A media platform's workload is bursty rendering plus steady serving. Which pricing models fit which parts, and what is the common mistake?

Matching model to workload shape

Steady baseline → commitments. Serving traffic has a predictable floor. Committing to it buys a substantial discount for capacity you would have run anyway. The risk is over-committing, which is why the commitment should cover the floor, not the average.

Interruptible batch → spot or preemptible. Rendering jobs that can be checkpointed and restarted are the ideal spot workload — heavily discounted capacity in exchange for the possibility of reclamation. The architectural requirement is checkpointing and idempotent restart, without which spot is unusable.

Unpredictable burst → on-demand. Expensive per unit, and the right choice for capacity used for hours. Paying a premium for something you use rarely is a bargain compared with owning it.

Serverless → low duty cycle, event-driven work. Excellent for asynchronous processing with idle periods; expensive for sustained high-volume serving, and cold starts land exactly when a spike begins.

The common mistake

Maximum commitment across the estate for the largest discount. It converts a variable cost into a fixed one, and the commitment outlives the architecture: a re-architecture that reduces usage means paying for capacity you no longer need, and — worse — the commitment becomes an argument against the re-architecture.

The second common mistake is the opposite: on-demand for everything to preserve flexibility, paying a large premium for capacity that has run continuously for two years and will continue to. Flexibility you never exercise is not flexibility.

The discipline

• Commit to the floor, measured over months, not to the average or the peak.
• Ladder commitments with staggered expiry, so the estate is never entirely locked at one point in time.
• Re-evaluate on architecture change, since a commitment made against an old design silently constrains the new one.
• Model the interruption cost for spot, including the wasted work and the scheduling complexity — spot is cheap only if restart is cheap.
• Measure the utilisation of committed capacity. A commitment at 60% utilisation may be more expensive than on-demand.

The judgement being tested

Whether you can decompose a workload by its demand shape rather than treating the estate as one purchasing decision. Most cost-efficiency in cloud comes from that decomposition, not from negotiating a better rate.