Meta Signs With AWS Graviton—Hundreds of Thousands of Chips for Agentic AI
Meta struck a deal with AWS to deploy hundreds of thousands of Graviton chips for agentic AI workloads. The agreement gives Meta access to Amazon's custom-designed processors as it competes with Nvidia.
- Meta will deploy hundreds of thousands of AWS Graviton chips across its infrastructure.
- The chips will power Meta’s agentic AI workloads and next-generation AI capabilities.
- The deal represents a major customer win for Amazon’s custom silicon against Nvidia.
Meta and Amazon Web Services announced a partnership to power Meta’s agentic AI ambitions with hundreds of thousands of AWS Graviton chips. The deal gives Meta exclusive early access to next-generation Graviton processors— reportedly Graviton4—while cementing AWS’s position as a serious alternative to Nvidia for AI infrastructure.
The announcement came Thursday, April 24. Meta will run its agentic AI workloads on Graviton chips, the custom-designed ARM-based processors Amazon has been developing since 2018. AWS says the deployment represents a significant expansion of Graviton adoption by one of the world’s largest AI companies.
For Meta, the deal solves a supply problem. The company has been scrambling for AI compute capacity, building out its own AI infrastructure while also relying on external cloud providers. Using Graviton lets Meta diversify away from Nvidia dominance—though Meta will certainly continue buying plenty of Nvidia chips too. The partnership gives Meta cost-effective inference capacity for the less demanding agentic AI workloads while keeping more expensive GPUs reserved for training.
AWS Graviton Gets a Validation Moment
For AWS, the Meta deal validates years of investment in custom silicon. Graviton chips—designed by Amazon’s Annapurna Labs—cost less than equivalent x86 processors while offering competitive performance for many workloads. Now a FAANG company is betting big on them. CNBC reported that Meta will adopt “hundreds of thousands” of Graviton chips, marking one of the largest Graviton deployments to date.
Graviton4 specifically was announced in late 2023, featuring improved performance and memory bandwidth over Graviton3. For agentic AI—AI agents that can take actions autonomously—inference efficiency matters almost as much as raw training throughput. Graviton’s power efficiency and cost structure make it attractive for 24/7 agent workloads that require constant inference.
The deal also signals something about the AI infrastructure market: hyperscalers are becoming comfortable with heterogeneous compute. Companies like Meta no longer default to Nvidia for every AI task. They’re increasingly willing to match workloads to whichever chip gives the best performance-per-dollar. AWS Graviton, while not a GPU replacement for training, has become legitimate for inference at scale.
Meta says it will be among the first companies to gain access to new capabilities in Graviton. The company specifically called out efficiency gains—Graviton chips consume less power than alternatives while offering competitive inference throughput. For Meta’s agentic AI rollout, which will require inference across billions of daily interactions, power efficiency translates directly to cost savings.
The announcement didn’t specify exact chip quantities or dollar amounts, but “hundreds of thousands” suggests a commitment worth hundreds of millions of dollars at minimum. For context, AWS launched Graviton4 with up to 96 Neoverse V2 cores and 536 GB/s memory bandwidth. Deployed at Meta’s scale, that’s substantial custom silicon volume. Amazon’s years-long bet on ARM servers appears to have found its validation moment.
