Atom Computing Demonstrates Quantum Error Correction With Toric Code, a First for Neutral Atoms

June 4, 2026 – Atom Computing has published the first demonstration of sustained quantum error correction using a toric code on a neutral-atom quantum computer. The preprint, posted to arXiv on June 4, reports up to 90 rounds of syndrome extraction on the company’s ytterbium-171 (¹⁷¹Yb) tweezer-array processor, with mid-circuit measurement and continuous reloading of lost atoms from a magneto-optical trap (MOT). The result makes Atom Computing the first neutral-atom platform to demonstrate many rounds of sustained quantum error correction (QEC) with true mid-circuit measurement.

The bottom line: this result validates neutral atoms as a viable platform for the kind of repeated, indefinite error correction that fault-tolerant quantum computing requires, and it does so using a code family that superconducting platforms physically cannot implement.

The team tested two twisted toric code variants: a smaller code with 16 data qubits and 16 ancilla qubits (code parameters [[16, 1, (4, 3)]]), and a larger code with 32 data qubits and 32 ancilla qubits ([[32, 1, (6, 4)]]). Comparing the two at up to eight rounds of syndrome extraction, the larger-distance code consistently produced a lower logical error rate: 0.56% per cycle versus 0.74% per cycle for the Z-type logical observable. At four cycles, this translated to an error suppression factor of Λ = 1.9 for the Z-type logical observable, meaning the larger code cut Z errors nearly in half relative to the smaller one. The X-type observable showed a smaller factor of Λ = 1.20, and the average across both observables was Λ = 1.30. This behavior, where adding more physical qubits actually reduces the logical error rate rather than increasing it, is the hallmark of below-threshold operation.

Atom Computing’s press release claims this places the company “among only two companies that have demonstrated many rounds of sustained quantum error correction,” with Google’s superconducting Willow processor being the other.

In a separate experiment, the team ran a distance-7 repetition code with an added 0.5-second idle period per cycle, stretching the total circuit duration to expose qubits to maximum environmental error. The logical memory persisted for over three minutes (a logical lifetime of 225 seconds), even though individual atoms survived in their traps for only about 10 seconds on average. The logical lifetime exceeded the physical atom lifetime by more than 20×.

The paper lists David Aasen from Microsoft Quantum as a co-author. Microsoft’s Dr. Matthias Troyer, Technical Fellow and Corporate Vice President at Microsoft Quantum, provided an endorsement in the press release, and Dr. Scott Aaronson of the University of Texas at Austin called the result “exciting progress toward fault-tolerance for neutral-atom quantum computers.”

The announcement comes two weeks after Atom Computing signed a Letter of Intent with the U.S. Department of Commerce for $100 million in funding as part of a $2 billion federal initiative spanning nine quantum companies. Atom Computing is also participating in Stage B of DARPA’s Quantum Benchmarking Initiative and is installing its Magne system, described as the world’s first commercial quantum computer with logical qubits, at QuNorth in the Nordic region in partnership with Microsoft.

My Analysis

Why the Toric Code Matters

Most QEC demonstrations to date have used the surface code, a planar variant of the toric code that works with nearest-neighbor connectivity. Superconducting systems are largely constrained to this topology because their qubits are fixed on a chip in a 2D grid. The toric code, by contrast, requires non-local connectivity: qubits on one edge of the lattice must interact with qubits on the opposite edge, as if the surface were wrapped into a torus.

Neutral atoms do this natively. Atom Computing’s zone-based architecture uses optical tweezers to shuttle atoms between spatially separated register, interaction, measurement, storage, and loading zones. Any atom can interact with any other atom, enabling codes that demand non-local connectivity without the routing overhead that would cripple a fixed-layout chip.

Why does this matter beyond a technical curiosity? Non-local codes, including the toric code and the broader family of quantum low-density parity-check (qLDPC) codes, offer higher encoding rates than surface codes. They pack more logical information per physical qubit. For fault-tolerant quantum computing at scale, this translates to fewer total physical qubits required. The Pinnacle Architecture paper from earlier this year estimated that qLDPC codes could reduce the qubit cost of breaking RSA-2048 to around 100,000 physical qubits, compared to millions with surface codes. Toric codes sit at the simpler end of this code family, but the principle is the same: flexible connectivity opens the door to more efficient error correction.

Atom Computing’s demonstration is the first experimental evidence that this theoretical advantage can translate into real hardware. That is a concrete differentiator for the neutral-atom modality.

What the Numbers Actually Show

The paper is careful about its claims, and readers should be too. The observed error suppression factor of Λ = 1.9 at four syndrome extraction cycles is encouraging, but the authors themselves note several caveats.

First, the per-cycle logical error rate for the larger code doesn’t follow a simple exponential decay model well, likely due to boundary effects and slow atom heating. The error suppression factor decreases with more rounds, and by the time the system operates with continuous atom reloading (beyond 10 cycles), the two code sizes produce nearly identical logical error rates (0.63% versus 0.64% per cycle for the Z-type observable). The paper states this directly: over arbitrary numbers of cycles, “the logical error rates do not increase with larger code distance.” That is a weaker statement than “the logical error rates decrease with larger code distance.” It means the system is operating near the threshold rather than clearly below it at scale.

Second, the error budget is dominated by two-qubit gate loss (atoms ejected during CZ gate operations) and various loss channels collectively. The stochastic two-qubit gate error is 0.43%, with an additional 0.96% per-gate atom loss rate. Loss is the dominant error source in neutral-atom systems, and while Atom Computing’s erasure conversion and loss-tolerant decoding handle it gracefully, it remains the primary bottleneck. The paper’s own simulations underestimate the observed logical error rates, suggesting unaccounted-for coherent errors and correlated noise.

Third, the comparison between Atom Computing’s result and Google’s Willow below-threshold demonstration requires context. Google achieved Λ = 2.14 with a distance-7 surface code at 0.143% logical error per cycle over millions of rounds. Atom Computing’s best comparable numbers are smaller codes (distance-4 and distance-6 for the Z-type observable), fewer rounds in the sub-threshold regime, and higher per-cycle error rates. The systems are at different maturity levels, and direct comparison across modalities is always tricky, but in raw error correction performance, the superconducting platform still leads.

The Real Engineering Achievement: Continuous Operation

Where this paper earns its significance is in the engineering of continuous operation. Neutral-atom quantum computing has always faced a fundamental challenge that superconducting systems do not: atoms leave. They get heated during gate operations, knocked out of traps during measurement, or simply boil off over time. Every lost atom is a lost qubit, and in an error correction circuit that needs stable participation from every qubit over many rounds, atom loss is a showstopper.

Atom Computing solved this by building a full atom-replenishment pipeline into the processor. Their zone-based architecture maintains a storage zone of spare atoms adjacent to the computation. Lost ancilla qubits are detected during mid-circuit measurement and replaced from this reserve. When the storage zone depletes (typically after about 15 cycles), fresh atoms are loaded from a magneto-optical trap located 30 cm below the tweezer array, transported via a moving optical lattice, and inserted into the storage zone without disrupting coherence in the computational qubits.

The result: 90 consecutive rounds of syndrome extraction at a roughly constant detection probability, with no fundamental limit on how many more rounds could follow. The detection probability spiked slightly during reloading cycles (a few percent higher for two to three cycles after each reload) but returned to baseline quickly.

This continuous-operation capability goes beyond an incremental improvement. It addresses one of the key engineering barriers in my CRQC Quantum Capability Framework, specifically Capability D.3: Continuous Operation. Any quantum computer that aims to run the deep circuits required for cryptographic relevance or practical chemistry simulation must maintain stable operation over hours, not microseconds. Atom Computing’s MOT-reloading architecture is the first neutral-atom demonstration that continuous operation is achievable in principle.

The role-swapping protocol addresses a second, subtler problem. By compiling periodic swap operations between data and ancilla qubits into the syndrome extraction circuit, every atom gets measured and reset after just two rounds. This prevents the accumulation of heating-induced errors on any single qubit and converts atom-loss events into detectable erasure errors, which are easier to decode than undetected Pauli errors.

Placing This in the Neutral-Atom Competitive Picture

This result changes the competitive picture within the neutral-atom modality. Until now, the Harvard/MIT/QuEra collaboration has dominated the neutral-atom QEC narrative. Their November 2025 Nature paper demonstrated a 448-atom fault-tolerant architecture with below-threshold surface code performance (Λ = 2.14), 96 logical qubits using high-rate codes, and transversal logical gates. It was, and remains, the most complete demonstration of fault-tolerant neutral-atom computing.

Atom Computing’s paper is more narrowly focused but makes three distinct contributions. First, true mid-circuit measurement and qubit replacement: the Harvard/QuEra surface code memory experiment retained all ancilla qubits and measured them at the end rather than performing round-by-round syndrome extraction with measurement and reset. Atom Computing’s experiment performs actual mid-circuit measurement, ancilla reset, and qubit replacement every round. Second, continuous operation via MOT reloading: although the Harvard/MIT collaboration demonstrated a 3,000-qubit system running for over two hours using a separate protocol, Atom Computing integrated atom reloading directly into the error correction circuit. Third, a non-local code: the toric code is inaccessible to planar architectures, and demonstrating it is a proof point for the code families that matter most for resource-efficient fault tolerance.

Atom Computing uses ¹⁷¹Yb atoms (nuclear-spin qubits with ~7-second T₂ coherence and atom lifetimes of 10–16 seconds in this system), while the Harvard/QuEra group uses ⁸⁷Rb (alkali atoms with faster gates but shorter coherence). These are substantively different hardware approaches within the same modality, and the fact that both are producing competitive QEC results strengthens the case for neutral-atom quantum computing as a whole.

Caveats and Open Questions

Several aspects of this result need further scrutiny.

The paper notes “significant variation” between datasets in detector frequency and loss, with possible explanations including drifting alignment between the imaging cavity and tweezer arrays. This suggests the system is sensitive to calibration drift on the timescale of hours, a challenge that will intensify as the system scales.

The error suppression factor of Λ = 1.9 was measured at the favorable four-cycle point. The paper’s own analysis shows this factor decreasing toward 1.0 as rounds increase, and the authors caution against extrapolating per-cycle error rates from limited data. They also note that per-operation error rates may grow with system size, a concern that applies more generally to any platform using dynamic atom rearrangement where more atoms means longer transport times and more opportunities for heating.

The “only two companies” claim in the press release requires qualification. Quantinuum’s trapped-ion system has demonstrated computing with many encoded logical qubits beyond break-even, and the Harvard/QuEra collaboration has shown below-threshold surface code performance. The specific claim hinges on the definition of “many rounds of sustained quantum error correction,” and Atom Computing defines this as repeated syndrome extraction with true mid-circuit measurement. That is a legitimate engineering distinction, but it is a narrower claim than the press release implies.

The paper is a preprint, not yet peer-reviewed. The core results appear solid and the authors are transparent about limitations, but the standard caveat applies.

What This Means for the Path to CRQC

This result does not change the Q-Day timeline. The system has 128 qubits in its register zone and runs codes at distance 4 to 6. A cryptographically relevant quantum computer would need thousands of high-quality logical qubits, each encoded in hundreds or thousands of physical qubits. The gap between here and there remains enormous.

What it does is add a data point to the engineering trajectory. In my CRQC Quantum Capability Framework, I track ten capabilities that a CRQC must achieve. Atom Computing’s result touches four of them: B.1 (quantum error correction), B.3 (below-threshold operation), B.2 (syndrome extraction), and D.3 (continuous operation). For B.1 and B.2, the demonstration is straightforward. For B.3, the sub-threshold evidence is suggestive but not definitive at the cycle counts tested. For D.3, the continuous atom reloading is a first for the neutral-atom modality.

I have updated my CRQC Scorecard to reflect this result, specifically in the neutral-atom QOT and continuous operation assessments, and to note that the qLDPC connectivity advantage now has its first experimental validation.

The broader significance is strategic. The quantum computing industry is converging on a consensus that non-local codes (qLDPC and related families) will be necessary for practical fault-tolerant systems, because surface codes require too many physical qubits per logical qubit. If that consensus holds, then platforms with native all-to-all connectivity have a structural advantage. Atom Computing’s toric code demonstration is the first experimental validation of that structural advantage at the level of actual error correction, going beyond gate-level benchmarks and theoretical arguments.

For organizations tracking quantum threats to cryptography: nothing here changes the urgency calculation. The deadlines are already set by regulators, insurers, and standards bodies, and they do not wait for any single hardware result. But for those of us watching the engineering trajectory, this is a meaningful data point from a platform that continues to demonstrate rapid progress.