Meta is deepening its compute infrastructure strategy through a latest agreement with Amazon Web Services to deploy tens of millions of AWS Graviton5 processor cores, marking another step in the company’s broader effort to diversify its hardware stack for agentic AI workloads. The partnership, announced on April 24, 2026, positions Meta as one of the largest customers of AWS’s custom silicon, which is designed to handle the sustained, complex reasoning tasks required by autonomous AI systems.
The deal builds on Meta’s existing relationship with AWS and reflects a growing industry shift toward heterogeneous computing architectures, where different types of processors are assigned specific roles based on workload demands. As agentic AI systems evolve to handle persistent, multi-step tasks such as real-time planning, code generation and deep research, the role of the CPU has shifted from a supporting component to a central orchestrator.
According to the official announcement from Meta and AWS, the first phase of the deployment will include tens of millions of Graviton5 cores, with each chip containing 192 cores. The agreement includes flexibility to scale further as Meta’s AI capabilities expand. This scale places Meta among the top global users of AWS Graviton processors.
The Graviton5 chip is built on the AWS Nitro System, which underpins the security, performance, and availability of AWS EC2 instances. AWS states that the chip is engineered to handle billions of interactions and coordinate complex, multi-stage agentic workflows — capabilities that are increasingly vital as AI systems move beyond simple pattern recognition toward sustained reasoning and decision-making.
Matt Kimball, VP and principal analyst at Moor Insights & Strategy, emphasized that the value of CPUs like Graviton5 lies not in raw scale but in system control. “Here’s really about control of the AI system, not just scale,” he said. As AI workloads become more stateful and less linear, the CPU takes on responsibilities such as memory management, task scheduling, and orchestration across accelerators like GPUs and specialized AI chips.
Kimball noted that Meta’s approach remains additive rather than replacement-based. The company is not moving away from GPUs or AI accelerators but is instead layering in general-purpose compute to create a more balanced, efficient infrastructure. “This is about assembling a heterogeneous system, not picking a single winner,” he said. “Heterogeneity is critical to long term success.”
This strategy aligns with Meta’s broader hardware diversification efforts. In recent months, the company has announced multiple generations of its in-house MTIA training and inference accelerator, secured a major deal with AMD for 6GW of CPUs and AI accelerators, expanded its partnership with Nvidia to access Blackwell and Rubin GPUs, and integrated Nvidia Spectrum-X Ethernet switches into its platform. Meta was also cited as one of Arm’s first major CPU customers in a push to challenge industry giants.
These moves collectively reflect Meta’s principle that no single chip architecture can efficiently serve every workload. By spreading investments across ARM-based designs like Graviton5, x86 accelerators from AMD and Intel, and Nvidia’s GPU lineup, Meta aims to match the right processor to the right task — whether it’s high-throughput training, low-latency inference, or persistent agentic reasoning.
Nabeel Sherif, principal advisory director at Info-Tech Research Group, highlighted the strategic implications of this expanded capacity. Whereas much of the compute will support internal experimentation and innovation, it also lays the foundation for Meta to potentially offer its own agentic AI services to external users, such as through APIs for its Llama family of models.
Sherif noted that the exact form of these future services — including the platforms, tools, and user guardrails involved — remains unclear. However, the availability of diverse, scalable compute gives Meta flexibility to explore various deployment models as the agentic AI market matures.
From a cost perspective, Kimball pointed out that as inference workloads become more persistent — especially in agentic systems that run continuously over time — the economic focus is shifting from peak performance metrics like FLOPS (floating-point operations per second) to sustained efficiency and total cost of ownership (TCO). Power-efficient, general-purpose CPUs like Graviton5 offer advantages for workloads that do not require the parallel throughput of GPUs but still demand reliable, always-on execution.
“At Meta’s scale, even small efficiency gains per workload compound quickly,” Kimball said. This principle underscores why the company is investing in layered architectures where tasks are routed to the most efficient processor based on behavior — such as distinguishing between prefill and decode phases in LLMs, or stateless versus stateful operations in agentic flows.
For enterprise IT teams and developers, the broader takeaway is that infrastructure decisions are becoming more workload-aware. Rather than asking simply “which cloud?” or “which chip?”, organizations are increasingly evaluating where specific parts of an application will run most efficiently based on their computational traits.
The Meta-AWS Graviton5 agreement was announced via Meta’s official news channel on April 24, 2026, and confirmed through AWS’s public statements. No financial terms of the deal were disclosed in the available sources.
As Meta continues to expand its compute footprint across multiple architectures, the company is positioning itself not just as an AI model developer but as a systems integrator shaping the underlying infrastructure for the next generation of intelligent agents.
For updates on Meta’s AI infrastructure partnerships and hardware developments, refer to the company’s official newsroom and AWS’s announcements on custom silicon.
What do you think about Meta’s push for heterogeneous computing in AI? Share your perspective in the comments or join the conversation online.