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.

Which input should lead the requirements brief without being treated as sufficient by itself?
The workload archetypeThe target PUEThe GPU vendorThe power price at the site
Which siting contrast is a useful starting hypothesis that still must be tested against the service and site requirements?
training needs better network peeringinference racks are heaviertraining requires colder climateslatency-bound inference may need user proximity, while latency-tolerant training can weight firm power more heavily
How many GPUs are connected in the rack-scale NVLink domain of the cited GB200 NVL72 configuration?
83672576
The depreciation controversy: what's the gap the guide warns about when underwriting GPUs?
GPUs appreciate in constrained markets5–6 yr book life can flatter earnings; estimates put useful life at 4–6 yr, while contested bear cases use 2–3 yrdepreciation is irrelevant for rented fleetsGPUs depreciate slower than book life suggests
Below roughly what utilization does a debt-financed GPU cluster start losing money on every rented hour?
~70%~30%~90%~50%
'LLMflation' refers to…
inference market prices falling roughly 10× per year for equivalent capabilityGPU prices rising with demandthe 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 ~$4 vs cloud ~$5 per GPU-hourspot ~$1.03 vs cloud ~$6.88–12.29 per GPU-hourspot is more expensive due to scarcitythey converged in 2025
Which time-to-power comparison across procurement paths is defensible?
Compare only construction durationUse actual provider availability, fit-out scope, tariff studies and agreements, generation and permit milestones, and equipment datesAssume rental is always fastest and self-build is always slowestUse one national duration for every path
An RL / post-training workload is best characterized as…
pure batch inferenceidentical to pre-training, just smallerinference-heavy training — rollouts dominate the compute mixstorage-bound rather than compute-bound
What decides whether inference belongs at an edge site or a centralized region?
always — edge is the stated end-statewhenever token volume is highwhen power is cheaper at the edgethe full requirement vector: latency, sovereignty/residency, availability and egress, data gravity, energy/carbon, capacity and model fit, and service geography

Check your understanding: Strategy, Workloads & Economics