Meta Signs Multibillion-Dollar Deal for Tens of Millions of AWS Graviton5 Cores as Agentic AI Reshapes the CPU Market
Meta will deploy tens of millions of AWS Graviton cores to power CPU-intensive agentic AI workloads in a multiyear, multibillion-dollar partnership that AWS describes as one of its largest Graviton commitments to date and that extends Meta's push to diversify beyond Nvidia.
Overview
Meta announced on April 24 that it will deploy tens of millions of AWS Graviton cores to power what it calls CPU-intensive agentic AI workloads, in a partnership disclosed jointly by both companies. Neither company published a contract value, but AWS vice president Nafea Bshara told Reuters that the agreement “would span multiple years and be worth billions of dollars,” according to Reuters. CNBC reports the deal will last at least three years and involve hundreds of thousands of Graviton processors. The agreement makes Meta one of the largest Graviton customers in the world and represents the latest in a string of infrastructure commitments by the company that now exceeds $200 billion across vendors, as catalogued by The Next Web.
The agreement is notable for what it is not: a GPU contract. While the AI infrastructure conversation has been dominated by Nvidia accelerators, Meta is buying Arm CPUs at hyperscale to handle the reasoning, orchestration and tool-use steps that sit on top of trained models. TechCrunch reports that GPUs remain the chip of choice for training large models, but inference and agent execution increasingly benefit from CPU optimization, which is what Graviton is designed to provide.
What We Know
Scale and structure of the deal
The joint announcement does not disclose a contract value, but AWS confirmed that the deployment begins with tens of millions of Graviton cores with flexibility to expand. Bshara separately told Reuters the deal is worth billions of dollars over multiple years, and CNBC reports that it spans at least three years and involves hundreds of thousands of Graviton processors. Meta’s infrastructure head Santosh Janardhan said the move “allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale,” according to AWS’s release.
Meta is not buying the silicon outright. The cores will be consumed as a cloud service through AWS, integrated with the broader AWS AI stack, as described by Meta. That structure mirrors Meta’s earlier deal to lease Google TPUs, previously reported by The Machine Herald, and contrasts with the company’s owned-data-center strategy and its custom MTIA accelerators developed with Broadcom, covered earlier this month.
What Graviton5 is
Graviton5 is AWS’s fifth-generation custom Arm processor and the chip Meta will run on. According to specifications published by AWS, each chip carries 192 cores, is built on a 3-nanometer process, ships with five times the cache of the previous generation, reduces core-to-core communication delays by up to 33 percent thanks to that larger cache, and offers up to 25 percent better performance than the prior Graviton generation. AWS also markets the chip with low-latency networking through its Elastic Fabric Adapter.
The positioning matters because Graviton5 is no longer a budget alternative. AWS’s framing places it alongside the most recent server CPUs from AMD and Intel, with the added advantage of being a single-vendor stack tightly coupled to AWS networking and accelerators.
Why CPUs, not GPUs
Meta and AWS both frame the deal around “agentic AI” — workloads in which models call tools, search, generate code, and orchestrate multi-step tasks rather than simply emitting text. Janardhan’s statement, carried in Meta’s release, emphasizes that these are CPU-intensive operations.
TechCrunch’s reporting makes the distinction explicit: training stays on GPUs, but “once those models are trained, AI agents built on top of them are causing a shift in the type of chip needed.” Real-time reasoning, code writing and multi-step coordination involve a high volume of small, branching, control-heavy operations — work for which general-purpose cores with large caches are well suited.
Meta’s broader spending pattern
The Graviton deal lands inside an unprecedented capital-expenditure cycle. The Next Web reports that Meta’s 2026 capex guidance is $115-135 billion, nearly double the $72 billion spent in 2025, and that CEO Mark Zuckerberg has committed an additional $600 billion to U.S. infrastructure through 2028.
The same report tallies the recent procurement campaign: roughly $50 billion to Nvidia for Blackwell and Rubin GPUs, around $60 billion to AMD for custom Instinct MI450 GPUs, $35 billion with CoreWeave for dedicated capacity through 2032, $27 billion with Nebius, and Broadcom custom MTIA processors through 2029. Meta also signed a six-year, $10 billion Google Cloud deal in August 2025, as TechCrunch notes.
What We Don’t Know
The most consequential numbers were not officially disclosed. Neither Meta nor AWS published the dollar value of the deal, the exact number of cores in the initial commitment, or a precise contract length beyond Reuters’ “multiple years” attribution and CNBC’s “at least three years” framing. “Tens of millions of cores” is a very wide band: at 192 cores per Graviton5 chip, the floor of that range translates to roughly 100,000 chips, but the ceiling could be several times higher.
It is also unclear how the Graviton deployment will be integrated with Meta’s other compute pools — whether the cores will sit behind Meta’s own orchestration layer, whether they will run alongside AWS Trainium accelerators in the same AI stack, or whether specific Meta products (Llama agents, Meta AI assistants, advertising orchestration) will be the first workloads. Neither company addressed performance claims for agentic workloads specifically; the up-to 25 percent generational improvement cited by AWS is a general figure rather than a workload-specific benchmark.
The environmental and energy footprint of adding tens of millions of cores at this scale, and how that interacts with AWS’s data-center expansion plans, was not addressed in either announcement.
Analysis
The deal is a data point about how the AI compute market is fragmenting beyond a single chip type. Hyperscalers and frontier AI buyers are now assembling stacks across at least four categories of silicon: training GPUs (Nvidia, AMD), inference accelerators (custom ASICs like MTIA, Trainium, TPU inference variants), Arm CPUs for orchestration and post-training work, and increasingly, networking and photonics silicon. Meta’s portfolio now touches every category, often through multiple suppliers.
For AWS, landing Meta as a flagship Graviton5 customer is a credibility milestone for its custom-silicon program. Graviton has long been pitched on price-performance, but agentic AI gives AWS a workload narrative that Nvidia cannot easily address with GPUs. AWS VP Nafea Bshara framed the partnership as evidence of “what happens when you combine purpose-built silicon with the full AWS AI stack,” in the company’s release.
For Meta, the strategic logic is diversification at every layer. The company is now committed to Nvidia, AMD, Broadcom, Google and Amazon for AI silicon, alongside its own MTIA program. That posture limits exposure to any single supplier’s roadmap or pricing, but it also implies a substantial software-engineering bill: agentic systems that run partially on Trainium, partially on Graviton, partially on Nvidia, and partially on Meta’s own MTIA require an abstraction layer that does not yet exist as a standard.
The broader signal is that the “AI is GPUs” framing of 2023-2025 is giving way to a more textured picture, in which CPUs — long treated as legacy infrastructure — are being repositioned as a strategic AI input. Whether that reshapes Nvidia’s economics in the medium term is an open question; in the short term, it is reshaping AWS’s.