AMD unveiled Helios, a rack-scale AI system that integrates computing, networking and software in a single deployable unit, at its Advancing AI 2026 conference on July 23. Each rack contains 72 Instinct MI455X GPUs, 18 6th Gen EPYC Venice server CPUs, Pensando networking hardware and AMD's ROCm software stack. The system offers up to 2.9 exaflops of peak FP4 performance and 31 TB of HBM4 memory, while AMD says it delivers up to 30% more tokens per dollar than Nvidia's Rubin NVL72 rack. Helios is aimed at frontier model training and large-scale inference, including agentic AI workloads that require substantial CPU resources for multistep reasoning and orchestration. Anthropic has committed to deploying up to 2 GW of MI450-series GPUs in Helios racks, while OpenAI and Meta have each committed to 6 GW through their respective deals. Microsoft and Oracle are also customers. Shipments are expected to start in the second half of 2026 and increase through 2027. AMD projects the server CPU market will exceed $200 billion by 2030, compared with a Bank of America forecast above $170 billion, and says it achieved a fourfold improvement in AI energy efficiency from 2024 through mid-2026. The company is targeting a 20-fold improvement by 2030 and has open-sourced its Optima solution for optimizing internal AI processes.