Microsoft is adding AMD’s Helios rack-scale AI systems to its Azure cloud, with the first shipments to the company and its customers expected to begin in the second half of 2026. The deal marks a major expansion of a decades-long partnership and gives Azure users a fresh accelerator option alongside Nvidia GPUs and Microsoft’s own Maia chips.

What Microsoft and AMD Actually Announced

The centerpiece is Azure’s deployment of the AMD Helios Rackscale Solution, built around the Instinct MI455X accelerator. AMD expects Helios-based racks to start shipping in the second half of 2026, and Microsoft plans to use them for frontier-model inference, Azure AI services, and customer workloads. The hardware will surface in the cloud as the ND MI455X v7 virtual machine family.

Helios isn’t a plug-in card or a single server. It’s a reference design for an entire rack—a blueprint that server makers and cloud operators can adapt. A typical configuration packs 72 MI455X GPUs, sixth-generation EPYC “Venice” processors, and Pensando data-processing units (DPUs) into a double-wide Open Rack form factor. Across the rack, AMD claims about 31 terabytes of aggregate HBM4 memory and 1.4 petabytes per second of memory bandwidth. Those numbers describe the rack as a whole—not a single pool with uniform access—but they underline the scale at which AMD wants to compete.

Microsoft also announced two additional Azure VM families based on the same EPYC Venice chips, but without the Helios acceleration:

  • HDv2: aimed at AI data processing, search, reinforcement-learning pipelines, and coordinating fleets of AI agents.
  • HXv2: designed for electronic design automation and other technical computing tasks that crave high per-core performance.

Finally, the partnership extends to networking. Microsoft already uses Pensando technology in parts of Azure, and the expanded deal puts those DPUs into AMD’s AI backend infrastructure and deeper into Azure services, with closer integration into Microsoft’s Azure Boost hardware-acceleration layer.

What This Means for You

The impact depends on who you are.

For everyday Windows users

Unless you’re an IT administrator or a developer, you won’t see a “Helios” badge in your settings. The benefit will arrive indirectly, through faster, more available, or less expensive AI services. Features like Microsoft 365 Copilot, Windows’ own AI-powered tools, and third-party apps running on Azure could become more responsive or less constrained during peak demand. Lower inference costs might also let Microsoft offer more capable models without hiking subscription prices—though don’t expect lower bills automatically; savings could be reinvested in larger models or longer context windows.

For IT professionals and cloud architects

Once the ND MI455X v7 series reaches general availability—likely after a limited preview period—you’ll have a new instance family for large-scale AI inference. The primary draw is cost. AMD’s public pitch is “best total cost of ownership and lowest per-token cost” for inference, and while a single Helios rack is estimated to cost between $5 million and $5.5 million (per industry analysts at Futurum Group), the operational savings could make it cheaper over time than some Nvidia alternatives.

But there’s a catch: to run directly on these VMs, your software must work with AMD’s ROCm stack. ROCm has matured—it now supports PyTorch, TensorFlow, JAX, ONNX Runtime, and vLLM—but it still trails Nvidia’s CUDA ecosystem in developer tooling, debuggers, and sheer breadth of community knowledge. Migration is rarely as simple as recompiling. You may need to rewrite custom kernels, validate numerical behavior, and re-benchmark every important model configuration.

For Azure-native services, Microsoft abstracts the accelerator. If you use Azure Foundry Managed Compute or similar offerings, your inference jobs might land on AMD hardware automatically, without any change on your part. That’s the smoothest path to benefiting from Helios.

For developers and data scientists

If you work directly with GPUs, start familiarizing yourself with ROCm now. Build container images that can run on both Nvidia and AMD hardware, and test your training or inference scripts on existing AMD instances (such as the current MI300X-based VMs) to gauge compatibility. The more portable your code, the sooner you’ll be able to experiment with the new instance family when it arrives.

Why Rack-Scale AI Systems Are Different

The AI industry has shifted from comparing individual GPUs to evaluating entire racks. Modern models spread across hundreds or thousands of accelerators, and the network, memory system, host processors, and software must all keep up. A fast GPU means little if the data can’t reach it quickly.

Rack-scale designs like Helios—and Nvidia’s competing Grace Blackwell or Vera Rubin systems—optimize the whole machine. They co-design accelerators, CPUs, interconnects, and power delivery. For inference providers, the metrics that matter are tokens generated per watt, tokens per dollar, end-user latency, and the ability to fit large models in fewer racks. By bundling its own GPUs, CPUs, DPUs, and software, AMD can argue on these system-level measures rather than just peak floating-point performance.

How We Got Here: A Growing Partnership and a Need for Choice

Microsoft and AMD have collaborated for decades, from Xbox consoles and Surface laptops to the first Azure instances based on MI300X GPUs in 2023. That history gave Microsoft the confidence to validate AMD hardware for hyperscale use. Today, Azure operates a carefully heterogeneous fleet: Nvidia GPUs, AMD Instinct accelerators, AMD and Intel server processors, and Microsoft’s own Maia AI chips. No single vendor is irreplaceable.

The Helios deal extends this strategy. AMD now counts eight of the world’s top ten AI companies as Instinct GPU users, and it expects tens of billions of dollars in data-center AI revenue starting in 2027, mostly from Helios. Microsoft’s commitment signals that AMD’s hardware has moved beyond a proof-of-concept and into serious deployment planning.

What You Should Do Now

For enterprise IT decision-makers: Identify AI inference workloads with significant recurring cloud costs. Monitor any Azure previews of ND MI455X v7 instances and, if your contract allows, request early access. Begin benchmarking existing models on current AMD instances to understand the effort required.

For developers: Ensure your AI stack is framework-rather than hardware-dependent. Test PyTorch, TensorFlow, or JAX code on ROCm. Containerize your workloads so you can switch compute backends without rebuilding everything. The goal isn’t to abandon Nvidia, but to keep options open.

For consumers: There’s nothing to do. You’ll start to notice the benefits only when Microsoft—and the thousands of businesses that rely on Azure—gets more efficient at serving AI models at scale.

What to Watch Next

The first major milestone is the transition from reference architecture to volume systems in the second half of 2026. After that, watch for regional availability, pricing announcements, and—most importantly—independent benchmarks that test real models under real traffic. Microsoft’s willingness to run high-profile internal services on Helios will say as much as any press release.

A second indicator is software portability. If enterprises can move models between Azure’s AMD, Nvidia, and Maia accelerators with minimal changes, the infrastructure market will become genuinely competitive. That would be good news for customers, even if it doesn’t dethrone Nvidia. The real contest is about giving you more choice, lower costs, and better AI experiences—without you ever having to think about the servers.