Meta signs multiyear AWS deal to run tens of millions of Graviton5 cores

Meta and Amazon Web Services said on April 24 that Meta will add tens of millions of AWS Graviton CPU cores to run agentic AI workloads. (aboutamazon.com) (about.fb.com) A CPU is the general-purpose chip that handles lots of smaller tasks, while a graphics processing unit, or GPU, is built to crunch huge batches of math in parallel. Meta and AWS said agentic AI needs more of that CPU-style work because systems must reason, call tools, and coordinate steps continuously. (aboutamazon.com 1) (aboutamazon.com 2) AWS said the agreement is multiyear, and Reuters reported AWS executives described it as worth billions of dollars. Reuters also reported Meta will use Graviton5 central processing units for those workloads. (msn.com) The first chip in focus is Graviton5, which AWS says has 192 cores, uses a 3-nanometer process, and delivers up to 25% better performance than the prior generation. Amazon also said its cache is five times larger than Graviton4's, cutting communication delays by up to 33%. (aboutamazon.com 1) (aboutamazon.com 2) Meta said the new capacity will support real-time reasoning, code generation, and multi-step task orchestration. Those are the parts of AI systems that decide what to do next, route requests, and manage tool calls after a model produces an answer. (about.fb.com) (aboutamazon.com) The deal adds AWS chips to a broader Meta infrastructure mix that already spans Nvidia graphics processors, Advanced Micro Devices accelerators, Arm designs, and Meta's own silicon. Meta said no single chip architecture can efficiently serve every workload. (about.fb.com) (networkworld.com) That matters because AI spending has become a capacity race as companies try to secure enough hardware for training and inference. This agreement gives AWS a marquee buyer for Graviton and gives Meta another way to expand compute without relying only on scarce GPUs. (reuters.com) (aboutamazon.com) Meta and AWS framed the announcement as an expansion of an existing relationship that already includes Meta's use of Amazon Bedrock. The next step is deployment, with the companies saying the initial rollout starts at tens of millions of cores and can expand from there. (aboutamazon.com) (about.fb.com)