Graviton
On June 10, 2026, AWS made its fifth-generation Graviton processor generally available, alongside the new EC2 M9g and M9gd instances that run on it. The headline number, 25% better compute performance compared to Graviton4, sounds like a routine product release. It is not.
If you run a company on AWS, this isn’t a technical news item to scroll past. Graviton is now the path AWS quietly pushes its customers down, and the gap between teams that follow it and teams that don’t is widening every release. Here is what’s actually in this announcement, what it changes for your bill, and what the catch is.
The M9g instances run on Graviton5, AWS’s latest custom ARM-based processor. The numbers AWS publishes, all sourced from their official documentation and benchmarks:
Up to 25% better compute performance compared to Graviton4-based M8g instances at the same size. Up to 35% faster for web applications and machine learning inference. Up to 30% faster for databases. 192 cores per chip (the highest core density in EC2). 5x larger L3 cache than Graviton4. DDR5-8800 memory (the fastest memory of any cloud processor) and PCIe Gen6 support.
Real customer benchmarks released alongside the launch are even stronger than the AWS-published averages: HubSpot reports MySQL query duration dropping by up to 60%. ClickHouse achieved a 36% performance boost with zero code changes. Honeycomb, after a six-month A/B test in production observability workloads, saw 36% better throughput per core. Airbnb measured up to 20% improvement over Graviton4 on production search workloads. Meta, the largest Graviton customer publicly known, has committed to deploying tens of millions of cores for its agentic AI infrastructure.
M9g and M9gd are available in US East (N. Virginia, Ohio), US West (Oregon), and EU (Frankfurt) at launch. For now, no UAE, no Bahrain, no Singapore, no other European regions. That matters if you’re running in Dubai or APAC.
There are three reasons this release is structural for decision-makers, not just a new bullet point in your AWS roadmap.
1. The price-performance gap is now too large to ignore. Each Graviton generation has narrowed the case for x86 instances. With Graviton5 delivering 25% more compute at the same price as Graviton4 (and 20-30% less than the equivalent Intel or AMD instance), staying on x86 by default isn’t a technical choice anymore. It’s an absence of choice. Over a year on a meaningful compute footprint, that gap translates into tens of thousands of dollars left on the table per fleet.
2. AWS is making its own bet visible. When the largest cloud provider builds its fifth-generation custom silicon and lines up Meta, Uber, Snowflake, Airbnb, HubSpot, Atlassian, and Cockroach Labs as launch testimonials, the message isn’t subtle. AWS is signaling where its investment, optimization, and future services will go. The agentic AI workloads it explicitly designed Graviton5 for, real-time reasoning, code generation, multi-step orchestration, are exactly the workloads scaling teams will run more of in the next two years. The platform you run on tomorrow is not the platform you run on today.
3. The Nitro Isolation Engine raises the security bar. Less discussed in the launch coverage, but important for any company tracking SOC 2, ISO 27001, or HIPAA: M9g introduces the Nitro Isolation Engine, the first formally verified cloud hypervisor. In plain terms, AWS now provides a mathematical proof that workloads on the same physical host cannot leak data to each other. For a fintech or healthtech preparing a certification audit, that’s the difference between “we trust the provider’s controls” and “the provider has proven its isolation in formal logic.”
Let’s be honest. Graviton migrations aren’t free, and three caveats deserve real attention.
First, ARM64 compatibility isn’t universal. Modern Java, Python, Node.js, Go, and Ruby workloads typically run on Graviton with zero code changes. Containerized applications using multi-arch images deploy without trouble. But proprietary compiled SDKs, legacy x86 cryptographic binaries, certain commercial databases, and anything depending on x86-specific assembly will not run. A pre-production validation window of 7 to 14 days is the standard practice, and skipping it is the most common reason migrations get rolled back.
Second, Windows workloads stay on x86. Graviton only supports Linux-based operating systems. If you run Windows Server workloads, this release changes nothing for that part of your fleet.
Third, availability is still limited. Three US regions plus Frankfurt at launch. If your production runs in eu-west-3 (Paris), me-south-1 (Bahrain), me-central-1 (UAE), or any APAC region, you wait. AWS typically rolls new instance types out across regions over six to nine months, but until your region gets M9g, the savings stay theoretical.
If your fleet is on m5, m6i, or m7i and your workloads are ARM-compatible, here is what a methodical move to M9g looks like in numbers.
On a 100-instance fleet of m5.xlarge running steady-state, the per-vCPU cost difference between m5 and m9g is roughly 30% in favor of m9g, once you factor in both the price drop from older generations and the 25% performance gain (meaning you can size down without losing throughput). On a workload billing $30,000 a month on m5 compute, that’s roughly $9,000 a month recovered, or $108,000 a year, on a migration that doesn’t change your application code.
Even teams already on Graviton4 (m8g) have a case to look at. The 25% performance gain at the same price means smaller instance sizes for the same throughput, which compounds with reserved capacity and Savings Plans.
That said, the savings are real only if the sizing is right. Migrating to m9g without rerunning Compute Optimizer first means committing to whatever your previous oversizing was. The right sequence is the same as for any quick win: rightsize first, then migrate to the current generation, then commit to a Savings Plan against the new, lower baseline.
Graviton5 is not the AWS announcement that will generate the most LinkedIn buzz this month. The agentic AI launches from the same Summit will absorb more attention. But for any company running real workloads on AWS, this is the release that moves the bill, quietly, the most.
The teams that act on it within the next two quarters will recover 20 to 30 percent of their compute spend with no architecture change. The teams that don’t will keep paying 2019 prices on 2026 infrastructure, while their competitors on Graviton5 ship faster and run cheaper. Like every Graviton generation before, the technology is in place. The only thing standing between most accounts and the savings is the decision to actually go look.
A project, an infra drifting, a bill spiraling. Tell us what's blocking you in Miami, we'll come back with a plan.
