Most teams treat RI, Savings Plans, and Spot as three competing choices when they're better understood as a portfolio. The right mix depends on how confident you are in your demand forecast — and demand forecasts are wrong more often than people admit.
This piece runs a simple portfolio simulation across three demand shapes (stable, growing, uncertain) and works out the rule of thumb that survives all three.
The three instruments
- Reserved Instances (RI) — commit to a specific instance family/region for 1 or 3 years. Up to ~70% discount. Highest lock-in.
- Savings Plans (Compute or EC2) — commit to a $/hour of compute spend. Compute SP works across EC2, Fargate, Lambda. Up to ~66% discount. More flexible than RIs.
- Spot — bid on spare capacity. Up to 90% discount. Can be reclaimed with 2-minute notice. No commitment.
A portfolio simulation across three demand shapes
Same baseline workload, three different forecasts. Total monthly compute floor: 1,000 vCPU.
| Strategy | Stable demand | +30% growth | −20% shrink (recession) |
|---|---|---|---|
| 100% on-demand | $87,000/mo | $113,000/mo | $70,000/mo |
| 100% 3-yr RI (no-upfront) | $30,500/mo | $30,500 + $26k on-demand | $30,500 (paying for 200 vCPU you don't use) |
| 70% Compute SP + 30% on-demand | $45,000/mo | $53,000/mo | $45,000 (still locked-in floor) |
| 50% SP + 30% Spot + 20% on-demand | $36,000/mo | $44,000/mo | $31,000/mo |
Notice what happens to the all-in-RI strategy under demand shrinkage. You committed for three years; you're paying for compute you no longer need. The hybrid portfolio is rarely the cheapest in any single scenario — it's the strategy that doesn't catastrophically lose in any of them.
A rule of thumb that holds
RIs / Compute SP
Cover your provable floor — the compute you're certain to use for the entire commitment term. Typically 50–70% of current spend.
Spot
Cover interruptible workloads — batch, CI, async jobs, dev environments. Often 20–40% of spend.
On-demand
Cover growth headroom. The remainder. Expensive per hour, but you only pay when you use it.
Three traps the simulation catches
- Right-sizing after committing. Teams buy RIs for their current instance fleet, then a quarter later right-size everything down. The RIs are now over-committed and effectively wasted.
- Spot for stateful workloads. Spot interruptions don't care about your in-flight work. Stateful, long-running, non-checkpointed workloads on Spot turn 90% savings into 90% rework.
- EC2 SP instead of Compute SP. EC2 SP locks you to a family/region in exchange for a slightly better discount. The "slightly" is rarely worth the lock-in.
Before committing to any RI/SP, ask: "If demand drops 30% next quarter, do I still want this commitment?" If the honest answer is "no," reduce the commitment. The discount is not worth the optionality cost.
Simulating the portfolio on pinpole
On the pinpole canvas, the compute nodes expose commitment mode (on-demand, Reserved, Savings Plan, Spot) and the cost simulator applies the appropriate rate. Wire a workload, set demand uncertainty by running three traffic-shape simulations, and the right portfolio mix usually emerges within four runs. Save the canvas as a "what-if" version snapshot for next quarter's review.
Commitments are a portfolio. Optimise like a portfolio.
Run RI, SP, Spot, and on-demand mixes through a simulation across three demand shapes — and pick the one that wins on the worst-case.
Start 14-day free trial →