Microsoft is significantly expanding its Azure AI infrastructure with AMD, committing to deploy the chipmaker's next-generation Helios rack-scale platform for large-scale inference workloads. The expansion, announced July 20, also brings new AMD EPYC “Venice” processors into Azure via two new virtual machine families, alongside deeper integration of AMD's Pensando networking hardware.
The centerpiece is the Azure ND MI455X v7 series—virtual machines purpose-built for production-scale AI inference, including reasoning, search, and agentic workloads. These instances will run on AMD Helios, a rack-scale system pairing Instinct MI455X GPUs with EPYC Venice CPUs and Pensando data processing units (DPUs). Microsoft says the initial focus is inference for frontier models and Azure AI services, not a wholesale replacement for existing Nvidia-powered GPU instances.
Alongside the Helios-powered VMs, Microsoft previewed two CPU-heavy offerings: Azure HDv2 and HXv2. Both rely on sixth-generation EPYC “Venice” processors, built on the Zen 6 architecture and manufactured on TSMC's N2 process using gate-all-around transistors. The HDv2 targets AI data preparation, search, and reinforcement learning, while HXv2 addresses traditional high-performance computing (HPC) workloads like semiconductor design and simulation.
A Trio of New Azure VM Families
Microsoft's announcement details three distinct upcoming Azure products, each targeting a specific slice of AI and HPC demand:
- Azure HDv2 – CPU-driven instances for agentic AI coordination, data pipelines, and search. Top configurations pack nearly 500 physical EPYC Venice cores, 4 TB of RAM, 32 TB of local NVMe storage, and 400 Gbps Azure Boost networking. This class is ideal for workloads that need high core counts and memory bandwidth but not dedicated GPU acceleration.
- Azure HXv2 – HPC-focused VMs that pair up to 176 Venice cores (running above 5 GHz) with 800 Gbps InfiniBand for MPI workloads. Memory options scale to roughly 2 TB or 4 TB, making them suited for electronic design automation, computational fluid dynamics, and other simulation-heavy tasks.
- Azure ND MI455X v7 – The GPU flagship, powered by AMD's Helios rack architecture. Microsoft positions this for “production-scale AI inference,” specifically calling out reasoning, retrieval-augmented generation, and agentic applications. No core or memory figures were shared, but the underlying hardware provides a clear picture.
Microsoft did not announce availability dates, pricing, or initial Azure regions for any of the new VM families. The company typically unveils previews months before general availability, so current estimates point to a late 2026 rollout.
Helios Rack Tackles AI Inference at Scale
AMD Helios is a rack-level AI platform that ships as an integrated system, not a collection of separate components. One Helios rack contains 18 compute trays, each with four Instinct MI455X accelerators and one EPYC Venice CPU. That totals 72 GPUs, over 4,600 CPU cores, approximately 31 TB of HBM4 memory, and a staggering 1.4 PB/s of aggregate memory bandwidth per rack.
The MI455X GPU itself carries 432 GB of HBM4 across four dies, with about 19.6 TB/s of bandwidth per chip. According to AMD, a single Helios rack delivers close to 2.9 FP4 exaflops of inference throughput—a number that matters most to cloud operators running thousands of concurrent inference requests on large language models.
Helios also embraces open standards. It uses UALink-over-Ethernet for inter-GPU communication, a consortium-backed alternative to Nvidia's proprietary NVLink. The UALink group includes AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft, signaling broad industry interest in a vendor-agnostic interconnect. For Azure, that could translate into more flexible networking topologies and easier integration with existing infrastructure.
Beyond GPUs: Networking and Offload
The Microsoft-AMD partnership extends beyond silicon for compute. Microsoft is expanding its use of AMD Pensando DPUs inside Azure, weaving them more tightly with Azure Boost—the company's system for offloading network and storage processing from host CPUs. In an AI cloud, moving model data, coordinating distributed nodes, and handling tenant isolation all consume valuable CPU cycles and network capacity. Offloading these tasks to dedicated DPUs can improve consistency and free host resources for paying workloads.
AMD's announcement notes that Pensando technology will be applied to AI backend networking and selected Azure services. Neither company provided performance figures, but the direction aligns with an industry-wide push to disaggregate infrastructure functions and treat networking as a first-class component of the AI stack.
What It Means for IT Pros and Developers
No client-side changes are required. The new Azure VMs will eventually appear in the portal, ARM templates, and CLI tools just like any other instance family. But the arrival of AMD's next-generation hardware in Azure carries practical implications for several audiences:
For cloud architects and IT decision-makers: The ND MI455X v7 series introduces a credible second-source GPU option for inference in Azure. Organizations currently running inference on Nvidia H100 or A100 instances may find competitive performance-per-dollar, especially if AMD's ROCm software stack has matured enough for their frameworks. The CPU-focused HDv2 and HXv2 families also broaden choice for data preprocessing and simulation, areas where EPYC's core density has historically shown an advantage.
For DevOps and MLOps teams: Before the hardware arrives, assess your AI inference pipelines for ROCm compatibility. Most major frameworks now support ROCm, but custom operators or older libraries may require porting work. Microsoft's Azure AI Foundry Managed Compute is mentioned as a likely landing zone for Helios-based VMs, which could abstract some of that heavy lifting.
For HPC users: The HXv2 instances, with up to 176 high-clock-speed cores and 800 Gbps InfiniBand, directly target MPI-based applications that are sensitive to single-thread performance and internode latency. If your organization currently runs simulation workloads on on-prem clusters or competing cloud instances, these may be worth monitoring once benchmarks emerge.
For budgeting and planning: While pricing is unknown, history suggests that AMD-powered Azure VMs often undercut comparable Nvidia and Intel instances on a per-core or per-GPU basis. The large memory footprints on HDv2 and HXv2 also mean that workloads bottlenecked by RAM capacity could see a cost shift by moving to configurations that avoid over-provisioning.
The Road to AMD's Azure Expansion
Microsoft's Azure has offered AMD EPYC instances for years, starting with the Da-series general-purpose VMs and gradually expanding to memory-optimized (Ea), compute-optimized (Fa), and HPC-focused (HB) families. But the gap between CPU-based and GPU-based AMD options remained wide. Nvidia dominates the Azure GPU catalog, from older T4 and V100 instances to the current A100, H100, and H200-based ND-series VMs.
The Helios commitment changes that dynamic. AMD's Instinct accelerators have been gaining traction among hyperscalers, with Meta also signing on for large-scale deployments. Research firm S&P Global Market Intelligence projects $7.2 billion in revenue for AMD's MI400-series accelerators in 2026 alone, reflecting growing demand for a non-Nvidia alternative in the AI training and inference markets.
Microsoft frames the partnership as part of a heterogeneous infrastructure strategy that also includes its own custom silicon (like the Azure Maia AI accelerator) and continued reliance on Nvidia. In practice, giving customers multiple hardware targets for their AI workloads hedges against supply constraints and allows workload-specific optimization.
The timing coincides with EPYC Venice's arrival. Built on Zen 6 and TSMC's N2 node, Venice marks AMD's shift to gate-all-around nanosheet transistors and a new SP7 socket. It scales to 256 cores, supports 16 memory channels with up to 1.6 TB/s of bandwidth, and introduces PCIe Gen 6 for faster CPU-to-GPU links. AMD says first Venice-based systems will ship in Q3 2026, which lines up with the Helios timeline.
Planning for the New Hardware
No immediate action is required. Microsoft will announce preview programs, region rollouts, and pricing through its usual channels. But organizations likely to adopt these VMs can take steps now:
- Audit your AI inference workloads for GPU architecture dependencies. Identify pipelines that rely on CUDA-specific libraries or Nvidia's Nsight profiling tools. Test those workloads on existing AMD ROCm environments—either through Azure's current AMD GPU instances (like the NVv4 series with Radeon Instinct MI25) or on-prem test beds.
- Engage with Microsoft about early access. If your organization runs large-scale inference or HPC workloads and might benefit from early testing, contact your Microsoft account team or AMD representative. Both companies typically invite key customers to private previews before wider availability.
- Monitor ROI models. While pricing is unknown, historical trends suggest AMD-powered instances often provide better price-per-core or price-per-GB ratios. If your usage involves bursty inference or variable HPC demand, new instance types could shift your total cost of ownership.
- Evaluate Pensando DPU benefits. If you manage Azure networking at the tenant level, the deeper Pensando integration may eventually surface in networking features or performance profiles. For now, it's a background infrastructure improvement, but one that could reduce latency or jitter in inter-instance communication.
The hardest step may be patience. Helios systems ship in the second half of 2026, and Azure VMs typically lag hardware shipments by weeks to months as Microsoft validates and integrates the platform. Realistically, general availability might slip into late 2026 or early 2027.
What to Watch Next
Availability announcements will be the next checkpoint. Expect Microsoft and AMD to disclose specific regions and instance sizes at Azure-focused events or AMD's Accelerated Data Center launch events. Independent benchmarks comparing ND MI455X v7 inference throughput against H100 and H200 instances will be crucial for enterprise evaluation, particularly for transformer-based models that dominate production AI today.
The partnership also tests how well AMD's UALink interconnect performs at cloud scale. If Helios delivers near-NVLink performance without the vendor lock-in, it could accelerate adoption of AMD Instinct across rival clouds. For now, the signal is clear: Azure's AI infrastructure is becoming a multi-vendor affair, and IT planners have new hardware to factor into their 2027 roadmaps.