AMD unveiled its Helios rack-scale AI system, integrating Instinct MI455X GPUs, Venice CPUs, ROCm software, and Pensando networking to form an end-to-end AI infrastructure designed to compete with Nvidia’s rack-scale solutions. Helios delivers around 2.9 exaflops FP4 per rack with 31TB of HBM4 and 1.4PB/s memory bandwidth, featuring 72 GPUs per rack and scalable interconnects. AMD also announced high-profile partnerships and deployments, including Microsoft’s plan to deploy Helios in Azure for frontier model inference and to add VM series powered by Venice CPUs and Pensando DPUs. In parallel, AMD expanded collaborations with Anthropic to deploy up to two gigawatts of MI450 GPUs, backed by a $5 billion investment, reinforcing its push into production-scale AI workloads. The company frames Helios within an open-standard effort via the UALink consortium, with Azure and Microsoft participating as co-authors of the interconnect standard to challenge Nvidia’s dominance. Overall, AMD signals a bold shift toward full-stack, rack-scale AI infrastructure with a growing ecosystem of large customers and partners.