2N vs N+1
2N duplicates the whole power or cooling path; N+1 adds one spare unit to each group of N. The choice comes down to what the load must survive: planned maintenance only, or also an unplanned fault on a path. Keep the cost comparisons apart: the whole-facility premium, the MEP premium and the electrical-path premium are different numbers.
| Axis | 2N | N+1 |
|---|---|---|
| Topology | two complete, independent power and cooling paths | one path plus one spare unit per N components |
| What it survives | maintenance or one component or path fault without load loss, as long as the paths share no common failure point | maintenance without load loss when the path is laid out for it; an unplanned fault outside the spared components can interrupt service |
| Whole-facility capex | ~25–40% facility premium for fault tolerance over concurrent maintainability (a single-source planning heuristic; re-estimate for your design) | lower whole-facility cost when fault ride-through is not required |
| MEP capex | Project-specific BOM/TCO premium for 2N over N+1 at the same usable load | the baseline for this comparison |
| Electrical-path capex | a project-specific delta; do not infer it from the whole-facility or MEP percentages | one distribution path with spare components avoids a fully duplicated electrical chain |
| Concurrent maintainability | yes: either path carries full load during work | yes, but only when the components and paths are arranged so work does not drop load |
| Failure exposure | a component or path fault can be invisible to IT load only if the paths share no controls, protection or cooling and the transfer has been tested; the label alone proves none of that | a fault during a maintenance window, or a correlated failure, can force load shedding; a fault in normal operation is covered by the spare |
| AI-era framing | protects the building, not cluster goodput; measure productive time with the named fleet, job, window, and event accounting | checkpointing, cordons and spare nodes target measured recovery loss under the same definition |
Quantitative cells are the guide's canonical figures — each is date-stamped and sourced in the numbers register and derived in the chapters below.
How the decision falls
Start from what the load must survive. If it only has to stay up through planned maintenance, N+1 on a concurrently maintainable path does that for less. Buy 2N on a path when the service must also ride through an unplanned fault there without dropping load, and the outage it prevents costs more than the premium; the planning heuristic for fault tolerance over concurrent maintainability is ~25–40% facility capex. Above that floor, a training fleet that resumes from checkpoints usually recovers more useful GPU-hours per dollar from faster checkpoint tiers, spare nodes and health checks than from a second complete path (Chapter 12.2).
What would flip it: A contract penalty or recovery loss larger than the 2N premium moves even a restart-tolerant service to duplicated paths. Paths that share controls, protection or cooling get less from 2N than the label suggests, so fix the shared dependency first.
Model this fork with your own numbers: Redundancy & availability calculator →
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)