One day after announcing 8,000 layoffs, Meta signed a billion-dollar chip deal with Amazon. The timing is jarring. But it tells you exactly what kind of company Meta is becoming: one that cuts people and buys compute, in that order. The deal with Amazon Web Services covers tens of millions of Graviton5 CPU cores, spans three to five years, and makes Meta one of the largest Graviton customers in the world.
Why CPUs are suddenly the hottest chips in AI
Everyone has spent the last three years talking about GPUs. Nvidia became the most valuable company on the planet because of them. Meta has already committed $48 billion to GPU rental companies CoreWeave and Nebius. But GPUs are for training AI models, pushing through terabytes of data until a model can actually do something useful.
Once a model is trained and out in the world, the job changes. Answering queries, generating code, running multi-step reasoning tasks: this is CPU-heavy work, and that is what Amazon’s Graviton5 is built for. Each chip carries 192 cores that can handle different tasks at once, runs 25% faster than the previous generation, and uses 60% less energy. AWS designs them in-house and sources the hardware through TSMC.
Intel’s CEO also noted this week that demand for Xeon server chips now exceeds supply. The CPU spent years getting talked over by GPU hype. Quietly, it’s back, and the biggest names in AI are moving fast to lock in capacity before it gets harder to come by.
Inside Meta’s Amazon Graviton deal
Before this deal, Meta had used Graviton at small scale. Now it ranks among the top five Graviton customers globally. Santosh Janardhan, Meta’s head of infrastructure, called chip diversification a “strategic imperative.” The goal is agentic AI: systems that can reason, plan, and execute complex tasks on their own. Think about what that means at Meta’s scale. Billions of users, millions of interactions per second, all requiring real-time reasoning. GPUs were never built for that kind of sustained, real-time load.
Meta keeps its GPU partnerships intact. It still works with Nvidia, AMD, and Arm. The compute stack just gained another layer. GPUs train the models. CPUs run them in production. Meta is trying to own both sides of that pipeline, and the Amazon deal puts the deployment side firmly in place.
Meta spending billions while cutting 8,000 jobs
Meta announced 8,000 layoffs the day before this deal. Roughly 10% of the workforce, gone. Another 6,000 open positions will stay unfilled. The company is building the AI future with one hand while dismantling the workforce that built everything else with the other.
Meta runs 32 data centers and serves 3.6 billion users daily. Zuckerberg has made clear that is not enough. He has staked the company on AI, and the hardware deals are his evidence. The 8,000 who lost their jobs to essentially fund it are a different story.
