Linear-optical fusion (heralded) g_fusion

A Bell-basis measurement — beamsplitter, phase shifters, number-resolving detectors — projecting two photons onto a Bell state or destroying both and saying so, handing the code erasure rather than Pauli error. Browne and Rudolph introduced type-I/II fusion in 2005 to grow cluster states without a deterministic gate S[451]; linear optics caps unambiguous discrimination at 50%, unentangled ancillae lift it to 75% S[419]. Attributes: flying photons, heralded entangling at ~100 ns; electro-optic room-temperature control, loss-dominated errors, photonic-IC fabrication.

The 50% ceiling is a linear-optics bound (Calsamiglia & Lütkenhaus 2001): passive optics with vacuum ancillae distinguishes at most two of the four Bell states, and the ancilla photons that raise it to 75% and beyond must themselves survive. The operating figure is success probability × transmission²; loss turns a success into heralded failure at best, an undetected error at worst. The gate is not the problem: 99.22 ± 0.12% on chip, 99.72 ± 0.04% chip-to-chip over 42 m at ~2 K D[169]fact. The floor is the path around it — SiN 1.8 ± 0.2 dB/m, BTO switch 100 mdB, fibre-to-chip 52 mdB, on-chip SNSPD 93.4% median D[169] — and failure itself: a 2026 re-analysis puts a logical-error floor at zero loss, so all-linear-optics designs need very large distance, while emitter-mediated fusion tolerates 7.0–7.3% S[264]fact.

Every dated fidelity is one fusion measured alone; no chained-fusion success rate is public.

No second group has replicated PsiQuantum's 99.22%/99.72% fusion fidelities as of 4 Sep 2026. Quandela's Lucy is described by its sources only as a 12-photonic-qubit machine, not as a fusion-based system. The main report's PsiQuantum waveguide loss of 0.5 dB/m conflicts with the Nature paper's 1.8 ± 0.2 dB/m for single-mode SiN, the figure used here.

No fusion-specific materials: the gate reuses the platform stack — SiN waveguides, BTO switches, on-chip SNSPDs on 300 mm silicon at GlobalFoundries Cfact. Manufacturing decides loss per element, and every element sits inside the attempt: switch 100 mdB, fibre-to-chip 52 mdB D[169]; best packaged coupling is Xanadu's 0.085 dB/facet Cfact. I/O is channels and latency: 2–4 number-resolving detectors per fusion plus feed-forward within the photon's flight time; Aurora runs 12 modes on a 1 MHz cycle D[173]. At 10³ fusions/s the constraint is accumulated switch loss; at 10⁴–10⁶, cryogenic channel count and GHz feed-forward, unbuilt. ECCN 4A906 covers the machine, 3A901 sub-4.5 K electronics fact.

Requires single photons, single-photon detection for heralding and electro-optic switching for feed-forward; provides the fusion measurements consumed by resource-state generation and fusion-based fault tolerance (PsiQuantum, Quandela, QuiX). It replaces CV Gaussian combination on the continuous-variable photonic GKP architecture (Xanadu), where it is an alternate — switching changes detectors, encoding and error model together. Clock contribution ~1.0 × 10⁻⁷ s per attempt, divided by success probability and multiplexing depth, so generation, not the gate, dominates the derived clock. Verification: the headline numbers are tomography of one fusion; no protocol reports success-at-depth, and no second group has replicated them.

Money.
2025-09-10 · PsiQuantum · Series E · USD 1 B at USD 7 B · BlackRock, Temasek · closed fact
2026-06-16 · Quandela · DARPA QBIT Stage A · undisclosed · announced fact
2026-07-22 · PsiQuantum · DARPA QBI Stage C expansion · USD 125 M · definitive fact

Market & supply chain. No separate market: fusion ships inside a photonic IC, so the foundry and packaging house set its usable success probability. Concentration risk sits with GlobalFoundries; only G3/G4 pay.

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

Roadmaps & track record. (2025-02 · Omega component metrics · delivered); (2026-07 · Stage C V&V of switches, packaging, cryogenics · in progress, funding validation not publication). Component credibility good, system-level fusion untested.

Strategic reading. Component fidelity is table stakes; the contest is success rate at depth, where the 2026 threshold argument favours emitter-mediated fusion — a challenge to PsiQuantum and QuiX, an opening for quantum-dot vendors. If linear optics holds, foundries capture the value.

Confirm by 2027: a published network of ≥3 chained fusions with end-to-end success rate and erasure statistics; demote if only single-fusion figures exist by 2028. Best case 2029: boosted fusion above 75% on chip. Worst case: the zero-loss failure floor holds and linear-optical fusion survives only inside emitter hybrids. Open: how correlated failures are across a chip; whether 93%-efficient number resolution supports boosting.