Microsoft has committed to deploying AMD’s first rack-scale AI architecture, Helios, inside Azure, introducing three new virtual machine families that will bring a 72-GPU accelerator system to the public cloud for the first time. The move, confirmed on July 20, 2026, and reported by CNBC, signals a major expansion of Microsoft’s AI hardware beyond Nvidia and gives cloud customers a fresh option for production inference, agentic AI data processing, and semiconductor design workloads.

Inside the Helios Rack: What Microsoft is Actually Getting

Helios is not another standalone GPU. It is a complete, pre-integrated rack-scale design that combines AMD’s Instinct MI455X accelerators, next-generation EPYC “Venice” CPUs, Pensando networking technology, and the ROCm software stack. Each full configuration packs 72 GPUs, each equipped with 432 GB of HBM4 memory, delivering roughly 31 TB of high-bandwidth memory per rack. AMD says the system can hit up to 1.4 exaFLOPS at FP8 precision and 19.6 TB/s of memory bandwidth per GPU—numbers aimed squarely at the memory-intensive demands of large language model inference and distributed training.

The architecture uses Meta’s Open Rack Wide (ORW) specification, submitted through the Open Compute Project. The double-wide format creates extra room for liquid cooling, dense trays, and networking, while the rack-scale design means AMD ships a blueprint that partners can build and customize. UALink over Ethernet connects the 72 GPUs into a single scale-up domain, and Pensando-based scale-out networking handles inter-rack traffic. It is a direct response to Nvidia’s Grace Blackwell and Vera Rubin platforms, and it turns AMD from a component supplier into a full-stack AI infrastructure vendor.

A Trio of New Azure VM Families

Microsoft isn’t simply dropping Helios into one generic machine type. The company is tailoring three distinct Azure offerings, each designed for a different class of workload. This granular approach reflects the growing specialization already common in general-purpose cloud instances.

  • ND MI455X v7 – The headline grabber. These virtual machines will host production AI inference, frontier-model serving, and Azure AI services. Microsoft explicitly positions them for large-scale, memory-hungry inference scenarios where cost per token is critical. The “v7” naming continues the ND-series lineage for GPU-accelerated instances, but with AMD silicon instead of Nvidia.
  • HDv2 – Co-designed with AMD for the “AI data pipeline.” Expect nearly 500 physical sixth-gen EPYC cores, 4 TB of RAM, 32 TB of local NVMe storage, and 400 Gb Azure Boost networking. This instance is built for the heavy CPU-and-storage tasks that feed accelerators: data preparation, search indexing, reinforcement learning, and agentic AI orchestration. Modern AI workloads demand more than just math throughput, and HDv2 addresses that.
  • HXv2 – Aimed at high-performance computing, especially electronic design automation. Microsoft plans 176 Venice cores clocked above 5 GHz, large per-core caches, 2–4 TB of memory, and 800 Gb InfiniBand. This is the instance that AMD itself may use to design future processors—a feedback loop that shows how deeply cloud hardware choices now influence the silicon supply chain.

All three families are expected to begin customer availability in the second half of 2026, with shipments of Helios systems to Microsoft starting by the end of the year. Initial deployments will land in data centers equipped for the substantial power and liquid cooling these racks require.

Why Microsoft is Diversifying Its AI Hardware

The decision isn’t a rejection of Nvidia. Azure will continue to operate a massive fleet of Nvidia GPUs, and the partnership with OpenAI keeps Nvidia central to Microsoft’s AI strategy. The real story is capacity and cost. Despite record spending, AI infrastructure remains supply-constrained by advanced packaging, high-bandwidth memory, power availability, and networking components. Adding a second credible accelerator supplier gives Microsoft access to an additional manufacturing pipeline and another roadmap.

It also strengthens Microsoft’s negotiating position. Hyperscalers prefer multiple suppliers to avoid single-vendor pricing and to influence product directions. AMD gets a marquee cloud validation; Microsoft gets leverage across its entire hardware ecosystem. And for workloads where memory bandwidth, not peak flops, determines performance, AMD’s 31 TB-per-rack design could offer compelling tokens-per-dollar economics. Even if AMD’s market share remains modest—industry estimates still put Nvidia above 95%—the sheer size of the AI infrastructure market means that capturing a few additional percentage points can translate into tens of billions of dollars in revenue.

What This Means for You—From the Desktop to the Datacenter

If you’re using Windows, you won’t see a “Helios” option in Settings. But you will feel its influence through the services Microsoft delivers. Copilot in Microsoft 365, GitHub Copilot, Azure AI services, and security tools all run on cloud infrastructure. If AMD hardware lowers inference costs, Microsoft can pack more capabilities into those products, expand context windows, support more concurrent users, or pass savings along—though subscription prices might not drop overnight. The key takeaway: a wider hardware pipeline reduces the risk that AI features degrade or get throttled when demand spikes.

For developers and IT teams already building on Azure, the new families represent practical new choices. If your model is memory-bound or you run high-volume inference, ND MI455X v7 could become a cost-effective alternative to Nvidia instances. The HDv2 and HXv2 families, meanwhile, fill gaps that pure GPU instances leave behind—data prep, agentic workflows, and CPU-intensive simulation. The catch is software. AMD’s ROCm stack supports PyTorch, TensorFlow, JAX, and popular inference engines, but it hasn’t yet reached the seamless maturity of CUDA. Microsoft will abstract much of that complexity in managed services, but if you plan to run raw VMs, budget time for porting and performance validation.

Enterprises should track not just availability dates but also which Azure regions come online. Power and cooling constraints mean Helios racks won’t land in every data center at once. If you operate in a region that relies on ample AI capacity, monitor Azure’s ND-series and H-series pricing pages for new SKUs later in 2026. The arrival of competitive hardware often improves existing instance pricing, too.

How We Got Here: From EPYC CPUs to AI Racks

AMD’s data center story started with EPYC server processors, which used chiplet designs and aggressive core counts to break Intel’s stranglehold. Rome, Milan, Genoa, and Turin established AMD as a credible alternative, and that CPU success gave the company the financial foundation to invest in Instinct GPUs. The MI300X accelerator, with its large HBM capacity, found traction in cloud inference workloads, and Microsoft was an early Azure customer.

The industry pivot to rack-scale systems changed the rules. A fast chip alone no longer suffices when clusters span thousands of accelerators. Nvidia realized this early with its Grace Blackwell and Vera Rubin platforms. Helios is AMD’s answer: a reference design that bundles compute, networking, cooling, and management into a single validated architecture. Acquisitions fueled the shift—Pensando for smart networking, Xilinx for adaptive compute, and ZT Systems for rack engineering. Helios is the product of that assembly.

What to Do Now (Actionable Steps)

For most Azure users, the immediate action is patience. The VMs won’t be generally available until late 2026, and Microsoft hasn’t announced preview dates. However, a few steps can put you ahead:

  1. Evaluate your model’s memory profile. If you run models that barely fit on current GPU instances, the 432 GB per GPU on MI455X may reduce the need for model sharding and simplify deployment. Start tracking whether your cost per inference makes a switch worthwhile.
  2. Keep an eye on ROCm updates. AMD’s software stack is improving rapidly. Key repositories—PyTorch, ONNX Runtime, vLLM, and Triton—now support AMD GPUs. Test your workload on smaller MI300X instances if available in your region to gauge porting effort.
  3. Watch Azure pricing updates. When Microsoft introduces new instance families, older GPU types sometimes drop in price. Even if you stick with Nvidia, the competitive pressure can benefit your budget.
  4. Factor Helios into capacity planning. If your organization anticipates needing large AI clusters in 2027, having another hardware option may ease procurement bottlenecks. Speak with your Microsoft account team about provisioning expectations.

Outlook: The Real Test Comes After the Announcement

AMD has assembled the pieces—processors, networking, rack design, and a growing customer list that includes Meta, OpenAI, Oracle, and TCS. Microsoft’s backing moves Helios from a vendor promise to a production reality. But delivery must follow. Shipments are expected by the end of 2026, meaning the first Azure clusters will light up around the same time or shortly after. Any slip could blunt momentum.

The true verdict will come from independent benchmarks measuring tokens per second, tokens per dollar, and multi-rack scaling efficiency on real models. Nvidia’s Rubin platform will also be advancing, so AMD is chasing a moving target. If Helios meets its specifications and ROCm proves production-ready, Microsoft wins a versatile new AI infrastructure option. If it falls short, the announcement will still have served to put pricing pressure on an industry that desperately needs it. For Windows users and Azure customers, that competitive tension alone is a reason to pay attention.