Microsoft will deploy the AMD Helios Rackscale Solution for large-scale AI inference. In addition, two new EPYC VM series will be introduced, and Azure will expand its use of Pensando DPUs.
The expansion covers the entire stack: GPUs, CPUs, networking, and software. At the heart of it all is Helios, an open rack-scale platform that combines AMD Instinct MI455X GPUs, EPYC “Venice” processors, Pensando networking, and ROCm software. Azure will use the platform for inferencing on frontier models, Azure AI services, and customer applications. AMD will begin shipping to customers, including Microsoft, in the second half of 2026.
Microsoft previously offered AMD’s MI300X GPU as an alternative to Nvidia within Azure.
A Helios rack contains 72 MI455X GPUs, delivering approximately 2.9 exaFLOPS of FP4 compute power and 31 TB of HBM4 memory. Each GPU features 432 GB of HBM4, which AMD says provides more memory capacity than Nvidia’s competing Vera Rubin platform.
Helios has deliberately opted for an open, Ethernet-based approach rather than a proprietary interconnect.
New VMs and network layer
In addition to Helios, Azure is introducing two new VM series based on the 6th-generation EPYC “Venice” processors. Azure HDv2 targets agentic AI and data pipelines, while HXv2 focuses on semiconductor design. Both further expand Azure’s EPYC offering.
The network layer is also receiving attention. Building on the existing rollout of Pensando DPUs, both companies are integrating Azure Boost with AMD technology. This is expected to improve network performance and connection processing at cloud scale.
Frontier model builders can use the infrastructure for training and inference, while enterprise customers can run production workloads via Azure Foundry Managed Compute.