Google Quantum AI said on 15 October 2026 that it has shown “magic state cultivation” working in practice on its Willow processor, reducing errors about 40x compared to without it, in work published the same day in the journal Science.
Google calls magic states one of quantum computing’s most severe resource bottlenecks. Its post says some estimates have shown that up to 95% of a quantum computer’s capacity might need to go into making them.
Business Pill · TOTAL COST OF OWNERSHIP
A one-minute explainer of total cost of ownership: why the full cost of running a system can far exceed its headline price. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the result targets cost, not raw qubit count: Google says magic states could take up to 95% of a quantum computer’s capacity, and cultivation cuts their errors about 40x.
Update, 16 Oct 2026: a second Google result on error-correction overhead
A related Google paper is now getting attention. In “Low-Overhead Quantum Error Correction with Boundary-Connected Planar Modules” (arXiv, posted 2 October 2026), researchers including Oscar Higgott, Craig Gidney and Johannes Bausch propose wiring flat chip modules together with sparse boundary connections to build “modular hyperbolic surface and color codes.” Their circuit-level simulations, the abstract says, “show that this modular memory design can reduce the physical qubit overhead by tenfold or more compared to the surface code,” and projecting to larger systems they find codes with rate k/n = 1/16 and distance d ≥ 22, “implying an over 30× overhead reduction relative to surface codes.”
These are simulation results and projections, not a hardware demonstration on Willow; Quantum Computing Report reported the work on 15 October. As we read it, the two results attack the same bottleneck from different ends: the magic-state work below cuts the cost of logic, the modular codes the cost of memory.
According to Google’s post, its protocol achieved a 40-fold reduction in errors relative to un-cultivated states on the Willow processor, and produced T-states with what it calls a record logical fidelity of 99.99%.
Google describes the result as “the first demonstration of a practical solution to one of quantum computing’s most severe resource bottlenecks.”
The post says the team also switched states between two kinds of error-correcting code, grafting them from a distance-3 color code to a distance-5 surface code, and that the purified states survived the transfer inside the hardware.
Google says magic state cultivation cut errors about 40x on its Willow processor and produced T-states with 99.99% logical fidelity; the team’s preprint says about 8% of attempts are retained (Google blog post, 15 October 2026; arXiv 2512.13908).
What Was Posted on X
Jensen Huang says quantum computers could help train AI by simulating biology and chemistry and generating synthetic training data.
“Quantum computer is not the next big thing, but it is a big thing.”
Wall St Engine (@wallstengine) posted on 25 Oct that Nvidia CEO Jensen Huang says quantum computers could help train AI by simulating biology and chemistry and generating synthetic training data: “Quantum computer is not the next big thing, but it is a big thing.” The post does not say when or where Huang spoke.
Why Magic States Are the Bottleneck
Google’s post explains that “universal” quantum computing needs a specific set of operations, and that these can be built by first creating so-called magic states.
The catch, the post says, is cost: creating high-quality, error-corrected magic states is resource-heavy, and once a magic state is used it cannot be reused, which makes it a scarce resource.
Cultivation works like a filter, according to the post: a noisy state is injected into a small region of the chip, and a series of coordinated measurements checks for errors and removes flawed attempts.
Google says the idea was developed in 2024 by its researchers Craig Gidney, Noah Shutty and Cody Jones, and that it could substantially reduce the size of the quantum computer needed for practical tasks if it worked on real hardware.
What the Paper Adds
The team’s preprint, posted to arXiv in December 2025, reports a state fidelity of 0.9999(1) while “retaining 8% of attempts”, meaning the filter discards most tries to keep the clean ones.
The paper says the work ran on a Willow processor that contains 105 tunable transmon qubits, and that adding cultivation improved the logical error probability by about a factor of 40.
It describes the fidelity as similar to recent demonstrations in trapped ions, while retaining the faster speeds of superconducting circuits.
Its conclusion calls magic state cultivation “a low-cost technique for generating high-fidelity magic states.”
What Google Says Is Still Missing
Google’s post says it still needs to demonstrate magic state cultivation at a larger scale to make it realistic for practical quantum computing, and that plenty of technical challenges remain.
The paper lists future directions including real-time implementation of T gates and grafting into larger codes to support high state fidelities.
Separately, on 14 October Nvidia named Cambridge quantum firm Riverlane among five UK startups it plans to invest in, as we reported.
The Structural Read
As we read it, the bottleneck is overhead: Google’s post says some estimates put up to 95% of a quantum computer’s capacity into making magic states, and that each one can be used only once.
As we read it, cultivation trades attempts for quality: the preprint reports a state fidelity of 0.9999(1) while retaining 8% of attempts, so most tries are discarded to keep the clean ones.
As we read it, this is a step, not a product: Google says it still needs to demonstrate cultivation at a larger scale, and the preprint lists grafting into larger codes among its future directions.
Google Quantum AI, in its 15 October 2026 blog post
“It demonstrates, on real hardware, that we can efficiently create these crucial magic states.”
Three Implications
MACHINE SIZEGoogle says cultivation could substantially reduce the size of the quantum computer needed for practical tasks, if it works at scale.
SUPERCONDUCTING CHIPS The preprint says the fidelity is similar to recent trapped-ion demonstrations while retaining the faster speeds of superconducting circuits.
THE NEXT TESTGoogle says plenty of technical challenges remain, and the preprint names real-time T gates and larger codes as next steps.
The framework argues: “It is a layered industrial stack — and at any given moment, the binding constraint sits at a different layer than it did six months ago.”
As we read it, quantum computing shows the same pattern inside a single machine: Google names magic-state production as one of its most severe resource bottlenecks, and the work attacks that layer rather than qubit count. The framework is about the AI stack and does not discuss quantum computing.
Google’s post says some estimates have shown that up to 95% of a quantum computer’s capacity might need to be dedicated to making magic states, which cannot be reused (Google blog post, 15 October 2026).
What Is Not Established
Neither the post nor the preprint gives a date for a quantum computer that uses cultivated magic states for practical work.
We read Google’s post and the arXiv preprint in full; the Science version did not load for us, so differences between the preprint and the published paper are not covered here.
Business Engineer Framework
The Nine Layers of AI
A Business Engineer framework on where the binding constraint sits in the AI stack.
Google says magic state cultivation cut errors about 40x on its Willow processor and produced T-states with 99.99% logical fidelity, in work published in Science on 15 October 2026.
The preprint shows the method keeps about 8% of attempts, and Google says it still has to show the technique at a larger scale.
A note on sourcing. We read Google’s 15 October 2026 blog post and the team’s arXiv preprint 2512.13908 (posted in December 2025) in full. The 40x figure, the 99.99% fidelity, the 95% estimate and the scale caveat come from Google’s post; the 0.9999(1) fidelity, the 8% retention rate, the 105-qubit processor and the trapped-ion comparison come from the preprint. The Science version did not load for us. The Riverlane context comes from our own earlier reporting.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
