Quiz: AI Data Center Strategy, Workloads & Economics (Part 1)

10 questions from Part 1 of the guide. Score instantly; every explanation links to the chapter that derives the answer.

What does the guide call the 'master variable' — the single input that determines almost every downstream design decision of an AI data center?
The power price at the siteThe workload archetypeThe GPU vendorThe target PUE
A training facility and an inference facility disagree most fundamentally about siting because…
training needs better network peeringinference must sit close to users for latency, while training chases the cheapest firm powertraining requires colder climatesinference racks are heavier
How many GPUs share one coherent NVLink domain in a GB200 NVL72 rack — the number that sets tensor- and expert-parallel ceilings?
83672576
The depreciation controversy: what's the gap the guide warns about when underwriting GPUs?
GPUs depreciate slower than book life suggestsbook life (5–6 yr) flatters earnings while economic life is closer to 2–3 yrdepreciation is irrelevant for rented fleetsGPUs appreciate in constrained markets
Below roughly what utilization does a debt-financed GPU cluster start losing money on every rented hour?
~30%~50%~70%~90%
'LLMflation' refers to…
GPU prices rising with demandinference market prices falling roughly 10× per year for equivalent capabilitythe token-length growth of promptsenergy cost inflation at AI campuses
In the 2026 GPU rental ladder, roughly how does spot pricing compare to major-cloud on-demand for an H100-class GPU?
spot ~$1 vs cloud ~$7–12 per GPU-hourspot ~$4 vs cloud ~$5 per GPU-hourthey converged in 2025spot is more expensive due to scarcity
Time-to-power across procurement archetypes: which ordering is right (fastest → slowest)?
greenfield self-build → colocation → neocloud rentalneocloud rental (days) → wholesale colo (6–12 mo) → greenfield self-build (24–36 mo)colo → self-build → rentalthey're all within a quarter of each other
An RL / post-training workload is best characterized as…
identical to pre-training, just smallerinference-heavy training — rollouts dominate the compute mixpure batch inferencestorage-bound rather than compute-bound
When does edge inference genuinely justify distributed micro-datacenters?
whenever token volume is highwhen a latency budget physically cannot be met from a centralized regionwhen power is cheaper at the edgealways — edge is the stated end-state

Check your understanding: Strategy, Workloads & Economics