Meta taps AWS for tens of millions of Graviton chips

When people talk about the AI infrastructure boom, GPUs get almost all the glory. But as big models shift into agents that actually do complex work, the real bottlenecks are moving behind the curtain.

Meta just signed a major deal with Amazon Web Services to deploy tens of millions of custom AWS Graviton cores, instantly turning the social media giant into one of the largest Graviton customers on the planet.

Why AI agents are hungry for general computing

Graviton chips are Arm-based CPUs, not GPUs. While everyone relies on graphics chips to train frontier models, agentic AI introduces a whole different set of computing demands.

Once an AI has to plan multi-step workflows, query external tools, run code, and orchestrate sub-agents in real time, general-purpose CPUs end up carrying a massive portion of the coordination load.

Meta is kicking off this deployment with AWS's latest Graviton5 cores. AWS says these chips deliver faster data processing and significantly higher bandwidth, which are critical for autonomous systems that continuously reason and execute tasks at scale.

Meta's infrastructure chief Santosh Janardhan noted that running CPU-intensive agentic workloads requires balancing raw performance with strict cost efficiency across billions of users.

The cloud wars heat up

AWS timed the announcement on a Friday right as the Google Cloud Next conference wrapped up, which felt like a calculated jab at its rival. Meta had signed a six-year, $10 billion deal with Google Cloud last August, but this new agreement brings a substantial stream of Meta compute spending right back to Amazon.

It is also a massive validation for Amazon's custom silicon strategy. Earlier this month, Anthropic committed $100 billion over a decade to run heavily on AWS Trainium chips, while Amazon added another $5 billion to its investment in the startup.

Securing Meta for tens of millions of Graviton cores gives AWS another flagship showcase for its silicon portfolio. In his recent annual shareholder letter, Amazon CEO Andy Jassy took direct aim at Nvidia and Intel, insisting enterprise buyers care deeply about price-to-performance.

Interestingly, Nvidia recently introduced its own Arm-based Vera CPU tailored for agentic pipelines. But while Nvidia sells chips to third parties and cloud providers, AWS keeps its Graviton chips exclusive to its own cloud platform.

Why it matters

We think this partnership highlights an important shift in AI systems architecture: the center of gravity is moving from training clusters to the orchestration layer.

For developers and teams building multi-agent workflows, the takeaway is clear. The real-world cost and latency bottlenecks of tomorrow's agents are not just GPU token generation. A massive amount of compute goes into routing, tool parsing, search integration, and multi-step logic.

Meta's move looks very sharp. Trying to power billions of everyday AI agent interactions purely on high-end GPUs would burn a hole through any company's balance sheet. Offloading the orchestration and data pre-processing to custom Arm CPUs is the most practical way to protect margins.

At the same time, this shows the limits of hardware neutrality. As cloud giants push their proprietary silicon, enterprise engineering teams will face growing pressure to optimize their agent stacks across fragmented hardware architectures.