Meta's Graviton bet

Meta said Friday it will use Amazon Web Services’ Graviton processors at scale, adding tens of millions of Graviton5 CPU cores for agentic AI. (about.fb.com) (press.aboutamazon.com) The companies said the first deployment starts with tens of millions of cores, with room to expand as Meta’s AI capacity grows. Amazon said the agreement makes Meta one of the largest Graviton customers in the world. (about.fb.com) (press.aboutamazon.com) CNBC reported the contract runs for at least three years and covers hundreds of thousands of Graviton chips, even as Meta described the footprint in cores rather than processors. AWS vice president Nafea Bshara told CNBC the deal puts Meta among the top five Graviton customers. (cnbc.com) A CPU is the general-purpose chip that handles the step-by-step work around an AI system, while a graphics processing unit, or GPU, is usually the engine for training large models. Amazon and Meta said agentic AI creates more demand for CPU-heavy jobs such as reasoning loops, search, code generation, and coordinating multi-step tasks. (press.aboutamazon.com) (about.fb.com) Meta’s infrastructure chief Santosh Janardhan said the company is diversifying compute sources, and Graviton gives it CPU capacity for those workloads at Meta’s scale. Amazon said Meta already uses Amazon Bedrock at scale, and framed the new agreement as an expansion of a long-standing AWS relationship. (about.fb.com) (press.aboutamazon.com) The Graviton deal lands two months after Meta announced a separate long-term agreement with Advanced Micro Devices for up to 6 gigawatts of Instinct graphics processing units. Meta said then that no single chip architecture can serve every workload efficiently. (about.fb.com) CNBC said Meta had also made $48 billion in recent AI infrastructure commitments with CoreWeave and Nebius, both tied to Nvidia-heavy capacity. The new AWS agreement shows Meta adding central processing capacity alongside those graphics-heavy bets. (cnbc.com) AWS has pitched Graviton on cost and energy efficiency as well as performance, and CNBC said Amazon claims the chips use 60% less energy for a given workload class. That gives Amazon another marquee customer as it pushes its own silicon against Nvidia, Advanced Micro Devices, and Intel in the AI buildout. (cnbc.com) For Meta, the immediate bet is not that CPUs replace GPUs. It is that building AI products for billions of users now requires both: GPUs to train and infer, and large fleets of CPUs to keep the agents working. (about.fb.com) (press.aboutamazon.com)