Fibre links between photonic modules ic_fibre

Commodity telecom single-mode fibre carrying single photons or heralded entangled pairs between photonic chips, racks or modules. PsiQuantum published chip-to-chip Bell fidelity 99.72 ± 0.04% over 42 m in Omega D[169]; Aurora networks 35 chips over fibre at ~14 dB total loss D[173]. The same physics underwrites metropolitan entanglement distribution, where rates are published. Attributes: flying photons, no deterministic timing, loss-dominated error; room-temperature electro-optic interfaces, photonic-IC fibre coupling.

Attenuation is not the binding term at machine scale. Forty-two metres at 0.2 dB/km is ~8 mdB, while the two fibre-to-chip transitions bracketing it cost 52 ± 12 mdB each D[169] — ~104 mdB, a fifth of a ~0.5 dB per-photon budget, spent on connectors rather than span. Modularity here is priced per interface, unlike matter-qubit links. What distance costs is time: 42 m is ~206 ns one way at group index 1.47, longer than the best 2026 feed-forward electronics, so separation eats the decision budget before the loss budget. Rate is the term no computing vendor publishes, and its exchange against loss is brutal: 0.7 coincidences/s through 66 dB of metropolitan fibre versus 132/s on the 93 km leg P[791]. At rack-scale loss the ceiling reverts to source brightness and multiplexing depth.

The 2026 benchmark claim does not survive comparison: PsiQuantum's 52 mdB is ~39% lower at the same interface a year earlier D[169]C[172]. Fidelity is past 99.7%; the missing number is throughput.

PsiQuantum's 99.72% chip-to-chip Bell fidelity has no independent replication. The ~206 ns per 42 m latency and the ~104 mdB two-interface cost are derived from published figures, not measured.

No new fabrication: fibre, connectors and edge-coupled arrays are commodity telecom, with Corning fibre arrays and DISCO singulation behind Xanadu's coupling figure C[172]. The scaling variable is attachment count: every module boundary is two attachments, each a yield event and a 52 mdB loss event, so at 10³ modules packaging throughput sets the build rate; at 10⁴ bundle volume and timing calibration bind; nothing addresses 10⁶. I/O burden sits upstream, in the source and the fusion measurement. The machine falls under BIS ECCN 4A906 only if its qubit count and C-NOT error land in one of the rule's bands, from 34–99 qubits at ≤ 10⁻⁴ to any error from 2,000 qubits fact.

Two photonic architectures depend on it — fusion (PsiQuantum, Quandela, QuiX) and continuous-variable/GKP (Xanadu) — to scale past one chip's mode count, contributing ~0.21 µs per 42 m to the derived clock, comparable to the whole feed-forward term. Verification: the 99.72% is a conditional Bell fidelity on heralded events, channel loss excluded D[169], so it says nothing about link efficiency, and it is single-source as of 2026-09-04.

Money.
2025-09-10 · PsiQuantum · Series E · USD 1 B at USD 7 B · BlackRock, Temasek, Baillie Gifford · closed Cfact
2026-03-26 · Xanadu · SPAC merger · ~USD 302 M gross, ~40% below plan · Crane Harbor · closed Cfact
2026-05-21 · PsiQuantum · CHIPS letter of intent · USD 100 M · US Commerce · LOI fact

Market & supply chain. Fibre and connectors are commodity; the concentrated input is precision fibre-array attachment, where Corning and a few packaging houses set loss and throughput alike. No cost per link is public. Pays for G3/G4, G6.

IP & standards. No dated patent family specific to inter-module photonic fibre linking surfaced as of 2026-09-04; ETSI QKD standards govern the physical layer, no computing-side interconnect standard exists.

Roadmaps & track record. PsiQuantum's 99.72% is unchanged since 2025-02, its 2026 progress visible only through DARPA V&V D[169]. Xanadu marketed a coupling benchmark it does not hold C[172]. No actor has published a dated inter-module rate target — a roadmap gap, not a slipped promise.

Strategic reading. While rate stays unpublished, every photonic logical-qubit roadmap rests on an unaudited number and buyers cannot compare vendors. Whoever publishes fidelity and rate together sets the benchmark; packaging houses, not chip designers, capture the margin, since loss and yield live at the connector.

Confirm by 2028: a vendor publishing an inter-module rate with its fidelity; demote if the field still reports fidelity alone. Best case 2029: multiplexed links close a stated rate gap at rack scale. Worst case: rate stays unmeasured. Open: is 99.72% reproducible outside PsiQuantum; does multicore fibre help inside a machine.