Meta and OpenAI commit 12 gigawatts to AMD's new AI chips as Nvidia's grip on data centers starts to slip - Startup Fortune

**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 of AI, you're in exactly the right place. Huge thanks to today's sponsors, AIEDA, helping businesses actually integrate AI into their real world workflows. Ago Consulting, that's Ago, your go-to for silicon development from AI to SoA. And Zen Semiconductor, the AI and Silicon venture group building the processors and AI fabric behind modern data centers, with incredible ventures like their Sierra RISC-V CPUs and Loom AI fabric. Today's episode is packed. We've got AMD coming out swinging with a major new accelerator, Meta and OpenAI making some jaw-dropping commitments, Nvidia doing something surprisingly interesting with CPUs, a deep dive into why TSMC remains absolutely irreplaceable, and a wild partnership between AMD and Cerebras that you're not going to want to miss. Let's get into it, alright? Let's kick things off with a story that really signals a shift in the competitive landscape of AI hardware. Meta and OpenAI, two of the biggest names in AI, have together committed to 12 gigawatts worth of AMD's new AI chips. 12 gigawatts. Let that sink in for a second. That is not a small pilot program or a hedge bet. That is a massive serious commitment to AMD as an alternative to Nvidia. Now for context, a gigawatt is a billion watts of power. When we're talking about 12 gigawatts committed to chip infrastructure, we're talking about enormous data center deployments at a scale that would rival some of the largest power grids in the world. This is the kind of number that gets utility companies on the phone and keeps power engineers up at night. So what does this mean for Nvidia? Well, for years, Nvidia has had what I'd call a near-monopoly grip on AI data center compute. Their H100s, their H200s, the Blackwell architecture, these have been the default, the go-to, the safe choice for anyone building serious AI infrastructure. If you were a hyperscaler and you needed AI compute, you bought Nvidia full stop, but that grip is starting to slip. And this Meta OpenAI commitment to AMD is probably the clearest sign yet. Think of it like this. For years, Intel owned the data center CPU market so completely that it almost became complacent. Then AMD came along with EPYC processors and started chipping away, slowly but surely. Now AMD is trying to run the exact same playbook in the GPU and AI accelerator space. The key here is that both Meta and OpenAI are sophisticated enough to evaluate alternatives seriously. These are not companies making decisions based on brand loyalty. They have thousands of engineers who benchmark chips, optimize software stacks and care deeply about performance per dollar and performance per watt. The fact that they're both landing on AMD in such a significant way tells you something real is happening on the technical side. And 12 gigawatts of commitment is also a supply chain signal. It tells AMD invest in manufacturing, invest in capacity. We're going to be here for a while. That kind of demand signal helps AMD negotiate with TSMC for wafer allocations and gives them the confidence to build out their ecosystem. It's a flywheel and it looks like it's starting to spin. Now Nvidia isn't going anywhere. Let's be clear about that. Their software ecosystem, CUDA, the developer tools, the years of optimization that Moat is still very real.
But when companies like Meta and OpenAI start making 12 gigawatt commitments to the competition, the Moat just got a lot narrower. This is a big deal and it's only going to get more interesting from here. Okay, story two, and this one caught me a little off guard, honestly. Nvidia, the company most famous for GPUs, has apparently shipped hundreds of thousands of grace standalone servers. And they're actively pivoting their messaging to put CPUs more front and center, particularly for agentic AI workloads. Let's unpack what's going on here. So grace is Nvidia's arm-based CPU. We've seen it before as part of the Grace Hopper Superchip. That's the combined CPU-GPU package. But now, Nvidia is shipping grace as a standalone CPU, without the GPU attached. And they're framing this around the rise of agentic AI. Agentic AI for anyone who needs the quick definition is basically AI systems that don't just answer a question and stop. They plan, they reason, they take sequences of actions, they call tools, they loop, and iterate.
Think of the difference between asking someone a question and hiring someone to manage a project. Agentic AI is that project manager. And here's the interesting hardware angle. Agentic workloads don't always need massive GPU horsepower for every single step. A lot of what an agent does is orchestration, routing, planning, calling APIs, managing memory, making logical decisions.