Microsoft has pulled back the curtain on three new Azure virtual machine families built with next-generation AMD silicon, but IT buyers hoping to spin up instances this quarter will be disappointed: none of the offerings have a release date, pricing, or regional availability yet. The July 20 announcement instead serves as a roadmap for infrastructure planners, signaling that AMD’s newest EPYC CPUs and Instinct GPUs will carve out distinct roles inside Azure’s AI and high-performance computing portfolios.

The three families—dubbed HDv2, HXv2, and ND MI455X v7—address different pressure points. HDv2 aims at the data-heavy work that feeds AI models. HXv2 targets cache-sensitive engineering simulations and tightly coupled scientific computing. ND MI455X v7 marks AMD’s Helios rack-scale design coming to Azure for production-scale AI inference. No one can order these VMs today, but the specs Microsoft provided make it clear they will change the conversation for anyone planning large-scale AI, chip design, or technical computing in the cloud.

What got announced, spec by spec

Microsoft’s blog post laid out concrete numbers for the CPU-based HDv2 and HXv2, while keeping the GPU-backed ND MI455X v7 mostly under wraps. All three will use AMD’s sixth-generation EPYC processors. The MI455X v7 instances add AMD Instinct MI455X accelerators and the Helios rack-level platform.

Azure HDv2 is the brute-force data handler. A single VM will pack nearly 500 physical EPYC cores, 4 TB of RAM, 32 TB of local NVMe storage, and 400 Gbps Azure Boost networking. That’s not a general-purpose server—it’s designed to feed and orchestrate AI systems. Microsoft explicitly called out data preparation, search, reinforcement learning, and agent coordination as ideal workloads. In production AI deployments, starved data pipelines often leave expensive GPUs idle; HDv2 is engineered to prevent that bottleneck.

Azure HXv2 takes over where the original HX series left off for electronic design automation (EDA), fluid dynamics, and distributed-memory HPC. Each VM will offer 176 AMD EPYC cores clocked above 5 GHz, with 50 percent more addressable cache per core than its predecessor. Memory scales to nearly 2 TB or 4 TB depending on configuration. The networking jump is significant: 800 Gbps InfiniBand, doubling the previous generation’s 400 Gbps NDR fabric. For MPI jobs that fan out across dozens of nodes, the wider pipe can directly shrink simulation turn-around times.

ND MI455X v7 is the wildcard. Microsoft says it’s built on AMD Helios, which packs Instinct MI455X GPUs, next-gen EPYC CPUs, and AMD networking into a rack-scale system. The target is inference—reasoning, search, and agentic AI workloads running at production scale. That’s a deliberate choice. Training might happen once, but inference runs continuously and racks up the real cloud bill. Microsoft did not disclose GPU counts, memory per GPU, cluster sizes, or supported regions. We also haven’t seen details about ROCm version compatibility or integration with Azure Machine Learning and Azure Kubernetes Service.

What the announcement means for different Azure users

Because availability and pricing are still unknown, the immediate takeaway is strategic, not operational. But the specs provide enough to start assessing fit.

AI and machine learning teams should pay attention to HDv2 and MI455X v7. Today, many Azure AI pipelines run on general-purpose Dv4 or Ev5 instances, which can buckle under large-scale data ingestion, vector indexing, or preprocessing. HDv2’s 4 TB memory and 32 TB local NVMe can shrink ETL windows and keep accelerators fed. For inference, the MI455X v7 introduces an AMD GPU option outside Nvidia’s ecosystem. That could matter for cost-sensitive deployments or organizations that want to avoid toolchain lock-in. But until Microsoft shares GPU specs and pricing, no one can compare it to existing NC or ND series offerings.

HPC and EDA engineers have a clearer upgrade path. Existing Azure HX VMs with AMD 3D V-Cache already cater to RTL simulation and memory-intensive EDA tools. HXv2 lifts both single-threaded performance (5 GHz clocks) and cache, which directly speeds up chip design workflows. The 800 Gbps InfiniBand also helps scale out MPI jobs that choke on network latency. Teams currently running HX or HB instances should pencil HXv2 into capacity projections but hold off on architectural changes until Microsoft publishes benchmark data and SKU lists.

Windows administrators and .NET shops may find the HDv2 spec sheet appealing—4 TB of memory and 500 cores can host large SQL Server databases or .NET-based data services that often sit alongside AI inference tiers. However, Microsoft hasn’t said a word about Windows Server support for any of the three new families. The current HX documentation shows only Linux-centric configurations, and the HDv2 and HXv2 announcements don’t break that pattern. Until Azure publishes official OS support statements, assume these VMs will require Linux workloads.

Procurement and FinOps teams get a heads-up that AMD silicon will be a bigger line item in future Azure bills. Microsoft is not simply adding one GPU option; it’s weaving AMD across CPU, HPC, and accelerator layers. That might increase negotiation leverage for large customers, but only when the services actually launch. For now, don’t factor these into committed-use discounts or reserved instances.

How Azure’s AMD collaboration reached this point

Azure’s relationship with AMD in the HPC space isn’t new. Microsoft launched its first HX-series VMs in 2023, using AMD EPYC 9V33X “Genoa-X” processors with 3D V-Cache technology. Those VMs offered up to 176 cores, 1,408 GB of RAM, and 400 Gbps InfiniBand, as documented in Microsoft Learn. They quickly found a niche in silicon design, where cache behavior trumps raw core counts. Synopsys, for example, has publicly discussed its Azure-based EDA workloads on HX instances.

The three new families expand the partnership in two directions: horizontally, by adding a data-plane workhorse (HDv2) and a dedicated inference GPU service (MI455X v7); and vertically, by pushing per-VM specs higher with next-gen EPYC and Helios. Microsoft’s “heterogeneous infrastructure” framing acknowledges that no single processor design optimizes for data prep, simulation, and model inference simultaneously. The product portfolio is splitting along workload lines.

AMD’s Helios platform also signals that the company wants to compete at rack scale, not just sell individual GPUs. For Azure customers, that could eventually mean pre-integrated GPU clusters tailored for inference, much like the ND H100 v5 series does for Nvidia. But Helios is still an unproven quantity in public cloud, and the missing details on MI455X v7 suggest the engineering and validation work is ongoing.

What to do now while you wait

Because these VMs aren’t orderable, immediate actions are all about preparation and monitoring.

  1. Audit your current workload mix. If you run large-scale data preparation, EDA simulations, or growing inference fleets on Azure, map out where existing VMs fall short. Are your GPUs frequently idle waiting for data? Do simulation runs hit memory walls? That will help prioritize which family matters most.
  2. Watch the Azure updates and product blogs. Microsoft typically rolls out new VM families via public preview in a few regions first. Bookmark the Azure Virtual Machines product page and the AMD partnership tag on the Azure blog. Look for “HDv2,” “HXv2,” or “ND MI455X v7” in the title—those will be your signal that deployment details are near.
  3. Engage your Microsoft account team if you have large-scale needs. Enterprise customers with significant AI or HPC spend can sometimes access early adoption programs. While nothing is guaranteed, expressing interest now can put you on the list for private previews when they open.
  4. Hold off on architecture redesigns. The specs are compelling, but until you see benchmark results, supported regions, and pricing, don’t re-platform existing services. Run a cost-benefit analysis on paper, but wait for hard data before committing.
  5. For Windows environments, assume Linux-only until proven otherwise. Given the HPC heritage of these families, it’s a safe bet that first releases will target Linux. If you need Windows support, flag that to Microsoft through support channels or user voice forums.

What to watch next

The real story will unfold when Microsoft moves from roadmap to reality. The most closely watched milestone will be the first public preview, likely starting with either HDv2 or HXv2 given their CPU-only nature. ND MI455X v7 may take longer, as it requires validating GPU drivers, ROCm, and rack-scale networking across multiple Azure regions. The competitive landscape will also heat up: Google Cloud and AWS are both expanding their own AMD EPYC and GPU options. For Azure customers, the net effect of this announcement is increased choice—just not yet. When the “sign up for preview” button appears, the practical work of testing and sizing can finally begin.

In the meantime, treat the July 20 blog post as a credible signal that your next inference fleet or simulation cluster might run on AMD silicon inside Azure. But until the regions are named and the bills are calculated, it stays in the planning column.