Reliability, Resilience & Standards

Tier and Rated assessments cover defined facility scopes; an AI operator must separately choose job fault domains, blast-radius limits and recovery, then buy concurrent maintainability, fault tolerance or both for the declared facility states at the lowest lifecycle cost.

An AI cluster earns its return on goodput — the share of bought GPU-hours doing useful work — and for checkpointed training the next redundancy dollar buys more goodput than facility nines; for SLA inference it does not, so price both above the facility states the contract requires.

DR for an AI factory is a per-workload decision about how much spare capacity to pre-pay and where; when power-bound activation would miss the RTO, energize the spare region before the disaster and reserve surviving state, compatible GPUs and promotion authority so detection, loading and verified restoration fit the same clock.

Promise facility availability to a tenant who is paying for goodput and the SLA tracks neither the customer's pain nor the provider's control; match the contracted metric to the workload.

RBD, fault trees, Markov chains and Monte-Carlo turn the redundancy debate into a model in which every availability result and percent of goodput has traceable event, exposure and recovery inputs; use one event inventory, reproduce the loss distributions and price the uncertainty before choosing the investment.