Amazon Web Services (AWS) has made its new Graviton5 processor generally available, positioning the Arm-based CPU as a key platform for the growing class of agentic AI applications that require real-time reasoning, code generation, and orchestration of complex tasks. Available through Amazon EC2 M9g and M9gd instances, Graviton5 delivers up to 25% higher compute performance than the previous generation while offering the highest CPU core density in Amazon EC2 with 192 cores per package.
AWS said Graviton5 incorporates a 5x larger L3 cache, DDR5-8800 memory, PCIe Gen 6 support, and up to 33% lower inter-core latency compared to Graviton4. The company also reports up to 35% faster web application performance, 35% faster machine learning inference, and 30% faster database performance. M9gd instances add up to 11.4 TB of local NVMe SSD storage and 30% higher IOPS for storage-intensive workloads. The processor is manufactured on a 3nm process node and is integrated with AWS’s sixth-generation Nitro System, including the new Nitro Isolation Engine, which AWS describes as a formally verified security component designed to provide mathematically proven isolation between virtual machines.
AWS executives highlighted growing adoption of Graviton as AI workloads drive increased demand for CPU infrastructure alongside accelerators. CEO Andy Jassy noted that approximately 98% of AWS’s top 1,000 EC2 customers use Graviton, with more than 120,000 customers now running workloads on the platform. He also disclosed that Meta plans to deploy tens of millions of Graviton cores to support agentic AI initiatives, while Uber and Snowflake are expanding their use of the processor family.
• Graviton5 delivers up to 25% higher compute performance than Graviton4
• 192 Arm-based CPU cores per package
• Up to 33% lower inter-core latency
• 5x larger L3 cache than the previous generation
• DDR5-8800 memory support
• PCIe Gen 6 connectivity
• Up to 15% higher network bandwidth and 20% higher EBS bandwidth
• Built on a 3nm manufacturing process
• Available in EC2 M9g and M9gd instance families
• Meta plans deployment of tens of millions of Graviton cores for agentic AI
“About 11 years ago, with our very talented Annapurna team and informed by the unusual scale and insight we had in operating the largest cloud infrastructure, we decided to design and build our own CPU chip,” said Andy Jassy, President and CEO of Amazon. “The reason customers are so excited about Graviton is that it offers about 30-40% better price-performance than comparable instances. When you layer on top of how much CPU customers normally use with the fact that AI’s growth is driving explosive CPU expansion given that post-training, reinforcement learning, and agentic actions use CPU, Graviton becomes even more compelling.”
🌐 Analysis: Graviton5 reflects AWS’s long-term strategy of vertically integrating its infrastructure stack, following a path similar to its custom Trainium AI accelerators, Nitro DPUs, and Annapurna networking silicon. While NVIDIA remains dominant for AI training and inference acceleration, AWS is emphasizing the growing importance of CPUs for orchestration, reinforcement learning environments, retrieval systems, agent execution, and AI infrastructure management. The company’s focus on agentic AI represents a notable shift from traditional cloud messaging centered on web applications and databases.
🌐 The announcement also underscores intensifying competition among hyperscalers developing proprietary silicon. AWS now fields Graviton CPUs, Trainium AI accelerators, Inferentia inference processors, Nitro infrastructure offload chips, and custom networking silicon. Meanwhile, competitors continue expanding their own silicon portfolios, including Google’s Tensor Processing Units (TPUs) and Microsoft’s Maia AI accelerator family. Meta’s commitment to deploy tens of millions of Graviton cores highlights the growing role of Arm-based server CPUs in large-scale AI infrastructure.
