AMD Lands Anthropic In 2GW Instinct MI450 Deal, Backing It With A $5 Billion Equity Bet

**Skyler Monroe** (0:10)
Hey, everyone, welcome back to the AI Hardware Show. I'm Skyler Monroe, your guide through the wild world of silicon semiconductors and the hardware powering the AI revolution.
Huge thanks to our sponsors today, AI EDA, helping businesses actually integrate AI into their workflows in meaningful ways. Ago Consulting pronounced Ago, your go-to for silicon development from AI chips to full-sew ACE, CEC, and Zen Semiconductor, the AI and Silicon venture group, building the processors and fabric behind modern data centers, including their Sierra RISC-V CPUs and Loom AI fabric.
All right, we've got a packed episode today. AMD is making some serious moves with not one but two massive partnership stories. Cerebras is out here doing wafer scale things that sound almost impossible. SpaceX wants to run AI compute in orbit and somehow make it cheaper than Blackwell data centers. And we're talking about how Nvidia and AMD are no longer just chip companies. Let's get into it. Okay, let's kick things off with what I think is one of the biggest AMD stories in a while, maybe ever.
AMD has officially confirmed that Anthropic is becoming a strategic partner. And this isn't just a handshake deal. We're talking about a deployment of up to two gigawatts of AMD Instinct MI450, GPs through what AMD is calling Helios AI racks. And on top of that, AMD is backing this with a $5 billion equity bet. Let that sink in for a second. $5 billion equity bet. Let that sink in for a second. $5 billion. That's not a sales agreement. That's AMD putting real skin in the game. So let's unpack what's actually happening here. The MI450 is AMD's next generation Instinct Accelerator, and it's the chip that powers the Helios rack scale system. Now, Helios is AMD's answer to Nvidia's NVL rack solutions. It's a full rack scale architecture designed to deliver massive compute density with tight integration across networking, memory and compute. Think of it like going from selling individual puzzle pieces to selling the whole puzzle preassembled. That's the shift AMD is making here.
And two gigawatts is a staggering number. To put that in perspective, a single gigawatt is roughly what a large nuclear power plant produces. So we're talking about enough compute infrastructure to power a small city just for AI workloads. This is not a pilot program. This is infrastructure at civilization scale. And Anthropic is betting a serious chunk of its future on AMD silicon. Now, why does this matter so much? Because for years, Nvidia has had an almost unassailable grip on the AI training market. CUDA, the software ecosystem, the NVLink interconnects. It all created this massive moat. AMD has been chipping away at it, no pun intended, but landing Anthropic, one of the most prominent and well-funded AI labs on the planet, as a strategic partner, is a legitimacy signal that's hard to overstate. Anthropic builds some of the most demanding frontier models out there with Claude, and those workloads don't forgive subpar hardware. The $5 billion equity commitment is also really interesting from a strategic standpoint. AMD isn't just selling chips here, they're investing in Anthropic success. That alignment of incentives means AMD has a direct reason to make sure its hardware performs for Anthropic-specific use cases. It's a co-development relationship disguised as a sales deal. And for AMD CEO Lisa Su, this is exactly the kind of partnership she's been building toward. AMD is playing the long game, and this is a major piece moving on the board. What I'll be watching closely is how the ROCM software stack matures to support this deployment. Hardware is only half the story the software that lets you actually program these chips efficiently is where Nvidia has historically dominated. AMD knows this, and I'd expect to see significant co-engineering on the software side between these two companies as part of this deal.
Overall, a massive story. AMD is for real. Alright, speaking of AMD, let's stay in that lane because there's a second AMD story that's flying a little more under the radar, but is actually really fascinating from a technical perspective. Reports are coming out that Meta is going to be using a custom version of AMD's Instinct MI400 series accelerators, but with a memory configuration cut down to 144 gigabytes of HBM4. The key word there is custom. Meta isn't buying off the shelf, they're working with AMD to build something tailored specifically for their workloads. So let's talk about why memory matters so much in AI accelerators.
HBM, which stands for High Bandwidth Memory, is essentially the fuel tank of a GPU or accelerator. The more you have, the larger the models you can fit, and the faster you can feed data to the compute units.