AMD Advancing AI 2026: AI Inference Surpasses Training, New Instinct MI455X Unveiled
At AMD's Advancing AI 2026 event, CEO Lisa Su announced that AI inference has surpassed training for the first time, now accounting for 60% of AI compute resources. AMD also unveiled the Instinct MI455X GPU featuring 432 GB of HBM4 memory and the EPYC "Venice" CPU designed for agentic AI. The shift of compute resources toward inference signals that AI is transitioning from development to widespread deployment and real-world application. AMD's new hardware directly challenges NVIDIA's dominance in the high-performance AI data center market, backed by partnerships with major AI labs like Anthropic and OpenAI. The Instinct MI455X is built on TSMC's 3nm and 2nm processes, featuring 320 billion transistors, and is integrated into the Helios rack solution which AMD claims outperforms NVIDIA's "Vera Rubin" by 15%. Additionally, Anthropic plans to deploy up to 2 GW of Helios infrastructure to scale its Claude AI models.
## BACKGROUND
AI training is the process of teaching a machine learning model using large datasets, which requires massive, one-time computational power. In contrast, AI inference is the execution phase where the trained model processes new data to make predictions or decisions in real-time. As agentic AI and consumer-facing applications grow, the ongoing demand for inference compute naturally scales beyond the initial training phase.