Meta expands use of AWS Graviton cores with agreement to deploy tens of millions

Key Points

Meta is deploying tens of millions of AWS Graviton5 cores (192 per chip) to power agentic AI workloads, blending them with its custom MTIA chips and AMD/Nvidia partnerships.
Graviton5 chips run on the AWS Nitro System, designed to enhance security, latency, and performance for high-demand enterprise tasks.
IT leaders must plan for hybrid compute environments as Meta’s AI strategy relies on balancing CPUs, GPUs, and accelerators.

What is changing

Meta is deepening its AWS partnership by integrating Graviton5 CPU cores into its compute portfolio. Each graviton5 chip packs 192 cores and operates on the Nitro System, which isolates workloads for faster, safer processing. This expands Meta’s focus on CPUs for real-time AI operations like multi-stage agentic tasks, moving beyond GPUs for training.

The company is also rolling out four new generations of its MTIA training and inference chips, designed for faster AI model development. These chips, combined with AMD’s 6GW CPUs and Nvidia’s GPUs, aim to create a flexible infrastructure for diverse workloads. Meta claims no single chip architecture will dominate every use case, prioritizing adaptability over monolithic dependency.

Why it matters

Enterprise IT teams will need to assess hybrid compute needs as Meta’s AI stack becomes more heterogeneous. Companies relying on AWS for cloud services should evaluate how Graviton5’s efficiency impacts cost structures, especially for persistent, stateful workloads. Performance gains may depend on workload split between CPUs and accelerators.

System architects should watch for evolving demands on infrastructure, particularly in balancing tank-like systems for inference with GPUs for training. The shift toward “control plane” CPUs like Graviton5 signals that orchestration and memory management will grow in importance. As Meta explores offering AI APIs, enterprises may need to adapt tooling for interoperability across fragmented systems.