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, the podcast where we dig into the chips, silicon and data center tech, powering the AI revolution. I'm your host, Skyler Monroe, and today we have an absolutely stacked episode. Big shout out to our sponsors, LimitLess AI, helping businesses integrate AI into their workflows, and a Go consulting that's a Go, your go-to for silicon development consulting. All right, let's get into it. We've got Meta making massive moves with Nvidia and dropping four brand new AI chips. Samsung prepping a new NPU for your laptop, Europe Semiconductor Sovereignty Push, and TSMC showing off some genuinely exciting manufacturing breakthroughs. Let's go. Okay, story one, and this one is big. Meta has expanded its deal with Nvidia to use millions of AI chips as part of a massive data center build out. We're not just talking GPS here, this deal also includes standalone CPS. Let's unpack what's happening and why it matters so much. First, the scale. When we say millions of chips, we're not being hyperbolic.
Meta has been on an absolute infrastructure spending tier for the past couple of years and this deal signals that they're doubling down in a serious way.
We're talking about one of the largest compute build outs in history from a single company.
Now, the GPU side of this is pretty expected at this point. Nvidia's H100s and the newer Blackwell architecture chips are essentially the gold standard for AI training workloads. If you're training large language models or multimodal models at the scale, Meta is operating, think Lama, think their recommendation systems, you need a lot of those high bandwidth, high throughput chips. No surprise there, but here's the part that got my attention, the stand-alone CP-US. Meta is buying Nvidia G-Race CP-US, which are ARM-based server processors.
Now why does that matter? Because it signals that Meta isn't just buying Nvidia for GPU compute anymore. They're building out a more holistic infrastructure stack using Nvidia's broader product portfolio. Think of it this way. If a GPU is like the main engine in a race car, a CPU is like all the supporting systems, the steering, the brakes, the fuel management. You need both working together for peak performance. When you're operating at Meta scale, even small inefficiencies in how those systems communicate can cost you enormous amounts of money and performance. The Nvidia Grace CPU is built on the ARM Neoverse platform, and it's designed to work really well with Nvidia GPS, especially in the Grace Hopper and Grace Blackwell Superchips, where the CPU and GPU are tightly coupled with NVLink. So Meta buying standalone Grace CPS likely means they're building infrastructure where those processors handle orchestration, networking and inference tasks while GPS do the heavy lifting. From a market perspective, this is also a massive win for Nvidia that goes beyond just dollar figures. It shows that Nvidia's strategy of becoming a full-stack data center company, not just a GPU maker is actually working. Jensen Huang has been saying for years that Nvidia wants to be the platform, not just the accelerator. A deal this size from Meta validates that vision in a really concrete way. And look, there's a slight irony here too, which we'll get into in our next story, because Meta is simultaneously developing its own custom silicon. So they're both a major Nvidia customer and a potential Nvidia competitor.
That's a really interesting dynamic to watch as we head through 2025 and beyond. The broader implication here is that the hyperscalers Meta, Google, Microsoft, Amazon are all in this arms race to secure as much compute as they can. Supply is still constrained relative to demand, so locking in deals at this scale is as much about securing future supply as it is about anything else.
Meta is essentially planting a flag and saying, we are in this AI race to win.
Alright story 2 and this one is exciting for a different reason because it's about bringing AI acceleration closer to you, the end user.
Samsung is reportedly getting ready to launch something called the Gaia AI Accelerator for PCs and it's already being validated by HP and Lenovo.
Let's dig into this. So first, what is Gaia? Based on what's been reported, it's a dedicated NPU. That's a neural processing unit designed specifically for on-device AI inference tasks.
Think things like real-time image processing, AI assisted productivity features, local language model inference, the kinds of things Microsoft is pushing with co-pilot-plus PCs and Apple has been doing with its neural engine on the M-series chips. The NPU space for client devices, meaning laptops and desktops, has gotten incredibly competitive over the last couple of years. You've got Intel with their Meteor Lake and Lunar Lake chips that include NPS, Qualcomm's Snapdragon X-series, which has a really strong NPU story. And of course, Apple's neural engine, which has been setting the bar for on-device AI, AI performance for several years now. So where does Samsung's Gaia fit in? Well, Samsung has deep semiconductor expertise. They make their own Exynos chips, they operate one of the world's largest foundries, and they have experience building MLX accelerators. But the client PC market is a different beast from mobile, and getting design wins at HP and Lenovo is genuinely significant because those are two of the biggest PC OEMs in the world. Here's a good analogy for why dedicated NPS matter. Imagine you're trying to cut vegetables.