AMD Brings Helios to Microsoft Azure: The Rack-Scale Platform Enters the Challenge with NVIDIA

AMD and Microsoft have announced an expansion of their strategic collaboration in the field of artificial intelligence, extending the partnership far beyond traditional processors intended for servers. The agreement involves GPU, CPU, networking, and software and includes the adoption of the new AMD Helios rack-scale platform within the Azure infrastructure, where it will be used for inference of cutting-edge AI models, Azure AI services, and cloud customer workloads.

The first deliveries of Helios to customers, including Microsoft, will begin in the second half of 2026. The Redmond company thus becomes one of the main users of the new platform alongside Meta, OpenAI, Oracle, and Tata Consultancy Services, all committed to rapidly increasing computing capacity intended for the development and execution of large-scale AI models.

Helios represents AMD's first complete rack-scale offering designed to compete directly with NVIDIA's Grace Blackwell and Vera Rubin platforms. The solution integrates AMD Instinct MI455X GPUs, sixth-generation "Venice" AMD EPYC processors, networking based on Pensando technologies, and the ROCm software stack into a single system, offering an open infrastructure designed for both training and inference of AI models.

According to Lisa Su, AMD's president and CEO, the collaboration with Microsoft represents an important step in expanding the company's AI offerings. Satya Nadella, Microsoft’s CEO, emphasized how Azure needs an infrastructure capable of adapting to different workloads, from training to inference, to data preparation, research, and reinforcement learning, providing customers with greater flexibility in choosing hardware.

The announcement also includes the arrival on Azure of two new families of virtual machines based on EPYC Venice processors. The Azure HDv2 series will be aimed at agentic AI workflows and data processing pipelines, while Azure HXv2 will focus on semiconductor design activities. The goal is to further expand the Azure instance portfolio based on AMD processors for AI, scientific, and engineering applications.

Among the new features announced is the new ND MI455X v7 instance, designed specifically for large-scale AI inference workloads. The new family of virtual machines is intended to handle reasoning applications, search engines, and agentic AI systems, areas that require high performance in executing models rather than in their training.

The agreement also extends to the networking infrastructure. Microsoft will continue to increase the use of Pensando DPUs within Azure services and will integrate AMD technologies with Azure Boost, the proprietary platform dedicated to accelerating cloud networking. The integration aims to improve efficiency, throughput, and connection management on a large scale in data centers.

For AMD, Helios represents one of the most ambitious projects in its recent history. Each rack uses 18 compute trays, each composed of four Instinct GPUs paired with an EPYC processor, while the networking subsystem integrates up to twelve Pensando chips per tray. The platform is the result of a strategy built in recent years through the development of the EPYC family and a series of acquisitions, including Xilinx, Pensando, ZT Systems, and several software companies that contributed to the evolution of ROCm, the open-source alternative to NVIDIA's CUDA ecosystem.

From a commercial perspective, the challenge remains complex. According to estimates from Futurum Group, NVIDIA still controls over 95% of the data center GPU market, while AMD stands at around 4.5%. Analyst Daniel Newman, however, believes that if Helios proves competitive in early implementations, AMD could eventually capture a market share between 20% and 25%.

Forrest Norrod, head of AMD's Data Center division, asserts that one of the platform's main strengths is the overall operational cost, with particular attention to the "cost per token" during AI model inference. Lisa Su also highlighted months ago how Helios could offer advantages over NVIDIA's rack-scale systems, especially in inference performance and bandwidth and memory management capabilities.

However, the competition on the software front remains open. Several analysts recognize that EPYC CPUs and Instinct GPUs have reached a competitive level compared to NVIDIA solutions, but believe that the main advantage for the competitor continues to be the CUDA ecosystem, still widely dominant among developers and companies. For this reason, the first implementations of Helios at Microsoft and other hyperscalers will be closely watched, as they could represent a decisive test for AMD’s ambitions in the artificial intelligence market aimed at data centers.