hardware advanced · 19 min read · April 29, 2026

Neutral Atom Quantum Computing: The Platform That Closed the Gap in Three Years

Neutral-atom quantum computers — Rydberg arrays of laser-trapped atoms — went from research curiosity to flagship-tier platform between 2022 and 2025. Atom Computing crossed 1,000 qubits in 2023; QuEra and Pasqal demonstrated logical qubits in 2024-2025; the 2025 Sales Rodriguez logical magic-state distillation experiment was on neutral atoms. This tutorial covers the optical-tweezer / Rydberg architecture, current 2026 numbers, and why neutral atoms are the most-improved platform of recent quantum-computing history.

Prerequisites: Tutorial 34: Trapped Ion Quantum Computing

In 2022, neutral-atom quantum computing was a research curiosity. By 2025, it was the platform behind the first peer-reviewed demonstration of logical magic-state distillation (Sales Rodriguez et al. 2025, on the QuEra/Harvard/MIT system, tutorial 24). Atom Computing crossed 1,000 atoms in 2023 and 10,000 in 2025. Pasqal demonstrated multi-thousand-qubit analog simulation. QuEra and Atom Computing have moved from research-only to commercial cloud access.

Three years from “interesting” to “flagship” is not normal in this field. Neutral-atom platforms are the most-improved hardware platform in quantum-computing history. The improvement is real and is driven by a combination of physics that turned out to be unusually scalable: optical tweezers to trap atoms with arbitrary 2D and 3D geometries, alkaline-earth species with hyperfine clock states giving long coherence, and Rydberg interactions that turn on and off with picosecond-scale laser pulses.

This tutorial covers the platform end-to-end: the optical-tweezer architecture, Rydberg gate physics, the alkaline-earth qubit revolution, current 2026 numbers, and an honest take on what the platform’s remaining bottlenecks are.

The qubit: a neutral atom in an optical tweezer

A neutral atom — typically Rubidium-87 ($^{87}\mathrm{Rb}$), Strontium-87/88 ($^{87}\mathrm{Sr}$/$^{88}\mathrm{Sr}$), or Ytterbium-171 ($^{171}\mathrm{Yb}$) — is held in a tightly focused laser beam called an optical tweezer. The tweezer’s intensity gradient produces a small confining potential that holds the atom against gravity and ambient motion.

The qubit is encoded in two atomic energy levels:

• Hyperfine qubit (Rb): two ground-state hyperfine levels, $\sim$ GHz separation. Long coherence ($T_2 \sim$ seconds), microwave-driven.
• Alkaline-earth clock qubit (Sr, Yb): the metastable optical clock transition $|^1S_0\rangle \to |^3P_0\rangle$, $\sim$ 100 THz separation. Even longer coherence ($T_2 \sim$ tens of seconds), laser-driven. Used in the most modern Atom Computing and QuEra systems.

The defining feature of neutral-atom systems compared to ions: the atoms are not charged. They don’t repel each other strongly, so you can pack them densely. They don’t need elaborate Paul-trap electrodes; you just need lasers. And critically, you can move them by moving the lasers — much faster and with finer geometric control than the QCCD shuttling of ions.

The array: arbitrary 2D and 3D geometries

A neutral-atom array is built by:

1. Generate many tweezers in parallel. A spatial light modulator (SLM) or acousto-optic deflector (AOD) takes one laser beam and splits it into hundreds or thousands of independently positioned tweezers.
2. Load atoms into tweezers. Cool a cloud of atoms to ~10 µK; the tweezers stochastically capture atoms one at a time with ~50% probability per tweezer per loading attempt.
3. Detect which tweezers are filled. Image the atoms with a camera; each filled tweezer shows up as a bright spot.
4. Rearrange. Use AOD-controlled tweezers to physically move atoms from filled tweezers to empty ones, producing a perfect target geometry. The rearrangement takes 10-100 ms and is done before computation starts.

This rearrangement step is one of the platform’s biggest advantages: the array geometry is arbitrary and dynamic. You can configure a 2D rectangular grid for surface-code error correction, a 1D chain for chemistry simulation, a Kagome lattice for condensed-matter physics, or a custom topology for a specific algorithm — all from the same hardware.

Atom Computing and QuEra both use this approach. Pasqal’s analog-simulation systems use a similar setup.

Two-qubit gates: Rydberg blockade

The core gate physics is the Rydberg blockade. A Rydberg state is an atomic state with a very high principal quantum number ($n \sim 70$), giving a huge electron orbital ($\sim$ µm scale) and correspondingly enormous interatomic dipole-dipole interactions when two atoms are excited to Rydberg states simultaneously.

The blockade mechanism: if atom 1 is in a Rydberg state, the energy required to excite atom 2 to a Rydberg state is shifted by the dipole-dipole interaction. If the shift is large enough, the laser pulse meant to drive atom 2 to its Rydberg state is detuned away — atom 2 cannot be excited. Two atoms cannot both be in Rydberg states simultaneously.

This blockade enables a controlled-Z gate. The pulse sequence:

1. Apply a $\pi$-pulse on atom 1, exciting it to Rydberg if it’s in $|1\rangle$ and leaving it alone if in $|0\rangle$.
2. Apply a $2\pi$-pulse on atom 2. If atom 1 is not in Rydberg, atom 2 cycles through Rydberg and back, picking up a $-1$ phase. If atom 1 is in Rydberg, atom 2 cannot enter Rydberg, and the pulse does nothing.
3. Apply a $\pi$-pulse on atom 1 to bring it back to ground.

The net effect is a phase of $-1$ when atom 1 is in $|0\rangle$ and atom 2 is in $|1\rangle$, otherwise unchanged. With single-qubit corrections this is a CZ gate.

Rydberg gates take ~ 100 ns - 1 µs — orders of magnitude faster than ion-trap gates, comparable to superconducting CZ. They are also “structurally fast” in the sense that the gate time is set by the Rydberg-state lifetime ($\sim 100\,\mu s$ to $\sim$ ms) and the laser intensity, not by trap frequencies or shuttling.

Two-atom Rydberg-gate fidelities have been steadily improving:

• 2018: $\sim 96\%$ (early demonstrations)
• 2022: $\sim 99\%$ (Lukin group, Browaeys group)
• 2024: $\sim 99.5\%$ (Atom Computing, Pasqal)
• 2025: $\sim 99.6\%$ (QuEra logical demos imply this fidelity at the physical level)

The fidelity is now competitive with — though not yet matching — trapped ions, and has been improving rapidly.

Connectivity and rearrangement during computation

A unique neutral-atom feature: you can rearrange atoms during a computation. AOD-driven tweezers can pick up an atom from one location and place it next to a different atom for a Rydberg gate, then move it back. This effectively gives all-to-all connectivity at the cost of a few-ms move time per non-local pair.

Pasqal calls this “atom shuttling”; QuEra calls it “atom transport”; the techniques are similar. The capability is what enables fault-tolerant algorithms with non-planar codes: a qLDPC code that needs degree-6 connectivity is straightforward on neutral atoms because you can move atoms to wherever the code calls for.

The cost is finite: each move takes 1-10 ms, and moving atoms through the optical-tweezer array can introduce decoherence if the move profile is not engineered carefully. But unlike fixed-connectivity superconducting hardware, the connectivity itself is programmable per-circuit.

Current 2026 numbers

The state-of-the-art neutral-atom systems and their published numbers:

System	Atoms	2-qubit fidelity	Coherence	Public access
Atom Computing (Sr)	10,000	~99.4%	$T_2 \sim 40$ s	Cloud (early access)
QuEra Aquila (Rb)	256	~99.5%	$T_2 \sim 2$ s	AWS Braket
QuEra Gemini-class (research)	~1000	~99.6%	$T_2 \sim 4$ s	Research
Pasqal Orion-class	~1000	~99%	$T_2 \sim 1$ s	Pasqal cloud
Infleqtion	~100	~99%	$T_2 \sim$ s	Cloud (research)

Three observations:

• Qubit count is large. Atom Computing’s 10,000-atom system has more qubits than any other platform. The qubit count itself is no longer the bottleneck for many algorithms.
• Fidelity is good and improving fast. $99.5\%$ two-qubit fidelity is now standard, just behind ions. The trajectory of improvement (year over year) is steeper than any other platform.
• Coherence is excellent. Alkaline-earth clock qubits have $T_2$ in the tens of seconds — better than ions, much better than transmons.

These numbers as a whole make neutral atoms the most balanced platform on the market in 2026: large qubit count, high fidelity, long coherence, programmable connectivity. The remaining gap is on system maturity: ions and transmons have decade-plus track records on commercial cloud delivery; neutral atoms are 2-3 years into commercial-grade deployment.

The Sales Rodriguez 2025 result

The most consequential single neutral-atom paper of recent years is Sales Rodriguez et al. 2025, published in Nature: experimental demonstration of logical magic-state distillation on a QuEra/Harvard/MIT neutral-atom platform. Tutorial 24 covered the importance of magic-state distillation; this paper was the first time it was demonstrated on logical qubits on any platform.

The technical content:

• Encoded magic states on a neutral-atom processor at modest code distance.
• Performed a single round of distillation with a Reed-Muller-style protocol.
• Measured the output magic-state error rate, showed it was lower than the input rate.

This is not a deployed fault-tolerance system. It is the first crossing of an experimental milestone that until 2025 lived only in resource-estimate slides. The fact that neutral atoms got there first, beating both ions and transmons to the demonstration, was a significant signal about the platform’s maturity.

Strengths and bottlenecks

Strengths:

1. Qubit count. 10,000 atoms is real and works. The platform scales to numbers no other technology has reached.
2. Programmable connectivity. Algorithms with non-2D-grid connectivity (qLDPC, certain quantum simulations) compile naturally.
3. Long coherence. Alkaline-earth clock qubits give $T_2$ in tens of seconds, essentially infinite for any practical algorithm.
4. Fast improvement trajectory. The platform went from 100 to 10,000 atoms and from 96% to 99.5% fidelity in three years.
5. Architectural flexibility. The same hardware can do digital gate-model computation, analog quantum simulation, and adiabatic algorithms.

Bottlenecks:

1. Atom loss. Atoms occasionally get lost from tweezers due to background gas collisions, off-resonant scattering from gate lasers, or rearrangement errors. Atom-loss rates of $10^{-3}$ to $10^{-4}$ per gate are typical. This is a unique error type that error-correction codes have to handle separately. Erasure-conversion techniques (treating known atom losses as erasures) help but add complexity.
2. Crosstalk during global laser pulses. Many gate operations use a laser that addresses many atoms simultaneously; differential intensity and phase across the array introduces crosstalk that must be calibrated.
3. Mid-circuit measurement immaturity. Reading out a subset of atoms without disturbing others is harder on neutral atoms than on ions — measurement requires ejecting the atom from the tweezer or using species-selective imaging, both of which are slower than ion-trap mid-circuit measurement.
4. Cycle time. Optical tweezers can address atoms in microseconds; rearrangement and reloading are millisecond-scale. For algorithms with fast inner-loop demands, this is slower than transmons.
5. Vacuum and laser engineering. The atomic-physics infrastructure (lasers, vacuum systems, magnetic-field shielding) is sophisticated and a significant fraction of the cost.

Companies and roadmaps

• Atom Computing (Berkeley CA): the qubit-count leader. 1,180-atom system in 2023; 10,000+ in 2025. Their roadmap targets fault-tolerant computation on Sr-87 atoms, with strong emphasis on erasure-conversion architectures.
• QuEra (Boston MA): the QuEra/Harvard/MIT collaboration produced the 2025 magic-state-distillation demo. Aquila (256 atoms) on AWS Braket; Gemini (research) at higher scale. Strong academic ties; deeply involved in fault-tolerance protocol design.
• Pasqal (Paris, France): European leader. Orion-class systems for analog simulation and gate-model work. Strong on algorithm-applications for chemistry and optimization.
• Infleqtion (formerly ColdQuanta) (Boulder CO): smaller scale but commercially aggressive in defense and government markets. Rb-based platforms.
• Planqc (Munich, Germany): Sr-based research platform; emerging.

The commercial landscape is still consolidating. Atom Computing and QuEra are the most visible cloud-access providers; Pasqal is most-visible in European research; Infleqtion has the broadest defense-sector footprint.

Common misconceptions

“Neutral atoms is just analog quantum simulation.” Wrong. Pasqal’s early focus was analog simulation (Rydberg Hamiltonians as quantum simulators), but Atom Computing, QuEra, and Pasqal all now build digital gate-model systems. The platform spans both regimes.

“Atom loss is a fatal flaw.” It is a real error mode, but erasure-conversion techniques convert atom-loss errors into well-handled erasures with known location. Modern error-correction analyses for neutral-atom platforms account for this and show that the loss-error overhead is comparable to other error modes.

“More atoms is meaningless without fidelity.” Atom Computing’s 10,000 atoms paired with 99.4% fidelity is the largest high-fidelity system in any quantum platform. The combination of qubit count and fidelity is what matters; both are present.

“Rydberg gates are slow because Rydberg states are exotic.” Rydberg gates are ~100 ns to 1 µs, comparable to superconducting CZ gates. The gate is exotic-physics-driven but not slow.

“Neutral atoms still have to prove they can scale to fault tolerance.” Increasingly false — the 2025 magic-state-distillation demo established the technical feasibility. Production fault-tolerance on neutral atoms is the next milestone, with QuEra and Atom Computing both targeting it for 2027-2029.

Decision rule

Pick neutral atoms over other platforms when:

1. Qubit count is binding and you can tolerate $\sim 99.5\%$ two-qubit fidelity. 1,000+ qubit demonstrations are routine on neutral atoms; not on ions or transmons.
2. Connectivity flexibility matters. Non-2D-grid algorithms, qLDPC codes, certain quantum simulations compile better on neutral atoms.
3. Long coherence is a real budget. Algorithms that interleave compute with significant idle wait times benefit from neutral atoms’ tens-of-seconds coherence.
4. You are exploring fault-tolerance protocols early. The platform has the most published logical-qubit experimental milestones.

Pick other platforms when:

1. Maximum gate fidelity is non-negotiable (chemistry to chemical accuracy, etc.) → ions.
2. Maximum gate speed is non-negotiable (real-time control) → transmons.
3. Mid-circuit measurement-heavy algorithms → ions still have an edge.
4. Production maturity matters (a customer wants a 2-year service contract, etc.) → ions or transmons have longer track records.

The platform is most clearly the right answer for the next-generation algorithms — qLDPC error correction, large-scale fault-tolerance demos, hybrid analog-digital simulations — that take advantage of qubit count, fidelity, and connectivity together.

Exercises

1. Atom-loss budget

A neutral-atom system has atom-loss rate $10^{-3}$ per atom per gate. For a 1,000-gate circuit on 100 atoms, what fraction of runs experience at least one atom loss? How does erasure-conversion help?

Show answer

Per-gate atom-loss probability across all atoms: $10^{-3} \cdot 100 = 0.1$. Per-circuit (1000 gates): expected number of losses $\sim 100$ — so essentially every run has many losses. Without erasure-conversion, the system is unusable. Erasure conversion: each loss is detected (by imaging) and reported to the decoder as a known-location error. The decoder treats the atom as erased rather than guessing what happened. Codes can handle erasures more cheaply than unknown errors; for surface-code-like architectures, the qubit overhead penalty for erasure errors is roughly $2\times$ vs ordinary errors, much better than treating loss as undetected error. Erasure-conversion turns a fatal-flaw error mode into a manageable one.

2. Why programmable connectivity matters

You have a Hamiltonian-simulation problem with all-to-all 2-body interactions on 50 qubits. On a transmon 2D grid you compile to ~150,000 SWAPs. On a neutral-atom system with rearrangement, what is the cost?

Show answer

On neutral atoms with rearrangement: each “non-local” gate becomes a few-ms atom-rearrangement followed by a 1 µs Rydberg gate. For 50 qubits, ~$\binom{50}{2} = 1225$ pair interactions; each costs ~5 ms of rearrangement plus ~1 µs of gate, total ~6 seconds. On the transmon: 150,000 SWAPs × 90 ns = 13.5 ms. Transmon is much faster, but each SWAP has ~3% error, so the transmon has cumulative error $\sim 1 - (1 - 0.003)^{150000} =$ effectively certain failure. Neutral atoms succeed at high probability but take 6 seconds. The platform choice depends on whether wall-clock or success-rate is the binding constraint.

3. Coherence-budget calculator

A circuit takes 100 ms wall-clock on a neutral-atom system with $T_2 = 10$ s. What is the dephasing-induced error budget?

Show answer

Dephasing error in the small-time limit: $\epsilon_\text{deph} \approx t_\text{circuit} / T_2 = 0.1 / 10 = 10^{-2}$ — a 1% dephasing error budget over the whole circuit. For most circuit depths, this is negligible compared to gate-error budgets. Compare to a transmon at $T_2 = 100\,\mu s$ for the same 100 ms circuit: $\epsilon_\text{deph} = 1000$, i.e., the qubit fully dephases many times over. Neutral atoms can run circuits two-three orders of magnitude longer than transmons before idle-time decoherence becomes the bottleneck. This is a real structural advantage.

4. Picking between QuEra and Atom Computing

You have a quantum chemistry research problem requiring 200 logical qubits at chemical accuracy. Pick one neutral-atom vendor and justify.

Show answer

Atom Computing has more atoms (10,000+) and somewhat better $T_2$ from clock-state qubits. For 200 logical qubits at chemical accuracy via fault-tolerant computation, you’d need $\sim 200 \times \text{overhead}$ physical atoms; even at the lowest published per-logical-qubit overhead numbers (~150 atoms), this is $\sim 30{,}000$ atoms — beyond Atom Computing’s current 10,000 but on their published roadmap. QuEra’s logical-qubit experiments are more advanced (the 2025 magic-state demo) but their array sizes are smaller (~1000 atoms). For 2026 research, neither vendor has demonstrated 200 logical qubits at chemical accuracy. The right answer is to use one of them for smaller demonstrations and wait 12-24 months for the scaling story to mature, while planning for the eventual 200-logical-qubit-class machine. If forced to pick today: QuEra for fault-tolerance research (more published logical-qubit work); Atom Computing for largest qubit-count demonstrations.

Where this goes next

Tutorial 36 covers photonic quantum computing — a fundamentally different fault-tolerance picture (fusion-based quantum computing, GBS, measurement-based architectures) that does not fit the same gate-model framework as transmon, ion, or neutral-atom systems. Tutorial 37 will start the journey beyond hardware into emerging-tracks territory: quantum networking, quantum sensing, or back to the algorithms track depending on which gap is most-cited.

---

Weekly dispatch

Quantum, for people who already code.

One serious tutorial per week, plus the industry moves that actually matter. No hype, no hand-waving.

Email Subscribe

Free. Unsubscribe anytime. We will never sell your email.