Compute that runs the same workload for thirty percent less.

AL2 was sliding toward end-of-life and the team had been deferring the migration because "ARM is a separate problem." We rolled them together. AL2023 itself is a relatively clean migration; the ARM transition needed careful workload classification.

Each workload class got benchmarked on Graviton in a staging environment with production-shape traffic. About two-thirds passed cleanly — same or better performance, same or better latency, sometimes a configuration tweak. The third that didn’t (older Java workloads with native bindings, some Rust workloads with x86-specific intrinsics) stayed on x86 with AL2023.

On the migrated 64% of the fleet the compute bill dropped 31%. Performance came out slightly ahead at the median. The AL2 end-of-life exposure is gone proactively rather than reactively. The team plans to revisit the held-back workloads in six months as ARM ports of their dependencies mature.