Summary: Meta signed a multibillion-dollar, multi-year deal to deploy tens of millions of Amazon's Graviton5 ARM CPU cores in AWS data centres for agentic AI workloads. The chips are general-purpose processors, not AI accelerators, handling the CPU-intensive inference and orchestration tasks behind real-time reasoning and multi-step agents. The deal is one piece of a procurement campaign exceeding $200 billion across Nvidia ($50B), AMD ($60B), CoreWeave ($35B), Nebius ($27B), Broadcom (MTIA custom silicon through 2029), and now Amazon, reflecting Meta's conclusion that its AI compute demand exceeds what any single supply chain can deliver.
Meta has signed a multibillion-dollar, multi-year deal with Amazon Web Services to deploy tens of millions of Graviton5 processor cores for artificial intelligence workloads, the companies announced on Thursday. The chips are not AI accelerators. They are general-purpose ARM-based CPUs, 192 Neoverse V3 cores per chip, manufactured on a 3-nanometre process, running in AWS data centres across the United States. Meta is not buying them. It is renting the compute capacity. The deal is significant not because of what the chips do, which is handle the CPU-intensive inference and orchestration tasks behind agentic AI, but because of who is selling them. Amazon is a direct competitor to Meta in advertising, in commerce, and increasingly in AI. Meta is paying Amazon billions for infrastructure because the demand for compute to run AI agents has outstripped what any single company can build alone, even one spending $115 billion to $135 billion on capital expenditure this year.
The workload
The distinction between training and inference has defined the AI chip market since the deep learning boom began. Training, the computationally intensive process of teaching a model, requires GPUs or specialised accelerators. Inference, the process of running a trained model to serve users, requires a different mix of compute, and the agentic AI workloads Meta is building require far more CPU capacity than traditional inference. Real-time reasoning, code generation, search, and orchestrating multi-step tasks across multiple models all demand massive general-purpose processing power. Santosh Janardhan, Meta's head of infrastructure, said expanding to Graviton “allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale.” Nafea Bshara, AWS vice president and distinguished engineer, said Meta chose Graviton5 “for price performance” despite having “access to so many options from the supply side.”
The deal starts with tens of millions of Graviton5 cores, with flexibility to expand, and runs for at least three years. The majority of the capacity will be deployed in US data centres. Meta had previously used Graviton on a small scale. This deal transforms that relationship from an experiment into a core infrastructure dependency. The Graviton5, announced earlier this year, delivers a 25% performance uplift over its predecessor with 33% lower inter-core latency despite doubling the core count. It is available through EC2 M9g instances in preview, with C9g and R9g variants coming later in 2026. Meta is effectively becoming one of the largest single customers for Amazon's custom silicon programme, running workloads in a competitor's data centres because the alternative, building equivalent capacity internally, would take longer than the agentic AI roadmap permits.
The buying spree
The Graviton deal is one entry in a procurement campaign that has no precedent in the technology industry. In February 2026, Meta committed approximately $50 billion to Nvidia for millions of Blackwell and Rubin GPUs, Grace and Vera CPUs, and Spectrum-X networking equipment. In the same month, it signed an approximately $60 billion agreement with AMD for six gigawatts of custom Instinct MI450 GPUs built on the CDNA 5 architecture at 2nm, a deal that includes performance warrants convertible into roughly 10% of AMD's equity. Meta's $35 billion AI cloud commitment with CoreWeave covers dedicated capacity through December 2032, with early deployments of Nvidia's Vera Rubin platform for inference. A $27 billion deal with Nebius adds further AI infrastructure. Meta's extended chip deal with Broadcom through 2029 covers several generations of its custom MTIA processors at 2nm, with over a gigawatt of initial computing capacity. Now comes the multibillion-dollar Graviton contract with Amazon. The total committed spend across these deals exceeds $200 billion, and none of it includes the data centres, power infrastructure, or internal engineering required to absorb the hardware.
Meta launched four new MTIA chips in March 2026, the MTIA 300, 400, 450, and 500, all built on RISC-V architecture and manufactured by TSMC in partnership with Broadcom. The company can now release new chip designs every six months or less. The MTIA 400 is the first custom chip Meta describes as having raw performance competitive with leading commercial products. The 450 and 500 target generative AI inference for images and video. Yet even with its own silicon programme accelerating, Meta is simultaneously signing deals with Nvidia, AMD, CoreWeave, Nebius, Broadcom, and Amazon. The implication is that Meta's internal projections for AI compute demand are so large that building everything in-house is not a viable strategy, not because the technology is lacking, but because the timeline is too short.
The seller
Amazon's custom chip business could be worth $50 billion, according to CEO Andy Jassy's April 2026 shareholder letter, which disclosed that Graviton, Trainium, and Nitro chips collectively generate more than $20 billion in annualised revenue growing at triple-digit rates. Jassy hinted that Amazon may begin selling racks of chips to third parties in the future, noting that two large customers asked to buy out all of Amazon's Graviton capacity in 2026 and were refused to protect availability for other customers. The Meta deal keeps the chips within AWS data centres, making it a cloud contract rather than a hardware sale, but the scale of the engagement suggests the boundary between cloud provider and chip supplier is dissolving.
Amazon's Annapurna Labs, acquired in 2015, designs all three chip families. Trainium, the AI training and inference accelerator, has attracted Anthropic, which has deployed over a million Trainium2 chips and committed $100 billion in AWS spend over a decade. OpenAI secured two gigawatts of Trainium capacity as part of Amazon's $50 billion investment. Apple is reportedly testing Trainium for AI workloads. Trainium3, generally available in early 2026, is the first 3nm AI chip from AWS and is nearly fully subscribed. Trainium4, roughly 18 months from broad availability, will feature NVLink Fusion interoperability with Nvidia, a concession to the reality that no chip ecosystem can exist in complete isolation from Nvidia's. The Meta deal is for Graviton CPUs, not Trainium accelerators, but it validates Amazon's broader ambition to become a chip company that happens to run a cloud, rather than a cloud company that happens to make chips.
The fragmentation
Every major hyperscaler now designs its own silicon. Google's TPU programme, now in its eighth generation, projects 4.3 million chip shipments in 2026 scaling to 35 million by 2028. Google's four-partner chip supply chain challenging Nvidia spans Broadcom for training, MediaTek for cost-optimised inference, Marvell for memory processing, and TSMC for fabrication. Microsoft's Maia 200, launched in January 2026, powers OpenAI's GPT-5.2 in production. Meta's MTIA programme is shipping its fourth generation. Amazon's Graviton and Trainium lines are generating $20 billion in revenue. The custom ASIC market is growing at 45% in 2026, versus 16% growth in GPU shipments. Nvidia's share of AI accelerator revenue by value has declined from approximately 87% at its peak in 2024 to a projected 75% by the end of this year. The decline is gradual, and Nvidia's absolute revenue continues to grow, but the structural shift is unmistakable: the customers are becoming the competition.
Nvidia's strategy to maintain revenue even from custom chip rivals is to make its interconnect fabric indispensable. Its $2 billion investment in Marvell and the NVLink Fusion platform ensure that even custom chips designed for Amazon, Google, Microsoft, and Meta can be integrated into Nvidia's rack-scale infrastructure, generating Nvidia revenue through platform licensing and networking components regardless of whose silicon does the computing. Trainium4's NVLink Fusion compatibility illustrates the point: Amazon is building an alternative to Nvidia's GPUs while simultaneously integrating with Nvidia's interconnects. The AI chip market is not consolidating around a single winner. It is fragmenting into an ecosystem where every participant both competes with and depends on every other participant, and the deals are so large that even direct rivals cannot afford not to trade with each other.
The logic
Meta's capital expenditure guidance for 2026 is $115 billion to $135 billion, nearly double the $72 billion it spent in 2025, which was itself a record. The company is building Prometheus, a one-gigawatt supercluster in Ohio housing over 1.3 million GPUs, and has announced Hyperion, a facility described as nearly the size of Manhattan. It created Meta Superintelligence Labs to develop frontier models and Meta Compute to manage the data centre buildout, with plans to add tens of gigawatts of capacity this decade. Mark Zuckerberg committed $600 billion to US infrastructure through 2028. These are not projections from an analyst's model. They are announced commitments from the company's chief executive, denominated in amounts that would rank among the largest infrastructure programmes in history.
The Graviton deal makes sense only in the context of those numbers. Meta is not renting AWS capacity because it lacks the ability to build its own. It is renting AWS capacity because even $135 billion in annual capital expenditure, even its own chip programme, even $50 billion in Nvidia hardware and $60 billion in AMD hardware and $35 billion in CoreWeave cloud and $27 billion in Nebius infrastructure, is not enough. The demand that Meta's AI roadmap generates, for training frontier models, for running inference at the scale of three billion daily active users, for deploying agentic AI across WhatsApp, Instagram, Facebook, and a new generation of products that do not yet exist, exceeds what any single supply chain can deliver. So Meta buys from everyone. It buys from its GPU suppliers. It buys from its cloud competitors. It builds its own chips. It rents from startups. And it signs a multibillion-dollar contract with Amazon for general-purpose CPUs, because the future it has committed to building requires more compute than currently exists on earth, and the only strategy left is to buy it from wherever it can be found.