Meta expands Nvidia deal to use millions of AI chips in data center build-out, including standalone CPUs - CNBC

Skyler Monroe (0:10)

Hey everyone, welcome back to the AI Hardware Show. I'm your host Skyler Monroe, and if you're into the silicon and semiconductors powering the AI revolution, you have found the right place. Big thanks to our sponsors keeping this show running. Ada, helping businesses integrate AI into their workflows seamlessly.

Ago Consulting, that's Ago, your go-to for silicon development consulting from AI to SoC.

And Zen Semiconductor, the umbrella company building the processors and AI fabric behind modern data centers, with incredible ventures like their Sierra RISC vCPUs and Loom AI fabric, plus Ada for deploying AI at scale in any business. Today we've got a packed episode. Meta is going on a serious Nvidia shopping spree. ByteDance found a creative workaround to get their hands on Blackwell chips. TSMC is absolutely printing money. We're talking about the unsung hero company that both Nvidia and AMD desperately need, and we're diving into why two identical AI clusters can perform completely differently. Let's get into it. All right, first up, Meta is expanding its already massive deal with Nvidia, and we're talking millions of AI chips for a major data center build out. And here's the part that really caught my attention. This includes standalone CPUs, not just GPUs. So let's unpack what's actually happening here and why it matters. Meta has been on an infrastructure tear for a while now. They've been pouring tens of billions of dollars into AI compute. And this latest Nvidia expansion is a signal that they are not slowing down. If anything, they're accelerating. When we talk about millions of chips, that's not a rounding error. That's a fundamental commitment to owning and operating some of the most powerful AI infrastructure on the planet. Now the GPU side of this is kind of expected. At this point, Meta needs massive compute for training their Lama models, for ranking and recommendation systems on Facebook and Instagram, for all the generative AI features rolling out across their apps. That all needs GPUs and Nvidia is still the dominant supplier. No real surprise there, but the CPU piece, that's interesting. Nvidia has been pushing into the CPU space with their Grace CPU, which they've paired with the Hopper and Blackwell GPU architectures in systems like the GB200. When we see Meta pulling in standalone CPUs from Nvidia as part of a big deal, it suggests they might be looking at Grace as a serious player in their infrastructure stack, not just as a companion ship to the GPU. Think about it this way, in a modern AI data center, the GPU does the heavy lifting for matrix math and model training, but you still need a CPU to orchestrate everything. Data pre-processing, managing IO., running the control plane of your infrastructure, that's CPU work. If your GPU is an Nvidia product, there are real efficiency gains to having a CPU from the same vendor with tight interconnects between them. Nvidia's NVLink and their grace CPU superchip architecture are designed specifically for this kind of tight CPU-GPU coupling. You get dramatically higher bandwidth between processor types compared to a traditional PCE.

For Meta, this could also be a strategic diversification play. They've invested heavily in their own custom silicon, their MTIA chips for inference, but for raw training scale, they still need Nvidia. Locking in a multi-year, multi-million chip deal probably comes with favorable pricing, priority allocation, and technical collaboration that you just can't get as a smaller customer. And from Nvidia's perspective, this is exactly the kind of deal Jensen Huang has been talking about. Not just selling GPI-US, but becoming the foundational compute platform for hyperscalers.

CPUs, GPI-US, networking through Mellanox software through CUDA, the full stack.

Meta going deep on Nvidia's CPU offering is a validation of that whole strategy. The scale here also tells us something about where AI infrastructure investment is heading. We're not in a phase where companies are dipping their toes in. Meta, Microsoft, Google, Amazon, they are all making decade scale bets on AI infrastructure.

And the chip suppliers are the ones benefitting enormously from that. If you needed any more evidence that the picks and shovels play in AI is very, very real, this is it. The other thing worth noting is supply chain dynamics. When Meta locks in millions of chips, those chips have to come from somewhere.

TSMC is manufacturing them. Co-O's packaging is involved, HBM memory from Samsung or SK Hynix or Micron is stacked on top. A deal this size ripples through the entire semiconductor supply chain. It's not just a Meta and Nvidia story, it's an industry story. Bottom line on this one, Meta is doubling down on Nvidia in a serious way. And the inclusion of CPUs tells us Nvidia's platform strategy is working. They're not just the GPU company anymore, they're positioning to be the compute platform of the AI era. And Meta is buying in. Okay, story 2 And this one has a little bit of geopolitical intrigue to it.