LayerA quick scan findssmalBlu goes further
CommitmentsA Savings Plan, based on the last few weeks.Rightsize first, then commit, so you never lock in waste. Plans staggered to follow your usage down, with coverage and utilization tracked daily.
ComputeServers with low average CPU.Memory and peak load over months, so busy seasons stay safe. Newer and cheaper server types, Spot with automatic fallback, test systems off at night.
StorageUnattached disks and old disk types.Paid-for speed nobody uses, orphaned snapshots and images, data tiered by how it is really read, old versions and logs kept forever.
Network and data transferIdle NAT gateways and unused IP addresses.Traffic taking the expensive route: storage traffic through a NAT gateway instead of a free endpoint, chatty services split across zones, replication nobody reads.
DatabaseOversized instances, for a few engines.Capacity modes and table classes, old versions paying for extended support, replicas nobody reads, production-size databases in test, and the slow queries that force a bigger instance.
KubernetesUsually nothing. Native tools do not size workloads inside clusters.Requests and limits set from real usage, pods packed onto fewer nodes, Spot node pools, forgotten preview environments, clusters paying for extended support.
GPUIdle endpoints, if GPU metrics are switched on.GPUs sliced and shared for small jobs, idle notebooks off, the right chip for each workload, reservations sized to real demand.
Tagging and allocationSpend by service and account, if teams tagged things.Untagged and shared costs split by rule, without retagging everything. Cost per team, product and customer, with alerts sent to the owner.