Atom transport by AOD tweezers (zoned architecture) cx_aod

An rf tone in a crystal makes an acoustic grating whose first-order diffraction angle tracks the tone, so sweeping the rf sweeps a tweezer and the atom in it; crossed x/y deflectors turn a comb of tones into a grid of movable traps. The zoned architecture partitions the plane into storage, entangling and readout regions, taking connectivity from moving atoms rather than fixed couplings — introduced by Bluvstein et al.'s 2022 coherent-transport processor (Harvard/MIT).

Attributes. Mobility: transport — atoms relocate hundreds of µm between zones, the platform's only long-range coupling. Time ~10⁻³ s per move, classical repositioning, neither deterministic nor heralded entanglement; the error is loss, not phase.

Two clocks bound a move: acoustic transit across the beam aperture sets how fast a tone can change (µs for millimetre apertures), and trap radial frequency sets how fast the atom can accelerate without heating out. The second dominates — transport error is release-and-recapture loss driven by trajectory jerk, so minimum-jerk profiles and deeper transit traps, not faster electronics, buy fidelity. Measured: 610 µm in 1.6 ms at 99.95%, 270 µm in 400 µs at 99.8% D[138].

Coherence is not binding here: hyperfine T₂ is 12.6 s in the same 6,100-atom array D[138] and 1.09(3) s under dynamical decoupling in a continuously reloaded 3,000-atom system D[144]. So the signature is loss — detectable by imaging, hence erasure-convertible, and over 80% of all leakage; used as a supercheck with ML decoding it gave 2.14(13)× suppression at d=3→5 D[4]fact. What moves the floor: minimum-jerk trajectories, deeper transit traps, and an acousto-optic lens adding the axial axis — so far a design study with no atoms S[503].

Best demonstrated: 610 µm in 1.6 ms at 99.95% D[138]. Typical at scale is the 448-atom zoned processor, where transport is interleaved with gates, not benchmarked alone D[4].

Dominant error term: atom loss at release and recapture — the channel that makes the platform erasure-friendly.

No fab here — free-space optics bolted to a vacuum cell, so "yield" is array filling fraction: 99.3% after rearrangement D[144], 60.5% before it in a metasurface array D[145]. Named hardware in the 448-atom system: crossed AODs (DTSX-400, AA Opto-Electronic), Hamamatsu modulator and camera, Spectrum Instrumentation AWGs (arbitrary waveform generators), Rohde & Schwarz microwave sources D[4]; Gooch & Housego supplies deflectors in the 3,000- and 6,100-qubit papers P[504]. That is the concentration risk: two acousto-optics houses serve the published field — Gooch & Housego, quoted on the London Stock Exchange's AIM market, and the privately held AA Opto-Electronic. Cost per channel is not quotable. Export exposure is nil — the BIS rule of 2024-09-06 names no tweezer optics or deflectors fact; it controls only the resulting machines (4A906), and those only when qubit count and C-NOT error fall in the same band, from 34–99 qubits at ≤ 10⁻⁴ to any error from 2,000 qubits fact.

The burden is one rf tone per trap per axis, synthesised by AWG; count scales with atoms, not zones, and intermodulation between tones — not laser power — caps simultaneous traps. The cycle is the real cost: transport 0.4–1.6 ms plus imaging 0.5–1 ms gives QEC rounds of 1–4.5 ms, ~10³× a superconducting cycle, with the decoder obliged to keep up. At 10³ atoms a few AWG channels suffice; at 10⁴ the wall is beam steering — a deflector resolves as many spots as its rf bandwidth times the acoustic transit across the beam, a transit that also bounds how fast it re-points — a trade no source consulted here quantifies against atom number; at 10⁶ the only design point puts RSA-2048 at 19 M atoms and 5.6 days at 1 ms cycles S[31].

It requires laser plus AOD/SLM optical control and provides for bivariate bicycle qLDPC codes, high-rate concatenated codes with transversal gates, and algorithmic fault tolerance — none reachable with fixed nearest-neighbour coupling, so it is a prerequisite, not a convenience. Architectures served: the alkali (Rb/Cs) and alkaline-earth(-like) (Yb/Sr) Rydberg tweezer arrays and the neutral-atom analogue simulator. Nothing replaces or conflicts with it in the graph; pressure comes from other transport physics — the 3,000-qubit continuous system moves atoms 0.5 m on optical-lattice conveyor belts, using tweezers only for local rearrangement D[144]. Contribution to the derived clock: dominant, at ~2.0×10⁻⁴ s per move it is 0.80 ms of the 1.31 ms sum of the syndrome round: gate layers + transport + readout + reset on both gate-model architectures, since CZ runs at 270 ns.

Move fidelity is post-recapture survival by fluorescence imaging: a loss metric, not a state fidelity, so 99.95% certifies the atom arrived, not that its phase did. No source isolates transport-only error at the logical level; the 448-atom results fold transport, gates and readout together D[4]. The sharpest negative evidence is Atom Computing's toric-code run, where suppression over four cycles vanished once continuous reloading was included (0.63% versus 0.64% per cycle over 90 rounds) D[148] — reloading and transport are the operations this node owns. Harvard and Caltech report 99.8–99.95% survival on different deflector hardware, consistent but not a replication.

The axial range, speed and survival in [503] come from an optical design study with no trapped atoms, not from an experiment. Not published: the channel count, or the deflector's trade between spot count and re-pointing speed, against atom number; a dated patent family (assignee and year) for AOD-based zoned atom transport; and the capacity and scale-up plans of AA Opto-Electronic and Gooch & Housego.

Money.

• 2025-07-17 · QuNorth · order for Magne (Atom Computing/Microsoft, 50 logical) · €80 M · EIFO + Novo Nordisk Foundation · delivery turn of 2026/27 G[11]
• 2025-09-09 · QuEra · financing round · $230 M+ · Google, SoftBank Vision Fund 2, NVentures · closed C[153]
• 2026-06-16 · Atom Computing · Series C plus CHIPS LOI · >$300 M total · Third Point · closed + LOI C[154]fact
• 2026-08-12 · Infleqtion · Q2 revenue, FY guidance · $12.6 M, ~$43 M FY2026 · — · reported G[332]
• 2026-08-28 · Pasqal · SPAC merger completed (Nasdaq PSQL) · ~$360 M cash · — · closed P[156]

Market & supply chain. The equipment is generic photonics with one exception: crossed AODs, where AA Opto-Electronic and Gooch & Housego are the only vendors named in sourced material D[4]P[504]. Deflector-channel demand grows linearly with atom count, so a small supplier could gate a platform. Unit economics are not quotable. G3 and G4 pay directly; G1 partly; G2, G5, G6, G7 indirectly.

IP & standards. No dated patent family with assignee and year was found for AOD-based zoned transport, and no standards body governs tweezer hardware. The one vendor specification with numbers is QuEra's Gemini page: 260 physical qubits, 99.2% global two-qubit fidelity, one shot per second Cfact.

Roadmaps & track record. QuEra (promised Jan 2024 · 100 logical qubits in 2026 · status 2026-09-03: replaced by Libra 2028, >256 logical — a two-year slip) R[162]. Atom Computing/Microsoft (2025-07 · Magne, 50 logical, turn of 2026/27 · installing) G[11]. Pasqal (2024 · 10,000 physical in 2026 · slipped to 2028) C[156]. Google entered March 2026 with no hardware result C[9]. Experimental groups deliver what they publish; vendor logical-qubit dates have slipped about two years each.

Strategic reading. If zoned transport scales, Harvard's lineage companies and the two deflector houses win, and fixed-connectivity Rydberg machines lose the code families that matter. The substitution threat is other optics, not another node: metasurface arrays (11,022 atoms, no gates) D[145] and photonic-chip trap delivery (4 atoms, 27.5 s, claimed 50× footprint cut) C[330] would displace bulk AOD/SLM stacks if they reach gate quality.

Confirm by end-2027 a zoned system holding >1,000 atoms under fault-tolerant control; demote if by end-2028 no group shows long-memory suppression (Λ > 2) with continuous reloading — the case where suppression has already vanished once D[148]. Best case by 2029: transport loss cut several-fold and the axial axis in use, unlocking 10⁴-atom zones. Worst case: channel count and deflector bandwidth cap zones near 10³–10⁴ atoms. Open questions: does move fidelity hold at QEC duty cycles under reloading; can channel count scale past the deflector's bandwidth–transit limit; who else supplies deflectors at volume?