While AWS’s Trainium and Inferentia chips get most of the attention as the hardware behind Amazon’s AI ambitions, a quieter custom-silicon program has been reshaping something more fundamental: the general-purpose CPU inside ordinary EC2 instances. That’s Graviton — AWS’s Arm-based processor family — and by 2026 it’s running a meaningful share of all new compute AWS customers deploy. Here’s what Graviton actually is, how its five generations compare, and what AWS’s own published numbers say about the price-performance case for switching from x86.
At a Glance
| Spec | Detail |
|---|---|
| What it is | AWS’s custom Arm-based CPU family for general-purpose EC2 compute |
| Current generation | Graviton5 (192 cores, Arm Neoverse V3, 3nm process, GA June 2026) |
| Cost claim vs. comparable x86 EC2 | Up to 20% lower cost (AWS’s own published figure) |
| Energy claim vs. comparable x86 EC2 | Up to 60% less energy use (AWS’s own published figure) |
| Adoption | More than 120,000 AWS customers building on Graviton, per AWS |
| Key instance families | T4g, M6g/M7g/M8g/M9g, C6g/C7g/C8g, R6g/R7g/R8g, X2gd/X8g, I4g/I8g, and more |
| Conceptual competitors | Google Axion (Google Cloud), Nvidia Grace (hybrid CPU/GPU systems) |
What AWS Graviton Is
Graviton is AWS’s own line of server CPUs built on the Arm instruction set rather than the x86 architecture that Intel and AMD chips use. Unlike Trainium (built for AI training) or Inferentia (built for AI inference), Graviton is a general-purpose processor — it runs ordinary EC2 virtual machines, the kind that host web servers, databases, microservices, and increasingly, agentic AI application logic that isn’t itself doing the heavy model math. AWS designs Graviton in-house and only offers it inside its own infrastructure; you can’t buy a Graviton chip for your own data center the way you can buy an Nvidia GPU or an AMD EPYC chip.
The pitch, repeated consistently across every Graviton generation, is “best price-performance” rather than “fastest at any cost.” AWS backs that with two headline figures on its official Graviton page: Graviton-based instances cost up to 20% less than comparable x86-based EC2 instances, and use up to 60% less energy for the same work.
Graviton Generations Compared
| Generation | Announced | Cores | Core architecture | Representative EC2 families |
|---|---|---|---|---|
| Graviton (1st gen) | 2018 | 16 | Arm Cortex-A72 | A1 |
| Graviton2 | 2019 | 64 | Arm Neoverse N1 | M6g, C6g, R6g, T4g, X2gd, I4g |
| Graviton3 / 3E | 2021 / 2022 | 64 | Arm Neoverse V1 | C7g, C7gd, M7g, M7gd, R7g, R7gd, C7gn, Hpc7g |
| Graviton4 | 2023 | 96 per socket | Arm Neoverse V2 | C8g, M8g, R8g, X8g, I8g |
| Graviton5 | 2025 (GA June 2026) | 192 | Arm Neoverse V3 | M9g, M9gd |
Graviton5 is the newest and largest jump in the lineup: 192 cores on a 3nm process, a roughly 5x larger L3 cache than Graviton4 (2.6x more cache per core), up to 33% lower inter-core latency, DDR5-8800 memory, and PCIe Gen 6 I/O. It’s built on AWS’s sixth-generation Nitro System, which handles virtualization, storage, and networking offload outside the main CPU — a security and performance architecture AWS has carried through every Graviton generation. According to AWS, Graviton5-based M9g instances deliver up to 25% better compute performance than Graviton4, with web applications and ML inference workloads seeing roughly 35% improvements and databases around 30%.
Which EC2 Instance Families Use Graviton
AWS maps each Graviton generation onto a parallel set of EC2 instance families alongside their x86 equivalents — a “g” suffix marks the Arm/Graviton version (for example, M7g is the Graviton3 counterpart to the x86 M7i). The main families span three traditional categories:
- General purpose (M-series): M6g through M9g — balanced compute and memory for web servers, app servers, and mid-size databases.
- Compute optimized (C-series): C6g through C8g — higher compute-to-memory ratio for batch processing, media encoding, and scientific modeling.
- Memory optimized (R-series/X-series): R6g through R8g, plus X2gd and X8g — for in-memory databases and caching layers.
- Storage/local-NVMe optimized: I4g/I8g and the “d” variants like M9gd, which pair Graviton compute with fast local disk for storage-heavy workloads.
- Burstable (T-series): T4g — the low-cost, burstable-performance option often used for dev/test environments and light web traffic.
AWS also extends Graviton into its managed services — Aurora, RDS, ElastiCache, OpenSearch, and EMR all offer Graviton-backed options, so adopting Graviton doesn’t require managing raw EC2 instances directly.
Real-World Adoption
AWS states that more than 120,000 customers are now building on Graviton. Specific examples AWS has cited include Pinterest, which runs Graviton-based infrastructure serving more than 500 million monthly active users; Meta, which has committed to deploying tens of millions of Graviton cores for agentic AI workloads; and Uber and Snowflake, both named as deploying Graviton for agentic AI use cases. Partner performance claims reported alongside the Graviton5 launch include Airbnb citing up to 25% better performance over comparable same-generation x86 architectures in production search workload testing, Atlassian reporting 30% higher performance and 20% lower latency running Jira on Graviton5-based M9g versus the prior generation, and SAP citing 35–60% improvement in OLTP query performance for SAP HANA Cloud on M9g.
How Graviton Compares to Google Axion and Nvidia Grace
Graviton isn’t the only custom Arm-based server chip in the cloud. Google Cloud’s Axion processor follows a similar playbook — a custom Arm Neoverse-based CPU built in-house and pitched on price-performance for general-purpose cloud workloads, analogous to Graviton’s role inside AWS. Nvidia’s Grace chip takes a different approach: it’s also an Arm-based CPU, but Nvidia designs it primarily to pair tightly with its own GPUs (as in the Grace Hopper and Grace Blackwell superchips) for AI and HPC workloads, rather than as a general-purpose EC2-style instance CPU competing on its own. In short, Graviton and Axion compete head-on as “cloud-provider house-brand Arm CPU for everyday compute,” while Grace plays a more specialized role bolted onto Nvidia’s GPU-centric AI stack — a distinction worth keeping in mind before assuming all three are interchangeable just because they’re all Arm-based.
Trade-Offs
Graviton’s biggest practical trade-off is software compatibility. Arm and x86 are different instruction set architectures, so any software with compiled x86-specific binaries, kernel modules, or dependencies needs an Arm-compatible build. Most mainstream Linux distributions, container base images, and managed-language runtimes (Java, Python, Node.js, Go) now ship native Arm builds, which is a big part of why Graviton adoption accelerated — but legacy applications, some proprietary third-party software, and x86-specific compiled extensions can still require rework or simply aren’t available on Arm. AWS’s own trial program (free t4g.small hours through the end of 2026, per its Graviton page) exists partly to lower the barrier to testing that compatibility before committing production workloads.
Bottom Line
Graviton has moved from a niche cost-saving option to a mainstream default across much of AWS’s EC2 lineup, backed by AWS’s own published claims of up to 20% lower cost and up to 60% less energy use versus comparable x86 instances, plus genuine adoption numbers in the hundreds of thousands of customers. Graviton5’s jump to 192 cores and a substantially larger cache puts clear distance between it and Graviton4, and the main real-world question for most teams isn’t whether Graviton is cheaper — AWS’s figures and partner testimonials are consistent on that point — but whether your specific software stack is Arm-ready. For anything running mainstream open-source software, the honest answer in 2026 is usually yes.
FAQ
Is Graviton the same as Trainium or Inferentia?
No. Graviton is a general-purpose CPU for ordinary EC2 compute. Trainium and Inferentia are AWS’s separate custom chips built specifically for AI model training and inference, respectively.
How much cheaper is Graviton than x86 on AWS?
AWS’s own Graviton page states Graviton-based instances cost up to 20% less than comparable x86-based EC2 instances, with up to 60% lower energy use for the same work. Actual savings vary by workload and instance family.
Do I need to rewrite my application to use Graviton?
Usually not a full rewrite, but you do need an Arm-compatible build of your software stack. Most modern Linux distributions, container images, and language runtimes support Arm natively; older or x86-specific binaries may need to be recompiled or replaced.
Can I buy a Graviton chip for my own servers?
No. Graviton is proprietary AWS silicon available only inside AWS’s own cloud infrastructure, unlike general-purpose chips from Intel, AMD, or Ampere.
Is Graviton5 generally available now?
AWS previewed Graviton5 at re:Invent 2025 and announced general availability in June 2026, with M9g (general purpose) and M9gd (local NVMe storage) as the first instance families built on it.
Sources
- AWS — Official Graviton product page
- About Amazon — AWS Graviton5 launch announcement and customer quotes
- Wikipedia — AWS Graviton generation history (cross-reference)
Related reading: AWS Trainium and Inferentia: Do Amazon’s Custom AI Chips Actually Beat Nvidia GPUs? and Inside AWS’s AI Networking Stack: The Nitro System, EFA, and How Amazon Builds Infrastructure for Massive GPU Clusters.
Discover more from Software Online Review
Subscribe to get the latest posts sent to your email.
