Neutral-atom quantum computers: computers built from single atoms

A neutral-atom quantum computer uses single atoms as its basic units of information. Each atom is held in place by a tightly focused laser beam, called an optical tweezer, and lasers are used to link neighbouring atoms together. The approach matters because it can hold very large numbers of atoms and rearrange them, which is useful if quantum computers are ever to become large enough for real calculations. Research groups have published a run of results since 2023, but none of the sources reviewed here reports a large, error-corrected calculation of practical use.

What a qubit is

A normal computer bit is either 0 or 1. A quantum bit, or qubit, can hold a blend of both states at once, which is known as a superposition. The Wikipedia article on qubits describes this as quantum mechanics allowing a qubit to be in a coherent superposition of multiple states at the same time.

In a neutral-atom machine, the qubit is stored in the internal energy states of an atom. Two of those states act as 0 and 1. Lasers can change the state, and they can also excite an atom into a very high-energy state, known as a Rydberg state, in which it interacts strongly with nearby atoms. That interaction is what links two qubits together so that their states depend on each other, which is needed for a computer to do useful work.

What has been demonstrated

The results below are research experiments. Each is listed with its conditions, because the numbers do not mean the same thing across experiments. Unless marked as a preprint, each has been published in the journal Nature.

• Two-qubit gates at 99.5%. A 2023 paper by Evered and colleagues (Nature 622, 2023) reports two-qubit entangling gates with 99.5% fidelity, performed on up to 60 atoms in parallel. The authors state that this surpasses the surface-code threshold for error correction (arXiv:2304.05420). The threshold is the level of error below which adding more physical qubits to a code reduces errors rather than increasing them. Their claim is about a benchmark gate, not a full calculation.
• Ytterbium-171 qubits. A 2023 paper by Ma and colleagues (Nature 622, 2023) uses the nuclear spin of a long-lived metastable state in ytterbium-171. It reports single-qubit gate fidelity of 0.9990(1) and two-qubit gate fidelity of 0.980(1). It also reports fast mid-circuit detection that turns many errors into errors whose location is known, with an induced error probability below 10 to the power of minus five on qubits still in the computational space during detection (arXiv:2305.05493). The 98.0% figure comes from a different experiment from the 99.5% figure, so the two should not be compared directly.
• A programmable logical processor. A paper by Bluvstein and colleagues (Nature 626, published online in December 2023) describes a processor using up to 280 physical qubits, running sampling circuits on up to 48 logical qubits. A logical qubit is one qubit encoded across several physical ones so that errors can be detected or corrected. The authors report that the encoding improved algorithmic performance with error detection. That is not the same as full error correction (arXiv:2312.03982).
• 24 entangled logical qubits (preprint). A 2024 preprint by Reichardt and colleagues reports a processor with 256 qubits, each an individual ytterbium atom. It reports 24 logical qubits entangled in 48 atoms, while catching errors and correcting for, on average, 1.8 lost atoms. It also reports a Bernstein-Vazirani algorithm on up to 28 logical qubits encoded in 112 atoms, with better-than-physical error rates (arXiv:2411.11822). The arXiv record lists no journal version, so we treat it as not yet peer reviewed.
• Over 3,000 atoms for more than two hours. A 2025 paper by Chiu and colleagues (Nature 646, 2025) reports an array of more than 3,000 atoms kept assembled for more than two hours. Atoms were reloaded at 300,000 atoms per second, giving more than 30,000 initialised qubits per second. The authors state that moving atoms into the tweezers did not affect the coherence of qubits stored nearby (arXiv:2506.20660). The array was kept full by continuous reloading, so this is a refilled array rather than one set of atoms held static for two hours.
• Repeated error correction on up to 448 atoms. A 2025 paper by Bluvstein and colleagues (Nature 649, 2025) reports reconfigurable arrays of up to 448 atoms. In a four-round characterisation circuit using surface codes, the authors report performance 2.14 times below the error-correction threshold, with an uncertainty of 0.13. The result relies on detecting atom loss and on machine-learning decoding (arXiv:2506.20661). The circuit is short, so this is a test of error correction rather than a computation.

What is reported but not yet checked

A company blog post from Microsoft, dated 19 November 2024, describes the 24 entangled logical qubits as the highest number of entangled logical qubits on record. It also says Atom Computing had announced a 99.6% two-qubit gate fidelity, calling it the highest fidelity of neutral-atom qubits in a commercial system (Microsoft Azure Quantum blog). The post links to technical papers through short links that no longer reach them (checked 23 October 2026). The 24-logical-qubit figure matches the 2024 preprint above, which is not yet peer reviewed. We have not found a source for the 99.6% figure. The "on record" claim depends on how entanglement is counted: the 2023 Bluvstein paper reports sampling circuits with up to 48 logical qubits entangled, using a different code and test. Treat both company figures as claims made in November 2024.

Limits and open problems

• Error rates. The fidelity figures above are for individual gates or short benchmark circuits. A full algorithm strings many gates together, and small per-gate errors add up. Many logical-qubit results so far use error detection or short circuits, not full error correction on long computations.
• Atom loss. Atoms can escape their traps during a run. The 24-logical-qubit preprint corrected for 1.8 lost atoms on average, and the 3,000-atom result depended on continuous reloading. Each loss has to be detected and handled without spoiling the calculation.
• Speed. The 2023 parallel-gate paper uses fast gates based on optimal control, and the 2025 Bluvstein paper reports raising experimental cycle rates by two orders of magnitude by reusing qubits mid-circuit. The sources reviewed do not give a speed figure for a full-scale machine, so it is not clear yet how long a large calculation would take.
• Control at scale. Thousands of tweezers, lasers and atoms all need to be steered and read out together. Keeping that stable for long runs is an engineering problem in its own right, and the sources reviewed do not show it solved at the size a useful machine would need.
• From demonstration to computation. The largest results in the sources reviewed are sampling circuits, short characterisation runs and test algorithms. None of them reports a large, error-corrected calculation of the kind a chemist or engineer would need. This is a fast-moving field, so this may change.

Speculative uses

The 2023 parallel-gate paper presents its results as groundwork for large-scale quantum algorithms, error-corrected circuits and digital simulations. The authors of the 3,000-atom paper also point to atomic clocks and sensors as possible uses of continuously operated atom arrays. These are hopes, not results, and none of the sources above shows them working.

What to watch next

• Logical error rates that keep falling as arrays grow, measured against a benchmark agreed in advance.
• A peer-reviewed version of the 2024 preprint, and a published source for the 99.6% company figure.
• Runs that keep atoms working for longer with fewer losses, and with fast, reliable readout.
• Results on error-corrected circuits large enough to run a real calculation, rather than sampling or test algorithms.