AWS Graviton5 continues a wave of Arm-based, purpose-built compute
At the start of this year, Arm made a bold prediction: close to 50% of the compute shipped to top hyperscalers in 2025 will be Arm-based.
After three quarters of confirmed shipments, we’re happy to say the market is right on track to meet that forecast. These servers aren’t just a statistical milestone. Much more importantly, they form the compute backbone of a new kind of infrastructure – the converged AI data center. From cloud-native services to the most demanding AI workloads, hyperscalers are standardizing on purpose-built compute on Arm as the way to balance performance, power, and scale.
AWS Graviton5, announced today, is a clear manifestation of this new model. Now in its fifth generation, it reflects years of compounding Arm efficiency and is designed for modern data center requirements – from high-density inference to scaled-out cloud-native workloads. For the third straight year, Graviton has accounted for over half of all new CPU capacity deployed at AWS, and today 98% of the top 1,000 Amazon EC2 customers rely on it in production.
Graviton is also part of an all-AWS silicon strategy in which all three AWS custom-designed chips are on the same unified compute sleds in the new AWS Trainium3 compute sled. In this design, Arm-powered AWS Graviton and AWS Nitro perform host CPU and high-performance networking functions respectively for AWS Trainium3 UltraServers.
Graviton5 is part of a broader industry shift as Arm increasingly powers the platforms defining the AI era:
- Google Axion: Now available with additional virtual machine (VM) options, the Axion family powered by Arm Neoverse processors are delivering new levels of performance for cloud and AI workloads.
- Microsoft Cobalt 100 and the newly announced Cobalt 200: Targeting cloud-native workloads in Microsoft Azure’s AI-optimized data centers, these processors power both Microsoft internal services and Azure customers.
- NVIDIA Grace Blackwell: Combining Arm-based CPUs with NVIDIA’s AI accelerators to create the most advanced AI computing platform yet.
The world’s leading hyperscalers, chipmakers, and system builders aren’t experimenting with purpose-built – they’re aligning their infrastructure strategies around it. Arm is the trusted foundation making that convergence possible.
The emergence of the converged AI data center
Taken together, these platforms show how quickly the data center is evolving. AI is collapsing traditional boundaries within infrastructure, turning what was once a collection of general-purpose servers and appliances into a tightly integrated, AI-optimized environment where compute, acceleration, networking, memory, storage, and software are co-designed and operated as a single system.
In the converged AI data center, performance and efficiency come from the cohesion of the entire stack – not from any individual component. Arm provides the common architecture that connects these layers, enabling providers to optimize holistically while maintaining the flexibility and efficiency modern workloads demand at massive scale.
Each layer of this system contributes differently to AI performance, and Arm’s architecture spans all of them:
- CPUs provide the control plane for AI systems – coordinating scheduling, data movement, memory, and services – while also executing key model logic that turns tokens into useful actions.
- Accelerators deliver dense compute where model math is most intensive, scaling training and inference efficiently across thousands of nodes.
- SmartNICs, DPUs and intelligent networking and storage systems offload and accelerate essential services such as security, networking, and data access. Increasingly built on Arm-based silicon – from AWS Nitro and NVIDIA BlueField to Intel IPU and others – these platforms ensure data flows securely and efficiently between every stage of AI computation.
The result is infrastructure that improves intelligence-per-watt across the board, enabling faster innovation while preserving software compatibility and ecosystem consistency. This collaboration-driven approach with partners like AWS, Google Cloud, Microsoft Azure, and NVIDIA is why purpose-built compute on Arm isn’t just gaining momentum – it’s accelerating.
Graviton5 and the new compute curve
With 192 cores and 5x larger cache compared to the previous generation, Graviton5 delivers up to 25% performance gains across the workloads that matter most to AWS cloud customers, illustrating the transformative effect of specialization while redefining the economics of cloud compute. When performance rises faster than power and cost, it changes the equation, and the Arm ecosystem is defining that inflection point.
This momentum extends well beyond hyperscale. Enterprises are now applying the same purpose-built principles to AI inference, autonomous systems, edge computing, intelligent networking, and much more. These use cases are being powered by the same Arm architecture that underpins the cloud.
The future: A compute foundation for everyone
It’s not just the world’s biggest companies that are co-designing the future on Arm. Through initiatives like Arm Total Design, leveraging the Arm Neoverse Compute Subsystems (CSS) platform, we’re enabling a broader ecosystem to design custom silicon with hyperscale-class efficiency and speed. Meanwhile, programs like Arm Cloud Migration make it easier for organizations of any size to adopt Arm-based compute and unlock its performance and efficiency advantages.
Purpose-built compute is winning. The industry is embracing it. And Arm is the compute platform at the heart of it all, powering what comes next.