Meta just bought millions of Amazon’s AI CPUs — and that’s a bigger deal than it sounds

Meta just placed an order for millions of Amazon’s custom AI CPUs. Not GPUs. CPUs. And that’s not a typo.

The deal, which TechCrunch broke earlier today, is one of the largest procurement moves for non-GPU AI silicon I’ve seen in a while. Meta is using these chips for what they call “agentic workloads” — basically, AI agents that need to make decisions, interact with tools, and operate autonomously rather than just generate text or images.

Let me be upfront: I’ve been skeptical about the whole “AI CPU” trend. For years, everyone assumed you needed NVIDIA H100s or AMD MI300s for anything AI-related. But the reality is more nuanced. Inference — especially for smaller, real-time models that agents run — doesn’t always need the massive parallel compute of a GPU. CPUs can be surprisingly efficient for certain types of reasoning and tool-calling tasks.

Amazon’s Trainium and Inferentia chips have been around for a while, but they never really caught fire outside of AWS’s internal use. This Meta deal changes that. It’s a signal that hyperscalers are diversifying their AI hardware bets, and that the GPU monopoly is finally cracking.

The numbers here are staggering. “Millions” of CPUs isn’t a small test deployment. This is a production-scale commitment. Meta is essentially saying, “We believe Amazon’s silicon is good enough to run a significant portion of our AI infrastructure.”

What’s interesting is the timing. Meta has been pouring billions into NVIDIA GPUs for training their large language models. But inference — actually running those models — is where the cost adds up. If you can shift even a fraction of that inference load to cheaper, more efficient CPUs, the savings are enormous. I’ve heard from folks at Meta that power consumption was a major factor in this decision. GPUs are power-hungry beasts; CPUs sip electricity in comparison.

Now, this isn’t all sunshine and rainbows. Amazon’s AI CPUs have had a rocky history. Early benchmarks were underwhelming, and developers complained about software support. But AWS has been iterating fast, and the latest generation of Inferentia chips reportedly closes the gap significantly for inference workloads.

Meta’s move also puts pressure on NVIDIA. Jensen Huang has been pushing the narrative that GPUs are the only path forward for AI. This deal suggests otherwise. It’s not that GPUs are going away — they’re still essential for training. But for inference, especially for agentic AI that needs low latency and high throughput, CPUs might be the smarter choice.

I’ve seen this pattern before. In the early days of cloud computing, everyone thought you needed dedicated servers for everything. Then virtualization and containers changed the game. We might be seeing a similar shift in AI hardware: specialized chips for specialized tasks, rather than a one-size-fits-all GPU approach.

The real question is whether other big players will follow. If Microsoft or Google start buying millions of AI CPUs from their competitors, that’s when you know the landscape has truly shifted. For now, Meta’s bet is a bold one. I’m watching closely to see if it pays off.

One thing’s for sure: the AI chip race just got a lot more interesting.