AWS Graviton5 reaches GA in EC2 M9g and M9gd instances six months after re:Invent preview

AWS opened general availability of EC2 M9g and M9gd on June 10, 2026, putting Graviton5 into a workload-shaped instance family for the first time. The new chip moves to a single-socket 192-core design on 3nm, with DDR5-8800 memory and PCIe Gen 6.

By Stackmaven

AWS opened general availability for EC2 M9g and M9gd instances on June 10, 2026, the first time Graviton5 appears in a workload-shaped product rather than a re:Invent preview slot. The launch lands six months after the December 4, 2025 preview announcement at re:Invent, and the rollout starts in US East (N. Virginia and Ohio), US West (Oregon), and Europe (Frankfurt) with additional regions to follow. The headline numbers AWS published center on 25 percent better compute against Graviton4 and a 192-core single-socket design built on 3nm.

What shipped

The M9g and M9gd families are general-purpose instances with up to 192 vCPUs and 768 GiB of memory at a 1:4 vCPU-to-memory ratio. The “d” variant adds local NVMe storage up to 11.4 TB and roughly 30 percent higher IOPS than the prior M8gd line. Network bandwidth tops out at 100 Gbps with EBS bandwidth up to 72 Gbps on the largest sizes. Each instance runs on the sixth-generation Nitro System, which now includes the formally verified Nitro Isolation Engine for mathematically proven VM-to-VM isolation.

The chip itself is the structural change. Graviton5 moves to a single 192-core Neoverse V3 socket on a 3nm process, replacing Graviton4’s dual-socket pair of 96-core Neoverse V2 dies. The cache hierarchy expands by roughly 5x at the L3 level, memory steps to DDR5-8800, and the PCIe controller jumps to Gen 6. AWS positions the package as the fastest memory subsystem of any cloud CPU it offers, which is the underlying lever behind the workload-specific gains the launch post quotes: up to 35 percent on web applications, up to 35 percent on ML inference, and up to 30 percent on databases.

The customer-result roster AWS published during preview is unusually specific. ClickHouse cited a 36 percent performance boost over M8g with no code changes, Honeycomb posted a 36 percent throughput-per-core gain across a six-month A/B test on production observability workloads, and HubSpot cut MySQL query duration by up to 60 percent. Meta has signed on to deploy “tens of millions of cores” of Graviton for its agentic AI work, and Uber and Snowflake disclosed Graviton commitments in the same window. C9g (compute-intensive) and R9g (memory-intensive) variants are on the roadmap for later in 2026.

Where this lands in the market

Graviton hit 120,000 customers in this window, and AWS reports that more than 50 percent of new CPU capacity it adds is now Graviton. That share is the relevant context for the M9g positioning. The competitive read is not Graviton5 vs. Graviton4 but Graviton5 vs. the AMD EPYC and Intel Xeon families that still anchor the M-family equivalents on M7i, M7a, and M8i. AWS is no longer pitching Arm-in-cloud as a price-performance experiment; the M9g launch frames Graviton as the default general-purpose substrate, with x86 increasingly positioned as the compatibility option.

The agentic-AI framing in the launch copy is doing two things at once. First, it is real: real-time inference, code generation, and multi-step orchestration are memory-bandwidth-bound workloads, and the DDR5-8800 plus 5x L3 cache directly addresses that profile. Second, it is positioning: Graviton5 follows Trainium3 UltraServers and the Nova 2 model family from re:Invent 2025 and the OpenAI-on-Bedrock GA from June 1. The throughline is AWS arguing that the full AI stack from silicon to model can live on first-party infrastructure, with third-party model access (Anthropic, OpenAI, Mistral) layered on top through Bedrock.

For platform teams already running Graviton, the rollout path is the same as every prior generation: targeted A/B tests against the new SKUs on the workloads that show the most memory or cache pressure, then a phased migration as the C9g and R9g variants land. The HubSpot and ClickHouse numbers suggest databases and analytic engines are the fastest payoff. For teams still on Intel or AMD M-family by default, the question is no longer whether to consider Graviton but whether the next reservation cycle still defaults to x86.

What’s worth watching

C9g and R9g cadence. AWS named both for later in 2026 without committing dates. The compute-intensive line is the one that will surface clearly in ML inference and HPC workloads; memory-intensive R9g is the one large in-memory database operators will want timed against renewal cycles.
Spot and Savings Plan pricing curves. The preview pricing skewed favorable to early adopters. The first Savings Plan commitment windows on M9g will set the operational economics for the rest of the year, and they will tell how aggressively AWS is pricing Graviton5 against the Intel and AMD comparables on its own floor.
Meta deployment proof points. A tens-of-millions-of-cores commitment from a hyperscale consumer is the kind of disclosure that typically lands at re:Invent. If Meta or Snowflake publishes a workload breakdown before then, it will be the strongest external read on Graviton5’s agentic-inference fit.

The plain frame is that Graviton5 is no longer the experimental custom-silicon story; it is the default. AWS is asking customers to treat the M9g launch as the new baseline for general-purpose compute, with x86 alternatives reserved for compatibility or specific proprietary-licensed workloads. The next six months will show whether the rest of the EC2 fleet lines up behind that framing as planned.

Sources cited