On July 20, 2026, Microsoft and AMD announced that they will expand Azure's AI and HPC infrastructure with AMD products. At the center of this is a new virtual machine, "ND MI455X v7," which uses AMD's rack-scale system "Helios." Microsoft will deploy it for large-scale inference, and AMD will begin shipping to customers—including Microsoft—starting in the second half of 2026. AMD's role in Azure is expanding from GPU-equipped VMs to rack-level infrastructure that integrates CPUs, networking, and cooling as a unified design.
At the same time, HDv2 and HXv2, which use the 6th-generation EPYC "Venice," are also being added. This configuration assigns inference, AI data processing, and semiconductor design to separate VMs. What exactly changes for the compute resources Azure users receive, and where does the uncertainty begin?
From 8-GPU VMs to Rack-Level Design
Azure's currently available AMD GPU VM, "ND MI300X v5," packs eight Instinct MI300X GPUs into a single VM. Each GPU has 192GB of memory, with a total of 896GB/s Infinity Fabric bandwidth within the VM and 3.2Tb/s networking between VMs. The host CPU is a 4th-generation Intel Xeon, configured with 96 physical cores.
In contrast, Helios runs behind ND MI455X v7. Helios is AMD's first rack-scale reference design, combining Instinct MI455X, EPYC Venice, and Pensando networking within a single rack. It marks a shift from the concept of mounting GPU cards onto existing servers, toward building compute, data movement, power, and liquid cooling as one integrated system.
The product name Microsoft has presented to customers is the VM called ND MI455X v7, and it has not been announced that this VM occupies all 72 GPUs of a Helios rack. The confirmed change is that Microsoft has adopted Helios for Azure's backend, using it for inference services and customer-facing infrastructure. The rack configuration and the VM's product specifications need to be considered separately.
How 72 GPUs and 31TB Change Inference Infrastructure
The Helios reference design packs 72 MI455X units, with a total of 31TB of HBM4 memory. AMD's stated theoretical peak performance is 2.9 exaFLOPS for FP4 and 1.4 exaFLOPS for FP8. Memory bandwidth per GPU is 19.6TB/s, total scale-up bandwidth within a rack is 260TB/s, and scale-out bandwidth between racks is 43TB/s.
The large memory capacity can be used to keep model weights and the KV cache accumulated during inference close to the GPU. In inference, what matters isn't just the speed of a single model run, but the ability to process many requests simultaneously without letting data stall between GPUs. Microsoft's mention of reasoning, search, and agentic workloads as use cases for ND MI455X v7 aligns with this design philosophy.
The 72 units are connected via UALink over Ethernet, and racks are linked together through Pensando's Ethernet network. The rack itself is based on the double-wide Open Rack Wide that Meta submitted to the Open Compute Project. AMD has chosen open standards from GPU interconnects all the way to the rack form factor. However, actual performance on Azure has yet to be demonstrated. Figures like 2.9 exaFLOPS represent AMD's design-level peak values and do not guarantee application processing speed or service quality.
HDv2 and HXv2 Split Up Venice Usage
Beyond the GPU infrastructure, this announcement also splits AI preprocessing and technical computing on the CPU side. The new services have distinct roles, as follows:
| Azure Product | Main Configuration | Intended Workload |
|---|---|---|
| ND MI455X v7 | Helios, Instinct MI455X | Large-scale inference, search, agentic processing |
| HDv2 | Approx. 500 physical cores, 4TB RAM, 32TB NVMe, 400Gb Azure Boost | Data preparation, search, reinforcement learning, agent orchestration |
| HXv2 | 176 cores, over 5GHz, 50% more cache per core, approx. 2TB or 4TB RAM, 800Gb InfiniBand | RTL simulation, EDA, scientific and engineering computing |
HDv2 prioritizes CPU density to organize large volumes of data and support search and agent execution. HXv2, meanwhile, leverages high clock speeds and 3D V-Cache, tailored for RTL simulation commonly used in semiconductor design and distributed computing via MPI. Rather than cramming the process of feeding data to AI accelerators and the process of designing the next generation of accelerators into the same CPU SKU, AMD has optimized them as separate VMs.
Network processing also falls under AMD's scope. Microsoft is expanding its existing Pensando DPU deployment, combining it with Azure Boost to offload communication, storage, and connectivity processing from the CPU. The adoption of Helios isn't limited to procuring MI455X units—it's a partnership that embeds AMD products across multiple layers of Azure's infrastructure.
Deployment Terms to Confirm Before Scale
AMD states that Helios shipments, including those for Microsoft, will begin in the second half of 2026. However, Microsoft has not disclosed how many racks it will deploy, nor the power consumption or contract value involved. For ND MI455X v7 as well, the number of GPUs per VM and memory capacity remain unannounced. Availability regions, pricing, and general availability dates have also not been provided.
While it has been announced that enterprises can use AMD-based infrastructure through Azure Foundry Managed Compute, compatibility and measured performance for migrations from the existing ND MI300X v5 will require further documentation. Even though ROCm supports major AI frameworks, at rack-scale inference, throughput, latency, and operational cost for each specific model will influence adoption decisions.
AMD will hold "Advancing AI 2026" on July 23. If the configuration of ND MI455X v7 and its Azure availability timeline are revealed there, a path will emerge for today's "large-scale deployment" to become an actual option for customers. Rather than the announced peak figures, it will be the performance and pricing once carved out as a VM that determines Helios's competitiveness.