Skip to main content
osky
Sección
AI

What's up, folks? If you think the AI hardware market is just NVIDIA getting filthy rich, I've got some good gossip for you. Turns out Zuck (yeah, the Meta guy) just signed a piggy-bank-breaking deal, but not with the green team—with Jeff Bezos and his AWS servers. The goal? To use Amazon's custom CPUs to train and run massive chunks of Meta's models.

To understand the magnitude of this shit, think of it like showing up to play at a fancy club and instead of the sacred Pioneer CDJ-3000s, the booth has pure Denon gear and a janky mixer. At first, you throw the tantrum of your life, but then you realize the Denon does the same mixing job if you know what you're doing, and the club saved thousands of dollars in rentals. That's Meta right now. NVIDIA GPUs (the "CDJs" of the AI world) are insanely expensive and there's a waitlist. Amazon told them: "Dude, I have these Graviton/Trainium CPUs that aren't as fast for certain things, but if you stack a shit-ton of them in parallel, we build the server."

The problem with current hardware is the bottleneck. Everyone wants GPUs to do tensor processing (the crazy math that makes your ChatGPT not look like an idiot). But Meta, burning millions of dollars daily training LLaMA, realized that if you balance the load—using CPUs for more logical, pre-processing tasks—you can free up the GPUs to focus on the heavy lifting. It's fucking microservices architecture applied to physical hardware. You take the heavy track, put a high-pass filter on the CPUs, and let the bass (the tensors) drop on the GPU.

What does this mean for us, the coding plebs? AWS is going to have a brutally optimized hardware ecosystem for LLaMA. If in a year you try to run an open-source model on AWS, it'll probably cost you three bucks because the infrastructure will be as oiled up as a techno promoter's ego.

Plus, this is a direct message that Meta doesn't want to be anyone's "technological vassal". If you depend on a single hardware provider, the day they raise prices, you die. It's a basic server lesson: never have a single point of failure. And Zuckerberg, like him or not, isn't stupid.

Server CPU

At the end of the day, this is going to make the iron cheaper for everyone. So go light your candles and prep your Terraform scripts, because the chip war just got more interesting than a 4 AM set.