What was shown
Quantum bits make mistakes constantly. Error correction spreads a piece of information across many physical qubits and keeps checking them against each other, much like a team of proofreaders who compare notes after every page. More proofreaders only help if they catch more slips than they add; with careless ones, a bigger team does worse.
The researchers ran this on a real 72-qubit superconducting device. They encoded a protected unit of information, called a logical qubit, in 49 physical qubits using the surface code, a layout in which qubits sit on a grid and check their neighbours. They compared it with smaller 17-qubit versions of the same code on the same device.
The larger code made slightly fewer errors per round of checking. The gain is small, but it points the right way: the bigger code did better, not worse. The measurement covers only these code sizes.
“We obtain a logical error per cycle ε5=(2.914±0.016) % ( 1σ statistical and fit uncertainty) for the distance-5 code, compared to an average of ε3=(3.028±0.023) % for the subset distance-3 codes, a relative error reduction of about 4%.”
— Section V, Decoding and logical error probabilities, in: Suppressing quantum errors by scaling a surface code logical qubit1
- How it was done: run on a physical quantum processor
- Qubits: A 72-qubit superconducting device; the larger code used 49 of its qubits, the smaller codes 17 each.
- Error correction: Error correction ran on real hardware, with no runs discarded for unusual disturbances, so the result includes realistic noise.
- Compared against classical computing: no comparison stated in the paper
Who it affects
This is groundwork for those who build quantum computers and for those who plan to use them for long, reliable calculations. For a company in any industry today, nothing changes: the device runs no useful task, and no business problem is solved faster or cheaper. What changes is the evidence base for planning. The scaling behaviour on which estimates of future machine size are built now has a first data point from real hardware.
Quantum computing vendors and their customers: McKinsey & Company estimates $43B to $71B by 2035 — quantum computing market size (vendor side) (Quantum Technology Monitor, April 2026, p. 8).
This figure is an estimate, not a measurement: McKinsey & Company is a consultancy with its own quantum advisory business. It covers a whole use case, not the result of this paper.
What is still missing
The device is not yet below threshold, the point at which adding qubits reliably lowers errors. The paper estimates that component performance must improve by at least 20% to get there, and significantly beyond that for practical scaling. The measured gain is small: 2.914% against 3.028% error per cycle, a difference of only about 4%. Correlated errors, such as high-energy impacts, must also be mitigated before larger codes can deliver the low logical error rates that computation needs. Useful error rates with manageable resources will require a suppression factor much larger than the break-even value.
Who is working on it
The paper (July 2022) lists its authors' affiliations as: Google Quantum AI; Google Research; Department of Physics, Columbia University, New York, NY; Department of Electrical and Computer Engineering, University of Massachusetts, Amherst, MA; and 8 further institutions. That is who signed this work — it is not a ranking of who is ahead.
This strengthens superconducting transmon qubits, a kind of superconducting qubit, running surface code error correction as a path to scalable logical qubits. It is the first hardware evidence that growing the code lowers the error per cycle, though only narrowly and only near threshold.
How this brief was made
This brief was written by software, not by a person: a language model read the full text of the paper, and code then checked every quotation, number and name in this brief against that text before it was published. The rules are set out under Method. Responsible for the rules and for this page: Georg Singer. The paper is peer-reviewed, published in Nature 614 (2023) 678-681. Georg Singer holds no shares in any company or institution named in this brief (as of October 2026).