Microsoft is bringing AMD’s Helios rack-scale AI design to Azure, introducing a trio of new virtual machine families built around the upcoming Instinct MI455X GPU, sixth-generation EPYC Venice CPUs, and Pensando networking. The centerpiece: the ND MI455X v7 series, aimed squarely at production inference for agentic AI and search workloads, backed by racks packing 72 GPUs and liquid cooling. The announcement, made jointly by the two companies, signals a deepening partnership that gives Azure a credible non-Nvidia accelerator option while extending AMD’s reach from individual chips to complete, cloud-ready systems.

What Exactly Did Microsoft Announce?

Microsoft detailed three new Azure instance families, all powered by AMD silicon that won’t enter customer previews until sometime after AMD begins shipping Helios racks “later this year” – meaning a 2026 or early 2027 availability window is realistic. Alongside the hardware news, Microsoft committed to optimizing Azure Boost for AMD platforms, offloading virtualization overhead to dedicated processors and freeing more host CPU for tenant workloads.

The three families break down as follows:

  • ND MI455X v7 – GPU-accelerated inference instances built around the MI455X, a CDNA 5 architecture chip with 432 GB of HBM4 memory and 19.6 TB/s of bandwidth per GPU. These VMs target large-model serving, agentic systems, and retrieval-augmented generation.
  • HDv2 – A CPU-heavy series for AI data preparation and orchestration. Top configurations offer nearly 500 physical EPYC Venice cores, 4 TB of RAM, 32 TB of local NVMe storage, and 400 Gb Azure Boost networking.
  • HXv2 – A successor to the existing HX family for electronic design automation (EDA) and high-performance computing. It delivers 176 physical cores running above 5 GHz, 50% more addressable cache per core, up to 4 TB of memory, and 800 Gb InfiniBand for tightly coupled MPI simulations.

A Closer Look at the Three New Families

ND MI455X v7: Inference at Scale

The Helios rack houses 72 MI455X GPUs, connected via UALink – an open-standard scale-up fabric that AMD implements as “UALoE” over Ethernet. Each GPU’s 432 GB of HBM4 means a single rack holds roughly 31 TB of high-bandwidth memory. That’s a deliberate bet on large-context, memory-hungry AI workloads, where fitting a model’s weights, key-value cache, and intermediate activations entirely on-package can dramatically cut communication overhead and boost throughput. Microsoft says the ND MI455X v7 is optimized for inference, placing it right where the industry is shifting: away from one-off training jobs toward persistent, always-on AI services that need competitive cost-per-token and low latency.

HDv2: The CPU Workhorse for AI Data

AI doesn’t live on GPUs alone. Data preparation, tokenization, reinforcement-learning orchestration, and retrieval pipelines often rely on heavy CPU lift. HDv2 addresses that with a massive core count (up to ~500 physical cores, not just threads) and abundant local NVMe storage for temporary datasets and caches. The 32 TB of local flash is particularly useful for shuffling large corpora without repeatedly hitting remote storage. Admins should note that local VM storage is ephemeral; critical data still needs durable Azure storage.

HXv2: Upgraded HPC and Chip Design

HXv2 takes direct aim at chip-design workflows and other latency-sensitive simulations. Moving from the original HX series’ 3D V-Cache to a more advanced Venice-based design, Microsoft quotes a frequency above 5 GHz and a 50% per-core cache increase. That matters because EDA tools like place-and-route and timing analysis often reward single-thread speed and fast memory access more than raw thread count. Adding 800 Gb InfiniBand extends the machine’s usefulness to traditional HPC workloads such as computational fluid dynamics and weather modeling, where low-latency message passing is essential.

Why AMD’s Whole-Rack Approach Matters

Microsoft isn’t simply plugging a new GPU into an existing server blueprint. Helios is a reference design that treats the entire rack – compute trays, switching, power delivery, liquid cooling, and serviceability – as a coordinated unit. Each compute tray pairs one EPYC Venice CPU with four MI455X GPUs; networking is handled by Pensando “Vulcano” network interfaces and “Salina” data processing units. The double-wide Open Rack form factor and mandatory liquid cooling reflect the immense power density of modern accelerators, but they also mean Azure can only deploy Helios in data centers equipped with the necessary coolant distribution and power infrastructure – likely limiting initial regional availability.

For Azure, the payoff is strategic. With Helios, Microsoft gains an alternative to Nvidia’s tightly integrated DGX and NVLink systems at a time when inference workloads are ballooning and supply diversification is critical. The open-standards design (UALink, Ultra Ethernet, Open Compute Project) also aligns with Azure’s own push toward Azure Boost offload – dedicating host CPU to tenants while moving networking, storage, and security to programmable DPUs and NICs.

What This Means for Your Workloads

For AI Developers

If you’re building or deploying generative AI models in Azure, the MI455X’s 432 GB of HBM4 per GPU is the standout figure. It can reduce the number of accelerators needed for a given model size, cutting cross-GPU communication. However, that benefit hinges on software maturity. The ROCm stack supports PyTorch, JAX, and ONNX Runtime, but many AI pipelines still assume CUDA at a deep level. Before betting on MI455X, teams should:

  • Inventory all GPU-specific dependencies in containers, custom kernels, and inference servers.
  • Validate functional compatibility on current MI300X instances as a stepping stone.
  • Benchmark end-to-end inference with realistic context lengths, not just short synthetic prompts.
  • Compare total cost per completed request, factoring in networking, storage, and CPU charges, rather than raw per-second GPU pricing.

For IT Admins and Enterprises

HDv2 and HXv2 fill specific roles that Nvidia-focused announcements often ignore. If your organization runs massive data prep jobs, reinforcement learning with agentic tool-calling, or traditional EDA/HPC simulations, these CPU-heavy instances may offer better price-performance than general-purpose VMs. Key considerations:

  • HDv2’s physical core count and local NVMe suit highly parallel data processing, but plan for ephemeral storage and checkpoint workflows.
  • HXv2’s cache-rich, high-frequency cores are tailored for latency-bound codes; migrating from the older HX series could yield a meaningful uplift if your application is cache-sensitive.
  • Regional availability will be constrained at launch because of liquid cooling and power requirements. Enterprises with data residency needs should monitor Azure’s rollout carefully.

For Windows and End Users

Consumers won’t see an “AMD Helios” toggle in Windows or Copilot. The impact will be indirect: if inference costs drop and capacity expands, services like Microsoft 365 Copilot could become more responsive or handle longer contexts. Power users and IT pros may eventually interact with HXv2 instances for specialized Windows-based HPC workloads, but the immediate benefits are all on the back end.

From EPYC Milan to Venice: The Path to Helios

AMD’s relationship with Azure has been building for years. EPYC Milan and Genoa already power numerous Azure VM families, and the MI300X accelerator offered an early alternative for memory-hungry AI models. Helios represents a leap in integration: instead of selling standalone GPUs, AMD is delivering a complete rack blueprint that Microsoft can adapt for its fleet. The sixth-generation EPYC Venice processors at the heart of all three new VM families use TSMC’s 2 nm-class process and advanced chiplet stacking, delivering up to 256 cores per socket. That density, combined with Azure Boost offload, should give tenants a larger share of the hardware’s potential.

Microsoft’s multi-vendor strategy hasn’t slowed its Nvidia purchases, but it has created a more balanced portfolio. The company now mixes Nvidia Hopper and Blackwell systems, AMD MI300X and Helios, and its own Maia accelerators and Cobalt CPUs. For customers, that means more choice – but also a steeper learning curve to match workloads with the right hardware.

Next Steps: How to Prepare for AMD-Powered Azure VMs

While you can’t yet spin up an ND MI455X v7 instance, there are moves you can make now:

  • Test the waters with MI300X: If your model runs on today’s AMD instances, you’re in good shape for Helios. Use the time to profile memory usage and identify any ROCm-specific optimizations.
  • Containerize for portability: Use PyTorch, JAX, or ONNX Runtime with standard Docker images. Avoid hard-coding CUDA versions or Nvidia-specific extensions unless you must.
  • Watch Microsoft’s VM pricing page: As dates firm up, Microsoft will publish instance sizes, GPU counts per VM, and hourly rates. Those details will determine whether MI455X is cost-competitive for your workload.
  • Plan for regional constraints: If your data must stay in a specific Azure region, check whether that region can support liquid-cooled racks. A limited launch could delay your adoption.
  • Evaluate HDv2 and HXv2 alongside GPU instances: Many AI projects overlook CPU-bound steps. Matching HDv2’s large memory and NVMe to your data pipeline could boost end-to-end throughput even before running inference.

The Outlook: More Choices, More Questions

Microsoft’s Helios bet is ambitious. AMD gains a marquee cloud customer for a rack-scale platform that challenges Nvidia’s system-level lead. Azure gets another high-capacity inference option and strengthens its negotiation position. But the real test will be in the details: pricing, per-core performance, the maturity of the ROCm ecosystem under production loads, and how quickly Microsoft can deploy thousands of liquid-cooled racks across global regions. The ND MI455X v7 series may not be available everywhere on day one, and even when it is, many teams will need time to validate their AI stacks. Still, the direction is clear – Azure’s AI hardware menu is growing, and that’s good news for anyone running at cloud scale.