AMD Launches MI400 GPUs and Helios Rack to Target AI Inference

As generative AI shifts from model training toward large-scale inference and agentic workloads, memory bandwidth, data movement and tokens produced per dollar are becoming decisive constraints for data-center operators. AMD is positioning its Instinct accelerators, EPYC processors, Pensando networking and open-source ROCm software as a full-stack, standards-based alternative to Nvidia’s largely proprietary AI infrastructure. Helios packages those components into a rack-scale reference design aimed at frontier models and high-volume inference, where deployment economics increasingly matter as much as peak compute.

At Advancing AI 2026 in San Francisco on July 23, AMD launched the MI400 series and Helios, a rack integrating 72 MI455X GPUs, 18 sixth-generation EPYC “Venice” CPUs and 31 terabytes of HBM4 memory. AMD said the rack delivers 2.9 exaflops of FP4 compute and as much as 30% more inference tokens per dollar than Nvidia’s Vera Rubin NVL72, while MI455X token throughput is 34 times that of MI355X. OpenAI expects to bring Helios online in the fourth quarter of 2026, Meta has begun validation, and Microsoft plans large-scale Azure deployment.