2N vs N+1
2N duplicates the whole path; N+1 adds one spare per group. The classic tradeoff is availability versus roughly double the electrical/mechanical capex — but AI changes the frame, because the cluster's goodput losses dwarf the facility's downtime.
| Axis | 2N | N+1 |
|---|---|---|
| Topology | two complete, independent power/cooling paths | one path plus one spare unit per N components |
| Availability class | Tier IV territory — 99.995% (~26 min/yr) | Tier III territory — 99.982% (~1.6 h/yr) |
| Capex | up to ~+40% facility premium vs baseline; electrical/mechanical roughly doubled | the cost-efficient default for most colo and enterprise builds |
| Concurrent maintainability | yes — either path carries full load during work | yes for Tier III designs, with tighter operational choreography |
| Failure exposure | single-component faults invisible to IT load | correlated failures or maintenance windows can force load shedding |
| AI-era framing | protects the building — but a Tier IV hall still loses ~10% of GPU-hours to cluster-level failures | many AI operators take N+1 + goodput engineering (checkpointing, cordons) over 2N + nines |
How the decision falls
Buy 2N where the workload genuinely cannot ride through (revenue inference with tight SLAs, mixed-tenant colo); for training fleets the marginal dollar usually buys more goodput in checkpointing, burn-in, and spares than in a second complete power path. Contract the cluster on goodput SLAs, the building on tier language.
Full derivations, worked examples, and the numbers behind this matrix: Redundancy topologies and fault domains (Ch 12.1) · Goodput vs facility availability (Ch 12.2) · SLAs and goodput contracts (Ch 12.4)