AMD Launches MI400 GPUs and Helios to Target AI Inference

Generative AI infrastructure is shifting from a race to train ever-larger models toward the economics of serving them at scale, making memory bandwidth, data movement and tokens per dollar critical. AMD is positioning Helios as an open, rack-scale alternative to Nvidia-led systems, combining Instinct accelerators, EPYC server CPUs, Pensando networking and ROCm software. The strategy matters because hyperscalers increasingly buy complete, tightly integrated AI factories rather than standalone chips.

AMD unveiled the Instinct MI400 series and production-stage Helios platform at Advancing AI 2026 on July 23. Each rack combines 72 MI455X GPUs, 18 sixth-generation EPYC “Venice” CPUs and 31 terabytes of HBM4 memory. AMD said MI455X delivers 34 times the token throughput of MI355X, while Helios offers up to 30% more inference tokens per dollar than a leading rival. OpenAI plans to bring Helios online in the fourth quarter, Meta is validating racks, and Microsoft will deploy the platform at scale on Azure in the second half of 2026. No contract values were disclosed.