Alkaline-earth(-like) atom (Yb/Sr) — erasure-native ae_atom

¹⁷¹Yb and ⁸⁷Sr have two valence electrons: a seconds-scale ³P₀/³P₂ metastable triplet above the ¹S₀ ground singlet, on a narrow clock line (578 nm Yb, 698 nm Sr). A ground-manifold imaging beam is dark to a metastable-manifold qubit, so decay and loss scatter photons while the code space stays quiet — the "omg" (optical–metastable–ground) scheme, from a 2021 ion-trap blueprint D[333]. Errors then arrive located, as erasures, not silent Pauli flips. Demonstrated in Yb (Princeton) and Sr (Caltech), May 2023 D[141], [143].
Attributes: natural carrier; deterministic Rydberg-blockade entanglement ~250 ns; fluorescence readout ~0.5 ms, non-destructive, mid-circuit-capable.
Optical control from room temperature; error structure erasure and loss, not native Pauli; manufacture by optical assembly, not lithography.

Erasure conversion relabels error rather than reducing it; the payoff is in the decoder — at 1% erasure the erasure-aware surface code tolerates Pauli error to 0.51%, 5.2× the standard protocol D[334]. The gate floor sits elsewhere: Rydberg-blockade CZ runs in a few hundred nanoseconds against Rydberg spontaneous plus blackbody decay (0.05–0.1%), laser phase noise (0.1–0.2%) and thermal motion (~0.1%), which a 2026 review argues caps the standard scheme near 99.9% absent a new gate mechanism P[140]. Atom loss is over 80% of leakage events D[4] — precisely what omg sees. Moving the floor needs a cryogenic enclosure against blackbody-driven Rydberg decay, lower trap-light scattering out of the metastable manifold (it sets the check interval), and imaging that spares neighbouring sites.

The dominant term is the gate, not the erasure machinery: Princeton's metastable CZ error 0.016(1) is ~6× Atom Computing's raw ground-state Yb CZ of 0.0060 D[135], [142], and only ~38% of it returns as flags. The 17.6 µs record is spinless ¹⁷⁴Yb, not ¹⁷¹Yb readout, so 0.5 ms stays the honest cycle.

No lithographic chip: the fab is a UHV chamber, a high-NA objective, SLM/AOD steering and clock, Rydberg, imaging and shelving lasers; yield is per-site loading probability plus SLM-aberration trap non-uniformity. Optics is the cost driver — Fraunhofer ILT's 2026 system makes 2,000 controllable Rydberg tweezers from four beams totalling 20 W Dfact, about 10 mW per site, so 10⁴ sites is a 100 W-class laser problem S. The thin points are not foundries: two crossed-AOD vendors appear anywhere in the literature (AA Opto-Electronic, Gooch & Housego) Dfact and Hamamatsu's SLM and camera line has no identified alternative Pfact. Export exposure is at system level: ECCN 4A906 controls quantum computers above a qubit-count and error-rate threshold (in force 2024-09-06, Licence Exception IEC for partners); no ECCN names tweezer optics, AODs or UHV chambers G[301]fact.

Addressing is AOD/SLM tweezers plus global clock and Rydberg beams; readout is imaging at 0.5–1 ms. A mid-circuit erasure check must be manifold-selective and must not heat the neighbours it illuminates — the constraint holding demonstrated conversion at 38–56% rather than the 98% ceiling D[141], [142]. Global beams plus AOD addressing suffice at 10³ sites; at 10⁴ beam steering bottlenecks parallel local operations — 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, and liquid-crystal modulators reframe at tens of hertz — and trap power crosses 100 W; at 10⁶ no path exists short of metasurface or photonic delivery.

This carrier defines the "Rydberg tweezer array — alkaline-earth(-like) (Yb/Sr), erasure-native" architecture (Atom Computing/Microsoft, Princeton, Caltech, planqc, Yaqumo) and is the alternate carrier of the "Neutral-atom analogue simulator (Rydberg arrays, lattice gases)" (ICFO's QUIONE, an ⁸⁴Sr lattice gas): optical assembly only, no foundry and no cryostat, hence no CHIPS-style capital wall. It feeds Rydberg-blockade CZ, the omg encoding, imaging readout, fast (≤20 µs) array readout and the atom–photon cavity interface. It replaces alkali species; a Rb incumbent's switching cost is a new laser and imaging system. Derived clock on the tweezer array of alkaline-earth and alkaline-earth-like atoms (Sr, Yb) = sum of the syndrome round (gates 1 µs, transport 800 µs, 1Q 0.34 µs, readout 501 µs) = 1.3 ms, transport-limited, against the measured ~1–4 ms round. The salient empty slot is an erasure-aware real-time decoder fed by live array syndromes.

Fractions come from state-selective fluorescence after gates; CZ fidelity from randomised benchmarking or tomography, reported raw and post-selected on the flag. A code spends a correction cycle on every heralded erasure, so raw numbers govern cost. The 1.9(4)× vs 3.6× conflict resolves inside arXiv:2506.13724v2: both figures are in that paper and describe different quantities — decay is 3.6(1)× slower during hold for the post-selected logical qubit, 1.9(4)× slower under unconditional decoding with erasure information D[142]. Only 1.9(4)× is architecturally meaningful; the Nature Physics text is paywalled [142]. The 56% (1Q, 2023) and 38(6)% (CZ, 2026) figures are different metrics and must not be averaged. No independent replication was found.

Not published: Magne's installation status after the 2025-07-17 order announcement. The ≈10 mW per site trap power at 10⁴ sites is an extrapolation from the Fraunhofer figures, not a measurement S. No ECCN names tweezer optics, AODs or UHV chambers; only the finished system is export-controlled.

Who.
| Organisation | Role | Country | What they do | Evidence |
|---|---|---|---|---|
| Atom Computing | developer | US | Ships ¹⁷¹Yb arrays; 1,225-site Magne | C[154] |
| Princeton University | research | US | Originated Yb erasure conversion; [[4,2,2]] logical qubit | D[141], [142] |
| Caltech | research | US | Sr erasure excision; 99.71% CZ | D[136], [143] |
| Microsoft | developer | US | Co-sells Magne; QEC and decoder stack | C[11] |
| Hamamatsu Photonics | supplier | JP | SLM and camera; sole identified source | P[335] |
| US Dept of Commerce | regulator | US | CHIPS LOI; ECCN 4A906 | G[300], [301] |

Money.
2025-07-17 · QuNorth · "Magne" order, 1,225 physical / 50 logical, from Atom Computing/Microsoft · €80 M · EIFO + Novo Nordisk Foundation · ordered C[11]fact
2025-11-06 · Atom Computing · DARPA QBI Stage B · up to $15 M · DARPA · awarded G[65]fact
2026-05-21 · Atom Computing · CHIPS letter of intent · $100 M · US DoC, $2.013 B package · LOI G[300]fact
2026-06-16 · Atom Computing · Series C · $100 M · Third Point · cumulative > $300 M · closed C[154]fact

Market & supply chain. Specialty optics, not silicon: narrow-linewidth lasers, crossed AODs (two vendors named anywhere), SLMs and cameras (one), UHV chambers. No $/qubit is public; QuNorth's €80 M for 1,225 physical / 50 logical implies ≈ €65 k per site and ≈ €1.6 M per logical qubit C[11]. G3 pays for this node, G4 conditionally, G1 shared with Rb arrays.

IP & standards. No dated patent family specific to alkaline-earth(-like) erasure encoding was found; the nearest grant is Amazon's US 11,748,652 B1 (2023-09-05), heralding amplitude-damping decay on dual-rail superconducting embodiments, not atoms G[336]fact. No standards activity identified.

Roadmaps & track record. Magne, 50 logical (promised 2025-07-17 · for turn of 2026/27 · no completion announcement as of 4 Sep 2026) R[11]; QuEra, 100 logical (2024-01 · for 2026 · became Libra, >256 logical, 2028) R[162]; Pasqal, 10,000 physical (2024-03 · for 2026 · slipped to 2028, SPAC-funded ~$360 M) R[156]. Princeton is peer-reviewed and self-consistent; Atom Computing ships hardware but saw suppression vanish once reloading was included D[148].

Strategic reading. If erasure bias holds at code distance, Yb/Sr gains a structural overhead edge toward G3/G4 over Rb incumbents, who must pivot lasers and imaging or invent native-Rb heralding; Google's undeclared species is the tell C[9]. Optics suppliers win either way, so the AOD and camera duopolies hold the durable bargaining power. The substitution threat is superconducting dual-rail erasure on a ~500 ns clock, D-Wave's $550 M purchase of Quantum Circuits being the bet Cfact.

Milestones (12–24 months): Magne shows 50 logical qubits under erasure-biased decoding in situ, not post-selected (confirm); an outside group replicates >50% 1Q conversion (confirm); a real-time decoder consumes array erasure flags inside the cycle budget (confirm) — or conversion vanishes under reloading as Λ did (demote). Best case 2029: erasure-native arrays underpin a >100-logical-qubit machine with validated overhead reduction; worst case, conversion stays a single-gate effect and the millisecond cycle decides the platform. Open questions: does conversion survive ms-scale reloading; can the metastable encoding recover its 6× gate deficit; does trap power or AOD bandwidth bind first above 10⁴. Watch: Magne acceptance, Google's species choice, the first non-Princeton replication.