The Three Bottlenecks That Will Decide Scale

Executive summary

Neutral-atom quantum computers are usually described through qubit counts, gate fidelities or Rydberg interaction strengths. The Vault tells a different story. The densest and most strategically revealing activity sits in the machinery required to keep a large atomic processor usable for long computations: loading and replacing atoms, controlling individual sites without an explosion in optics, and making atomic loss visible to the decoder. These are not secondary engineering details. They determine practical clock rate, error budget, capital intensity and ultimately the fault-tolerance overhead of the platform.

The report therefore treats the three as an integrated stack. The most important competitive question is not “who has solved each problem?” but “which architecture minimises the combined time, error, hardware and software cost of solving all three, repeatedly?”

1Scope, method and evidence discipline

The evidence base is deliberately asymmetric. Roughly 80% of the substantive observations, patent examples, portfolio signals, publication leads and white-space hypotheses originate in the Strata IP Quantum Vault. The remainder is external corroboration: peer-reviewed articles, recent preprints, patent gazette records and official company research updates, used to verify key numbers, update developments after the Vault cut, or correct attribution.

1.1 Evidence hierarchy

1. Vault patent and publication findings are the analytical spine. Counts are treated as signals, not as a legally complete patent census.
2. Vault-indexed publications are linked to the underlying journal or preprint so the reader can inspect the primary source directly.
3. External-only sources are used sparingly, for corroboration, current-state updates or corrections, never to replace the Vault analysis.
4. Patent to paper relationships are described as mechanistic or topical unless a formal citation, common inventorship, assignment, licensing relationship or judged relationship analysis is established.
5. Apparent white space means “not seen in this Vault slice.” It is not a patentability, novelty, validity, infringement or freedom-to-operate conclusion.

1.2 Data-quality cautions that materially affect interpretation

Assignee aliases matter. ColdQuanta and Infleqtion must be consolidated. QuEra is materially undercounted by first-assignee-only filtering, because university co-assignment is common.

Publication year is not filing year. A sharp publication spike may reflect older priority dates and examination timing rather than a sudden research or financing event.

Harvard’s headline count is an upper bound for the neutral-atom cluster, because some records relate to adjacent quantum technologies.

2Bottleneck one: atom logistics

The problem has shifted from “can we load atoms?” to “can we keep computation running?”

The Vault’s strongest hardware theme is atom logistics: source preparation, stochastic loading, defect removal, rearrangement, transport between functional zones, replacement of lost atoms and continuous refill. A logistics-focused search returned 63 patent hits spanning direct competitors, universities, telecom companies and national-lab ecosystems. The breadth of assignees is itself evidence that the problem is architectural rather than company-specific.

2.1 Five sequential handoffs

2.2 Loading: the 50% occupancy problem is being attacked at multiple layers

The conventional optical-tweezer workflow accepts probabilistic loading and then rearranges atoms into a defect-free target. The Vault shows an emerging alternative: reduce the amount of rearrangement needed in the first place. Pasqal and CNRS’s EP-4649506-A1 detects trap occupancy and introduces a repulsive optical barrier once the desired occupancy is reached. That is a qualitatively different bet from post-loading rearrangement: improve the loading physics rather than make the scheduler ever more sophisticated.

The Vault also surfaces a structurally unusual idea in EP-4273889-B1 (AtomQL): use Coulomb repulsion to self-organise ions into an ordered array, then neutralise them after placement. It is isolated in the supplied landscape, but strategically interesting because it bypasses the stochastic optical loading loop instead of optimising it.

2.3 Rearrangement: patent density is thinner than publication density

Rearrangement is where the patent and publication asymmetry becomes visible. The Vault’s patent examples span image-feedback dynamic tweezers (CN-120338130-A), binary-coded parallel movement decisions (US-12462946-B2) and internal-state-dependent collective movement (WO-2026091637-A1). The publication corpus is much richer: Tetris-like packing, reduced-path planning, highly parallel composite moves, loss-aware replanning and square-root-time reconfiguration all appear in recent literature. Neutral-atom compilation is becoming partly a robotics and scheduling problem, in which movement cost and failure probability must be optimised alongside the quantum circuit.

2.4 Transport: separating dirty and clean zones becomes an architectural primitive

Atom Computing’s US-20250378972-A1 is revealing because its problem statement is architectural: loading and laser cooling are hostile to stored coherence, so the loading region and the computation region must be separated. The invention is not merely “move an atom”; it is “make zoned operation possible without sacrificing trap depth or coherence during the move.” ETH Zurich’s US-20260045383-A1 addresses a related handoff between different spatial periods.

2.5 What to measure when comparing companies

3Bottleneck two: individual optical control

The field is choosing between scaling the channels and deleting the requirement.

The second dense Vault theme concerns how to control one atom, or one selected subset, inside a dense array without disturbing its neighbours. The search returned roughly 40 addressing-optics hits. Pasqal and Atom Computing dominate the corporate side after alias consolidation, while the academic layer includes MIT, Wisconsin and Tsinghua. The trade-off is stark: local control buys generality and circuit depth, but the required optics can grow with qubit count; global control suppresses hardware scaling but constrains the gate set and shifts complexity into compilation, atomic species, frequency selectivity or interaction geometry.

3.1 Path A: scale the optics

The most important signal is not a single patent but repeated filings around the same optical limitations. Pasqal appears across addressing architecture, phase and intensity control, MEMS and interferometric concepts, integrated optics and alternative laser sources. Multiple filings against one engineering cost centre generally mean the problem is on the critical path.

3.2 Path B: avoid local addressing

The strategic point is that Pasqal patents both sides. It invests in better addressing hardware and in circuits that require less addressing. That is architectural hedging rather than contradiction: a company does not need to predict which extreme wins if it can own useful hybrid territory between them.

3.3 The publication frontier is moving into modulator physics

Device research is moving faster than many platform patents. Christen et al. demonstrated a 16-channel thin-film lithium-niobate photonic engine with visible-wavelength operation, CMOS-compatible drive voltages and multi-GHz modulation, reaching a PCB-limited 7 GHz switching bandwidth at 780 nm (Christen et al., 2025). Vanackere et al. then demonstrated a silicon-nitride, piezoelectrically actuated spatial light modulator fabricated on 200 mm wafers, with above 100 MHz modulation and above 20 dB extinction (Vanackere et al., 2025). A 2026 Stanford preprint encodes spatial information into optical frequency bins and decodes it with a two-dimensional spectrometer, reporting a 44 ns rise time, more than 10 million frames per second, while retaining reconfigurable two-dimensional beam control (Wei et al., 2026).

These results do not prove that any given processor will adopt one modulator platform. They validate the deeper point: optical control is becoming a semiconductor and photonics integration problem, not merely a laser-alignment problem.

3.4 The delete-addressing camp is accelerating

The Vault flagged a thinner patent landscape around global-control alternatives, while recent publications are far more explicit. Kazemi et al. use globally phase-modulated Rydberg driving for multiqubit parity operations; Doultsinos and Petrosyan propose a stationary Rydberg excitation bus for gates between distant atoms; a 2026 quantum-actuators proposal uses auxiliary systems to activate selected interactions under global drive; and Wang et al. propose a toric-code memory using three atomic species and global species-selective pulses, explicitly avoiding atom movement, local addressing and mid-circuit fluorescence measurement (Kazemi et al., 2025; Doultsinos and Petrosyan, 2025; Menta et al., 2026; Wang et al., 2026).

These are research proposals, not production systems. Their importance is directional: the cost of local optics is high enough that credible groups are redesigning the architecture to remove it. That creates counter-pressure against movement-heavy, individually addressed machines.

3.5 The likely outcome is hybrid

The most recent Vault patents already cluster in the middle: deflector plus SLM, temporal plus spatial modulation, one laser with multiple functions, digital plus analog computation. Pure global control struggles with arbitrary programmability; pure local control struggles with channel count, calibration and cost. A scalable machine will likely use global operations wherever symmetry permits, local operations only where information flow requires them, and integrated photonics to compress the remaining optical bill of materials.

4Bottleneck three: erasure-aware error correction

Turning a platform weakness into decoder-friendly information.

4.1 The anchor patent is unusually direct

The Vault identified US-12657505-B2, “Efficient quantum error correction in neutral atoms by conversion to erasure errors,” as the clearest claim set directed at erasure conversion itself. The claims broadly cover an array of encoded neutral atoms or ions, optical detection of departure from the computational space, and use of the known location in error correction. The claim analysis also highlights negative limitations around detecting that departure without disturbing qubits that remain in the computational space, plus species-specific dependent claims around ytterbium transitions.

External cross-checking adds an attribution detail: the June 2026 US grant is assigned jointly to The Trustees of Princeton University and the Wisconsin Alumni Research Foundation, not Princeton alone. That matters when mapping licensing or collaboration pathways, though it does not change the technical reading of the claims (USPTO Patent Gazette, 2026).

4.2 The adjacent landscape splits into three philosophies

This is more useful than asking who owns erasure correction. The Vault shows a stack: physical-state engineering at the qubit layer, maintenance and replacement at the hardware layer, and leakage-aware decoding at the classical layer. A production system can combine all three, so the competitive moat may lie in cross-layer orchestration rather than in any isolated erasure mechanism.

4.3 Why erasure bias is valuable

The scientific sequence is unusually coherent. Wu et al. proposed that alkaline-earth Rydberg qubits could convert a large fraction of dominant gate errors into detectable erasures. Ma et al. then demonstrated high-fidelity gates and mid-circuit erasure conversion in metastable ¹⁷¹Yb qubits. Scholl et al. independently demonstrated erasure conversion and detection in a high-fidelity Rydberg simulator. Sahay et al. showed that codes tailored to biased erasure noise tolerate much higher error rates than generic noise models (Wu et al., 2022; Ma et al., 2023; Scholl et al., 2023; Sahay et al., 2023).

The significance is not simply better error correction. It is error-model engineering. Rather than forcing the decoder to assume a generic Pauli channel, the hardware transforms dominant physical failures into a structured channel the code can exploit. That changes the resource equation: fewer ambiguous errors can mean lower code distance, fewer ancillas, fewer syndrome rounds, or more forgiving thresholds for the same target logical error rate.

4.4 From component demonstrations to logical circuits

The Atom Computing and Microsoft processor work used 256 ytterbium atoms and designed operations so that important error sources become detectable atom loss. The team demonstrated entanglement of 24 logical qubits encoded into 48 atoms, and reported improved performance when erasure information was used (Reichardt et al., 2024). In June 2026, Zhang et al. reported logical qubits and logical circuits using metastable ¹⁷¹Yb with noise biased toward erasures, including mid-circuit erasure measurements and conditional recovery (Zhang et al., 2026).

The 2026 Atom Computing toric-code result adds the systems dimension: lost atoms are not only detected but replaced during repeated syndrome extraction, and the reservoir itself is reloaded. This is the clearest evidence that the three bottlenecks have merged. Erasure-aware QEC is only useful at depth if logistics can replace erased qubits and optical control can measure and reset the necessary subset without decohering the rest (Atom Computing et al., 2026).

4.5 The next branch creates erasures on purpose

The Vault flagged loss-biased fault-tolerant error correction as an emerging unpatented direction. Pecorari et al. propose fast mid-circuit ionisation to convert spurious Rydberg excitation into atom loss, deliberately turning correlated excitation errors into erasure-like noise and enabling shorter QEC cycles (Pecorari et al., 2026). Once located loss is easier to correct than hidden coherent error, losing an atom can become preferable to keeping a corrupted one.

4.6 A counter-direction removes the erasure check

Another branch avoids mid-circuit leakage detection and local corrective action altogether. Measurement-based or globally controlled architectures push error handling into cluster-state geometry, species-selective global control, or coherent correction. The 2026 measurement-free toric-code proposal is the extreme case: no atom movement, no local addressing, no mid-circuit fluorescence measurement. Whether such architectures win experimentally is uncertain, but their existence signals that conventional QEC-cycle overhead is now large enough to motivate architectural redesign (Wang et al., 2026).

4.7 What to measure when comparing companies

5The three are converging into one architecture problem

Why optimising each layer separately can produce the wrong machine.

The strongest conclusion from the Vault is cross-layer. A company can improve each bottleneck independently and still lose at the system level. Faster atom movement can increase heating or control complexity. More local addressing reduces movement but increases channel count and calibration. Erasure checks improve decoding but increase measurement latency, reset traffic and replacement demand. Fault tolerance multiplies these interactions, because the same operations repeat thousands or millions of times.

5.1 Clock rate is becoming more important than qubit count

A credible 10,000-atom machine is not necessarily a better computer than a 3,000-atom machine if every QEC round is slowed by transport, imaging, reset, refill and optical calibration. The next useful industry metric will look more like “logical operations sustained per second under continuous maintenance” than “physical atoms trapped.” That metric forces logistics, control and QEC into one denominator.

5.2 The most valuable IP may sit at the interfaces

• Scheduler to rearrangement hardware: choose moves using real-time defect maps and loss probability.
• Optical controller to compiler: exploit global operations where possible, reserve local beams for information-theoretically necessary steps.
• Measurement to reservoir manager: turn detected erasures into replacement tasks automatically.
• Decoder to physical control: use erasure confidence to alter recovery, not merely report an error.
• Reservoir to QEC cadence: refill before inventory depletion becomes a logical failure mode.

5.3 Competitive context: a large incumbent enters

In March 2026 Google Quantum AI announced that it is adding neutral-atom research alongside superconducting qubits, describing neutral atoms as complementary and emphasising very large arrays while superconducting systems retain far faster cycle times (Google Quantum AI, 2026). That does not alter the three bottlenecks. It raises their strategic importance, because the framing mirrors the Vault thesis exactly: neutral atoms have a scale and connectivity opportunity, but they must close the cycle-time, control and fault-tolerance gap to use it.

6Emerging white spaces and watch items

Research hypotheses from the Vault, not legal conclusions.

7Questions to ask neutral-atom companies

Use the bottlenecks to move conversations beyond qubit counts.

1. How many initialised replacement atoms per second can the system supply during an active computation, and what happens to coherence elsewhere while the reservoir refills?
2. What is the end-to-end latency from detecting a lost atom to having a replacement reinitialised and available to the next QEC round?
3. How many physical atom moves are required per logical two-qubit operation in the target architecture?
4. At full target array size, what part of the optical stack scales approximately linearly with qubit count?
5. What is the worst-case addressing crosstalk and calibration overhead when many sites are driven in parallel?
6. What percentage of the intended fault-tolerant circuit can run on global or broadcast operations rather than site-selective control?
7. What fraction of dominant physical errors is converted into located erasures, and how is that measured experimentally?
8. How often does erasure detection itself disturb a good qubit or produce a false flag?
9. Does the classical decoder consume erasure confidence in real time, or is it used only for post-selection?
10. What is the complete QEC cycle time including transport, measurement, decode, feed-forward, reset, cooling and replacement, not only gate time?
11. Which key patents are owned directly, which are university co-assigned, and which are licensed? How are legacy names such as ColdQuanta and Infleqtion normalised?
12. Which system-level metric do you expect to improve fastest over the next 24 months: atom availability, control parallelism, erasure bias, or logical error per second, and why?

8Outlook: what to watch through 2028

From impressive arrays toward continuously maintained logical machines.

Neutral atoms will continue to post larger physical arrays, but raw size will become less differentiating. The frontier moves toward sustained logical operation under continuous maintenance. The groups that matter will increasingly disclose cycle-level metrics: how much movement occurs, how often atoms are lost, how quickly they are replaced, how many control channels operate in parallel, how much error becomes erasure information, and whether logical error decreases when all of those operations repeat.

• Atom logistics will be judged by sustained throughput, not pre-run fill fraction.
• Optical control will migrate toward hybrid global and local architectures with more integrated photonics.
• Erasure conversion will be judged at the logical level, not by isolated state-detection demonstrations.
• Software will move closer to hardware: scheduling, loss maps, calibration and decoding sharing real-time state.
• University and spinout IP structures will remain strategically important, making co-assignee and inventor normalisation essential.

The central thesis is therefore broader than “neutral atoms can scale to many qubits.” The more defensible version: neutral atoms can become a fault-tolerant platform if their physical flexibility is converted into lower system overhead rather than higher system complexity. The three bottlenecks in this report are where that conversion will be won or lost.

Appendix: Vault findings carried forward

188 filings across six consolidated entities, by publication year. Two rows carry known distortions and are drawn dashed: Harvard’s 52 is its whole quantum portfolio, including diamond-NV and nanophotonics work, so it is an upper bound on neutral atoms. QuEra at 3 is not credible for one of the leading companies in the field, and is discussed below.

What the grid says

• Pasqal is a different kind of filer, not merely the largest. 81 patents, 53 of them published in 2025 alone, across three legal entities including a Dutch one. That is not organic research output; it is a deliberate portfolio push, and the Netherlands B.V. appearing in 2024 and 2025 suggests a European filing structure built for it. Nobody else shows a fivefold year-on-year spike.
• The 2026 column belongs to Harvard. 15 in 2026 against 8 in 2025, its largest year in a twenty-year record, while Atom Computing held flat at 12 and Pasqal fell from 53 to 18. 2026 is a partial year and publication lags filing by around 18 months, so those counts will grow. The direction is what matters: the academic anchor is accelerating while the commercial leader digests its 2025 wave.
• ColdQuanta and Infleqtion peaked in 2024 and have declined since (11, then 5, then 3). It is the only falling trend in the grid, and the only entity whose record starts before 2021. Read alongside the rebrand, that is either a company that filed heavily on its cold-atom base and moved output elsewhere, or one filing under a name this Vault is not catching.
• The field is startlingly young. Outside Harvard there is essentially nothing before 2021 and little before 2024. Every commercial portfolio here is under three years old in publication terms, which means few granted rights, thin prosecution history, and priority dates from 2022 and 2023 deciding the next decade.
• planqc appears only in 2026, with two filings. A European entrant arriving as Pasqal’s wave crests is a watch item, not a conclusion.

Tag-based scoping was tested as an alternative population definition (“neutral atom”, “Rydberg”, “optical tweezers”, “cold atoms” and variants). It returns only 54 patents vault-wide and misses Atom Computing entirely, so the tag layer is too sparse to build this grid on. The company axis is the sounder basis.

• Erasure conversion: one particularly direct anchor grant (US-12657505-B2) plus a broader set of adjacent inventions covering detection, loss repair, de-excitation and leakage-aware decoding.
• Optical addressing: approximately 40 hits, Pasqal and Atom Computing highly visible after consolidating name variants, in two architectural camps.
• Atom logistics: approximately 63 hits across loading, rearrangement, transport, repair and refill, distributed across companies, universities and other institutions.
• Relationship caveat: several patent-paper pairings were topical or mechanistic matches rather than judged relationship determinations.

References and source provenance

Harvard-style author–date referencing; hyperlinks point to the primary public source. Approximately 80% of the analysis originates in the Strata IP Quantum Vault, with external sources used to verify publication details, validate performance numbers, correct attribution, and update competitive context.

Vault-indexed publications

1. Atom Computing et al. (2026) Quantum error correction with the toric code. Available at: arXiv:2606.04079
2. Chiu, N.-C. et al. (2025) ‘Continuous operation of a coherent 3,000-qubit system’, Nature, 646, pp. 1075–1080. doi: 10.1038/s41586-025-09596-6
3. Christen, I. et al. (2025) ‘An integrated photonic engine for programmable atomic control’, Nature Communications, 16, 82. doi: 10.1038/s41467-024-55423-3
4. Doultsinos, G. and Petrosyan, D. (2025) ‘Quantum gates between distant atoms mediated by a Rydberg excitation antiferromagnet’, Physical Review Research, 7, 023246. Available at: journals.aps.org
5. Kazemi, J. et al. (2025) ‘Multiqubit parity gates for Rydberg atoms in various configurations’, Physical Review Research, 7. Available at: journals.aps.org
6. Ma, S. et al. (2023) ‘High-fidelity gates and mid-circuit erasure conversion in an atomic qubit’, Nature, 622, pp. 279–284. doi: 10.1038/s41586-023-06438-1
7. Manetsch, H. J. et al. (2025) ‘A tweezer array with 6,100 highly coherent atomic qubits’, Nature. doi: 10.1038/s41586-025-09641-4
8. Menta, R. et al. (2026) ‘Global control via quantum actuators’, arXiv. Available at: arXiv:2603.23362
9. Norcia, M. A. et al. (2024) ‘Iterative assembly of ¹⁷¹Yb atom arrays with cavity-enhanced optical lattices’, PRX Quantum, 5, 030316. Available at: journals.aps.org
10. Pecorari, L. et al. (2026) ‘Loss-biased fault-tolerant quantum error correction’, arXiv. Available at: arXiv:2604.21876
11. Reichardt, B. W. et al. (2024) ‘Fault-tolerant quantum computation with a neutral atom processor’, arXiv. Available at: arXiv:2411.11822
12. Sahay, K. et al. (2023) ‘High-threshold codes for neutral-atom qubits with biased erasure errors’, Physical Review X, 13, 041013. Available at: journals.aps.org
13. Scholl, P. et al. (2023) ‘Erasure conversion in a high-fidelity Rydberg quantum simulator’, Nature, 622, pp. 273–278. doi: 10.1038/s41586-023-06516-4
14. Vanackere, T. et al. (2025) ‘Piezoelectrically actuated silicon-nitride-based high-speed spatial light modulator’, Nature Communications, 16, 11637. doi: 10.1038/s41467-025-66718-4
15. Wang, H. et al. (2026) ‘Measurement-free toric-code memory in a globally controlled Rydberg array’, arXiv. Available at: arXiv:2606.12030
16. Wei, X. et al. (2026) ‘A 10 megahertz spatial light modulator’, arXiv. Available at: arXiv:2601.08906
17. Wu, Y. et al. (2022) ‘Erasure conversion for fault-tolerant quantum computing in alkaline-earth Rydberg atom arrays’, Nature Communications, 13, 4657. doi: 10.1038/s41467-022-32094-6
18. Zhang, B. et al. (2026) ‘Logical qubits with erasure conversion using metastable neutral atoms’, Nature Physics, 22, pp. 910–916. doi: 10.1038/s41567-026-03309-0

External corroboration and corrections

1. Google Quantum AI (2026) Building superconducting and neutral atom quantum computers, 24 March. Available at: blog.google
2. USPTO Patent Gazette (2026) US 12,657,505 B2, Efficient quantum error correction in neutral atoms by conversion to erasure errors, 16 June. Used to verify grant and joint assignment. Available at: patentsgazette.uspto.gov

Selected Vault patent records

• US-12657505-B2 Efficient quantum error correction in neutral atoms by conversion to erasure errors Princeton / WARF
• US-20260080295-A1 Methods and systems for error correction in neutral atom quantum computers Atom Computing
• US-20260066148-A1 Techniques for detection of Rydberg excitations ColdQuanta
• WO-2026120174-A1 Reducing particle loss in a Rydberg-based quantum processing unit Pasqal
• US-20240185113-A1 Fault-tolerant quantum computation Harvard / QuEra
• US-20250328806-A1 Leakage-aware decoding family Riverlane
• US-20240249173-A1 Continuous operation of a cold-atom device using a separate reservoir array Atom Computing
• WO-2025255376-A1 Continuous reloading of atomic qubits including two optical conveyor belts Harvard
• US-12462946-B2 Methods and apparatus for arranging atoms in arrays QuEra
• US-20250356236-A1 Individual qubit control for atom-array processors University of Chicago
• US-12548688-B2 Parallel implementation of multi-qubit quantum gates Harvard
• CA-3274556-A1 Qubit addressing Pasqal
• WO-2024112864-A1 Multi-scale architecture for optical addressing and control of qubit arrays WARF

Strata IP Quantum Vault (2026) Neutral-atom quantum computing patent and publication extracts. Internal research dataset, accessed 16 August 2026. Primary source for the patent counts, assignee clusters, publication leads and white-space hypotheses in this paper. This is technology and competitive-intelligence analysis, not a freedom-to-operate, validity, infringement or legal opinion.