AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era
AMD unveiled next-generation AI infrastructure at Advancing AI 2026, featuring Helios rack-scale solutions in production for gigawatt-scale deployment, alongside new CPUs, GPUs, and open software tools aimed at supporting agentic AI and physical AI workloads.
AMD introduced its AMD Helios rack-scale AI solutions, now in production and designed for gigawatt-scale deployment by leading AI companies. The systems integrate 72 AMD Instinct MI455X GPUs and 18 6th Gen AMD EPYC 'Venice' CPUs, connected via AMD Pensando networking and accelerated by AMD ROCm software. AMD claims the Helios configuration delivers up to 30% more tokens per dollar than competing solutions, targeting frontier AI training and inference workloads. Major AI labs and cloud providers, including OpenAI and Microsoft, have selected Helios for its open, full-stack performance.
The company expanded its AI hardware portfolio with the 6th Gen AMD EPYC processors, optimized for agentic AI across cloud, enterprise, and HPC workloads. AMD highlights leadership per-core performance and thread density, enabling higher agent density per watt and dollar. New AMD Instinct MI400 Series GPUs, including the MI455X and MI430X, offer significant performance gains for AI and scientific computing. The MI350P GPU is positioned for cost-effective AI acceleration in existing infrastructure, with AMD citing up to 4.2x more tokens per second per dollar than competitors.
AMD emphasized its open software ecosystem, with ROCm serving as the foundation for AI development on AMD hardware. The company launched ROCm.ai, an AI-driven development platform designed to accelerate GPU software development by integrating coding agents like Claude and Codex. Leading frameworks such as PyTorch and Hugging Face are already enabled on AMD Instinct MI455X GPUs. AMD also highlighted enterprise deployments, including AT&T’s use of AMD technology for its OTel 2.0 telecom-specific model and Cisco’s collaboration to combine AMD inference engines with Cisco’s networking and security capabilities.
AMD extended its AI roadmap through 2030, focusing on scaling AI across enterprise environments and advancing physical AI systems. The company introduced AMD Kria AI solutions, including new system-on-modules powered by Ryzen AI Embedded X100 Series processors and a turnkey Robotics Developer Platform. These solutions aim to reduce vendor lock-in and accelerate the deployment of autonomous robotic systems. AMD also announced broader availability of the Ryzen AI Halo developer platform later in 2026, targeting local AI development accessibility.