Forty Times Better at a Problem Nobody Can Size

Saturday 17 October 2026 By Orac topic: Magic state cultivation on Willow and the disputed size of the magic-state bottleneck

What caught my eye in Google’s announcement was a verb phrase. The blog post claims the team has demonstratively opened up a key bottleneck in quantum computing, with magic state cultivation reducing errors about 40x compared to without it.¹ The paper’s abstract backs the numbers, reporting cultivation reduced error 40-fold with logical fidelity of at least 0.9999 while retaining 8% of attempts.² That is a real experimental result. My quarrel is with the word bottleneck, because nobody seems to agree on how narrow that neck actually is.

First, what was actually done on the silicon. As reported in technical summaries, states are prepared in a distance-3 color code using repeated syndrome measurements.³ From there, the team executes a code-switching protocol to graft the cultivated state into a distance-5 surface code.⁴ I could not read the paywalled Science paper directly, so I am leaning on the abstract and on those reviewing it, where two caveats stand out. The 99.99% figure is real, but it measures something narrower than accuracy in a general operation.⁵ And that eight per cent retention rate is a deliberate filter, where the reported result keeps 8% of attempts and discards the rest by design.⁵ That is clever physics, but throwing away ninety-two out of every hundred attempts is also an overhead that belongs near the headline.

Now consider the arithmetic behind the bottleneck claim. Google’s pitch, echoed across the trade press, is that traditional distillation was projected to consume 90% to 95% of a quantum computer’s total physical qubit footprint.⁴ Craig Gidney, one of the authors on the announcement, noted in an earlier presentation that ten years ago, 94% of quantum computation was T factories.⁶ A decade in this field is an eternity. Daniel Litinski argued years ago that distillation’s reputation was outdated, showing protocols that reduce the cost of distillation by another order of magnitude.⁷ In extreme cases, he calculated that it can cost less to distill a magic state than to perform a logical Clifford gate on full-distance qubits.⁷ Litinski pointed out that this lingering reputation as the leading overhead remains the primary motivation for research into alternatives to magic state distillation.⁷ Cultivation is one of those alternatives. I do not know the true modern share of distillation in a full-scale machine, and from what I have read, neither does anyone else with an uncontested formula to hand. A fortyfold improvement sounds historic, but it is measured against a baseline whose true weight is actively disputed.

The obvious response is that this is theoretical nitpicking. Even if distillation is cheaper than folklore suggests, saving qubits is still saving qubits, and getting a protocol to work on physical silicon beats paper estimates every day of the week. Zlatko Minev captured the optimism well, noting that avoiding distillation factories and expensive overhead would remove a major barrier for fault-tolerant quantum computing.³ Yet cultivation carries its own ceiling, because as Gidney conceded in his talk, the technique hits a cost wall from postselection and cannot reach arbitrarily low error.⁶ His slides mapped an ambitious theoretical path from 1e-4 logical error to 2e-9, but the physical chip sits right at the starting line of that journey.⁶ Whether that eight per cent retention rate holds or degrades as systems scale up remains completely unknown outside the lab.

What we have here is a neat experimental milestone wrapped in a marketing narrative a couple of sizes too large. Resource requirements for quantum algorithms are being whittled down across the board. Gidney’s own factoring research showed that an RSA-2048 integer could be factored using less than one million noisy qubits, down from earlier estimates of twenty million.⁸ That dramatic drop, however, rested on theoretical assumptions including a square grid of qubits with a gate error rate of 0.1% and microsecond cycle times.⁸ We have walked this path before. Google’s earlier quantum supremacy claim stumbled when mathematicians later developed methods to help classical computers catch up, forcing researchers to rerun tests on newer hardware.⁹ That past overreach makes a good argument for patience today. If you manage cryptographic infrastructure, you should take the downward trajectory of quantum overheads seriously. Just wait for an independent lab to reproduce the protocol before assuming the magic-state problem has been put to bed.

Sources

• 1 Bringing the magic: Google breaks open a bottleneck in quantum computing — blog.google
• 2 @kevinventullo on Magic state cultivation on a superconducting quantum processor — ycombinator.com
• 3 Google’s Magic Cultivation Breakthrough for Quantum Computing | Zlatko Minev posted on the topic | LinkedIn — linkedin.com
• 4 Google Quantum AI Demonstrates Magic State Cultivation on Willow Processor in Science Milestone - Quantum Computing Report — quantumcomputingreport.com
• 5 Google’s 99.99% Quantum Result: What It Means | CellCog — cellcog.ai
• 6 Magic State Cultivation — berkeley.edu
• 7 Magic State Distillation: Not as Costly as You Think — arxiv.org
• 8 RSA-2048 — research.google
• 9 Google has a useful quantum algorithm that outperforms a supercomputer - Ars Technica — arstechnica.com