AMD

AI Infrastructure · Semiconductors · Cloud Computing

Launch

AMD launches Helios rack-scale AI system to challenge Nvidia head-on

Bundling GPUs, CPUs, networking and software into one open rack lets AMD sell a full data-center system, not just chips.

• AMD officially launched its Helios rack-scale AI infrastructure system at its Advancing AI 2026 conference in San Francisco on July 23, 2026.
• AMD designs CPUs and GPUs; Helios is its first fully integrated rack combining Instinct GPUs, EPYC CPUs, Pensando networking and ROCm software rather than selling components separately.
• Each Helios rack packs 72 AMD Instinct MI455X GPUs, EPYC 'Venice' CPUs and Pensando 'Vulcano' networking over UALink, delivering up to 2.9 EFLOPS of FP4 compute and 31TB of HBM4 memory with 19.6TB/s of bandwidth.
• On the DeepSeek-V4-Flash model, AMD says Instinct MI455X GPUs deliver up to 34x higher token throughput and up to 18x lower token cost versus the prior MI355X generation.
• Microsoft announced it will deploy Helios across Azure to power frontier model inference and AI services, joining existing customers Meta, OpenAI, Oracle and Anthropic; shipments begin in the second half of 2026.
• An industry analyst noted Helios trades some raw speed against Nvidia's Vera Rubin for about 50% more memory per rack and open standards instead of Nvidia's proprietary interconnect, appealing to buyers wanting flexibility.
• By packaging compute, memory and networking as one deployable system, AMD is shifting from selling chips to competing directly with Nvidia's full AI data-center stack, its most serious challenge in years.