At the Advancing AI conference in San Francisco, AMD unveiled Helios, a rack-level AI system designed for training and running large-scale frontier models. The system integrates 72 Instinct MI455x GPUs, 18 sixth-generation EPYC “Venice” CPUs, Pensando networking infrastructure, and ROCm software. AMD CEO Lisa Su described it as having the highest performance of any AI rack; internal company estimates indicate that under specific inference workloads, Helios can process up to 30 percent more tokens per dollar than Nvidia’s Vera Rubin NVL72. Companies including OpenAI, Anthropic, Meta, Microsoft, and Oracle plan to adopt Helios. Among them, OpenAI expects to launch its systems starting in the fourth quarter of 2026; Anthropic plans to deploy up to 2 gigawatts worth of MI400-series GPUs via Helios, with its first 1-gigawatt project expected to begin in the first half of 2027. Meanwhile, Meta is testing related equipment to prepare for scaled deployment. Simultaneously, AMD launched the MI400 series GPUs and sixth-generation EPYC processors, while announcing the Venice-X CPU aimed at high-performance computing in data centers, which is scheduled for release in 2027. Ms. Su stated that agent-based AI tasks require repeated reasoning, tool invocation, and data access, significantly driving up compute demands; AMD projects that by 2030, the global market size for AI accelerators will reach approximately $1.4 trillion, with GPUs continuing to hold the dominant share.