AWS Graviton in 2026: Five Generations of ARM Dominance in Cloud Computing

Updated (February 2026): Complete rewrite replacing the original Graviton3 overview with current coverage of Graviton4 GA instances, Graviton5 announced at re:Invent 2025, ARM server market dynamics, Azure Cobalt and Google Axion competition, AI/ML inference optimization, migration tooling, and real-world customer cost savings data.

The ARM Takeover Is No Longer a Prediction

When AWS launched Graviton in 2018, custom ARM chips in the cloud were an experiment. In 2026, they are the default. Over 90,000 AWS customers run on Graviton. 98 percent of AWS's top 1,000 EC2 customers have adopted it. More than 50 percent of new CPU capacity added by AWS has been Graviton-based for three consecutive years. During Prime Day, Graviton powered over 40 percent of EC2 compute used by Amazon.com itself.

The broader ARM server market tells the same story. Arm-based server shipments jumped 70 percent in 2025, reaching 21 percent of global shipments according to IDC. Arm's own claim is bolder: roughly 50 percent of compute shipped to top hyperscalers is now ARM-based. Every major cloud provider — AWS, Azure, Google Cloud, Oracle — now offers custom ARM instances. The x86 monopoly in the data center is over.

Graviton4: The Current Generation

Graviton4 — announced at re:Invent 2023 and reaching general availability through 2024 — represents a major architectural leap. Built on Arm Neoverse V2 cores (ARMv9.0-A ISA with SVE2 extensions), it is the first Graviton to support multi-socket configurations.

Specifications

• 96 Arm Neoverse V2 cores with 2 MB L2 cache per core (192 MB total)
• 12 DDR5-5600 memory channels delivering 537 GB/sec bandwidth per socket — 75 percent higher than Graviton3's 307 GB/sec
• 96 lanes of PCIe Gen5 — tripled from Graviton3
• Multi-socket support enabling up to 192 vCPUs in dual-socket configurations
• Full hardware encryption on all high-speed physical interfaces

Instance Families

Graviton4 powers a broad set of instance types, with regional expansion continuing through late 2025:

• R8g (memory-optimized): Up to 192 vCPUs, 1,536 GB memory. GA July 2024
• M8g (general-purpose): Balanced compute, memory, and networking
• C8g (compute-optimized): HPC, gaming, video encoding
• X8g (high-memory): Up to 192 vCPUs, 3 TB DDR5 memory with 16:1 memory-to-vCPU ratio
• I8g (storage-optimized): Up to 22.5 TB NVMe SSD with 65 percent better storage performance per TB versus I4g
• M8gn/M8gb: Network-optimized and EBS-optimized variants, GA December 2025

Performance Over Graviton3

The generational improvements are substantial across every workload category:

• 30 percent better general compute (SPECintrate2017)
• 75 percent higher memory bandwidth
• 40 percent faster for databases
• 45 percent faster for large Java applications
• 2.8x improvement in cryptographic workloads
• 29 percent better overall price-performance

Phoronix called Graviton4 "easily the most competitive AArch64 server processor" in their Ubuntu 24.04 benchmarks, noting "terrific generational uplift" across web serving, database, compilation, and HPC workloads.

Graviton5: 192 Cores at 3nm

Announced at re:Invent 2025, Graviton5 doubles the core count and moves to a cutting-edge 3nm process node.

Architecture

• 192 Arm Neoverse V3 cores on a single socket (ARMv9.2 ISA)
• TSMC 3nm process — the same node used for Apple's M4 chips
• 180-192 MB L3 cache — 5x larger than Graviton4
• DDR5-7200 memory (upgradeable to DDR5-8400/8800), delivering up to 806 GB/sec bandwidth per socket
• 25 percent better compute performance versus Graviton4
• 33 percent lower inter-core latency despite doubling the core count
• 15 percent higher network bandwidth, 20 percent higher EBS bandwidth

M9g instances are available in preview, with C9g and R9g configurations expected through 2026. First production Graviton5 servers are already running in AWS data centers.

SAP has been an early validator — on Graviton5, SAP observed a 35 to 60 percent increase in OLTP query performance compared to Graviton4.

Graviton4 vs. x86: Benchmarks

Graviton4 does not just compete on price. In many workloads, it matches or exceeds AMD EPYC and Intel Xeon on raw performance:

Compilation (Gem5): Graviton4 completed in 186.8 seconds versus AMD EPYC 9R14 at 213.3 seconds versus Intel Xeon 8488C at 244.9 seconds.

Machine learning (XGBoost training): 53 percent faster than AMD, 34 percent faster than Intel at equivalent configurations.

In-memory data stores (Redis): Up to 93 percent more operations per second versus x86.

HPC (MiniFE): Graviton4 outperformed both Graviton3 and AMD EPYC 9R14.

Cost per benchmark run: Graviton4 at approximately $0.186 versus AMD at $0.288 — the best value regardless of raw throughput.

Where x86 still leads: single-threaded workloads where Intel and AMD hold a 6 to 14 percent per-thread advantage, and workloads requiring AVX-512 instructions for specific HPC algorithms.

The ARM Cloud Competition

AWS no longer competes only with x86. Every major cloud provider now has custom ARM processors.

Azure Cobalt

Cobalt 100 (GA October 2024): 128 Arm Neoverse N2 cores at TSMC 5nm. Available across 32+ Azure regions with Dpsv6, Dplsv6, and Epsv6 instance types. Microsoft reported that 32.9 percent of new Azure CPUs shipped in Q4 2024 were Cobalt 100.

Cobalt 200 (announced November 2025): 132 Arm Neoverse V3 cores at TSMC 3nm — the same core generation as Graviton5. Features 3 MB L2 cache per core, 192 MB L3 system cache, and 12-channel DDR5 per socket. Microsoft claims 50 percent performance improvement over Cobalt 100. First production servers are already running; GA expected in 2026.

Google Axion

Built on Arm Neoverse V2 (the same core as Graviton4), Axion offers up to 72 cores with DDR5-5600 and up to 576 GB memory. Available as C4A, N4A, and C4A metal instances. Google claims up to 50 percent better performance than x86 and up to 60 percent more energy efficiency.

In benchmarks, Axion outperformed Graviton4 in 5 of 7 LLM inference tests, with advantages ranging from 9.9 to 47.8 percent. Graviton4 showed advantages in XGBoost training and Redis throughput.

Ampere Computing

The independent ARM server chip company, acquired by SoftBank for $6.5 billion, offers AmpereOne with up to 192 cores at TSMC 5nm. A 256-core 3nm chip was announced for 2025. Ampere powers ARM instances at Oracle Cloud, and supplies processors to Google, Microsoft, Alibaba, Tencent, HPE, and Supermicro.

AI and Machine Learning on Graviton

Graviton is not a GPU replacement — it is a cost-effective inference platform for workloads that do not require CUDA acceleration.

Inference Optimization

PyTorch 2.0 on Graviton delivers up to 3.5x faster inference for ResNet50 and up to 1.4x faster for BERT versus the previous PyTorch release. Key optimizations include:

• Arm Compute Library (ACL) with SVE and MMLA hardware accelerators
• bfloat16 GEMM kernels via OneDNN fast math mode — up to 2x inference speedup without model quantization
• INT8 quantization via TorchAO with a dedicated Arm quantizer
• SageMaker Graviton containers with pre-optimized TensorFlow and PyTorch runtimes

At up to 50 percent cost savings for PyTorch inference on C7g/C8g versus comparable x86 instances, Graviton is compelling for classification, NLP, recommender systems, and any model that fits in CPU memory.

Where Graviton Does Not Fit

• CUDA-dependent GPU training workloads
• Large language model training (use Trainium or NVIDIA GPUs)
• Workloads requiring GPU tensor cores for matrix multiplication at scale

AWS Custom Silicon Strategy

Graviton exists within a broader custom silicon ecosystem:

• Graviton: General-purpose compute (5 generations, annual cadence)
• Inferentia2: Up to 4x throughput, 10x lower latency versus Inferentia1 for inference
• Trainium2: Used by Amazon Bedrock; customers including Anthropic report up to 50 percent training cost reduction
• Trainium3: Announced at re:Invent 2025 on 3nm — 4.4x compute performance, 4x energy efficiency versus Trainium2. EC2 Trn3 UltraServers support up to 144 Trainium3 chips

AWS is building toward a "converged AI data center" where Graviton, Inferentia, and Trainium run on unified compute sleds.

Cost Optimization

Pricing Advantage

Graviton instances are 15 to 35 percent cheaper per hour than comparable x86 instances at on-demand rates. When factoring in performance gains, the total cost advantage reaches 40 to 50 percent for most workloads.

Graviton4 pricing dropped an additional 7 percent in us-east-1 in September 2025, reflecting AWS's commitment to aggressive pricing as competition increases.

Commitment Discounts

• 3-year EC2 Instance Savings Plans: Up to 72 percent discount
• 1-year Compute Savings Plans: Up to 66 percent discount
• Spot instances: Up to 90 percent savings; Graviton Spot instances often have better availability than x86 Spot due to lower competition

Real-World Customer Results

The customer savings data is compelling:

• Pinterest: 47 percent cost savings, 38 percent compute reduction, 62 percent reduction in carbon emissions
• Datadog: 40 percent better price-performance across Java, Ruby, and Go workloads; 50 percent boost in CI pipeline speed
• CyberAgent: 50 to 60 percent total cost reductions from Graviton2 adoption
• Fluid Attacks: 70 percent cost reduction combining Graviton with Spot instances
• SAP: R8g instances deliver 25 percent improvement for analytical and 40 percent for transactional HANA workloads versus Graviton3

When Graviton Is Not the Right Choice

Graviton runs Linux only. It is not suitable for:

• Windows Server workloads — no Windows support on Graviton
• Legacy .NET Framework applications (Windows-dependent; .NET 5+ works)
• CUDA/GPU-accelerated workloads requiring NVIDIA GPUs
• x86-only binary dependencies — closed-source software without ARM builds
• Specific Intel instruction sets like AVX-512 for certain HPC algorithms
• FPGA-dependent workloads

Software Ecosystem and Migration

Language Support

The software ecosystem is mature. Most workloads run without modification:

• Interpreted languages (Python, Node.js, Ruby, PHP): Zero code changes required
• Java: JDK 11 minimum, JDK 17+ recommended. Amazon Corretto delivers best performance. Graviton4 is 45 percent faster for large Java apps versus Graviton3
• Go: Recompile with GOARCH=arm64 — straightforward
• Rust: Full ARM support, recompile for target
• .NET: Supported on Linux with .NET Core/.NET 5+. Legacy .NET Framework (Windows-only) is not supported
• C/C++: Recompilation required; most modern builds work natively

Container Support

Container migration is the most common path to Graviton adoption:

• Docker multi-arch images via docker buildx — the single most important migration step
• Amazon ECR supports multi-arch image manifests
• Amazon EKS supports mixed x86/arm64 node pools in the same cluster
• AWS Fargate supports Graviton2 and later
• Multi-arch images automatically pull the correct architecture-specific layers

Migration Tools

AWS provides dedicated tooling for migration:

• Porting Advisor for Graviton: Python-based scanner that detects inline assembly without arm64 equivalents, x86-specific library links, architecture-specific intrinsics, and preprocessor issues
• AWS Graviton Getting Started Guide: Language-specific migration guides on GitHub
• AWS Graviton Savings Dashboard: Identifies workloads most suitable for migration and tracks adoption progress
• AWS Graviton Fast Start Program: AWS-supported migration assistance

Migration Strategy

The recommended approach:

1. Start with interpreted languages — Python, Node.js, and Ruby services need zero code changes
2. Build multi-arch container images to support gradual rollout
3. Mix architectures in EKS — run Graviton and x86 nodes side by side
4. Test in staging with Graviton-based development instances before production
5. Avoid cross-architecture emulation in CI/CD — it causes 2.5x build time increases. Use native arm64 CodeBuild runners

The trajectory is clear: AWS is deploying over 1.2 million ARM CPUs in servers in 2025 alone. Graviton has evolved from an experiment to the backbone of AWS compute, and custom ARM silicon from every major cloud provider ensures this shift is permanent. For engineering teams, the question is no longer whether to evaluate Graviton but which workloads to migrate first — starting with the interpreted languages and containerized services where migration requires zero code changes and delivers immediate 15 to 50 percent cost savings.