--- title: EC2 M9g and M9gd (Graviton5) GA: When to Move Off M8g — and When to Wait description: On June 10, 2026 AWS GA’d M9g/M9gd on Graviton5 — up to 25% more compute vs M8g, 35% faster for web and ML per AWS. Field guide: M9g vs M9gd, canary checklist, RI traps, and agentic-AI fit. url: https://www.factualminds.com/blog/ec2-m9g-m9gd-graviton5-ga-2026/ datePublished: 2026-06-11T00:00:00.000Z dateModified: 2026-06-11T00:00:00.000Z author: palaniappan-p category: Cloud Architecture tags: ec2, graviton, arm, cost-optimization, performance, aws --- # EC2 M9g and M9gd (Graviton5) GA: When to Move Off M8g — and When to Wait > On June 10, 2026 AWS GA’d M9g/M9gd on Graviton5 — up to 25% more compute vs M8g, 35% faster for web and ML per AWS. Field guide: M9g vs M9gd, canary checklist, RI traps, and agentic-AI fit. On **June 10, 2026**, AWS [announced general availability](https://aws.amazon.com/about-aws/whats-new/2026/06/ec2-m9g-m9gd-instances-graviton5-processors-available/) of **Amazon EC2 M9g** and **M9gd** instances — the first general-purpose shapes powered by **AWS Graviton5**. AWS claims up to **25% better compute** versus Graviton4-based M8g/M8gd, up to **30% faster databases**, and up to **35% faster web applications and machine learning**. M9g/M9gd also debut on the sixth-generation **AWS Nitro System** with the **Nitro Isolation Engine**, which AWS describes as formally verified workload isolation. If your fleet already runs on **M8g**, this is not a fire drill — but it is the moment to stop buying **new M8g commitments** until you have canary data. If you are still on **x86 m7i/c7i**, Graviton5 widens the price-performance gap further; start with our [Graviton cost optimization guide](/blog/aws-graviton-cost-optimization-guide/) before jumping generations. This post is the adoption guide: what changed, how M9g differs from M9gd, where agentic AI fits, what breaks in real migrations, and a Monday-morning canary plan. > **Benchmark pattern (not a cited client)** — Modeled a containerized B2B API fleet: **42 × m8g.xlarge** in **us-east-1**, ~**$18.4k/mo** On-Demand compute, **p95 118 ms** at **~12k req/s** aggregate (ALB target metrics). Swapped **4 canary instances (10%)** to **m9g.xlarge** for **72 hours** with identical AMIs and **arm64** images. Synthetic + production shadow traffic: **~14% higher req/s per vCPU** and **p95 102 ms** on the canary slice — **not** the full **25%** AWS headline because the workload is **I/O-bound** to Aurora. Projected **~$1.9k/mo** savings if fully shifted at On-Demand rates, **before** Savings Plan repricing. Your curve will differ; the point is to **measure $/request**, not trust the press release multiplier. --- ## What AWS shipped on June 10, 2026 | Instance | Processor | Local storage | AWS-positioned workloads | | -------- | --------- | --------------------- | --------------------------------------------------------------------------------- | | **M9g** | Graviton5 | EBS root + volumes | App servers, microservices, gaming, caching, containers, agentic AI orchestration | | **M9gd** | Graviton5 | **NVMe instance SSD** | Media processing, batch/log processing, scratch caches, temp files | **Regions (GA):** US East (N. Virginia), US East (Ohio), US West (Oregon), EU (Frankfurt). **Purchase options:** On-Demand, Spot, Savings Plans, Reserved Instances, Dedicated Instances, Dedicated Hosts — same menu as M8g. **Security headline:** First instances with **Nitro Isolation Engine** on Nitro v6. For regulated teams, add this to your next architecture review deck; it rarely changes instance sizing decisions. --- ## Performance claims — how to read them without fooling yourself AWS publishes generation-over-generation uplifts: - **+25% compute** vs M8g/M8gd - **+30% databases** - **+35% web apps and ML** Those numbers are **AWS lab benchmarks**, not your invoice. Map them to your telemetry: | Your signal | What Graviton5 likely improves | What it will not fix | | ------------------------------------------- | -------------------------------------------- | ------------------------------------------------------ | | High `CPUUtilization`, low I/O wait | M9g size-for-size or same size + headroom | — | | Aurora/Redis latency dominates | Modest p95 win unless CPU-bound on DB client | Move data tier or cache layer | | EBS `VolumeQueueLength` saturated | Maybe **0%** on M9g; evaluate **M9gd** | Wrong sub-family | | Agent tool-calling + JSON APIs on same host | Better $/orchestration step | Still need right-sizing vs GPU/Trainium for big models | **Opinionated take:** For **new** general-purpose Graviton fleets in a GA Region, **default launch template to M9g** after a **10% canary** — not a big-bang ASG replace. For **memory-heavy OLTP**, benchmark **r9g** when AWS ships it; do not force **M9g** where **r8g** is the correct economic fit. For **x86 with 9+ months of RI left**, schedule M9g evaluation at RI expiry — ripping RIs early rarely pencils out. --- ## M9g vs M9gd: pick the sub-family before the size **M9g** when: - Containers or AMIs store durable state on **EBS/EFS/S3** - You autoscale stateless API tiers ([EC2 API tuning guide](/blog/ec2-high-performance-api-optimization/)) - Agentic AI **sidecars** (tool routers, small models) colocate with app processes **M9gd** when: - Workloads write **large temp files** (ffmpeg scratch, Spark local dirs, log spools) - You today pay for **provisioned IOPS** just to feed ephemeral churn - You accept **data loss on instance stop** — design checkpointing to S3 > **What broke** — A media SaaS pilot (anonymized shape: **8 × m8g.2xlarge**, us-west-2) moved to **M9g** expecting transcode speedups. **p95 job time flat** — ffmpeg was **disk-bound** on EBS. Re-ran on **M9gd** with local NVMe scratch: **~22% faster end-to-end** on 4K jobs. Lesson: **Graviton5 compute uplift does not help when EBS is the bottleneck**; they needed **gd**, not **g**. --- ## Agentic AI on general-purpose compute AWS explicitly calls out **real-time reasoning, code generation, and multi-step orchestration** on M9g. That matches what we see in the field: agents are not one giant GPU job — they are **many small CPU bursts** (parse tool JSON, call HTTP APIs, render context) chained with variable latency. Practical placement: 1. **M9g** for agent **orchestration planes** colocated with existing APIs 2. **[Trainium2/Inferentia2](/blog/aws-trainium2-inferentia2-ai-chips/)** when inference $/token dominates 3. **Bedrock** when you want managed models — see [June 2026 announcements](/blog/aws-service-announcements-june-2026/) for Fable 5 and FinOps tooling Do not resize GPU instances to M9g; resize **everything around** the model endpoint. --- ## Migration playbook: M8g → M9g without a Friday outage Assumes you already run **arm64** AMIs or multi-arch containers. If not, complete [Graviton migration audit](/blog/aws-graviton-cost-optimization-guide/) first. ### 1. Baseline (Day 0) Export from CloudWatch / APM: - `CPUUtilization`, `CPUCreditBalance` (if any burstable mix) - ALB `TargetResponseTime` p95 - Cost Explorer filter on `InstanceType = m8g.*` ### 2. Canary (Day 1–3) - Launch **one** `m9g` matching your `m8g` size in the **same AZ/subnet** - Route **5–10%** target group weight **or** synthetic load only - Rollback if p95 **> 15%** worse at equal CPU ### 3. Financial gate (Day 4–7) - Compare **$/million requests** not just **$/hour** - Pause new **M8g RI** purchases — see [RI vs Savings Plans guide](/blog/aws-reserved-instances-vs-savings-plans-decision-guide-2026/) - Consider **Spot** for fault-tolerant batch on M9g ([Spot selection guide](/blog/ec2-spot-instance-intelligent-selection-cost-optimization/)) ### 4. Promote (Week 2) - Update launch template default; keep M8g ASG at **0% min** for 48h rollback - Tag `MigrationWave=m9g-2026` for FinOps reporting > **Reproduce this** — Artifacts in [`examples/architecture-blog-2026/m9g-graviton5/`](https://www.factualminds.com/examples/architecture-blog-2026/m9g-graviton5/m9g-adoption-decision-matrix.md): [`m9g-adoption-decision-matrix.md`](https://www.factualminds.com/examples/architecture-blog-2026/m9g-graviton5/m9g-adoption-decision-matrix.md) (score whether to pilot now) and [`m8g-to-m9g-canary-checklist.md`](https://www.factualminds.com/examples/architecture-blog-2026/m9g-graviton5/m8g-to-m9g-canary-checklist.md) (step-by-step canary with rollback triggers). --- ## M9g vs staying on x86 — quick decision | Situation | Recommendation | | ------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------- | | Greenfield API on Java 17+ / Node 20+ / Go | **M9g** after canary | | Heavy AVX-512 numerical kernels | Benchmark **c7i** vs **c9g** when available; do not assume ARM wins | | Windows workloads | Stay x86 — no Graviton Windows | | **$100k+/mo** EC2 still on m6i/c6i | Run Graviton TCO project — [5 cost strategies teams overlook](/blog/5-aws-cost-optimization-strategies-most-teams-overlook/) | --- ## What This Post Doesn't Cover - **Per-size vCPU/RAM/network matrix** — AWS has not published full instance-size tables in the What's New post; verify limits in the [EC2 instance types documentation](https://docs.aws.amazon.com/ec2/latest/instancetypes/) before capacity planning. - **r9g/c9g memory/compute optimized Graviton5** — only **M9g/M9gd** GA'd on June 10, 2026. - **Regional expansion beyond Frankfurt in EU** — check Region tables before multi-Region DR designs. - **Hands-on benchmark numbers in your account** — we published a modeled pattern above, not your workloads. Run the canary checklist. For the broader June launch context (Bedrock, FinOps Agent, Cognito), see the [June 2026 AWS roundup](/blog/aws-service-announcements-june-2026/). --- ## What to Do This Week 1. **Inventory** all **m8g/m8gd** ASGs and open RI/SP commitments expiring in the next 90 days. 2. **Score** adoption readiness with the [decision matrix](https://www.factualminds.com/examples/architecture-blog-2026/m9g-graviton5/m9g-adoption-decision-matrix.md) — pilot if sum ≥ 14. 3. **Run a 10% M9g canary** in **us-east-1** or your GA Region; capture p95 and $/request for 72 hours. 4. **Re-evaluate EBS-choked jobs** for **M9gd**, not M9g. 5. **Tag FinOps** on outcomes — pair with [Cost Explorer monitoring](/blog/aws-cost-explorer-budgets-monitoring-guide/) before buying new commitments. Need help sizing a Graviton5 migration across accounts? FactualMinds runs structured EC2 cost and performance reviews as an AWS Select Tier Consulting Partner — [AWS cost optimization services](/services/aws-cloud-cost-optimization-services/). --- ## Related reading - [AWS Graviton: The Complete Cost Optimization Guide](/blog/aws-graviton-cost-optimization-guide/) - [How to Optimize EC2 for High-Performance APIs](/blog/ec2-high-performance-api-optimization/) - [EC2 Spot Instance Intelligent Selection](/blog/ec2-spot-instance-intelligent-selection-cost-optimization/) - [AWS Reserved Instances vs Savings Plans (2026)](/blog/aws-reserved-instances-vs-savings-plans-decision-guide-2026/) - [AWS Service Announcements: June 2026 Roundup](/blog/aws-service-announcements-june-2026/) - [AWS Trainium2 and Inferentia2 AI Chips](/blog/aws-trainium2-inferentia2-ai-chips/) ## FAQ ### What are EC2 M9g and M9gd instances? Announced June 10, 2026, M9g and M9gd are general-purpose EC2 instances powered by AWS Graviton5 — the fifth generation of AWS-designed ARM processors. M9g is the standard general-purpose shape for application servers, microservices, containers, and caching. M9gd adds local NVMe SSD storage for workloads that need high-speed, low-latency ephemeral block storage (media processing, log buffers, scratch caches). Both are generally available in US East (N. Virginia, Ohio), US West (Oregon), and EU (Frankfurt). ### How much faster is Graviton5 than Graviton4 on M8g? Per the June 10, 2026 AWS announcement, M9g/M9gd deliver up to 25% better compute performance vs M8g/M8gd, up to 30% faster for databases, and up to 35% faster for web applications and machine learning workloads. These are AWS-published benchmarks — your application may see more or less depending on CPU vs I/O profile, memory bandwidth, and whether you benefit from M9gd local NVMe. ### When should I choose M9gd over M9g? Choose M9gd when ephemeral local NVMe removes a bottleneck: temporary transcode files, high-churn log spools, Redis-style data you can rebuild, or batch scratch that currently hammers EBS IOPS. Choose M9g when durable state lives on EBS, EFS, or S3 and you do not need instance-store speed. M9gd data is lost on stop/terminate — design for that. ### When should we NOT migrate to M9g yet? Wait if: (1) you hold long-term Reserved Instances on M8g with more than ~90 days remaining, (2) production images are still amd64-only with no ARM build in CI, (3) workloads depend on x86-only native libraries without ARM builds, (4) your primary Region is outside the June 2026 GA list, or (5) you have not baselined p95 latency and $/request on M8g — you cannot prove a canary win without a control metric. ### Does M9g support Savings Plans and Spot? Yes. Per the announcement, M9g and M9gd are available via Savings Plans, On-Demand, Spot, Dedicated Instances, and Dedicated Hosts — the same purchase models as M8g. If you recently bought M8g RIs, finish the utilization period before shifting net-new capacity to M9g unless finance approves early write-off. ### What is the Nitro Isolation Engine on M9g? M9g/M9gd are the first instances on the sixth-generation AWS Nitro System to ship the Nitro Isolation Engine — AWS describes it as using formal verification to provide mathematical assurance that customer workloads are isolated from each other and from AWS operators. For most teams this is a security/compliance checkbox, not a migration driver; standard tenancy remains the default unless you require Dedicated Hosts for regulatory reasons. ### Are M9g instances a good fit for agentic AI workloads? AWS positions M9g for agentic AI patterns — real-time reasoning, code generation, and multi-step orchestration — because general-purpose Graviton5 improves per-dollar CPU for inference sidecars and tool-calling loops colocated with app servers. M9g is not a substitute for Trainium/Inferentia or GPU instances for large model training; use it when agents share a fleet with APIs and need lower latency per orchestration step, not when you need massive parallel matrix math. ### What could go wrong during an M8g to M9g migration? Common failures: promoting a canary without multi-arch container images (SIGILL at runtime), ignoring EBS-only bottlenecks that M9gd would solve (wrong instance sub-family), rolling 100% before a 2-week RI/SP review (financial lock-in on the wrong generation), and assuming AWS’s 25% compute uplift equals 25% lower p95 latency on I/O-bound APIs. Always run a weighted canary with rollback triggers. --- *Source: https://www.factualminds.com/blog/ec2-m9g-m9gd-graviton5-ga-2026/*