Meta is making a major move to scale its AI infrastructure, announcing a significant partnership with Amazon Web Services that will bring tens of millions of AWS Graviton CPU cores into Meta’s compute portfolio. The goal: meet the rapidly rising demand for compute as agentic AI systems become more capable, more complex, and far more resource-hungry.
AI conversations often focus on GPUs, but agentic AI is increasingly becoming a CPU-heavy story too. As AI systems shift toward multi-step reasoning, tool use, planning, orchestration, and real-time execution, there’s a growing need for powerful, efficient general-purpose compute alongside accelerators. Meta says its compute needs are evolving in that direction, and the company is positioning CPUs as a central pillar of its next wave of AI capacity.
At the center of this deal is AWS Graviton, Amazon’s Arm-based server CPU line. Meta plans to deploy “tens of millions” of Graviton cores, a massive scale-up that underscores how quickly the AI infrastructure race is accelerating. For context, AWS’s latest Graviton5 chips pack up to 192 Arm Neoverse cores per processor, meaning these deployments translate into an enormous pool of CPU compute for AI-related workloads.
Meta frames this as part of a broader infrastructure strategy: no single chip architecture is ideal for every workload. In practice, that means building a diversified compute portfolio where the right hardware is matched to the right job—whether that’s training, inference, data processing, orchestration, or the behind-the-scenes services that keep agentic systems running smoothly at scale.
Amazon, for its part, is pitching this as more than a raw hardware expansion. The company’s message is that purpose-built silicon combined with a full AI platform can help teams build systems that understand, anticipate, and scale efficiently to billions of users. The scale of Meta’s deployment also positions Meta as one of the largest Graviton customers globally, signaling that Arm-based CPUs are becoming increasingly important in large-scale AI deployments.
This partnership also highlights a broader industry trend: AI companies are chasing compute wherever they can get it. Demand is surging across the board, and virtually every major CPU maker is seeing increased interest as AI firms look to secure large volumes of chips. At the same time, many of those AI firms—including Meta—are investing in custom silicon efforts as another route to secure long-term capacity and optimize performance for specific workloads.
Meta has already been working on purpose-built AI hardware, including its own accelerator initiatives and a partnership with Broadcom aimed at custom AI silicon to support a multi-gigawatt-scale ecosystem. But with advanced semiconductor capacity constrained across the industry, large AI players are increasingly balancing three paths at once: working with established chip designers, building custom chips where it makes sense, and expanding via cloud providers that can deliver compute at scale.
Importantly, this “tens of millions of cores” deployment appears to be only the beginning. Meta has signaled intentions to continue scaling its AI resources aggressively over the coming years. As agentic AI expands into more products and more real-time experiences, the need for massive CPU capacity—alongside GPUs and specialized accelerators—is expected to grow even faster.
