Meta picks Amazon's 3nm Graviton chips to power next AI wave

Meta has signed a deal to deploy tens of millions of AWS Graviton processor cores as it expands the computing infrastructure for its next generation of AI systems, deepening its relationship with Amazon Web Services and signaling a shift in how AI infrastructure is built.

The agreement highlights a growing recognition in the AI industry: while GPU clusters remain essential for training large models, CPU power is becoming the bottleneck for inference, real-time reasoning, search, coding tools, and multi-step AI agents. Meta's bet on Graviton — Amazon's custom ARM-based processors built on 3nm technology — reflects the calculation that inference compute demand will outpace training demand as AI systems move from development to deployment at scale.

Amazon said the rollout will begin with tens of millions of Graviton cores and can expand further. The scale is significant — Meta currently operates one of the world's largest custom AI infrastructure footprints, and supplementing it with AWS capacity suggests that even Meta's internal infrastructure can't keep pace with projected demand.

The deal also represents a competitive dynamic within big tech. Meta could have expanded its own data center capacity or turned to other cloud providers, but chose AWS and its custom silicon. This is a vote of confidence in Amazon's chip design capabilities — Graviton has evolved from a cost-saving experiment into a performance-competitive processor that major AI companies are willing to bet on.

For the broader market, the deal underscores that AI infrastructure spending is accelerating, not plateauing. Every major tech company is simultaneously building and buying compute capacity, and the supply constraints that characterized 2024-2025 have not eased — they've shifted from GPUs to the supporting infrastructure needed to run trained models at scale.

**What This Means For You:** For investors, this deal validates Amazon's custom silicon strategy and positions AWS as the inference infrastructure layer for AI — potentially more durable revenue than training compute, which can be done once. For anyone working in cloud or AI infrastructure, ARM-based processors for inference workloads are the growth area. The GPU shortage narrative is evolving into a broader compute infrastructure story, and the companies building that layer — Amazon, Google with TPU, Microsoft with Maia — are where the long-term value accumulates.