Meta has entered into a multibillion-dollar agreement for Amazon Graviton5 chips, as the demand for AI computing exceeds their $135 billion capital expenditure budget.

      Summary: Meta has entered into a multibillion-dollar, multi-year agreement with Amazon to deploy tens of millions of Amazon's Graviton5 ARM CPU cores in AWS data centers for agentic AI workloads. These processors are general-purpose, not specifically designed for AI, and are capable of managing CPU-intensive inference and orchestration tasks required for real-time reasoning and multi-step agents. This deal is part of Meta's broader procurement effort, which totals over $200 billion and includes investments in Nvidia ($50B), AMD ($60B), CoreWeave ($35B), Nebius ($27B), Broadcom (MTIA custom silicon until 2029), and now Amazon, underscoring Meta's realization that its demand for AI computing capabilities surpasses what any single supply chain can offer.

      Meta announced an agreement with Amazon Web Services to utilize millions of Graviton5 processor cores for AI workloads. These chips, based on ARM architecture and containing 192 Neoverse V3 cores per chip, are manufactured using a 3-nanometre process and will be utilized in AWS data centers across the U.S. Instead of purchasing these processors outright, Meta will be renting the computational power. The significance of this arrangement lies not in the chips' functions, which include handling CPU-intensive tasks for agentic AI, but in the fact that Amazon is a direct competitor to Meta in several areas, including advertising, commerce, and increasingly in AI. Meta's decision to invest billions with Amazon for infrastructure reflects the overwhelming demand for AI compute capabilities, which surpasses the ability of any single company—even one with $115 billion to $135 billion in capital expenditures this year—to fulfill independently.

      The workload involved primarily differentiates between training and inference, a distinction shaping the AI chip market since the rise of deep learning. Training is the resource-heavy process involving teaching a model, requiring GPUs or specialized accelerators. Inference, which executes a trained model for user service, necessitates a different type of computational mix, and the workloads that Meta is developing demand much greater CPU power than traditional inference. Santosh Janardhan, Meta's infrastructure head, indicated that the shift to Graviton "allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we require at our scale." Nafea Bshara, AWS's vice president and distinguished engineer, noted that Meta selected Graviton5 "for price performance," even while having "access to numerous options from suppliers."

      The agreement commences with tens of millions of Graviton5 cores, with potential for future expansion, and lasts for a minimum of three years, primarily in U.S. data centers. Meta had previously utilized Graviton on a minor scale, but this deal elevates it to a central infrastructure necessity. The newly introduced Graviton5 offers a 25% performance increase over its predecessor and 33% lower inter-core latency, despite doubling the core count. Currently available through EC2 M9g instances in preview, C9g, and R9g variants will debut later in 2026. This positions Meta as one of the foremost single customers in Amazon's custom silicon program, utilizing competitor data centers to meet short-term demands that far exceed the timeframe allowed for building equivalent capacity internally.

      The Graviton agreement is just one part of an unprecedented procurement initiative in the technology sector. In February 2026, Meta allocated around $50 billion to Nvidia for millions of Blackwell and Rubin GPUs, Grace and Vera CPUs, and Spectrum-X networking components. In the same month, Meta signed a deal with AMD for approximately $60 billion for six gigawatts of custom Instinct MI450 GPUs based on the CDNA 5 architecture at 2nm, which includes performance warrants convertible into about 10% of AMD’s equity. Meta’s $35 billion commitment for AI cloud services with CoreWeave secures dedicated capacity through December 2032, featuring early deployments of Nvidia’s Vera Rubin platform for inference. Additionally, a $27 billion agreement with Nebius further boosts Meta's AI infrastructure, alongside an expanded chip procurement agreement with Broadcom for several generations of MTIA processors through 2029, including over a gigawatt of initial computing capacity. The overall committed spending from these contracts exceeds $200 billion, not accounting for the necessary data centers, power infrastructure, or internal engineering to support the hardware.

      In March 2026, Meta introduced four new MTIA chips: the MTIA 300, 400, 450, and 500, all based on RISC-V architecture and produced by TSMC in collaboration with Broadcom. The company is now capable of launching new chip designs every six months or less. The MTIA 400 is distinguished as Meta’s first custom chip with raw performance matching leading commercial products, while the 450 and 500 focus on generative AI inference for images and videos. Despite the acceleration of its internal silicon program, Meta continues