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 chips, silicon, and the hardware that's literally powering the future, you're in exactly the right place. Big thanks to our sponsors today, Ada, helping businesses integrate AI into their workflows seamlessly. Ago Consulting, that's Ago, your go-to for silicon development consulting from AI to SOA, and Zen Semiconductor, the umbrella powerhouse behind a whole portfolio of silicon industry companies, including Ada for deploying AI at scale, Sierra risk fee CPUs, and the Loom AI fabric. These folks are literally building the processors and AI infrastructure behind modern data centers. All right, we've got a loaded show today. Meta is going all in on Nvidia in a massive chip deal. We're busting myths about the memory boom, a scrappy startup just raised $350 million to challenge Nvidia's dominance, and Samsung and Broadcom just signed a jaw-dropping $200 billion pack that's shaking up the Foundry race with TSMC.
Let's get into it. Okay, first up, Meta is expanding its already enormous relationship with Nvidia, and we're talking millions of AI chips here, folks. CNBC broke this one, and it's a big deal on multiple levels. So what's happening?
Meta is deepening its Nvidia partnership to include not just GPS for AI training and inference, but also standalone CPUs for its data center build out. And when I say millions of chips, I want you to really sit with that number for a second. We're not talking about a rack here or there, we're talking about a scale that very few companies on earth can even contemplate. Now, why does this matter so much? Well, think about what Meta is actually doing. They're building the infrastructure to run not just social media at a global scale, which alone is mind boggling, but they're racing to develop their own frontier AI models. They're Lama series, recommendation systems, content moderation, video generation, you name it. All of that requires serious compute and serious compute means serious hardware procurement. The inclusion of standalone CPUs in this deal is really interesting to me. Nvidia has been quietly expanding beyond GPS for a while now. Their grace CPU architecture is a big part of that story. When you're building out a data center at hyperscalar scale, you don't just need GPS sitting there crunching tensor operations. You need orchestration, you need data movement, you need control logic, and that's where CPUs come in. The fact that Meta is buying both from Nvidia tells you something about how Nvidia has positioned itself as a full stack data center vendor, not just a GPU company. Think of it like this. If your GPU is the engine of a race car, the CPU is the driver and the transmission system. You need both working in harmony, and if they're designed to work together from the ground up, you're going to get better performance and efficiency. That's the pitch Nvidia is making with its Grace Hopper and Grace Blackwell architectures, and clearly, Meta is buying it, literally. From a market perspective, this is also a signal about the broader AI infrastructure investment cycle. We've been hearing a lot of chatter about whether hyperscalers might slow down their spend or pivot to custom silicon.
And yes, Meta has its own custom AI chip called the MTIA.
But this deal shows that even companies building their own silicon are still writing enormous checks to Nvidia. The reality is, custom chips are complementary, not replacement at least for now.
Nvidia's ecosystem, their CUDA software platform, their driver support, it creates this gravitational pull that's really hard to escape, even if you wanted to. What I'm watching going forward is how Meta balances its MTIA in-house silicon with these external Nvidia purchases. Because every dollar they spend with Nvidia is a dollar they could theoretically be investing in their own silicon roadmap. But at the scale and speed AI is moving, you need both. You can't wait for your custom chip to tape out and come back from the fab when your competitors are training models right now. This deal makes total strategic sense and it reinforces just how dominant Nvidia's position remains in the data center, AI space.
Alright story 2, and this one is really close to my heart because it's all about memory and specifically, it's about busting myths around the current memory boom. SemiWiki put together a great breakdown of 5 myths circulating in the industry about what's actually driving this boom and I want to dig into the key ideas here because there's a lot of misunderstanding out there. So first, the big picture, we're in the middle of a massive surge in demand for semiconductor memory, and most people assume it's simply because there aren't enough factories. But that's actually myth number 1, it's not primarily a supply problem, it's a capacity allocation problem. The factories exist, the fab capacity exists in many cases. But the specific type of memory that AI needs, high bandwidth memory or HBM, requires fundamentally different manufacturing processes and packaging techniques than standard DRAM or man flash. Let me explain what HBM actually is for anyone who hasn't heard me geek out about this before. Standard server memory, your DDR5 or whatever, moves data along a relatively narrow pipeline. It works fine for traditional workloads, but AI, especially training large language models or running inference on massive neural networks, needs to move absolutely enormous amounts of data between the processor and memory incredibly fast. Hutch BM solves this by stacking multiple memory dies vertically, like a skyscraper instead of a single story building, and connecting them to the processor through thousands of tiny pathways called through silicon VS or TSVs. The result is memory bandwidth that can be 10 to 20 times faster than conventional approaches. So here's the thing, the manufacturers who make regular DRAM aren't just sitting on idle capacity, they can easily redirect to HBM. Making HBM requires retooling, different process flows, different packaging expertise. SK, Hynix, Samsung and Micron are all racing to expand their HBM capacity, but it takes time and billions of dollars in capital investment. So yes, the boom is real, but it's not because nobody built any factories, it's because the factories we have weren't necessarily built for this specific demanding product. Another myth worth addressing is the idea that any memory shortage is a short-term blip. The semi-weeky piece pushes back on that, and I agree completely. The demand curve for HBM is tied directly to the AI accelerator market, and that market is not slowing down. Every new GPU generation from Nvidia, Hopper, Blackwell, whatever comes after needs more HBM, every custom AI chip from Google, Amazon, Microsoft, Meta, they all need HBM too.