AI-RAN ↔ Neocloud Resiliency: the last mile, priced in nines

When an operator hosts tenant AI at a radio site and calls it carrier-grade, what does the last mile actually deliver — and which investment buys the next nine? This repository answers with a Monte Carlo fault-injection model of the converged AI-RAN/edge-neocloud site: calibrated fiber cuts, microwave rain fades, LEO flaps calibrated against Starlink's measured 15-second reconfiguration cadence, grid outages, storms that correlate all of them, and GNSS jamming against holdover clocks — run against a five-rung resilience policy ladder, P0 (single fiber) to P4 (multi-transport bonding + generator + rubidium + local autonomy).

TL;DR: run airan-resiliency ladder and the ordering the industry's pitch implies comes out backwards. (1) A single-fiber edge site is a ~2.7-nines site (~17 h/yr of tenant downtime; ~19 h/yr for the radio class) — "carrier-grade edge AI" on that footing is marketing. (2) Multi-transport bonding is the cheapest rung ($400/mo) and it works — but only for tenant traffic: the fronthaul/timing class R0 cannot ride 5G or LEO, so the class that justifies the site gains nothing. (3) You cannot bond your way out of a power outage: every transport rung — second fiber, E-band, bonded 5G+LEO — converges on the same ~3.0-nines wall (the site's own power chain), worth ~0.3 nines over P0; the generator rung buys ~1.2 more — buy the generator (and the $60/mo rubidium clock) before the third link. (4) E-band microwave matches a second fiber — within simulation noise on availability — at $700/mo less: the shared-duct SRLG tax eats part of what "diverse" fiber should buy, and rain mostly degrades a properly engineered E-band link rather than dropping it. (5) Availability nines are cheaper than capacity nines: no transport rung moves R2 usable-capacity retention off ~99.8% — rain fades and lost bond links are charged by capacity share — and even P4 only reaches ~99.88%.

Part of the DIMAGGI series on turning GPU capital into usable compute — the recovery factor of usable = nominal × network × scheduling × recovery × placement, taken to the network edge. Full analysis in docs/study.md (including the Deliverable-1 convergence blueprint: split planes, R0/R1/R2 classes, DPU-enforced zero trust, and what control-plane machinery is actually mature); all calibrations trace to REFERENCES.md. Companion repo: edge-continuum-placement — where workloads belong on the continuum this repo keeps alive.

What each rung buys

Transport rungs (P1–P3) move R1/R2 from 2.7 to ~3.0 nines and stall — the $400 bonding rung, the $800 E-band rung, and the $1,500 second fiber all land within noise of each other, because what remains is the site's own power chain. And P3's R0 equals P0's: the ladder holds everything else fixed (same 4 h battery, same clock), and bonded 5G/LEO paths cannot carry a ~100 µs fronthaul budget. The jump to ~4.2 nines across every class is the generator rung.

Where the hours live

Decomposed by cause: bonding (P3) eliminates the connectivity share of tenant downtime almost entirely — what remains is pure power, plus the GNSS tail for R0 that only a better holdover clock (rubidium: 36 h inside ±1.5 µs, vs ~6 h for OCXO) removes. That is consistent with the FCC's disaster record — more than half of cell-site outages in major events are power failures — which the storm process was calibrated to reflect. What survives at P4 is the shape carrier engineers will recognize: roughly one ~21-hour event every ~40 years, the generator that didn't start in the storm that outlasted everything else.

The price of a nine

P0→P4 is ~1.5 additional nines of tenant availability for $2,120/month. The best-value transport rung is the cheapest one (bonding, $400); the decisive rung is the generator; the worst $/nine on the board is the second fiber — $1,500 for what E-band delivers at $800 and bonding approximates at $400.

Quickstart

pip install airan-neocloud-resiliency        # or: pip install -e .
airan-resiliency ladder                       # P0->P4, availability + cost
airan-resiliency policy -p P3                 # one policy: ETTR, retention
airan-resiliency transports                   # the calibrated failure catalog

Every rate, repair time, battery-hour, and dollar is a dataclass field — recalibrate to your region and rerun:

from resiliency.policies import P2
from resiliency.sim import StormConfig, simulate

hurricane_coast = StormConfig(rate_per_year=2.5, mean_duration_h=36.0)
simulate(P2, years=2000, storm=hurricane_coast).classes["R2"].nines

Reproduce

make test        # 24 invariant tests: eligibility, monotonicity, findings
make figures     # regenerates figures/ (needs matplotlib)

Python 3.10+, stdlib only; matplotlib only for figures. The findings the study reports are asserted as test invariants — they must fall out of the simulation, not be asserted by it.

Series — turning GPU capital into usable compute

• GPU Cluster Networking (network-vs-more-gpus) · GPU Cluster Scheduling (scheduler-vs-more-gpus)
• Compute↔Power Placement (compute-power-placement) · Edge↔Neocloud Placement (edge-continuum-placement) — the companion: where workloads belong
• Chaos Fidelity Standard (ai-cluster-chaos-fidelity) · Reliability Economics (reliability-economics) · Governed Autonomy (governed-autonomy)
• AI-RAN ↔ Neocloud Resiliency (this work) — the last mile, priced in nines

Margaret (Maggie) Nanyonga — Founder & Principal Architect, DIMAGGI AI. Governed AI infrastructure: the control, reliability, and audit layer for autonomous systems operating production networks and compute.