Meta has expanded its long-standing partnership with Amazon Web Services (AWS) through a new agreement to deploy Graviton processors at scale. The move will see Meta utilize tens of millions of Graviton cores, with the potential for further expansion as it builds out its next-generation artificial intelligence infrastructure.
This development reflects a broader shift in how large technology companies are structuring compute resources, particularly as AI workloads evolve beyond traditional training tasks.
The agreement focuses on large-scale deployment of AWS’s custom-designed Graviton processors, which are optimized for high-performance and energy-efficient computing.
Meta’s decision signals a strategic diversification of its compute infrastructure, combining GPUs for model training with CPUs designed for specific AI workloads. The approach aims to balance performance, cost, and scalability.
Shift
A key driver behind this partnership is the growing importance of agentic AI workloads. These systems are designed to operate autonomously, handling complex, multi-step processes with minimal human intervention.
Unlike traditional AI training, which is GPU-intensive, these workloads rely more heavily on CPU performance.
Key Use Cases:
- Real-time reasoning and decision-making
- Code generation and automation
- Search, orchestration, and workflow management
These tasks require consistent, large-scale processing capabilities, making CPU-optimized architectures increasingly relevant.
Technology
Meta’s deployment will primarily rely on the AWS Graviton5 processor, the latest generation in AWS’s custom silicon lineup.
Core Specifications:
| Feature | Details |
|---|---|
| Core Count | 192 cores |
| Cache | 5x larger than previous generation |
| Performance Gain | Up to 25% improvement |
| Manufacturing | 3nm process |
The expanded cache reduces delays in data exchange between cores, improving efficiency in high-demand environments. This is particularly important for distributed AI workloads.
System
Graviton5 operates within the AWS Nitro System, a platform that integrates dedicated hardware and software for performance and security.
Key Components:
- Bare-metal access for direct hardware utilization
- Elastic Network Adapter (ENA) for high-speed networking
- Amazon Elastic Block Store (EBS) for scalable storage
This setup allows Meta to run workloads with minimal virtualization overhead while maintaining flexibility.
Network
To support large-scale AI operations, the deployment also leverages the Elastic Fabric Adapter (EFA).
EFA enables:
- Low-latency communication
- High-bandwidth data transfer
- Efficient distribution of workloads across processors
This is essential for coordinating billions of interactions across distributed systems, particularly in real-time AI applications.
Efficiency
Energy efficiency is a central consideration in this partnership. The Graviton5 processors are built using 3-nanometer fabrication technology, which allows for improved performance while reducing power consumption.
Because AWS controls the entire stack, from chip design to deployment, the system can be optimized more effectively than standard off-the-shelf hardware.
Key Benefits:
- Lower energy consumption per workload
- Improved performance per watt
- Reduced operational costs at scale
These factors are increasingly important as AI infrastructure expands globally.
Strategy
Meta’s infrastructure strategy reflects a hybrid approach to compute resources.
Key Objectives:
- Use GPUs for training large AI models
- Deploy CPUs like Graviton5 for inference and agentic tasks
- Optimize cost and efficiency across workloads
By diversifying compute sources, Meta reduces reliance on a single type of hardware while improving scalability.
Industry
The agreement also highlights a broader industry trend. Technology companies are increasingly investing in purpose-built silicon rather than relying solely on general-purpose processors.
This shift is driven by:
- Rising AI compute demands
- Need for cost efficiency
- Performance optimization for specific workloads
Custom processors like Graviton are becoming a key part of this transition.
Statements
Nafea Bshara, Vice President and Distinguished Engineer at Amazon, noted that the partnership goes beyond hardware, emphasizing the integration of infrastructure and AI services to support large-scale applications.
Santosh Janardhan, Head of Infrastructure at Meta, highlighted the importance of diversifying compute resources, stating that Graviton processors provide the performance and efficiency required for CPU-intensive AI workloads.
Outlook
Meta’s expanded use of AWS Graviton processors illustrates how AI infrastructure is evolving to accommodate new types of workloads. As agentic AI systems become more prevalent, the demand for CPU-optimized environments is expected to grow alongside traditional GPU-based systems.
The partnership positions both companies to address these changing requirements, combining custom silicon with cloud-scale infrastructure. Its long-term impact will depend on how effectively these systems support emerging AI applications and whether similar strategies are adopted across the industry.
FAQs
What is AWS Graviton?
It is AWS’s custom ARM-based processor.
Why is Meta using Graviton?
For CPU-intensive AI workloads and efficiency.
What is agentic AI?
AI systems that act autonomously on tasks.
What is Graviton5 performance gain?
Up to 25% better than previous generation.
How many cores does Graviton5 have?
It features 192 cores.
