Meta’s Compute Expansion: AWS Graviton5 Deal Strengthens Its Agentic AI Infrastructure Strategy

Meta has signed an agreement with Amazon Web Services (AWS) to deploy tens of millions of AWS Graviton5 processor cores into its compute infrastructure to support agentic AI workloads, marking a significant expansion of its long-standing partnership with the cloud provider.

The deployment begins with tens of millions of Graviton5 cores, with the flexibility to scale further as Meta’s AI capabilities grow, positioning the company as one of the largest Graviton customers globally. This move supports Meta’s broader strategy of diversifying its compute infrastructure across multiple architectures to match workload demands.

Agentic AI Drives Demand for CPU-Intensive Workloads

While graphics processing units (GPUs) remain essential for training large language models, the rise of agentic AI—autonomous systems that reason, plan, and execute complex multi-step tasks—is increasing demand for CPU-optimized infrastructure. Graviton5 is specifically designed for these workloads, enabling efficient handling of real-time reasoning, code generation, search, and orchestration of agent workflows at scale.

According to AWS, Graviton5 processors can manage billions of interactions while coordinating complex, multi-stage agentic tasks, leveraging the AWS Nitro System for performance, security, and availability. These capabilities make the chips well-suited for the control-plane functions required in persistent, stateful AI environments.

Part of a Heterogeneous Compute Strategy

The AWS agreement reflects Meta’s diversified approach to hardware, which includes partnerships with Nvidia, AMD, and Arm, as well as internal development of its MTIA series of training and inference accelerators. Meta emphasized that no single chip architecture can efficiently serve every workload, reinforcing its strategy of matching compute resources to specific AI tasks.

Recent actions underscoring this approach include the announcement of four new generations of Meta’s MTIA chips, a multi-year deal with Nvidia for Blackwell and Rubin GPUs and Spectrum-X Ethernet switches, a significant AI chip agreement with AMD, and early adoption of Arm-based processors in its infrastructure.

Industry Analysts See Strategic Value in Heterogeneity

Matt Kimball, VP and principal analyst at Moor Insights & Strategy, noted that Meta’s expanding portfolio of chip partnerships—including the Graviton5 deal—is not about replacing GPUs or accelerators but about assembling a heterogeneous system where workloads are assigned to the most efficient architecture.

Kimball explained that as agentic AI shifts focus from peak performance to sustained efficiency and total cost of ownership, CPUs like Graviton5 become critical for handling persistent workloads that do not require accelerators but still demand continuous, efficient operation. At Meta’s scale, even small efficiency gains per workload compound significantly over time.

He added that the trend points toward tighter coupling between CPUs, GPUs, and specialized accelerators, with infrastructure decisions increasingly based on workload characteristics—such as whether a task is stateless or stateful, bursty or persistent—rather than simply choosing a cloud provider.

Supporting Internal Innovation and External Services

The expanded compute capacity will primarily support Meta’s internal experimentation and innovation in agentic AI, while also laying the groundwork for offering its own AI services externally, such as through APIs for its Llama model. However, the specific form these services will take, including platforms, tools, and user guardrails, remains under development.

Nabeel Sherif, principal advisory director at Info-Tech Research Group, observed that the investment gives Meta flexibility to explore diverse use cases across architectures, particularly valuable in an environment marked by both innovation in CPU design and ongoing supply chain constraints.

Sherif and Kimball both agreed that the AWS deal is additive to Meta’s existing investments, not a substitution, and reflects a broader industry shift toward workload-aware infrastructure planning rather than reliance on a single compute paradigm.