AMD launches AI server, OpenAI adopts new chips
AMD just announced Helios, a powerful AI server platform set to launch in 2026. Built for massive AI workloads, Helios packs 72 MI400 GPUs, next-gen EPYC "Venice" CPUs, and speedy Pensando "Vulcano" networking—all aimed squarely at challenging NVIDIA's lead in the AI data center space.
Each rack has 72 MI400 GPUs
Helios is seriously stacked: each rack has 72 MI400 accelerators (with 432GB memory per GPU), up to 31TB total memory. For connecting everything together, it uses Ultra Accelerator Link (UALink) and Vulcano NICs. Cloud provider Crusoe is already betting big with $400 million in orders.
Helios will outperform NVIDIA's systems
Helios goes head-to-head with NVIDIA's Vera Rubin and Blackwell systems but stands out by using open networking standards instead of NVIDIA's proprietary tech—meaning more flexibility for partners. AMD says Helios will outperform NVIDIA on memory capacity and bandwidth, and big names like OpenAI, Meta, xAI, and Oracle are already on board.