Compute service, purchasing, and edge/local economics

Do not apply one purchasing model to an entire platform. Separate noninterruptible baseline, uncertain interactive peaks, checkpointable batch, scheduled development, licensed hosts, and site-local processing. Rightsize each segment first; a discount on an oversized resource preserves waste. On-Demand capacity retains flexibility for new or uncertain demand. Spot fits work that can tolerate interruption, retry from a checkpoint, and diversify across instance pools and Availability Zones. Savings Plans exchange a measured hourly commitment for eligible usage discounts, while EC2 Reserved Instances have scope-specific billing behavior. Capacity Reservations solve capacity assurance and are not themselves the same as discounts. Dedicated Hosts earn their cost only when host affinity, socket or core licensing, or tenancy visibility is contractual. Estimate the conservative baseline over the term and leave forecast error or one-off peaks elastic. Committing to the observed maximum is not optimization when demand can fall. Likewise, an all-Spot design is invalid if losing capacity violates the service objective.

Lambda fits bounded invocation work when supported runtime, resource, duration, and event semantics match. Its per-request and duration model can avoid idle fleets, but sustained high utilization, downstream limits, logging, and concurrency still require modeling. Fargate runs containers without managing the underlying hosts, trading host operations for task-resource pricing and platform constraints. EC2 remains appropriate when the workload needs OS, kernel, specialized hardware, custom images, persistent processes, or detailed host control. Container packaging alone does not guarantee lower cost; oversized task requests or an idle EC2 cluster can both strand capacity. Compare successful work per dollar together with patching, deployment, observability, and failure headroom. A service boundary can reverse as demand stabilizes: an intermittent function may later justify a continuously utilized container, while an infrequent administrative service may move away from an always-on instance. Avoid selecting a managed option merely because it is labeled serverless, and avoid self-management merely because the raw instance rate looks lower.

Regional compute is the baseline when connectivity, latency, residency, and upstream transfer are acceptable because pooled capacity avoids duplicating operations at every site. Local processing becomes valid when a measured response must occur during disconnection, raw data must be filtered before transfer, a residency rule keeps data on premises, or round-trip latency violates the application. Lightweight device or gateway processing and an Outposts deployment represent different operating models; Outposts should not be chosen as a generic transfer appliance or discount. Model site count, local steady capacity, spare parts or service commitment, deployment, patching, observability, configuration rollout, reconnection synchronization, and the Regional processing that remains. For a factory safety loop, continuing through a WAN outage can outweigh duplicated capacity. If connectivity later becomes reliable and only small aggregates are processed, Regional services may regain the advantage. Preserve independent controls and logs so local autonomy does not become unmanaged drift. The purchase decision follows this placement boundary: commitments belong only to stable measured capacity in the chosen locations, while uncertain bursts remain flexible.