Quiz: AI Data Center Reliability & Goodput (Part 12)

The guide's core critique of legacy tier standards for AI clusters is that they…

are too expensive to certifyrate a building, not a job — facility nines say nothing about GPU-hours converted to useful workwere written for European gridsignore cooling entirely

Tier III vs Tier IV availability, in downtime terms?

~1.6 h/yr vs ~26 min/yr (99.982% vs 99.995%)~9 h/yr vs ~1 h/yridentical in practice~26 min/yr vs ~1.6 h/yr

What does 2N cost over N+1 in MEP construction terms?

~5%+30–50%+300%It's cheaper at scale

Industry-average vs best-in-class training goodput?

~50% vs ~70%~90% vs ~96%~99% vs ~99.9%~75% vs ~85%

On a 16,384-H100 training run, how often did unplanned interruptions occur?

about one per weekabout one every 3 hours — 419 in 54 daysabout one per monthtwice, total

The single highest-leverage goodput fix per the guide is…

buying a Tier IV facilitymulti-tier checkpointing — cutting recovery from 15–30 min to under 2doubling spare GPUsslowing training to reduce stress

What sets the RPO floor for a training workload?

the storage vendor's replication SLAthe checkpoint interval — typically 2–4 hours of acceptable lossthe facility tierzero — training must never lose work

Active-active geographic failover for an AI service costs roughly…

~10% extra~2× — you carry a second live fleetnothing if automated~20% extra

A GPU-cloud SLA baseline in 2026 typically commits to…

five nines on every GPU~99.9% node uptime / ~99% rack uptime — with goodput terms still raregoodput floors of 95%+no uptime terms at all

What share of impactful data-center outages does power cause?

~10%~25%~45%~80%

Check your understanding: Reliability, Resilience & Standards