Fusion-based fault tolerance code_fusion

A fault-tolerant architecture with no deterministic two-qubit gate: pre-made photonic resource states are consumed by destructive linear-optical fusions, and the code lives in the pattern of which fusions succeeded, failed or reported loss. Bartolucci, Birchall, Bombín and colleagues at PsiQuantum formalised it with thresholds in 2023 S[179]. It replaces GKP-bosonic concatenation as photonics' route to fault tolerance. Attributes: flying photons, no native control modality and no clock of its own; pure-loss errors, no manufacturing precursor beyond fusion and resource-state generation.

Loss is the design variable: a lost photon turns a fusion into a heralded erasure at a known location; an intrinsic failure erases one of its two outcomes, and the decoder routes around both. Hence unusually high tolerances — 2.7% loss per photon with a boosted 6-ring, 17.4% with a {7,4}-encoded state of ~168 photons nobody can make S[179]. The catch sits below threshold. A 2026 re-analysis shows fusion failure alone, at zero physical loss, leaves a logical-error floor shrinking only with code distance, so being under threshold says nothing about the overhead for 10⁻¹⁰; emitter-mediated fusion raises the threshold to 7.0–7.3% against an unencoded 6-ring's 0.38–0.82% S[264]fact. Moving the floor takes lower-loss photonics, boosted fusion, or emitters.

On the report's cross-platform table photonics scores zero on every fault-tolerance primitive: no below-threshold scaling, no logical qubit past break-even, no logical two-qubit gate, detection-only decoding D[687].

The 6-ring loss thresholds are not comparable: 2.7% per photon for a boosted 6-ring [179] against 0.38–0.82% (static bias) for an unencoded 6-ring [264], under different boosting, bias models and decoders. Every fusion threshold is a simulation, none is measured, and no independent group has reproduced them. As of 4 Sep 2026 PsiQuantum has published no response to the 2026-06 subthreshold critique and no 2026 hardware result; its progress is visible only through DARPA V&V contracting.

Nothing is manufactured for this code; it inherits the photonic-IC stack of its inputs (300 mm silicon at GlobalFoundries) Cfact. No yield, cost or energy figure exists because nothing runs it. The burden it imposes is synchronisation: a {7,4}-encoded state means ~168 photons per round, each with a source, switch path and detector channel, against a record of 8 fused photons at 0.4–2.3 per minute D[688]. At 10³ fusions/s the constraint is switch loss; at the 10⁶ a useful machine needs, cryogenic detector channels and GHz feed-forward, neither built. ECCN 4A906 covers the assembled machine fact.

Requires linear-optical fusion and a resource-state generator (RSG), neither at code-relevant scale, and provides nothing downstream — it tops the photonic fusion architecture (PsiQuantum, Quandela, QuiX). It replaces GKP-bosonic concatenation, and the switch is total: different carrier, detectors, error model. Its MHz cycle is a design claim, not a measurement: the real cycle is resource-state supply, ~0.4–2.3 per minute D[688]. Verification: the thresholds are decoder simulations under stated noise models, never measured or independently reproduced; the 2026 re-analysis is the first published challenge to the framework's relevance, unanswered as of 4 Sep 2026.

Money.
2024-04 · PsiQuantum · Australian federal and Queensland commitment · A$940 M · announced Pfact
2025-09-10 · PsiQuantum · Series E · USD 1 B at USD 7 B · closed fact
2026-05-21 · PsiQuantum · US CHIPS letter of intent · USD 100 M · non-binding fact
2026-07-22 · PsiQuantum · DARPA QBI Stage C expansion · USD 125 M · definitive fact

Market & supply chain. No product, no market: the asset is a roadmap plus government commitments; the suppliers that matter sit upstream. Pays for G3/G4 alone.

IP & standards. No dated patent family specific to the fusion-lattice code as of 4 Sep 2026; nearest is ORCA's US 12,437,225 fact.

Roadmaps & track record. (2024-04 · Brisbane machine useful by end-2027 · at risk: site moved to Moreton Bay, groundbreaking slipped to 2026-06, Linde cryoplant not due for delivery until 2H 2027 Cfact) Pfact. Strong paper record, weak schedule.

Strategic reading. If the subthreshold critique holds, the largest privately funded photonics bet needs a far lower-loss network or a pivot to emitter hybrids — a redesign for a company already late. Winners if it works: PsiQuantum, its foundry, detector vendors; losers: bosonic GKP and anyone selling a fast gate rather than cheap erasure.

Confirm by 2028: an end-to-end fusion-lattice demonstration at any code distance with measured loss and logical error; demote if only component numbers persist. Best case 2029: a small working lattice. Worst case: it stays a paper and emitter hybrids inherit it. Open: does anyone rebut the 2026 critique; which threshold survives hardware; does Stage C publish a code-level number.

Fusion networks need not carry a whole computation: a 2026 construction confines them to the seam of distributed surface-code lattice surgery between matter-qubit processors, where 'hybrid checks' combine local ancilla readouts with the fusion outcomes of photonic resource states P[893]. In simulation with no local noise, the seam tolerates erasure of 50% of fusion outcomes (38% and 32% with resource chains truncated to eight and five qubits) and the merge observable's interface distance rises from d + 1 to 2d + 1; at circuit and resource-state error rates of 10⁻³, local errors largely mask that advantage, and resource-state generation is not modelled P[893].