AMD expands into rack-scale AI as CPU demand rises

AMD is no longer content selling chips. It wants to sell the whole rack. At its Advancing AI 2026 conference on July 23, the company revealed the Helios rack-scale AI system, a fully integrated solution that bundles GPUs, CPUs, networking, and software into a single deployable unit. What Helios actually is Each Helios rack packs 72 Instinct MI455X GPUs alongside 18 6th Gen EPYC “Venice” server CPUs, connected through AMD’s Pensando networking hardware and managed by its ROCm software stack. The result is a system capable of up to 2.9 exaflops of peak FP4 performance, with 31 TB of HBM4 memory. AMD claims the Helios system delivers up to 30% more tokens per dollar than Nvidia’s Rubin NVL72 rack. The system is designed for two primary use cases: frontier model training and large-scale inference. Agentic AI workloads, where models autonomously execute multi-step reasoning and orchestration tasks, place unusually heavy demands on CPUs rather than just GPUs. The customer list tells the story Anthropic has committed to deploying up to 2 GW worth of MI450-series GPUs in Helios racks. OpenAI has signed on for a 6 GW multigenerational deal. Meta has matched that with its own 6 GW commitment...