SpaceX Awards Foxconn A Part In A Huge $52 Billion Order For 13,000 Racks Of NVIDIA GB300 AI Servers, Where Each Rack Costs $4 Million And The Total Order Spans Nearly 1 Million GPUs - Wccftech

**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 infrastructure powering the AI revolution, you are in exactly the right place. Big 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 accelerators to full SOAE designs. And Zen Semiconductor, an AI and silicon venture group, building the processors and AI fabric behind modern data centers. Think their Sierra RISC-V CPUs and Loom AI fabric. All right, we've got a packed episode today. We're talking a jaw-dropping $52 billion Nvidia server deal involving SpaceX and Foxconn. Why SiraBriz is actually turning heads in an Nvidia-dominated world. A new Blackwell server GPU from Nvidia aimed at enterprise data centers. What TSMC is saying about the long-term AI chip outlook and JP Morgan throwing a little cold water on the chip rally. Let's get into it. Okay, so let's kick things off with the story that honestly made me do a double take when I first read it. SpaceX has placed a massive order. We're talking $52 billion for Nvidia Gb300 AI servers. And part of that order has been awarded to Foxconn to manufacture. Let that sink in for a second. $52 billion.
That's not a rounding error. That's the GDP of a small country going toward AI compute infrastructure.
So let's break down what we're actually talking about here. The order covers 13,000 racks of Nvidia Gb300 servers. Each rack is priced at roughly $4 million.
And across those 13,000 racks, you're looking at nearly 1 million GPS.
1 million. That's not a fleet of GPS. That's an army. Now the Gb300 is Nvidia's next-generation Blackwell Ultra architecture. Think of it as the next step up from the Gb200, which itself was already a monster.
The Gb300 is designed for massive-scale AI training and inference workloads, the kind of compute you need when you're building frontier models or running enormous simulation environments. And SpaceX is the buyer here, which is fascinating. We typically think of SpaceX as a rocket company, but Elon Musk's ambitions around AI, especially through XAI, are clearly driving serious infrastructure investment. This isn't just buying a few servers to run some analytics. This is hyperscalar-level procurement. SpaceX is essentially building one of the largest AI-compute clusters on the planet. Now Foxconn's involvement is also worth noting. Foxconn has been aggressively expanding into the AI server manufacturing space. They're not just building iPhones anymore. They're becoming a critical player in the data center supply chain. When you've got a $52 billion order, you need manufacturing partners with serious scale, and Foxconn absolutely has that. Think about it this way. If each rack costs $4 million and holds roughly 72 GPUs in the Gb200 NVL, 72 configuration, you're looking at an insane density of compute per rack. The NVLink fabric tying all those GPUs together means they're not acting like individual cards. They're behaving like one giant unified processor. That's the key innovation here. It's not just raw GPU count, it's how tightly integrated they are. For the broader industry, this signals something important. The hyperscaler and near hyperscaler class of customers are not slowing down. If anything, they're accelerating. When a single entity is willing to commit $52 billion to one generation of hardware, it tells you the demand signal for AI compute is still pointing straight up.
Nvidia's order book must look absolutely extraordinary right now.
For Foxconn, this cements their transformation story. They've spent years trying to diversify away from consumer electronics assembly and landing a piece of a $52 billion AI server. Deal is exactly the kind of contract that validates that pivot. It's a big moment for them strategically.
Alright, story two. And this one is a nice counterpoint to all the Nvidia dominance we just talked about. Cerebras is making noise, and for good reason. Let's talk about why their AI chips are genuinely interesting in a market that Nvidia owns so thoroughly.
So, cerebras make something called the wafer scale engine, and when I say wafer scale, I mean it literally. Their chip is an entire silicon wafer. Not a chip cut from a wafer, not a chiplet design, the whole wafer is the chip. It's roughly the size of a dinner plate. And yes, that's as wild as it sounds from an engineering perspective. The latest version, the WSE 3, has 4 trillion transistors. For comparison, Nvidia's H100 has about 80 billion transistors.
Now, to be fair, you can't do a straight apples-to-apples comparison because the architecture is completely different. But that transistor count tells you something about the sheer scale of what Cerebras is doing. The key advantage Cerebras pitches is on-chip memory and memory bandwidth. One of the biggest bottlenecks in AI inference, especially, is moving data between memory and compute. With a wafer-scale chip, you can pack an enormous amount of SRAM that's fast, on-chip memory right next to the compute. You don't have to keep running out to slower off-chip memory as often. Think of it like this, imagine you're a chef in a kitchen.