Reserved Instances vs Spot — Commitment Trade-offs

The Three Pricing Models#

Every major public cloud — and most GPU clouds — offers three broad pricing models for compute. The trade-off is always between commitment, discount, and risk of interruption.

Reserved Instances#

A Reserved Instance (RI) is a commitment to pay for a specific instance type in a specific region for either one or three years. In exchange, the provider gives a substantial discount against on-demand pricing. RIs are useful for workloads with predictable steady-state demand — production inference services, baseline training capacity, always-on databases.

RIs are rigid: if your workload moves to a different instance type or region, the reservation either has to be exchanged (where supported) or sits unused. AWS Convertible RIs offer some flexibility at a slightly lower discount; standard RIs are non-convertible.

• 1-year RI — typically 30-45 % off on-demand depending on payment option (no upfront, partial, all upfront).
• 3-year RI — typically 50-65 % off on-demand; suitable for long-lived stable workloads.
• All-upfront RIs give the deepest discount but tie up the most capital.
• Reservation must match instance family, size, region, and tenancy to apply.

Savings Plans#

AWS Savings Plans are a commitment to a level of compute spend per hour (in dollars), rather than to a specific instance. The discount is similar to RIs, but the commitment is portable: as long as you spend at least your committed dollars per hour on covered services, the discount applies regardless of which instance type or region you actually use.

Compute Savings Plans cover EC2, Fargate, and Lambda across all regions and instance families. EC2 Savings Plans are cheaper but lock you to an instance family and region. GCP Committed Use Discounts and Azure Reservations behave similarly to RIs; Azure also offers Savings Plans for Compute.

Spot / Preemptible Instances#

Spot capacity is unused server capacity that providers sell at a steep discount in exchange for the right to reclaim it. AWS gives a two-minute interruption notice; GCP Preemptible gives 30 seconds; Azure Spot gives 30 seconds.

Spot is exceptional value for workloads that can tolerate interruption — fault-tolerant distributed training with checkpointing, batch inference, CI/CD runners, large-scale data processing. It is a bad fit for stateful single-instance services or workloads that cannot recover quickly from a kill.

• Use checkpointing for training jobs so a reclaim costs minutes, not hours.
• Mix spot with a small on-demand or reserved baseline so service continues during a reclaim wave.
• Diversify across instance types and zones — spot capacity is per-pool.
• Set a maximum price to cap exposure if spot rates spike.

Workload-to-Pricing Mapping#

Commitment Strategy#

A defensible commitment strategy starts from a clear understanding of baseline utilisation — the GPU-hours you know you will consume regardless of business variability. Cover the baseline with reservations or savings plans, the middle with on-demand, and the volatile top of the curve with spot where tolerable.

Yobitel's reserved-capacity model is closer to a Savings Plan than a classic RI: customers commit to a compute envelope across H100, H200 and B200 capacity at a published discount tier, with capacity guaranteed under the sovereign-residency contract.