Meta Spent Billions on AI Chips — Not From Nvidia

AI CPU chips in a modern data center with emerald green lighting

KEY TAKEAWAYS

Meta signed a multi-year, multi-billion-dollar deal to deploy tens of millions of AWS Graviton CPU cores for agentic AI workloads.
Graviton5 features 192 cores with 5x larger cache and 33% lower latency, purpose-built for AI agent tasks like reasoning, code generation, and multi-step orchestration.
The deal marks a strategic shift: AI agents need CPUs, not just GPUs, creating a new front in the chip war between Amazon, Nvidia, and Google.
AWS timed the announcement to coincide with the end of Google Cloud Next, as Meta had previously signed a $10B deal with Google Cloud.

Forget the GPU gold rush for a moment. The next AI hardware battle is being fought over a chip most people have never heard of. Meta just committed billions of dollars to Amazon’s homegrown Graviton processors, ARM-based CPUs designed not for training giant models, but for running the AI agents that are about to reshape how businesses operate. It is a deal that quietly rewrites the rules of the AI chip market.

The Deal: Tens of Millions of Cores, Billions of Dollars

What Meta Actually Bought

Amazon announced on Friday that Meta has signed a multi-year agreement to use millions of AWS Graviton chips to power its growing AI operations. Reuters confirmed the deal spans multiple years and is worth billions of dollars. Critically, these are CPUs, not GPUs. While GPUs handle the heavy lifting of training frontier models, CPUs are what power the real-time reasoning, code generation, search, and multi-step orchestration that AI agents require.

“As we scale the infrastructure behind Meta’s AI ambitions, diversifying our compute sources is a strategic imperative,” said Santosh Janardhan, Meta’s head of infrastructure. “Expanding to Graviton allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale.”

Why Graviton5 Matters

The latest Graviton5 chip is built on 3-nanometer technology and packs 192 cores with a cache five times larger than the previous generation. That design reduces inter-core communication delays by up to 33%, a crucial advantage for agentic AI workloads where large-scale tasks must be distributed across many processors working in tight coordination. AWS vice president Nafea Bshara called the deployment a demonstration of what happens “when you combine purpose-built silicon with the full AWS AI stack to power the next generation of agentic AI.”

Business Insight — The Graviton deal signals that the AI hardware market is bifurcating. GPUs will remain essential for training, but the operational layer of AI, the agents that actually do work, runs on CPUs. Companies building AI strategies should evaluate CPU-optimized infrastructure alongside their GPU investments.

The Cloud Chess Match: AWS vs. Google vs. Nvidia

A Strategic Counter to Google Cloud

The timing of this announcement was no coincidence. AWS dropped the news just as Google Cloud Next 2026 wrapped up, where Google had showcased its own next-generation AI chips. Last August, Meta had signed a six-year, $10 billion deal with Google Cloud, shifting workloads away from its traditional AWS home. This Graviton agreement claws a significant portion of Meta’s AI compute spending back toward Amazon.

The Anthropic Factor

Amazon’s own AI GPU, the Trainium, is largely spoken for. Earlier this month, Anthropic signed a massive agreement to spend $100 billion over 10 years on AWS infrastructure, with Trainium chips at the center of that deal. In exchange, Amazon invested another $5 billion into Anthropic, bringing its total stake to $13 billion. With Trainium capacity effectively reserved, Graviton becomes Amazon’s primary weapon for winning new AI workloads from hyperscale customers like Meta.

Nvidia’s New Competitor

The Graviton deal also puts Amazon in direct competition with Nvidia’s new Vera CPU, an ARM-based chip designed for similar AI agentic workloads. AWS CEO Andy Jassy took direct aim at Nvidia and Intel in his recent annual shareholder letter, arguing that enterprises want better price-performance ratios. The difference: Nvidia sells its chips to anyone, while Amazon only offers Graviton through its cloud service, locking customers into the AWS ecosystem.

Business Insight — The AI chip market is no longer a single race. It has split into two parallel competitions: GPUs for training and CPUs for inference and agentic operations. Amazon is betting that whoever controls the CPU layer controls the economics of AI deployment at scale.

What This Means for the AI Industry

The Rise of Agentic Infrastructure

This deal is a strong market signal: the AI industry is transitioning from model-building to model-deploying. As organizations move from training frontier models to deploying autonomous agents that handle customer service, write code, manage workflows, and coordinate multi-step tasks, the infrastructure requirements are shifting fundamentally. Agents need sustained, low-latency compute, which is CPU territory, not burst GPU compute.

The Multi-Cloud AI Strategy

Meta’s approach reflects an emerging pattern among large enterprises: distributing AI workloads across multiple cloud providers and chip architectures to avoid vendor lock-in and optimize for different workload types. With Google Cloud for some workloads, AWS Graviton for agentic compute, and its own custom silicon for internal needs, Meta is building what may become the template for enterprise AI infrastructure in 2026 and beyond.

Business Insight — For enterprise leaders planning AI budgets, this deal is a wake-up call. The cost of running AI agents at scale will increasingly be determined by CPU performance and cloud pricing, not just GPU availability. Evaluating ARM-based CPU options should now be part of every AI infrastructure roadmap.

Sources

AI Biz Insider · AI Business EN · aibizinsider.com