Building a Neutral-Atom Quantum Computer

Introduction

In February 2026, a 140-qubit neutral-atom quantum computer arrived at CINECA in Bologna, Italy’s largest public supercomputing operator. It was co-funded by the EuroHPC Joint Undertaking and the Italian Ministry of University and Research. The installation joined the Leonardo supercomputer at the DAMA Emilia-Romagna Technopole, and CINECA’s Sara Marzella described it as one of the first European supercomputers supporting hybrid HPC-QPU workloads in a standard Slurm environment.

What she did not need to mention: the installation required no dilution refrigerator, no helium-3, no chilled-water plant for pulse-tube compressors, no 750 kg floor loading remediation, and no Faraday cage. The system runs at room temperature and draws approximately 3 kW. It fits in a standard server rack.

This is what makes neutral-atom quantum computing different from every other modality covered in this series. The superconducting build and silicon-spin build are dominated by cryogenic infrastructure that accounts for 30-50% of capital cost and a disproportionate share of operational complexity. The trapped-ion build replaces cryogenics with precision laser systems and ultra-high vacuum. Neutral-atom systems still require lasers and vacuum, but the scale of both is smaller, the packaging is more compact, and the result is a system that deploys into existing data center environments without the facility remediation that superconducting demands.

For the physics of how neutral-atom qubits work, see my Quantum Computing Modalities series. For who makes the components, see What It Takes to Build a Quantum Computer.

The vendor picture

Three companies sell or provide access to neutral-atom quantum computers in 2026, each with a different architecture and market position.

Pasqal (Paris) is the deployment leader. The Orion series (Alpha, Beta, Gamma generations) uses rubidium-87 atoms trapped in optical tweezer arrays, operating in both analog and digital modes with over 140 physical qubits in the current Gamma generation. Pasqal has the broadest on-premises deployment footprint of any quantum hardware vendor across any modality: systems are running at GENCI/TGCC in France, the Jülich Supercomputing Centre in Germany, CINECA in Italy, Aramco in Saudi Arabia, and DistriQ in Canada. Cloud access is live on OVHcloud, Scaleway, Azure Quantum, and Google Marketplace. Over 275 employees, over 25 clients, over $300 million in total funding. Roadmap: Vela (200+ qubits, 2027, with PIC-enabled parallel gates from the Aeponyx acquisition), Centaurus (early FTQC, 2028), Lyra (impactful FTQC, 2029).

Pasqal’s systems are sold as integrated appliances, not component kits. But they are designed for data-center integration from the ground up: rack-mountable, room temperature, standard power, Slurm-schedulable via QRMI. This is the closest any quantum computer vendor has come to making deployment feel like installing an HPC accelerator rather than commissioning an industrial facility.

QuEra (Boston) builds the Aquila (256-atom analog mode, on AWS Braket) and Gemini (260-qubit digital gate-model) systems using rubidium-87 with Dynamic Qubit Arrays (DQA). QuEra holds the world record for verified logical qubits: 96 logical qubits on a 448-physical-qubit processor using a [[16,6,4]] high-rate code, published in Nature in November 2025. A May 2026 paper simulated 580 logical qubits on 1,152 physical qubits and 1,156 logical on 2,304 physical, achieving an encoding rate above 50%, roughly 2 physical qubits per logical qubit. Over $230 million in funding. On-premises installations at AIST in Japan alongside NVIDIA-powered systems. Roadmap: third-generation systems with large numbers of continuous logical qubits in 2026-2027, targeting 100 logical qubits from 10,000 physical atoms.

Atom Computing (Boulder, Colorado) demonstrated 1,180 neutral atoms trapped in a 1,225-site array using strontium-88, with a 40-second coherence time record on the Phoenix platform. Partnership with Microsoft for logical qubit development. Strontium offers a nuclear spin qubit with exceptionally long coherence, at the cost of more complex laser requirements than rubidium (the cooling transition at 461 nm is straightforward, but the clock transition at 698 nm used for qubit operations requires careful stabilization).

Other players: Infleqtion (formerly ColdQuanta, US) with the Hilbert platform (cesium/rubidium, sensing and computing), planqc (Munich, strontium).

Architecture differences that matter for integration. Pasqal and QuEra both use rubidium-87 but differ in how they handle the atom array. Pasqal’s Orion uses relatively static tweezer configurations loaded from a MOT, optimized for analog-mode quantum simulation and increasingly for digital gate-model operation. QuEra’s Dynamic Qubit Array physically moves atoms during computation, separating storage zones from gate zones (conceptually similar to QCCD shuttling in trapped-ion systems, but using optical tweezers instead of electric fields). This architectural choice affects gate parallelism, connectivity, and the trade-off between circuit depth and atom loss. Atom Computing’s strontium approach trades rubidium’s simpler laser requirements for the nuclear spin’s natural isolation from environmental noise, enabling the 40-second coherence times that dwarf any other platform.

For a buyer, the practical distinction: Pasqal delivers packaged, data-center-ready appliances with the broadest deployment track record. QuEra leads in error correction demonstrations and logical qubit density. Atom Computing offers the longest coherence times. All three are viable, but they serve different stages of the quantum computing adoption curve.

What you are actually building

A neutral-atom quantum computer traps individual uncharged atoms in an array of tightly focused laser beams (optical tweezers), then uses a second set of lasers to execute quantum gates by briefly exciting atoms to high-energy Rydberg states where they interact strongly with their neighbors. The system is conceptually simpler than superconducting (no Josephson junctions, no microwave resonators, no millikelvin temperatures) but optically complex.

The vacuum system

Like trapped ions, neutral atoms require ultra-high vacuum to prevent collisions with background gas. The required pressure is typically below 10⁻¹¹ mbar. The vacuum chamber is smaller and simpler than a typical trapped-ion system: a compact glass or steel cell with optical access ports, magneto-optical trap (MOT) coils for initial atom cooling, and an alkali metal dispenser (rubidium or strontium source) for loading atoms into the trap. Ion and non-evaporable getter (NEG) pumps maintain the vacuum. The bake-out procedure is similar to trapped-ion (150-200°C for days), but the chamber itself is smaller, and in commercial systems like Pasqal’s Orion, the vacuum subsystem is fully enclosed within the rack-mounted unit.

The trapping laser system

The trap laser (typically 1064 nm for rubidium, ~813 nm for strontium, 5-10 W) creates a grid of tightly focused spots via a spatial light modulator (SLM) or acousto-optic deflectors (AODs). Each spot holds one atom. SLMs create static or slowly reconfigurable tweezer patterns; AODs enable fast, dynamic rearrangement of the atom array (QuEra’s DQA architecture). A high-numerical-aperture objective lens (NA 0.6-0.8, working distance 200-500 µm) focuses the tweezer beams into the vacuum chamber.

The atom loading sequence: a magneto-optical trap (MOT, using cooling and repump lasers near the atomic resonance) captures atoms from the background vapor and cools them to microkelvin temperatures. The tweezers then pick up individual atoms from the MOT. Because atom loading is stochastic (each site has roughly 50% probability of capturing an atom), the array must be rearranged after loading to fill all desired sites, either by moving atoms with AODs or by selectively pushing away extras. This rearrangement step is unique to neutral-atom systems and adds a loading overhead of seconds to each experimental cycle.

The Rydberg excitation system

Two-qubit gates use the Rydberg blockade mechanism: when one atom is excited to a high-lying Rydberg state (principal quantum number n = 50-100), the strong van der Waals interaction shifts the energy levels of neighboring atoms, preventing them from being excited simultaneously. This conditional excitation implements a controlled-Z or controlled-phase gate.

The Rydberg mechanism gives neutral atoms a capability that no other modality offers natively: parallel, non-local two-qubit gates. Because the Rydberg interaction extends over several micrometers (the “blockade radius”), any pair of atoms within that radius can interact. By illuminating multiple atom pairs simultaneously with the Rydberg beam, dozens of two-qubit gates can execute in parallel in a single time step. Superconducting systems, constrained by nearest-neighbor connectivity on a fixed chip, cannot do this without SWAP routing overhead. Trapped-ion systems can do it within a single trap module (all-to-all connectivity) but not across modules. Neutral atoms can address arbitrary subsets of the array in each gate layer.

This parallel gate capability is one reason why neutral-atom systems achieve high logical qubit counts with fewer physical qubits. The high-rate [[16,6,4]] code that QuEra used for 96 logical qubits requires non-local connections that map naturally onto the neutral-atom interaction graph but would require significant routing overhead on a nearest-neighbor superconducting lattice.

The Rydberg excitation requires either a direct UV laser (approximately 297 nm for rubidium, requiring frequency-doubled or frequency-tripled solid-state lasers) or a two-photon path using two lower-energy lasers (approximately 420 nm + 1013 nm for rubidium). The two-photon path is more common in commercial systems because the component lasers are easier to source and stabilize. Laser vendors: Toptica Photonics, M-Squared Lasers.

Single-qubit gates use either microwave pulses (global rotations, addressing all atoms simultaneously) or focused Raman beams (individual addressing via AODs or MEMS mirrors). The choice affects the addressing resolution and the laser infrastructure complexity. Pasqal’s acquisition of Aeponyx (photonic integrated circuits for beam steering) targets the Vela generation (2027), where PIC-enabled parallel gate operations would reduce the number of free-space optical components and improve beam switching speed.

The imaging and readout system

Qubit readout uses fluorescence imaging: a resonant laser illuminates the atom array, and atoms in one qubit state scatter photons (bright) while atoms in the other state do not (dark). A high-sensitivity camera (EMCCD or sCMOS) captures the fluorescence image. State discrimination fidelity depends on the photon collection efficiency and the camera noise performance.

The imaging system doubles as the atom rearrangement verification step: after loading and rearranging, an image confirms which sites are occupied before computation begins.

Why room temperature changes the integration calculus

The contrast with superconducting builds is stark enough to be worth stating explicitly.

A superconducting quantum computer at the 20-qubit scale requires: a dilution refrigerator ($700K-$2.5M, 4-12 months lead time, 750 kg), a helium-3 charge (~$100K), a closed-loop gas handling and recovery system (~$80K), a chilled-water plant, a vibration-isolated slab rated to 1,000+ kg/m², a Faraday enclosure or EMI shielding, a dedicated 3-phase 63A power panel with online UPS, a 3-meter ceiling with overhead crane access, and a 100-meter exclusion zone from vibration and EMI sources. Total facility preparation cost: $200K-$400K. Annual operational overhead for the cryogenic subsystem alone: $60K-$100K (pulse-tube service, helium management, maintenance windows with 5-10 days of downtime each).

A Pasqal Orion neutral-atom system at 140+ qubits requires: a standard 19-inch server rack with standard air cooling, a single-phase power supply at approximately 3 kW, network connectivity (10/25/100 GbE), and laser safety protocols for service access. Facility preparation cost: negligible to minimal, assuming the data center meets standard temperature (20-24°C) and humidity (30-60% RH) requirements. No cryogenic operational overhead.

The infrastructure cost advantage is measured in hundreds of thousands of dollars for a first system and millions for multi-system deployments. The operational simplicity advantage is harder to quantify but equally significant: no risk of a five-figure unplanned warm-up event, no helium-3 supply chain dependency, no annual pulse-tube service windows.

This does not mean neutral-atom systems are free of integration challenges. The laser and optics subsystem inside the Pasqal Orion, while enclosed and vendor-managed, is a precision instrument that requires periodic calibration, optical alignment verification, and laser replacement. Atom loading success rates, Rydberg laser stability, and vacuum quality all affect performance. But these are challenges contained within a vendor-supported appliance, not challenges that the customer’s facility team must solve.

The HPC integration advantage

Neutral-atom systems currently lead all modalities in production-grade HPC integration.

The QRMI demonstration by Pasqal with NVIDIA CUDA-Q in March 2026 made neutral atom the first modality with a validated path to Slurm-native QPU scheduling. A researcher submitting a hybrid quantum-classical job to an HPC cluster sees the neutral-atom QPU as another schedulable accelerator alongside CPU and GPU nodes. No special quantum-aware scheduler is needed; standard HPC workflows apply.

This integration readiness is partly a consequence of the room-temperature architecture. A superconducting QPU needs a dedicated NVQLink connection between its control electronics and a GPU node, with careful consideration of latency budgets for real-time QEC decoding. A neutral-atom QPU in a standard rack can sit on the same network fabric as everything else. The physical integration is simpler because the hardware looks like a server, not an industrial installation.

The CINECA deployment illustrates this: the Pasqal system joined the Leonardo supercomputer’s HPC environment through the HPCQS hybrid project, with IBM’s QRMI providing the orchestration layer. GENCI’s TGCC deployment follows the same pattern. At Pasqal Thoughts 2026 in April, more than 150 industry and government leaders convened to discuss how these installations are being used for real workloads in finance, energy, and telecommunications.

The integration sequence: from procurement to first computation

For a Pasqal Orion on-premises deployment (the most documented neutral-atom installation pathway):

Weeks 1-4: Procurement and facility verification. Order the system from Pasqal (lead times vary; contact vendor for current delivery schedules). Verify that the destination data center meets standard environmental requirements: temperature stability (20-24°C), humidity (30-60% RH), standard rack power, network connectivity. No structural or EMI remediation typically required. Arrange laser safety officer designation and training for service access personnel. Prepare the HPC integration: Slurm configuration, QRMI setup, network VLAN allocation, user authentication (OIDC/SAML).

Weeks 4-8: Delivery and installation. The system arrives as a rack-mounted unit. Vendor field-service personnel install it: rack mounting, power connection, network connection, initial system verification. First atom loading and trapping verification. Basic qubit characterization: single-qubit gates, Rabi oscillations, T1/T2 measurements. The vendor handles the internal laser alignment, vacuum verification, and atom loading optimization. For Pasqal Orion, installation takes approximately 1-2 weeks of on-site vendor work, followed by 1-2 weeks of commissioning and performance verification.

Weeks 6-10: Commissioning and HPC integration. Two-qubit gate calibration (Rydberg blockade gates). Atom rearrangement optimization for the target array geometry. System-level benchmarking: measure achievable circuit depth given atom loss rates, characterize two-qubit gate fidelity across different atom separations, and establish the recalibration cadence for the Rydberg laser system. QRMI/Slurm integration testing: submit test jobs from the HPC scheduler, verify hybrid quantum-classical workflow execution, measure job submission latency and throughput. Cloud connectivity setup if the system will also serve remote users. User onboarding and training on Pulser (Pasqal’s pulse-level Python SDK for programming analog and digital neutral-atom sequences).

Total timeline: 6-10 weeks from order to first useful computation, assuming standard facility conditions. This is roughly half the superconducting timeline (5-9 months) and comparable to the trapped-ion timeline (10-16 weeks), but with far less facility preparation work. The bottleneck is vendor delivery and commissioning, not facility remediation.

For QuEra on-premises installations (reference: AIST Japan deployment), the timeline is comparable. QuEra’s Gemini systems are larger (260 qubits) and may require more calibration time for the dynamic qubit array rearrangement, but the facility requirements remain standard data-center conditions. QuEra has demonstrated integration alongside NVIDIA-powered systems and Dell HPC infrastructure, showing that the neutral-atom QPU functions as a standard accelerator in the HPC fabric.

What goes wrong

Neutral-atom systems have their own failure modes, different from both superconducting and trapped-ion.

Atom loss during computation. Neutral atoms are not charged, so they are held in the trap solely by the optical tweezer’s intensity gradient. Atoms can be lost through collisions with background gas (vacuum quality), through heating from photon scattering during gate operations, or through Rydberg excitation events that eject atoms from the trap. Atom loss rates set a practical limit on circuit depth: if you lose 1% of your atoms per gate layer, a 100-layer circuit will lose roughly 63% of the array. The Harvard/MIT team demonstrated a solution in 2025: continuous qubit replenishment during computation, running a 3,000-qubit neutral-atom array for more than two hours by reloading atoms mid-computation. This capability is not yet standard in commercial systems but is on the roadmap.

Loading stochasticity and rearrangement overhead. Each tweezer site captures an atom with roughly 50% probability. After loading, the array must be rearranged to fill all desired sites, adding seconds of overhead per experimental cycle. For applications requiring many repeated measurements (shot-based quantum algorithms), this loading overhead reduces effective throughput. Commercial systems optimize this with fast AOD-based rearrangement, but it remains a throughput constraint compared to superconducting (where qubits are fabricated in fixed positions and available immediately after cool-down).

Rydberg laser instability. The two-qubit gate fidelity depends critically on the Rydberg laser’s intensity, frequency, and pointing stability. Fluctuations in any of these parameters degrade gate fidelity. The UV or near-UV wavelengths used for Rydberg excitation are more susceptible to thermal drift and optical component degradation than the near-IR wavelengths used for trapping. Periodic recalibration of the Rydberg laser system is required, though this is handled by the vendor’s calibration software in commercial systems.

Crosstalk from Rydberg interactions. The van der Waals interaction that enables two-qubit gates can also cause unwanted interactions between non-target atom pairs if they are close enough. Managing this crosstalk requires careful array geometry design and gate sequencing. In reconfigurable arrays (QuEra’s DQA), atoms can be physically moved to increase separation during gate operations, but this adds shuttling time. In fixed arrays, the minimum inter-atom spacing must be large enough that the blockade radius does not overlap unintended pairs. This constraint limits the qubit density achievable in a given array size.

Laser-induced decoherence. The trapping laser itself can cause decoherence through off-resonant photon scattering. This sets a fundamental trade-off: tighter traps (higher laser power, stronger confinement) improve atom retention but increase scattering rates that degrade coherence. Strontium-88 systems (Atom Computing) mitigate this by using a “magic wavelength” trap where the differential light shift on the qubit states is zero, but this constrains the trap laser wavelength and power. Rubidium systems operate at wavelengths where this cancellation is approximate, making trap depth optimization a routine calibration task.

Team and skills

For a vendor-supplied system like Pasqal Orion, the customer team is simpler than for any other modality:

One HPC/DevOps engineer for QRMI/Slurm integration, network configuration, API surface, and user management. This is the primary on-site role. One to two application scientists or quantum software engineers for algorithm development, Pulser programming (Pasqal’s pulse-level SDK), and user support. Vendor support contract for hardware maintenance, laser servicing, and calibration. Laser safety officer designation (can be a shared role, not a dedicated FTE).

For a single system, 2-3 FTEs plus vendor support. For a multi-system deployment serving external users, 4-6 FTEs.

This is roughly half the staffing requirement of a superconducting system (3-8 FTEs) and reflects the fact that the vendor handles the most specialized subsystems (laser, vacuum, atom loading) through the support contract. The customer’s team focuses on HPC integration and application development rather than hardware operation.

For research groups building neutral-atom systems from components (academic labs, not commercial deployments), the skill set shifts to AMO physics: laser cooling and trapping expertise, SLM/AOD optical engineering, UHV assembly, Rydberg spectroscopy. This is the domain of PhD-level experimental physicists, not IT operations teams.

Where neutral atom fits in the procurement decision

Neutral-atom quantum computers occupy a specific position in the procurement decision framework. They are the right choice for organizations that want on-premises quantum hardware with minimal facility disruption, strong HPC integration, and a path to fault-tolerant computing through high-rate error correction codes.

The facility advantage is decisive for organizations that cannot modify their data centers to accommodate superconducting-class requirements: reinforced flooring, EMI shielding, chilled-water plants, helium-3 storage. If the facility is the constraint, neutral atom is the modality that can deploy without structural remediation.

The logical qubit advantage is significant. QuEra’s 96 verified logical qubits on 448 physical qubits, and the simulation of 580 logical qubits on 1,152 physical qubits at a 2:1 encoding ratio, suggest that neutral-atom architectures may reach useful fault-tolerant computation with fewer total physical qubits than superconducting surface codes (which typically require 1,000:1 physical-to-logical ratios for useful error rates). This is an active area of research, not a settled comparison, but the direction of the data favors neutral atom for encoding efficiency.

The gate speed disadvantage is real. Neutral-atom gate times (microseconds to milliseconds for Rydberg gates, seconds for atom loading and rearrangement) are slower than superconducting (tens of nanoseconds) and trapped-ion (tens of microseconds). For applications where clock speed dominates, superconducting may retain an advantage. For applications where logical qubit count and error-correction efficiency dominate, neutral atom may be faster to practical utility despite slower physical gates.

For the cost comparison across modalities, the cryogenic infrastructure that neutral atom eliminates, and the broader context of the Quantum Open Architecture model that makes modular quantum computing possible, see the other articles in this series.