Who Wins If Neutral Atoms Win: Mapping a Modality’s Supply Chain

In December 2023, a group from Harvard, MIT, and QuEra reported 48 logical qubits running on an array of roughly 280 rubidium atoms (Bluvstein et al., Nature, December 2023). The paper has been picked over ever since, mostly for the error-correction result. What gets less attention is what the machine was made of. There was no processor in it, in the sense the word usually carries. Nothing had been fabricated to hold the qubits. They were individual atoms suspended in a vacuum chamber, each one pinned by a focused beam of light, in a pattern defined by software rather than by lithography.

For anyone running diligence on quantum hardware, that missing chip changes what the diligence has to examine. A supply chain map answers one question a vendor list doesn’t: which subsystem accounts for most of the cost and most of the engineering risk, and who supplies it? For superconducting machines the answer is cryogenics and Josephson junction fabrication. For trapped ions it is microfabricated traps plus a large laser bill. For photonics it is foundry access and single-photon detectors. For neutral atoms, almost everything hard is optical, and that single fact runs through cost, scaling, geography, and industrial policy.

We use neutral atoms here as a worked example because the supplier base behind them looks unlike any other modality’s, and because the answers are unusually clean. The five questions at the end transfer to superconducting, trapped ions, photonics, or whatever arrives next. The neutral-atom answers are the demonstration, not the point.

This is technology and market analysis rather than investment advice, and we have cut every figure from our source material that we could not trace to a published result.

What the Machine Actually Needs

A neutral-atom quantum computer runs at room temperature. The atoms inside it are laser-cooled to a few millionths of a degree above absolute zero, but the apparatus around them sits at ambient conditions on an optical table or in a rack. There is no dilution refrigerator and no fabricated qubit chip.

The qubits are held by optical tweezers. A tweezer is a laser beam focused so tightly that the intensity gradient at its waist forms a microscopic well, deep enough to trap a single atom and hold it against gravity and thermal motion. Split one beam into a thousand focused spots and you have a thousand trap sites. Rearrange the spots and you have rearranged the machine’s connectivity, in milliseconds, without touching hardware.

What a working system requires:

• Cooling and trapping lasers tuned to the atom’s resonance lines, to slow a room-temperature vapour into a cloud cold enough to load
• A high-power trapping laser, split into the tweezer array
• One or two Rydberg excitation lasers, which drive the two-qubit gates
• An ultra-high vacuum chamber with an atomic source inside it
• A high-numerical-aperture objective and a low-noise camera, to see and read out individual atoms
• Real-time control electronics, plus classical compute for sorting atoms and decoding errors
• A compiler and orchestration layer

Read that list as a bill of materials and the shape of the supply chain becomes obvious. Five of the seven items are photonics, optics, or the electronics that drive them. Nothing on it was invented for quantum computing. All of it was, until recently, sold in single-digit unit volumes to physics departments.

The Laser Stack

Neutral-atom machines need more distinct laser wavelengths than any other modality, and each one has to be tuned to a specific atomic transition and held there.

Take a rubidium-87 system, the choice of QuEra, PASQAL, and several others. Cooling and imaging run near 780 nanometres, on rubidium’s D2 line, with a second repump beam a few gigahertz away to return atoms that fall into the wrong ground state. The tweezer light sits far from resonance, typically between 810 and 1064 nanometres, so it traps without scattering photons off the qubit. Then there is the Rydberg excitation: either a single ultraviolet beam near 297 nanometres, or a two-photon scheme combining roughly 420 and 1013 nanometres.

A Rydberg state is an atom excited so far that its outer electron orbits at a great distance, giving the atom an enormous electric dipole. Two atoms a few micrometres apart, both pushed toward that state, interact strongly enough that exciting one prevents the other from following. That is the Rydberg blockade, and it is the entire mechanism for two-qubit gates in this modality. No blockade, no computer.

Switch atomic species and the laser bill changes. Cesium cools near 852 nanometres. Strontium, used by planqc and Atom Computing, needs a blue transition near 461 nanometres and a narrow red line near 689 nanometres, which is harder light to generate and harder to stabilise. Species choice is a supply chain decision as much as a physics one.

Two requirements make these lasers specialist rather than commodity. First, frequency stability: the light has to stay locked to the atomic transition within a small fraction of the line width, which means reference cavities, frequency combs, and locking electronics as part of the delivered system. Second, power that scales. Every additional tweezer site needs its own share of the trapping watts, so the laser requirement grows roughly in proportion to qubit count. A ten-thousand-site array is not a software upgrade to a thousand-site array.

The suppliers are a short list. TOPTICA Photonics in Munich sells tunable diode laser systems across essentially the full wavelength range these machines need, and appears in nearly every atomic physics lab in the world. M Squared in Glasgow builds titanium-sapphire tunable systems widely used for Rydberg excitation, and has gone further into building a machine of its own with the University of Strathclyde. Menlo Systems, also in Munich, supplies the optical frequency combs used as absolute references, and partners with planqc. NKT Photonics supplies fiber lasers, including lines aimed at strontium transitions. Exail in France came into the market from a different direction: it industrialised cold-atom systems for quantum gravimeters, and now supplies the rack-integrated laser stacks behind PASQAL’s deployments. Coherent and Thorlabs cover the broader component layer.

Unlike superconducting, where dilution-refrigerator lead times gate the whole industry, neutral atoms depend on a handful of laser houses, most of them European and most of them private, none individually critical but collectively load-limited. The binding constraint is industrialisation rather than laser physics: mean time between failures, field service, and the ability to ship a rack that a computing centre can operate without a postdoc attached.

The Optical Engine

If a superconducting machine’s processor is a fabricated chip, a neutral-atom machine’s processor is a pattern of light. Two components generate that pattern.

A spatial light modulator (SLM) is a liquid-crystal chip whose pixels each delay the light reflecting off them by a controllable amount. Program the right phase pattern and a single incoming beam emerges as a hologram: hundreds or thousands of individually focused spots, in whatever geometry the algorithm wants. Pixel count and phase resolution set how many traps the array can hold and how cleanly they can be placed.

An acousto-optic deflector (AOD) does the complementary job. Radio-frequency sound waves driven into a crystal create a moving diffraction grating, so changing the drive frequency steers the beam. Drive it with several tones at once and one beam becomes several, each independently steerable within microseconds. AODs provide the mobile tweezers that pick atoms up and put them down.

That sorting step is worth understanding, because it drives a surprising amount of the system design. Loading is probabilistic. Each trap site catches an atom roughly half the time, so a freshly loaded thousand-site array holds about five hundred atoms scattered at random. The machine images the array, computes a rearrangement, and physically walks atoms into a defect-free register, all before they warm out of their traps. It is a real-time classical scheduling problem with a hard deadline measured in tens of milliseconds, and it happens before every single computation.

The objective lens is the third piece. It has to be diffraction-limited across the whole array, work at several wavelengths from ultraviolet to near-infrared simultaneously, and sit within millimetres of the atoms behind a vacuum window. The camera is the fourth: detecting fluorescence from one atom demands single-photon sensitivity and fast readout, which in practice means scientific CMOS sensors from a small group of makers.

Supplier concentration here is genuinely low. Meadowlark Optics in Colorado, Hamamatsu in Japan, and Holoeye in Berlin cover most SLM demand. AA Opto-Electronic in France, Gooch & Housego in the UK, and Isomet in the United States cover deflectors. Thorlabs, Mitutoyo, and specialist optics houses cover objectives. Not one of these companies depends on quantum computing for the bulk of its revenue, which cuts both ways: the layer is resilient, and it is also nearly impossible to invest in as a quantum thesis. Today’s quantum orders move these order books by a rounding error.

Vacuum, Cells, and the Atoms Themselves

Atoms held by light alone are knocked out of their traps by collisions with stray gas molecules. Chamber pressure therefore sets how long an atom survives in a trap, which sets how long a computation can run before the array has to be reloaded. Neutral-atom machines run near ten to the minus eleven Torr, roughly the pressure of interplanetary space, and getting there means ion pumps, getter pumps, bakeouts, and careful materials selection to keep the walls from outgassing.

One design constraint separates this from trapped-ion vacuum work. The objective has to get close to the atoms, so the chamber tends toward a compact glass cell with wide optical viewports rather than a steel vessel with small windows. Several companies have developed proprietary cell designs for exactly this reason, and PASQAL has described architectures that bring fiber interfaces into the vacuum envelope so that separate processors can eventually be linked without breaking it.

The suppliers are the ordinary ultra-high vacuum industry: Pfeiffer Vacuum, Edwards, Kurt J. Lesker, VACOM, and a scattering of specialist glass cell makers. Mature, competitive, multiply sourced.

Then there are the atoms. Rubidium, cesium, and strontium are abundant, non-radioactive, non-strategic, and consumed in milligram quantities from resistive dispensers. Any chemical supplier will sell them. Nobody is going to embargo rubidium. Set that against helium-3 for dilution refrigerators, which comes from tritium decay and sits in government-controlled stockpiles, or against high-purity niobium, or the specialised substrates behind photonic switching. At the qubit layer, this modality carries essentially no critical-materials exposure, and that finding survives any plausible export-control shock.

Control Electronics and the Classical Layer

The control stack is the most modality-agnostic part of the machine. Field-programmable gate arrays sequence laser pulses and camera triggers with nanosecond precision. Arbitrary waveform generators produce the multi-tone radio-frequency signals that drive the deflectors, where each tone corresponds to one moving tweezer, so waveform complexity grows with the number of atoms being sorted. Separate electronics hold the laser locks.

What is distinctive is how much classical computing sits inside the loop. Sorting solutions have to be computed between imaging and gating. Error-correction decoding has to keep pace with the physical cycle rate or the correction is useless. Both are latency problems rather than signal-purity problems, which is why the neutral-atom companies have gravitated toward accelerated computing partners rather than microwave engineering ones. QuEra’s integrations with NVIDIA, built around GPU clusters and the CUDA-Q platform, are the clearest example, and Riverlane’s decoder work and Q-CTRL’s control optimisation sell into the same gap.

Keysight, Zurich Instruments, and Quantum Machines supply the instrumentation, as they do for every other modality. An investor holding this layer holds quantum control demand generally, not a neutral-atom position.

Physical Qubits, Logical Qubits, and the Bill of Materials

A physical qubit here is one atom in one trap. A logical qubit is a block of physical qubits running an error-correcting code, so that errors can be detected and repaired faster than they accumulate. Below threshold means the property that makes the whole approach viable: enlarging the encoding block, which is what code distance measures, causes the logical error rate to fall rather than rise. The 2023 Harvard result and the work that followed it were chasing exactly that crossover.

The procurement consequence is arithmetic. The ratio of physical to logical qubits multiplies every line in the bill of materials. One more physical qubit means one more trap site, one more share of the trapping laser’s power, more pixels committed on the modulator, one more atom to sort inside the same millisecond deadline, and one more spot to resolve on the camera. So gate fidelity is not just a physics scoreboard number. Fidelity is the probability that a two-qubit operation produces the result it was meant to produce, and it sets how large the code has to be. A machine with lower two-qubit fidelity needs a larger code distance for the same logical result, which means more atoms, and more atoms mean more watts and more optics for identical computational output. Fidelity is a cost line.

That gives you a reading rule for hardware announcements, and we would apply it to every modality. When a number appears, establish four things before it enters a model. Whether the qubits counted were physical or logical. Which code and which distance. Whether correction ran in real time during the circuit or was applied afterwards in post-processing. And whether the figure was demonstrated in a published result or announced as a roadmap target for a future year. Roadmaps and results belong in separate columns, and the interesting companies are the ones whose columns converge over time.

Two further variables belong in the same model. Duty cycle is the share of wall-clock time the machine spends computing rather than reloading, and continuous replenishment of atoms in a running array is the engineering work aimed squarely at it. A machine that pauses every few seconds to reload is not the same product as one that doesn’t. Cycle rate is the other. Gate operations take microseconds, but a full cycle of load, sort, run, and image consumes a large fraction of a second, so these machines execute far fewer shots per second than a superconducting processor. For shallow, wide problems that hardly matters. For deep circuits with many sequential operations, it turns into wall-clock hours.

The European Position, and Its Limits

Run the supplier list back and a pattern appears. TOPTICA and Menlo in Munich. Exail in France. M Squared in Glasgow. Holoeye in Berlin. VACOM and Pfeiffer in Germany. AA Opto-Electronic in France. Gooch & Housego in England.

The machine builders follow the same geography. PASQAL was co-founded in 2019 near Paris by Alain Aspect, who shared the 2022 Nobel Prize in Physics, and has installed systems at European high-performance computing centres including Jülich in Germany, GENCI in France, and CINECA in Italy. planqc spun out of the Max Planck Institute of Quantum Optics in Garching in 2022, uses strontium in optical lattices rather than tweezer arrays, and holds contracts with the German Aerospace Center and the Leibniz Supercomputing Centre. Welinq in Paris builds the quantum memories intended to link separate processors together. ParityQC in Innsbruck supplies architecture and compilation. Q-PLANET, a PASQAL-led European program, exists specifically to organise this supplier base as an industrial base.

For a policymaker, this is the strongest sovereignty case in quantum computing, and it is stronger than the equivalent case for superconducting, where the fabrication and cryogenic dependencies point elsewhere. Europe holds the founding physics, the laser manufacturing, the optics manufacturing, the vacuum manufacturing, and the system integration. The atoms are commodity chemicals.

Three caveats keep that from being the whole assessment, and we would want all three in any brief that reaches a minister.

Ownership is not the same as address. Menlo Systems and NKT Photonics are German and Danish operations under a Japanese parent. An assessment that counts factories rather than shareholders will overstate the European position, and the same correction applies in reverse when American firms manufacture in Europe.

Capital markets move the centre of gravity even when the buildings stay put. The 2021 merger that formed Quantinuum put Cambridge Quantum Computing’s British software team inside a company majority-owned by Honeywell and headquartered in Colorado, without relocating a laboratory. A merger does not move a fab, but it does move a board, and boards decide where the next facility goes.

And the demand signal is public money. Most neutral-atom systems sold in Europe so far have gone to state-funded computing centres, which makes procurement policy the industrial policy. That works while the budget lines hold. It is a different business from selling to banks and pharmaceutical companies, and the transition between them has not happened yet in any modality.

What Would Change the Assessment

Five things would move our read of this supply chain, and each one is observable well before it becomes a headline.

Fidelity plateaus. The best published two-qubit gate fidelities for neutral atoms remain below those published for trapped ions. Improvement has been quick, and if it continues the code distance stays manageable. If it stalls, the physical-to-logical ratio grows, and every extra atom bills through the optical stack. Watch published gate fidelities, not qubit counts.

Photonic integration slips. Several companies, Infleqtion most publicly, are betting that today’s bulk laser systems can be replaced by photonic integrated circuits, which put laser sources, frequency references, and beam steering onto a chip. Success would shrink these machines by orders of magnitude in size and power, and would shift the bottleneck from European laser houses to photonic chip foundries, which is a very different sovereignty picture. Delay means machines stay optical-table sized, expensive, and awkward to service in a data centre.

Reliability doesn’t industrialise. Demonstrating continuous atom replenishment in a laboratory and running a machine for months in a computing centre with an availability commitment are different engineering problems. The first has been shown. The second is what a customer buys.

Wall-clock time bites. If the applications that reach commercial value require deep circuits, the cycle rate becomes the constraint that qubit count cannot fix. This modality’s advantage is width, and the market may pay for depth.

Incumbent software gravity holds. IBM, Google, Quantinuum, and IonQ have larger user bases, more mature tooling, and more enterprise contracts. A better array does not automatically inherit the workload, and the neutral-atom companies mostly reach customers through cloud partners rather than owning the channel.

A Method You Can Reuse

Strip the example away and five questions remain. We teach them in this order because the answers to the early ones constrain the later ones.

1. What is the machine mostly made of, by cost and by volume? Follow that subsystem, not the qubit. For neutral atoms it is lasers and optics. For superconducting it is the cryostat and the wiring inside it.
2. Is there a fabricated part with a single qualified supplier? Neutral atoms have none, which removes a whole class of risk that trapped ions and superconducting both carry.
3. What material does it consume, and who controls that material? Rubidium and cesium are catalogue chemicals. Helium-3 is not.
4. What multiplies the bill of materials as the machine scales? Usually the physical-to-logical ratio, which makes gate fidelity a procurement variable in every modality.
5. Who owns the suppliers, and who owns the owners? Two layers, because the first layer flatters whichever region you are writing the brief for.

Applied to neutral atoms, the answers make a coherent picture: a modality with almost no materials risk, no fabrication dependency, low supplier concentration in optics, moderate concentration in lasers, and one open engineering question, which is whether the optical stack can be shrunk and hardened enough to ship as a product rather than an instrument.

Where to Learn This Properly

Reading a supply chain is a skill that transfers across technologies, and it is mostly a matter of knowing which questions survive contact with a vendor briefing. The physics in this article is the part that makes the questions answerable: you cannot judge whether a laser supplier is critical without knowing why the machine needs seven wavelengths.

Quantum Academy’s programs cover this ground for people who have to make decisions rather than run experiments, working from published results and physical constraints instead of vendor decks. The current programs, their contents, and their assessment formats are listed at quantumacademy.com/.

Three places to go next. For the component-by-component technical treatment of this modality, including the material we cut for length, see PostQuantum.com’s neutral-atom supply chain analysis. For the cryptographic migration side of quantum readiness, which runs on a schedule set by regulators rather than by hardware roadmaps, see pqcframework.org. And for how the roles described here map onto actual jobs in the industry, see QuantumCareers.com.