Graviton instances sell more throughput per dollar for many server workloads. A JVM on aarch64 is a first-class OpenJDK port, not an emulator. The migration dies on JNI: compression libs, crypto accelerators, brotli, protobuf native, old libsigar, and that one vendor .so from 2017. If your Docker image is amd64 on an ARM node, you are paying qemu and you will not hit the price/performance slide.
Multi-arch is the real project
Build linux/amd64 and linux/arm64 images in CI. Pin a JDK that publishes AArch64 (Temurin, Corretto, and friends). Test C2 compilation, GC (G1/ZGC on ARM), and vector/crypto intrinsics — some AES paths differ. JMH on x86 is not a Graviton result. Run the canary on m7g with production GC flags.
FROM eclipse-temurin:21-jre-jammy
# image must be built for TARGETARCH, not copied from amd64Netty native (epoll, tcnative/boringssl) must match arch. Alpine musl vs glibc surprises exist on ARM too. Java agents (coverage, APM) need ARM builds; an x86 agent will fail the container.
Performance shape
More cores, often better watts. Single-thread turbo comparisons versus x86 SKUs vary by generation. Tail latency can improve or worsen depending on cache and NUMA (Graviton chips have their own topology). Disable UseContainerCpuShares folklore and set CPU requests honestly. Huge pages and transparent hugepages policies still matter.
JNI that falls back to pure Java is a silent tax. Look for unsatisfied link at boot, not in a sampling of happy logs. os.arch in a debug endpoint saves hours.
Rollout
Canary a stateless service first. Databases and anything with C++ extensions next. Spot/Graviton mix in a cluster needs a node selector, not hope. Some ISV licenses are still x86-only.
Graviton is not "the JVM is slower on ARM." It is "your native surface is the risk." Budget a week for dependency archaeology, not an afternoon for an instance type change.
Read AWS's Graviton getting-started and your JDK vendor's AArch64 release notes. Then file every .so in the image. The first x86 ELF is the migration.