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
In Chapter 1.8’s matched reserved-rental case, what happens to the rental bill when the installed fleet is idle within a committed annual period?
Every installed calendar hour still costs the assumed $3.0/GPU-hourOnly GPU-active hours are billedBook depreciation pays the rental billThe bill falls with model-FLOPs utilization
'LLMflation' refers to…
inference prices at fixed capability falling ~9–900× per year, depending on the capability thresholdGPU prices rising with demandthe token-length growth of promptsenergy cost inflation at AI campuses
Chapter 7.11’s illustrative eligible fleet uses 60 nodes at 960 kW. What must match before comparing a purchase with a spot or on-demand offer?
Only the accelerator brandUseful output, availability, support and exit terms for the 960 kW fleetOnly peak FLOPSOnly the lowest advertised hourly rate
Which time-to-power comparison across procurement paths is defensible?
Compare only construction durationRead each path off its own gate — queue position and tariff process for self-build, provider capacity and fit-out for colo, available inventory for neocloudAssume 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