AWS Graviton5-powered M9g and M9gd instances reach general availability

AWS has opened general availability for two new EC2 instance families — M9g and M9gd — built on the Graviton5 processor that was previewed at re:Invent 2025. The move takes the fifth-generation custom silicon out of early access and into standard purchasing channels across four AWS regions: US East (N. Virginia), US East (Ohio), US West (Oregon), and Europe (Frankfurt).

What changed in the silicon

Graviton5 is the first AWS processor to use PCIe Gen6 and DDR5-8800 memory, giving it what AWS describes as the highest memory bandwidth among processor instances currently offered in the cloud. The chip scales to 192 cores and carries five times the L3 cache of Graviton4, with inter-core latency reduced by up to a third. Those cache and latency figures are particularly relevant for workloads that shuttle data between threads frequently — databases, JVM-based services, and the orchestration layers common in multi-step AI pipelines.

On the I/O side, M9g and M9gd instances add roughly 15% more network bandwidth and 20% more EBS bandwidth on average across sizes compared to their M8g predecessors, with the largest 48xlarge configuration reaching 100 Gbps network and 72 Gbps EBS. The families also support Instance Bandwidth Configuration, which lets operators shift up to 25% of total bandwidth allocation between EBS and VPC networking without resizing the instance — useful for tuning database write-heavy vs. read-heavy phases.

A new hypervisor isolation model

The M9g launch introduces the Nitro Isolation Engine, a component AWS is positioning as an industry first. Rather than relying solely on conventional testing, the engine uses formal verification — mathematical proof techniques — to demonstrate that the hypervisor correctly enforces memory, CPU register, and I/O boundaries between virtual machines. AWS states this makes Nitro the first formally verified cloud hypervisor. For operators running multi-tenant or regulated workloads, this provides a qualitatively different class of assurance than coverage-based QA alone.

Customer benchmarks from the preview period

Three early adopters published production results during the preview. ClickHouse reported a 36% throughput gain over M8g with no code modifications. Honeycomb ran a six-month A/B test against Graviton4 production infrastructure and measured 36% better throughput per core. HubSpot deployed M9g for MySQL and saw query durations fall by as much as 60%. AWS also disclosed that Meta is adopting Graviton at scale — starting with tens of millions of cores — to support agentic AI workloads, making it one of the platform's largest customers.

Relevance for infrastructure teams

For teams already on M8g or earlier Graviton instances, the upgrade path is straightforward: Graviton5 maintains Arm binary compatibility, and AWS offers the Transform service to automate x86-to-Graviton code migration for Java applications. The Graviton Savings Dashboard can be used to model cost differences before committing.

The M9gd variant adds local NVMe storage for workloads that benefit from low-latency scratch space — caching layers, log pipelines, and key-value stores — ranging from 59 GB on the medium size up to three 3.8 TB drives on the 48xlarge and bare-metal configurations.

Both families now sit within the standard EC2 purchasing model. AWS notes Graviton currently underpins more than 350 instance types serving over 120,000 customers, giving the ecosystem a wide base of compatible AMIs, managed services, and third-party tooling.