Meta's decision to run its AI agents on tens of millions of AWS Graviton CPU cores marks a pivotal moment in its AI infrastructure strategy.
This deal formalizes a two-layer approach to AI computing. The first layer involves using a diverse mix of powerful, specialized accelerators from various vendors and its own in-house MTIA chips for the heavy lifting of training AI models and handling complex inference tasks. These are the 'brains' doing the complex math.
The second layer, which is the focus of this deal, is for running the vast, ever-growing world of AI agents. These agents, which will power experiences across Facebook, Instagram, and WhatsApp, require a different kind of computing. Their work involves orchestration—managing conversations, calling tools, and retrieving information. These are tasks that need to happen in parallel for billions of users, demanding massive scale, energy efficiency, and cost-effectiveness. This is where Arm-based CPUs like AWS Graviton excel.
Several recent events paved the way for this strategic move. First, the market validated the readiness of custom silicon at scale. Anthropic's massive $100 billion commitment to AWS, spanning both Trainium accelerators and Graviton CPUs, demonstrated that top-tier AI labs trust this hardware. Furthermore, Amazon's CEO revealed that its custom silicon business is already a $20 billion run-rate operation, with some customers even wanting to buy its entire 2026 Graviton capacity. This signaled to Meta that AWS had the supply and the right economics.
Second, Meta's own internal developments created the need for a robust CPU layer. As Meta expanded its custom MTIA chip roadmap with Broadcom for inference, the 'surround' computing tasks—like data routing and agent management—grew in importance. To avoid bottlenecks and vendor lock-in, Meta has pursued a multi-provider strategy, evidenced by its $21 billion GPU-centric deal with CoreWeave. Securing a massive fleet of AWS CPUs perfectly complements its GPU capacity.
Finally, product rollouts and regulatory pressures added urgency. The launch of Meta's advanced 'Muse Spark' model meant more concurrent agent usage, and looming EU regulations on WhatsApp interoperability required a resilient, geographically distributed infrastructure. Relying on AWS Graviton allows Meta to scale its agent layer quickly and flexibly across different regions, addressing both product demand and regulatory risk.
- AI Agent: An autonomous AI program designed to perceive its environment, make decisions, and perform specific tasks on behalf of a user, such as booking a flight or summarizing information.
- Orchestration: In computing, this refers to the automated configuration, coordination, and management of complex systems and services. For AI agents, it involves managing tasks like tool-calling, data retrieval, and state management.
- AWS Graviton: A family of custom-designed processors by Amazon Web Services, based on the Arm architecture. They are engineered for better price-performance and energy efficiency for cloud workloads compared to traditional x86 processors.