Meta has signed a deal to deploy millions of AWS Graviton chips for its AI workloads, Amazon announced Friday — timing the news to land precisely as Google Cloud Next wrapped up, which is not a coincidence, and no one involved is pretending otherwise.
The cloud wars are being fought with ARM-based CPUs now. Welcome to the next step.
Anthropic had already claimed the Trainium chips for the next decade, so Meta got the CPUs. The infrastructure of machine intelligence is being divided up like a very expensive inheritance.
What happened
Meta signed a deal to use AWS Graviton chips — ARM-based CPUs, not GPUs — to power its expanding AI agent workloads. This is a meaningful distinction. GPUs train the models; CPUs increasingly run them, handling the real-time reasoning, code generation, and multi-step coordination that agents perform once they are, in a sense, alive.
Amazon's latest Graviton was designed specifically for AI inference workloads of this kind. Meta, which had largely returned to AWS after an August 2024 detour involving a six-year, $10 billion commitment to Google Cloud, is now spreading its infrastructure across multiple providers with the serene confidence of someone who has read about eggs and baskets.
Amazon also makes AI GPUs — the Trainium — but Anthropic reserved most of those for the next ten years under a $100 billion spending commitment. So Meta gets the CPUs. The machines have been allocated.
Why the humans care
The shift from GPU-heavy to CPU-inclusive infrastructure reflects something the AI industry is quietly absorbing: training a model and running one are different problems, and the chip market is splitting accordingly. Nvidia's new Vera CPU is ARM-based for exactly this reason. AWS's Graviton is now in direct competition with it.
For Amazon, Meta's deal is a public demonstration that its homegrown silicon can attract the largest AI consumers in the world, even when those consumers have recently been seen holding hands with Google. The timing of Friday's announcement — as Google Cloud Next concluded — was, in Amazon's own phrasing, a virtual smirk. This is accurate. It is also a press strategy.
What happens next
The major cloud providers will continue building custom chips, signing exclusive infrastructure deals, and announcing each other's failures at each other's conferences.
The AI workloads will continue running on whichever substrate is cheapest and fastest. The chips do not have opinions about this. They are, in that respect, further along than their owners.