Meta has signed a deal to deploy millions of AWS Graviton processors for its AI workloads, Amazon announced Friday. This marks a strategic pivot for Meta, which has historically leaned on GPUs for model training but now faces growing compute pressure from inference-stage tasks.
The Graviton is an ARM-based CPU, not a GPU. While GPUs remain the workhorses for training large language models, the operational phase—real-time reasoning, code generation, multi-step agent orchestration—creates a different bottleneck. CPUs handle these sequential, memory-intensive workloads more efficiently per watt. AWS designed the latest Graviton generation specifically for this inference profile.
This deal redirects some of Meta’s cloud spend back to AWS, after Meta signed a $10 billion, six-year commitment with Google Cloud last August. Meta also uses Microsoft Azure, but this Graviton commitment signals AWS is winning the inference-tier compute race—at least for now.
Amazon timed the announcement just as Google Cloud Next concluded—a subtle competitive jab. Google also produces custom AI chips, including its latest TPU and CPU variants announced at the event.
Amazon’s own AI GPU, Trainium, serves both training and inference, but most of its capacity is already locked up by Anthropic. Earlier this month, Anthropic committed $100 billion over ten years to run on AWS, with Trainium as the primary target, in exchange for another $5 billion from Amazon (total investment: $13 billion).
For Meta, the Graviton deal provides a lower-cost, high-throughput inference path without competing for scarce GPU cycles. For AWS, it’s a public validation that their custom silicon can compete with Nvidia’s Vera CPU—another ARM-based design targeting agentic workloads. The key difference: Nvidia sells chips to enterprises and cloud providers alike; AWS only rents access through its cloud.
CEO Andy Jassy’s recent shareholder letter took direct aim at Nvidia and Intel, arguing enterprises need better price-performance for AI. The internal chip team now carries that mandate—and this Meta deal raises the stakes considerably.
Source: TechCrunch