Spot vs reserved vs on-demand: the three cloud pricing models, and when each wins
The same VM can cost three very different amounts depending on how you buy it. Here is how on-demand, reserved, and spot pricing work, the discount each offers, the risk attached, and how to mix them without overcommitting.
Ask “what does this instance cost?” and the honest answer is “which way are you buying it?” The same machine can run at full list price, at a 40–60% discount, or at a 70–90% discount — the difference is purchasing model, not hardware. Here is how the three work and when each is the right call.
On-demand: flexible and expensive
On-demand is the headline rate: pay per second/hour, no commitment, turn it off whenever. It is the most expensive per hour and the most flexible. Use it for unpredictable, short-lived, or spiky workloads — development, a launch you can't size yet, a job that runs for two days. Paying the premium for flexibility is correct when you genuinely can't predict usage.
Reserved / committed: cheaper for a promise
Commit to a baseline of usage for one or three years and the provider discounts it heavily — typically 40–60% off on-demand, more for longer terms and upfront payment. AWS calls them Reserved Instances / Savings Plans; GCP has Committed Use Discounts; Azure has Reserved VM Instances. The trade is commitment risk: you pay for the reservation whether or not you use it. Use reservations for the steady-state floor — the capacity you know you'll run 24/7 for the next year. The classic mistake is reserving your peak; reserve your baseline instead.
Spot / preemptible: cheapest, but can vanish
Spot instances sell spare capacity at 70–90% off, with one catch: the provider can reclaim them with little warning (often ~2 minutes) when it needs the hardware back. Use spot for interruptible, stateless, or restartable work — batch processing, CI runners, rendering, big-data jobs, fault-tolerant worker fleets. Never put a stateful database or a single point of failure on pure spot. With checkpointing and auto-replacement, spot is the single biggest lever on a compute bill.
The mix that actually works
- Reserved for the floor. Cover the capacity you run continuously with 1- or 3-year commitments.
- Spot for the elastic, interruptible middle. Scale batch and stateless workers on spot, designed to survive reclaim.
- On-demand for the unpredictable top. Absorb spikes and experiments at full price — it's cheaper than over-reserving or risking critical work on spot.
Two traps to avoid
Over-committing. A three-year reservation on a workload you'll re-architect in six months is a discount that becomes a liability. Match commitment length to how confident you are.
Ignoring the egress and storage lines. Compute discounts don't touch data-transfer or storage costs — a 70% spot discount on compute means little if egress dominates the bill.
Bottom line
There is no single “cheapest” price — there is the cheapest way to buy for each slice of your workload. Reserve the steady floor, run interruptible work on spot, and keep on-demand for the spiky top. Blend the three and the same infrastructure can cost a fraction of an all-on-demand bill.