Microsoft is bringing a new class of supercomputing-class hardware to its cloud, announcing on July 20, 2026, that it will deploy AMD’s Helios rack-scale AI systems alongside new CPU-focused virtual machines with nearly 500 cores. The move, aimed squarely at AI inference and data-heavy workloads, will begin rolling out in the second half of 2026, giving Azure customers access to some of the densest computing configurations ever offered in the public cloud.

The Hardware: What’s Actually Coming

At the center of the announcement is AMD’s Helios Rackscale Solution, a pre-integrated architecture that packs 72 Instinct MI455X GPUs, sixth-generation EPYC “Venice” CPUs, Pensando networking components, and liquid cooling into a single rack. Each MI455X GPU, built on AMD’s CDNA 5 architecture, carries up to 432GB of HBM4 memory and delivers 19.6TB/s of memory bandwidth. The full rack offers 31TB of total HBM4 and boasts up to 260TB/s of aggregate scale-up bandwidth between accelerators, plus 43TB/s of scale-out networking via Pensando technology and an Ultra Ethernet approach.

Microsoft positions Helios not for model training but for large-scale inference workloads—serving frontier AI models, powering Azure AI services, and handling customer applications that demand low latency, high throughput, and long-context processing. This focus on inference reflects the operational reality of many AI services, where every user query, agent action, or retrieval request consumes precious accelerator cycles.

Alongside Helios, Microsoft introduced three new Azure virtual machine families:

  • HDv2: A CPU monster designed for AI data pipelines, search, and reinforcement learning. It pairs nearly 500 physical sixth-generation EPYC cores with 4TB of RAM, 32TB of local NVMe storage, and 400Gb Azure Boost networking. This machine targets the often-overlooked preprocessing, indexing, and orchestration stages that feed GPU clusters.
  • HXv2: Built for electronic design automation, semiconductor simulation, and technical computing, it offers 176 EPYC cores clocked above 5GHz, AMD’s 3D V-Cache technology (50% more cache per core), up to 4TB of RAM, and 800Gb InfiniBand for distributed workloads.
  • ND MI455X v7: The GPU workhorse that directly leverages Helios racks for production AI inference, especially suited to reasoning models, multi-agent systems, and retrieval-augmented generation.

All three are forward-looking—Microsoft has not yet announced general availability or pricing—but the specifications signal a deeper integration of AMD silicon across Azure’s entire AI stack, from raw compute to networking offload via AMD Pensando DPUs and Microsoft’s own Azure Boost platform.

Who Stands to Benefit

For AI Developers and Data Scientists

If you’re building or deploying models, the HDv2 and ND MI455X v7 families could reshape how you architect production systems. The high memory capacity and fast interconnects of Helios make it feasible to serve very large models without aggressive sharding, keeping latency predictable. However, software compatibility is paramount. AMD’s ROCm stack now supports PyTorch, TensorFlow, JAX, ONNX Runtime, and vLLM, but you’ll need to validate your exact inference engine, quantization methods, and distributed communication libraries. Pro tip: start testing your containers on existing Azure AMD instances today to flush out ROCm-specific quirks before these new VMs arrive.

For IT Infrastructure Managers

The expansion means more diversity in Azure’s AI hardware, reducing vendor lock-in and potentially lowering costs. But you’ll face planning decisions: when to migrate from older AMD series like HBv2 (which Microsoft is retiring), how to budget for per-token inference costs instead of per-hour VM prices, and how to manage quotas for high-demand VMs that may launch in limited regions. Watch for reserved capacity options and benchmark results that measure real-world throughput, not just theoretical peak specs.

For Everyday Windows and Microsoft 365 Users

You won’t log into an HDv2 VM or tweak ROCm drivers, but the services you use—Copilot in Windows, intelligent search in Outlook, the cloud-based AI behind Microsoft 365—could run on this hardware. Tighter inference loops and more headroom for complex agent tasks may translate into snappier responses and more capable features over time. The impact will be indirect but pervasive.

The Bigger Picture: Why This Matters Now

Microsoft and AMD have collaborated on Azure EPYC VMs for years, but the Helios partnership represents a step change. It brings together CPU, GPU, networking, and software in a rack-scale design validated for Azure’s hyperscale environment. This is happening because inference demand is exploding: every AI-powered chat, code completion, or enterprise agent requires continuous, low-latency processing. Cloud providers can no longer treat training and inference as symmetric problems; the economics and hardware needs differ sharply.

Against the backdrop of an Nvidia-dominated GPU landscape, AMD’s open-standards approach—Helios aligns with the Open Compute Project’s Open Rack Wide format, UALink for scale-up, and Ultra Ethernet for scale-out—offers customers a path to portability. The retirement of Azure HBv2 VMs (based on older EPYC processors) also nudges orgs toward these next-gen options, making it timely to evaluate a refresh of your AI infrastructure.

What You Should Do to Prepare

  1. Check your pipeline’s CPU bottlenecks. If data preprocessing or search indexing is slowing your GPU cluster, an HDv2 VM could be a force multiplier. Map your end-to-end workflow to identify where you’re waiting on CPU-bound tasks.
  2. Validate ROCm compatibility now. Use existing Azure NDv4 or NVv4 instances (or on-prem AMD hardware) to test your model-serving stack. Pay special attention to kernel support for custom ops and the performance of distributed communication.
  3. Monitor preview announcements. Microsoft will likely open early access programs for HDv2, HXv2, and ND MI455X v7. Sign up for Azure updates and be ready to move when capacity becomes available—initial quotas will be tight.
  4. Plan for migration from HBv2. If you’re still running HBv2 instances, Microsoft’s retirement means you’ll need a new home. HXv2 is the natural successor for many HPC workloads; evaluate whether the higher clock speeds and cache outweigh any per-core licensing implications.
  5. Think in cost-per-task, not cost-per-hour. For inference, a VM that costs more per hour but delivers lower latency and higher throughput can actually reduce the cost of serving each request. Run your own TCO models once pricing is published.

Outlook

AMD says Helios systems will begin shipping in the second half of 2026, so expect Azure region rollouts to follow that timeline. Pricing, exact SKU sizes, and supported operating systems remain unknown. Independent benchmarks—especially for large-model serving and agentic workflows—will be critical. In the meantime, Microsoft’s expanded AMD footprint sends a clear signal: the cloud AI hardware market is diversifying fast, and that’s good news for anyone building the next generation of intelligent applications on Azure.