AMD Sells Meta GPUs

Meta has inked a deal to deploy "up to" (less than) 6 GW of AMD GPUs starting with Helios deployments in the second half of this year. Conforming to the Meta-backed OCP ORW standard, Helios will use AMD's next-gen MI450 GPUs that will compete with Nvidia's Rubin. We've long said that the AI accelerator pecking order is Nvidia, proprietary (e.g., TPU), AMD, and everyone else. The MI450 is AMD's best bid since the MI300 launch to move up a notch or at least narrow the spread. As for Meta, its homegrown MTIA accelerator has been "deployed at scale" according to the company, but it's uncelar what that scale is. The MTIA targets recommendation models, the backbone of Meta's AI-powered advertising and Facebook-feed software. Meta's Llama LLM family set the standard for open-weight LLMs but have lost relevance over the past one-year-plus, even as the company spends mightily on AI expertise. Only insiders know whether Meta chose AMD because its GPUs are as good as or better than what Nvidia offers, are a better deal, or just represent a supplier diversification strategy. AMD *does,* however, have the advantage of chiplet-level customization, being able to scale down the number of GPU chiplets, add CPU chiplets, and add FPGA dice. Nonetheless, 6 GW is a big endorsement, representing 3 million GPUs (assuming 2 kW per GPU). It's a notable success for AMD, which has a comparably-sized deal with OpenAI. That AI company went on to strike deals with other chip suppliers, tempering the significance of the AMD arrangement. Because Meta announced a deal with Nvidia last week (https://xpu.pub/2026/02/18/meta-nvidia/), including signing up to be the first non-GPU server customer of Nvidia Grace and Vera, the AMD-Meta tie-up is pre-disastered. What would show AMD has really arrived would be one of these companies running large training workloads on the MI450 because that would be an endorsement of AMD's software infrastructure and require a greater commitment of developer resources. As for AMD offering equity to customers, we see that as savvy, provided its aligned with AMD stock-price increases. Warrants are dilutive, but deal-driven appreciation offsets that.

Other contents

Capping GPUs?

Super Cores and Chiplets Define new Apple M5 Processors

Super Cores and Chiplets Define new Apple M5 Processors

US Faces Unintended Consequences of China Trade Policy

US Faces Unintended Consequences of China Trade Policy

Wimpy Cores Fight Back

Wimpy Cores Fight Back

Nvidia to Resolve Dilemma

Nvidia to Resolve Dilemma

Amazon Doubles Down on OpenAI

Amazon Doubles Down on OpenAI

AMD Sells Meta GPUs

AMD Sells Meta GPUs

AI-Chip Startup MatX Nabs $500 Million

AI-Chip Startup MatX Nabs $500 Million

Intel Funds SambaNova

Intel Funds SambaNova

Adopting Grace, Meta is Nvidia’s First Major Server-Processor Customer

Adopting Grace, Meta is Nvidia’s First Major Server-Processor Customer