A Nature paper from Microsoft and Quantinuum reports that error corrected logical qubits, the unit a useful quantum algorithm would run on, suppressed faults at up to 800 times the rate of the noisy physical hardware beneath them on a trapped ion machine.
A useful quantum computer will not run on raw hardware. It will run on logical qubits, error-corrected units composed of many noisy physical qubits, where faults are detected and repaired on the fly so that the computation survives. Quantum error correction is the engineering discipline that makes that possible, and a peer-reviewed Nature paper from Microsoft and Quantinuum now supplies one of the cleanest published measurements of how much that correction can buy you on a real machine.
The result: across several non-trivial circuit classes, the logical-qubit error rate was 11 to 800 times lower than the physical-qubit baseline beneath it. In one worked example reported in the paper, a Bell-state preparation that failed 0.8% of the time on the physical hardware dropped to roughly 0.001% under the logical-qubit code. The 800x figure is not a single number; it is a range that depends on which class of circuit you measure. That distinction matters: a quantum error correction result that suppresses faults on a carefully chosen sub-circuit is not the same as crossing the fault-tolerant threshold, and the paper is careful to position the work as hardware-codesign data rather than a fault-tolerance announcement.
A logical qubit is an error-corrected, algorithm-usable unit built from many physical qubits. A physical qubit is a single noisy hardware element: in Quantinuum's case, a trapped ion shuttled around a Quantum Charge-Coupled Device (QCCD), a chip-scale electrode array that moves ions through zones where gates and measurements happen. Quantum error correction works by entangling several physical qubits into a code, then running a syndrome extraction step, a measurement that identifies which physical qubit probably failed, without collapsing the underlying quantum state. The result is fed forward, the suspected error is corrected, and the logical qubit continues. Microsoft's qubit virtualization runtime compiled the logical circuits; Quantinuum's trapped-ion QCCD executed them.
The two code constructions in the paper were chosen to fit QCCD connectivity. One is a 12-qubit code inspired by Knill's 1998 teleportation-based error correction, a construction that uses a pre-shared entangled ancilla to detect faults in one round. The other is a companion code optimized for hardware-aware syndrome extraction on the trapped-ion architecture. The framing fits a broader pattern in the field: error-correcting codes are increasingly being co-designed with the hardware that will run them, rather than picked from a generic library.
What the paper is not claiming is that useful quantum computing is now close. Suppressing errors on a small logical-qubit demonstration is a prerequisite for fault-tolerant machines, not the same thing as running a useful algorithm at scale. The fault-tolerance question is a different one, and it remains open on every hardware platform. The Nature result is also a study by the same team that designed both the code and the machine, which is normal for the field but worth flagging for readers weighing how to compare it with prior announcements.
For context, the trapped-ion, hardware-codesign path is not the only one in the race. Superconducting groups at IBM and Google have published their own logical-qubit data and surface-code roadmaps, and neutral-atom efforts at QuEra and Atom Computing are running similar experiments on a different modality. The Microsoft-Quantinuum result is a real data point in that competitive landscape, and the peer-reviewed venue means it can be compared on the same terms as the others rather than read as a press-release claim.
The watch item is whether the 11x-to-800x range holds up as the circuits get larger, and whether the hardware-codesign approach extends from small demonstrations to algorithm-scale runs. The Nature paper is now the benchmark other teams will measure against, and the underlying data and circuits are available from the authors on request.