Meta has signed a deal to deploy millions of Amazon’s homegrown Graviton CPUs to power its growing AI workloads, Amazon announced Friday. According to TechCrunch AI, the agreement marks a significant win for AWS, pulling more of Meta’s infrastructure spending back into Amazon’s orbit after Meta inked a six-year, $10 billion contract with Google Cloud last August. The timing was no accident: AWS dropped the news just as Google Cloud Next wrapped up.
What stands out here is the chip type. Graviton is an ARM-based CPU, not a GPU. GPUs still dominate model training, but the rise of AI agents is reshaping demand for what runs after training. Agents handle real-time reasoning, code generation, search, and multi-step task coordination. Those workloads lean heavily on CPUs, and AWS says its latest Graviton was purpose-built for this category of compute.
Why Meta Picked Graviton
Meta has historically been an AWS customer that also tapped Microsoft Azure, before broadening to Google Cloud last year. Bringing Meta back as a flagship Graviton customer gives Amazon a marquee proof point for its in-house silicon strategy, per TechCrunch AI’s reporting.
The deal lands in a chip market that’s getting more crowded by the month:
- Nvidia’s Vera CPU is also ARM-based and aimed at agentic workloads, sold to enterprises and cloud providers (AWS included).
- Google’s custom AI chips got new versions unveiled at Google Cloud Next.
- Amazon’s Trainium GPU handles both training and inference, but much of its capacity is now spoken for by Anthropic.
Anthropic’s role matters here. Earlier this month, the Claude maker committed $100 billion over 10 years to run workloads on AWS with a heavy Trainium focus, while Amazon raised its total Anthropic investment to $13 billion. That left AWS needing another high-profile customer to validate its CPU lineup. Meta fits the bill.
The Bigger Strategic Picture
Amazon CEO Andy Jassy used his annual shareholder letter this month to take direct aim at Nvidia and Intel, arguing that enterprises want better price-performance ratios for AI compute and that AWS plans to win on that math. The Meta announcement is the first major receipt backing that pitch.
There’s a structural difference worth flagging. Nvidia sells chips and systems to anyone who’ll buy them. AWS only rents access to Graviton through its cloud. That makes every large customer commitment doubly valuable for Amazon: it’s revenue, and it’s also the only way the chips get used in the wild.
What This Signals for the Industry
Three things to watch:
- CPU demand is climbing fast. As more products ship AI agents into production, inference and orchestration workloads will push CPU spending up alongside the GPU race.
- Hyperscalers are doubling down on custom silicon. Amazon, Google, and Microsoft all want to reduce dependence on Nvidia. Meta’s choice signals that homegrown chips are now credible at the largest scale.
- Pressure on Amazon’s chip team is intense. With Anthropic locking in Trainium and Meta locking in Graviton, AWS has bet its AI roadmap on its own labs delivering competitive performance year over year.
The AI infrastructure war is no longer just about who has the most GPUs. It’s about who controls the full compute stack underneath the agents that businesses are starting to deploy at scale. Full details are available at the original source.