some gates are cheap and safe. Others, notably the T gate, are not. You cannot do a T gate directly on a protected qubit without breaking the protection.
The workaround is to prepare a special resource called a magic state ahead of time, then consume it to perform the gate. Without magic states, a quantum computer can only run problems a classical computer can simulate efficiently. With them, it is universal.
The problem is cost. The standard way to get a clean magic state is distillation: start with many noisy states and combine them to make a single cleaner one, over and over. Google's proposal, and press summaries of the new result, describe this as a bottleneck that can consume a very large share of a machine's resources. Treat the often-quoted "95 percent" figure as a rough social-media summary, not a number I could verify in the paper.
How does cultivation work?
The 2024 proposal, "Magic state cultivation: growing T states as cheap as CNOT gates," describes a different route. Instead of distilling many states, you grow one inside a small patch of the error-correcting code and keep checking it. If a check fails, you throw the attempt away and start over. If the checks pass, the state gets better as it grows.
Think of a filter. Cultivation is cheap per attempt, and it rejects bad runs early. The authors say it uses roughly the same number of physical gates as a lattice-surgery CNOT of equivalent reliability, which is the standard cheap operation. At a circuit noise of one part in a thousand, they projected logical error rates as low as 2 times 10 to the minus 9, and lower again at better hardware noise.
The new experiment, from a team including Emma Rosenfeld, Gabrielle Roberts, Nathan Lacroix and Alexandre Bourassa, tests this on a superconducting processor. Its reported result is a 40x error reduction and a fidelity of 0.9999, retaining about 8% of attempts.
What are people asking?
Is 99.99% fidelity enough to run useful algorithms?
Not by itself. Real algorithms need many billions of clean T gates, so the error rates must fall much further, and the machine must be large. The point of this result is that the cost per clean state is dropping, not that the finish line has arrived.
Why does the 8% number matter?
A 40x improvement is much less exciting if you throw away nearly every attempt, because you pay for the discarded runs in time and qubits. The right question is the overall cost per good magic state, compared with distillation. The preprint's authors frame cultivation as cheaper in total, but expect follow-up analysis from outside groups.
Does this threaten Bitcoin and TLS?
No, not now. Estimates for breaking widely used public-key cryptography need hundreds of thousands of high-quality qubits or more. Willow has 105. Still, it is a reason to keep migrating. Our coverage of Trump's quantum executive orders and post-quantum cryptography explains the policy clock, and some blockchains are already shipping post-quantum features, such as Algorand, per press summaries.
Is this an AI story?
Only indirectly. Quantum and AI share a hardware and funding landscape, and Stony Brook and Brookhaven's 21 km wireless quantum network shows the networking side. But the claim here is about error correction, so keep it separate from AI model progress.
Why did a New York mayor tweet about quantum the same day?
A separate announcement. New York City's mayor said the city negotiated board seats and a 15% stake of intellectual property from NYU's Quantum Fabrication effort. That deal is unrelated to Google's paper, but it shows quantum is now a municipal-budget topic.
Where does this fit on the road to a useful quantum computer?
Fault-tolerant quantum computing needs several pieces to work together, and each has been a separate research milestone.
- Good physical qubits. Google's Willow chip, announced in December 2024 with 105 qubits, was best known for showing that adding more qubits to a surface-code patch lowered the logical error rate, the "below threshold" result.
- A logical memory. Hold information for a long time inside an error-corrected block. Willow's earlier results addressed this.
- Logical gates. Perform operations between blocks, through techniques such as lattice surgery.
- Cheap non-Clifford gates. This is the magic state problem. Without it you do not get a universal machine.
- Scale. Millions of physical qubits for the hardest applications, with cooling, control wiring and decoding keeping pace.
Cultivation addresses step four. Showing it on hardware matters because proposals on paper often meet surprises in real devices: correlated noise, leakage, calibration drift and the classical electronics needed to decide in real time whether to keep or discard a run.
A worked example: why filtering can beat brute force
Imagine you need one good component and your factory makes a bad one most of the time. Two strategies exist. In distillation, you build many bad components and combine them in a complicated process to get one good one. In cultivation, you build one at a time and test it at each stage, throwing it away the moment a test fails.
The second approach wastes most attempts but each attempt is cheap and small. If only about 8% survive, you run roughly twelve attempts per accepted state, yet each attempt uses a small patch of qubits instead of a large distillation factory. Whether that wins depends on the cost per attempt, the time per attempt, and how many attempts can run in parallel on a chip. That is why the cost per accepted state, not the 40x headline alone, is the number to watch.
How should you read quantum headlines?
Quantum news is easy to over-read, so here is a short checklist that applies to this announcement and the next ones.
- Find the denominator. A 40x reduction compared with what? Here it is compared with not using cultivation, which is a fair internal baseline but not the same as beating the best distillation protocol.
- Look for postselection. Discarding most runs is legitimate in fault-tolerant protocols when the filter is part of the design, but you must count the cost.
- Separate qubit count from quality. One hundred and five good qubits are a different object than one hundred and five noisy ones.
- Check for peer review and replication. This one reached Science, which is a strong signal, and independent follow-ups will add more.
- Match the claim to the threat. Encryption risk depends on logical qubits and runtime for specific algorithms, not on a single component improving.
What should readers do with this?
- Developers: nothing urgent. If you build long-lived systems, start inventorying where you use RSA and elliptic-curve cryptography, because migration takes years.
- Learners: magic states are a good entry point to fault tolerance. Read the 2024 cultivation paper's introduction first.
- Skeptics: wait for independent replication and a full resource comparison against distillation.
Why the date and venue matter
Google's team posted the result the same day the paper appeared, which means outside experts have had hours, not weeks, to examine it. Expect commentary from other labs working on trapped ions, neutral atoms and other superconducting platforms, because each will want to know whether cultivation transfers to their hardware.
What we do not know yet
- The full cost per accepted state compared with distillation on real hardware.
- How the result scales to larger code distances.
- Whether other labs reproduce it on different qubit technologies.
Figures come from Google Quantum AI's announcement posts on October 15, 2026 and the public arXiv preprint and proposal. The Science paper text was not independently reviewed here.
Related reading
- Trump's quantum executive orders and post-quantum cryptography
- Stony Brook and Brookhaven 21 km wireless quantum network
- Tesla's flying Roadster: what was shown and what is unproven
- AI bubble reality check 2026
Sources: Magic state cultivation (arXiv 2409.17595), Magic state cultivation on a superconducting quantum processor (Google Research), and the Google Quantum AI announcement.