Arm or x86 cloud instances: what changes for your bill

By Rohit Athaide · Published 3 Nov 2026

Most cloud servers use x86 processors from Intel or AMD. Several clouds also offer Arm processors: AWS Graviton, Azure Cobalt and Ampere-based sizes, Google’s Axion and Tau T2A, and Ampere processors on Oracle Cloud. They matter for cost because they are often listed at a lower hourly rate than x86 instances of a similar size.

Why the price can differ

Vendors design or buy Arm processors for efficiency, and pass some of the saving on. Whether a given Arm size is cheaper than the x86 size you use now is not a rule, it is something to check in the price list, because families, generations and regions all differ. The architecture filter in the virtual machines catalog lists both side by side for the instances we have priced.

What you must check before moving

The price is the easy part. The work is making sure your software runs.

• Your code and its dependencies need an Arm build. Interpreted languages such as Python, Node.js, Java and Ruby usually run unchanged, but native libraries need Arm versions.
• Container images must be built for Arm, or built as multi-architecture images that carry both.
• Third-party software and agents. Monitoring agents, security tools and commercial software may not support Arm. Check each.
• Build pipelines. Your CI must produce Arm artifacts, either by building on Arm or by cross-compiling.
• Performance is workload specific. Some work runs faster on Arm per dollar, some slower. Measure your own, not a benchmark.

A safe way to try

1. Pick one stateless service.
2. Build it for Arm in your pipeline.
3. Run a small group of Arm instances beside the x86 ones, behind the same load balancer.
4. Compare latency, errors and CPU at the same traffic.
5. If it holds, move more traffic, and keep an x86 fallback until you are confident.

What not to compare

Two sizes with the same vCPU count are not the same capacity across architectures. Compare price against measured throughput, not against vCPU count alone. And do not compare a burstable size with a fixed-performance one. See burstable versus fixed performance.

Data from this site

In the catalog, the architecture is derived from the vendor’s processor name, because the vendor’s own architecture field only says 64-bit. Instances with no processor listed show as unknown and are left out of an architecture filter rather than guessed. See using the price catalog for how the filters behave, and choosing a cloud VM size for sizing in general.