Meta Signs Multibillion-Dollar Deal With Amazon to Use Its CPU Chips for AI

Amazon pulled off a quiet coup on Friday right as Google Cloud Next wrapped up by announcing that Meta has signed a multibillion-dollar deal to run its AI workloads on AWS Graviton chips. But this isn’t just another cloud contract. It signals something bigger: the AI chip wars are shifting from GPUs to CPUs, and Amazon just got one of the world’s most powerful AI companies to bet on its homegrown silicon.

What Is the Meta-Amazon Graviton Deal, Exactly?

Meta has agreed to deploy tens of millions of AWS Graviton CPU cores across its AI infrastructure, Amazon announced Friday morning. The deal is multibillion-dollar in scale and spans three to five years, with deployments concentrated in US data centers. Under the agreement, Meta ranks among AWS’s top five Graviton customers a distinction that tells you everything about the size of this commitment.

The Graviton chip is Amazon’s own ARM-architecture CPU, developed in-house and, until now, used exclusively to power Amazon’s own data center fleet. The latest iteration Graviton 5 was designed with AI-related compute tasks squarely in mind. What makes this deal notable isn’t just the dollar figure. It’s what Meta is using these chips for: not training large language models, but running AI agents at scale.

That’s a carefully worded sentence, and it says a lot. Meta isn’t betting on one chip vendor. The company already buys Nvidia GPUs, Broadcom processors, and AMD chips in enormous quantities a multi-vendor infrastructure strategy that mirrors how the biggest AI players are thinking about hardware risk in 2026. Adding Amazon’s Graviton to that stack reflects deliberate diversification, not a vendor preference.

Why Is Meta Using CPUs Not GPUs for AI?

This is the question that trips people up. GPUs like Nvidia’s Blackwell and Rubin chips are the workhorses of AI model training. Training requires massive parallelism, and GPUs are architecturally built for exactly that. But once a model is trained, something different happens: inference.

AI agents, which execute multi-step tasks, reason through problems, write code, run searches, and coordinate with other agents, create a different kind of compute demand. These workloads require fast, efficient general-purpose processing the kind CPUs handle better and more cost-effectively. As Meta scales its AI agent infrastructure across its apps (Instagram, WhatsApp, Facebook, Messenger), the volume of inference tasks dwarfs training cycles by orders of magnitude. CPUs become the economical choice.

Amazon built Graviton 5 with this exact use case in mind. The latest version of the chip is optimized for the real-time processing patterns that AI agents generate and it runs at a significantly better price-to-performance ratio than GPU inference setups. Amazon CEO Andy Jassy made this point explicitly in his annual shareholder letter earlier this month, where he took direct aim at Nvidia and Intel, arguing that enterprises want better price-performance ratios and that AWS intends to win on that basis.

What’s notable here is that this shift is broader than just Meta. The AI industry is maturing from a “train everything on GPUs” phase to a “run agents cheaply and at scale” phase. CPUs long overshadowed by the GPU gold rush are quietly returning to center stage.

Why This Deal Matters Right Now

Timing is everything in this story. Amazon announced the Meta deal on the exact same day Google Cloud Next 2026 concluded a conference where Google showed off its latest Trillium TPU chips and made a series of infrastructure announcements. That’s not a coincidence. It’s a calculated move by AWS to land a headline about a marquee customer while Google’s news cycle was winding down.

The cloud infrastructure market has become intensely competitive. Meta had previously signed a major six-year cloud deal with Google Cloud worth more than $10 billion back in August 2025, pulling business away from AWS where Meta had historically been a significant customer. This Graviton deal represents, at least in part, a partial return of spend to Amazon’s platform.

There’s also the Anthropic angle. Amazon had recently sealed a landmark deal where Anthropic committed $100 billion in AWS spending over 10 years in exchange for a further $5 billion Amazon investment bringing Amazon’s total Anthropic investment to $13 billion. A key part of that deal focused on Trainium, Amazon’s AI accelerator chip. With Trainium cores largely spoken for by Anthropic, Graviton CPUs became the natural chip to offer Meta. Two major AI players, two different AWS chips, one tightly coordinated strategy.

How AWS Graviton Works and Why Meta Chose It

AWS Graviton is Amazon’s custom silicon built on ARM architecture. Amazon began developing its own chips before 2018, launching the original Graviton processor as a cost-competitive alternative to Intel’s server CPUs. Over subsequent generations, Amazon iterated significantly Graviton 3 and Graviton 4 offered substantial performance improvements, and Graviton 5 is the first version built with AI workloads as a first-class use case.

The key design principle behind Graviton is price-performance optimization. Because Amazon builds these chips for its own massive internal workload and doesn’t need to sell them to external customers through traditional hardware channels it can optimize aggressively for its own cloud pricing model. The result is a chip that AWS claims offers better efficiency for many workloads than equivalent Intel Xeon or AMD EPYC server processors.

For Meta, the attraction is specific to AI agent inference. The rise of AI agents inside enterprise products has driven an explosion in continuous, low-latency compute requests. Every time someone asks Meta AI on WhatsApp to book a restaurant or Instagram’s AI to edit a reel, that’s an inference call. Multiply that by billions of users across Meta’s apps and you get an astronomical number of compute cycles that need to run cheaply, fast, and reliably.

Graviton’s ARM architecture also benefits from power efficiency important for Meta’s sustainability commitments and data center operating costs. At the scale Meta operates, even marginal improvements in energy efficiency per compute cycle translate into hundreds of millions of dollars in annual savings.

What This Means for Meta’s AI Agent Strategy

Meta has been building aggressively toward an AI-native product stack. The company’s $2 billion acquisition of AI agent startup Manus in December 2025 a startup focused on sophisticated multi-step AI task execution directly increases CPU demand. Manus-style agents don’t just chat; they research, plan, execute, and coordinate. That coordination overhead is precisely what makes CPUs more efficient than GPUs for the task.

Meta’s 2026 capital expenditure guidance of $115 to $135 billion nearly double its 2025 spending reflects how seriously the company is investing in AI infrastructure. You don’t spend that kind of money without having a chip strategy that spans every workload type. The Graviton deal fills the CPU-inference layer of that stack, sitting alongside Nvidia’s Blackwell GPUs (for training and heavy inference), AMD’s MI450s under a separate massive deal, and Meta’s own custom silicon projects.

What’s the part most people are missing in coverage of this deal? It’s not really about Amazon or Meta. It’s about what happens when the world’s most-used social platforms apps that collectively reach over four billion people become fully AI-agent-driven. The infrastructure that serves those agents at scale needs to be cheap, scalable, and diversified. That’s the actual demand signal driving deals like this one.

Challenges: The Risks Behind the Headline

No deal of this scale comes without complications. A few things are worth flagging.

First, Graviton chips are only available through AWS cloud services Amazon doesn’t sell them directly to companies for use in their own server farms. That means Meta is committing to running a portion of its AI workloads on Amazon’s cloud rather than its own owned infrastructure, which the company has traditionally preferred. The tradeoff between cost efficiency and infrastructure control is a real one, and Meta will need to manage the operational complexity of hybrid deployments across AWS, Google Cloud, Azure, and its own data centers.

Second, Amazon is reportedly exploring whether to sell Graviton chips externally to other companies’ server farms a potential expansion that would change the competitive dynamics significantly. If that happens, Nvidia’s Vera CPU (also ARM-based and designed for AI agentic workloads) would compete head-to-head with Graviton in the open market. Amazon CEO Andy Jassy’s semiconductor division is reportedly targeting $20 billion in annual revenue a target that implies external chip sales will eventually be part of the strategy.

Third, Meta’s multi-vendor chip strategy, while sensible for resilience, adds serious operational overhead. Managing different chip architectures, firmware updates, performance tuning, and cost optimization across Nvidia, AMD, Broadcom, AWS Graviton, and its own custom silicon requires engineering resources that even Meta’s scale may find stretched as the AI arms race accelerates.

What’s Next: Amazon’s Chip Ambitions Are Just Getting Started

The Meta deal is a validation moment for Amazon’s internal chip strategy but it’s also a signal of where the company wants to go. AWS has been building its silicon team for years, and the pressure to deliver is now extremely high. The Trainium AI accelerator (used by Anthropic, OpenAI, and reportedly Apple) and Graviton CPU are the two prongs of Amazon’s chip strategy one for training and heavy inference, one for scalable agentic compute.

Amazon’s ambition, as signaled by Jassy, is to become a credible alternative to Nvidia across the full AI compute stack. The Meta deal proves that the Graviton narrative better price-performance for inference at scale is resonating with buyers who can choose anyone. When one of the world’s largest AI spenders signs a multibillion-dollar commitment, it lends credibility that no amount of marketing can buy.

For the broader US cloud and AI infrastructure market, the trajectory is unmistakable. Microsoft, Alphabet, Amazon, Meta, and Oracle collectively committed more than $650 billion in capital expenditure in 2026 nearly double 2025 levels. The scale of this investment is now measured in gigawatts of compute, not server racks. The Meta-Amazon deal is one data point in a much larger story about how the US is building the physical infrastructure for the AI economy and who’s winning the race to supply it.

Expect more CPU-focused AI deals in the coming months. Nvidia’s Vera CPU, Intel’s AI server push, and Qualcomm’s ARM-based server efforts are all positioning for the same market. The GPU gold rush isn’t over but the inference compute market is now an equally large and separate battleground. Amazon just put Meta’s name on its scoreboard.

Frequently Asked Questions

What chips is Meta getting from Amazon in this deal?

Meta is purchasing access to AWS Graviton CPU cores Amazon’s ARM-based, in-house designed processors. The deal covers tens of millions of Graviton cores deployed primarily across US-based AWS data centers over a 3–5 year period. Graviton is a CPU, not a GPU, making it suited for AI agent inference rather than model training.

Why is Meta using CPUs instead of GPUs for AI?

GPUs like Nvidia’s Blackwell chips are best for training large AI models. But once a model is trained, AI agents running real-time tasks reasoning, searching, coding, multi-step coordination generate inference workloads that CPUs handle more efficiently and cost-effectively. At Meta’s scale of billions of users, CPUs offer better price-to-performance for that use case.

How much is the Meta-Amazon Graviton deal worth?

The deal is described as multibillion-dollar in scale across a multi-year timeframe, per The Wall Street Journal and corroborating reports from TechCrunch and other outlets. Exact financial terms have not been officially disclosed by either company as of April 24, 2026.

Does this deal replace Meta’s Google Cloud or other chip agreements?

No. Meta is running a deliberate multi-vendor chip strategy that includes Nvidia, AMD, Broadcom, AWS Graviton, and its own custom silicon. The Graviton deal adds CPU inference capacity alongside existing GPU and TPU agreements. Meta previously signed a 6-year, $10 billion deal with Google Cloud in August 2025, which remains active.

What does this deal mean for Nvidia?

In the short term, limited impact Nvidia dominates GPU training and still powers much of Meta’s largest model workloads. But the deal signals that the inference compute market, which is growing faster than training, is increasingly contested by ARM-based CPUs like Graviton and Nvidia’s own Vera CPU. Nvidia’s dominance in AI compute is real but narrowing at the inference layer.