TL;DR: Sort the fleet by predictability and interruption tolerance, then buy to match: commitments sized to the steady trough, on-demand absorbing variance, spot taking everything that survives eviction. The real skill is sizing commitments so being wrong about your own forecast costs little, which means laddered short positions and coverage targets rather than maximum discount.
How to approach it
Ask for the fleet split first: what share of compute is steady, bursty, and interruptible. Every recommendation follows from that ratio, so naming it signals method rather than memorised pricing. Then give the mapping, the commitment-depth arithmetic, and the failure modes.
A strong answer
Three questions place any workload. Can you forecast it twelve months out? Does it tolerate eviction? Is the underlying technology likely to change?
The steady baseline, meaning the usage floor beneath every daily peak, belongs on committed capacity. Bursty production traffic stays on-demand: paying full price for peak hours beats owning a commitment shaped for a peak that moved. Interruptible work such as batch, CI runners, Spark executors and stateless workers goes to spot, with an on-demand fallback tier behind it.
For an AWS example, compare the actual quote and eligible usage for each instrument. Compute Savings Plans and EC2 Instance Savings Plans have different flexibility. Savings Plans provide billing discounts rather than capacity reservations. Capacity assurance is a separate requirement. AWS Savings Plans FAQ.
Build scenarios from hourly eligible usage after rightsizing. In a simplified example, let one unit-hour cost $1 on demand and $0.70 under a commitment, paid whether used or not. If only 60% of committed unit-hours are consumed, the effective cost per used unit-hour is $0.70 / 0.60 = $1.17, worse than on demand. Break-even utilization is 70% under these assumptions. Real plans apply monetary commitments and service-specific discounts, so use their actual billing rules.
Compare forecast paths that include migration, growth and contraction. Ladder purchases only if doing so helps manage those risks. A low historical percentile is useful evidence, but does not guarantee future demand. There is no universal 75–85% coverage target or cross-cloud discount table.
Reversal condition: a hyperscale-stable estate under negotiated enterprise pricing justifies larger and longer positions, because forecast risk there is low. For everyone else, optionality outranks the last five points of discount.
What interviewers probe next
"Why not commit to 100 percent of usage?" Unused commitment is pure loss and anchors you to current shapes; coverage targets exist precisely because forecasts decay faster than terms expire.
"Spot prices spike mid-quarter." Diversify across types, sizes and zones, configure automatic fallback to on-demand, and let degradation be a cost event rather than an outage.
"Who owns commitments after purchase?" A named owner with weekly utilisation reporting; stranded commitments are the classic silent leak, and utilisation below roughly 90 percent gets investigated like an incident.
Common mistakes
Sizing from average usage instead of a low percentile, which contracts today's waste for three years.
Buying long reservations in an instance family scheduled for replacement within eighteen months.
Treating spot as free without a fallback tier, then discovering the failure mode during the first capacity crunch.
Presenting this as pricing trivia when the scored skill is matching commitment depth to forecast confidence.