Microsoft will bring AMD’s Helios rack-scale AI hardware to Azure in the second half of 2026, the company confirmed on July 20, giving Windows admins and developers several new cloud compute options for AI inference, data engineering, and high-performance computing (HPC). The announcement, detailed in a company blog post, describes three upcoming Azure virtual machine (VM) families—ND MI455X v7, HDv2, and HXv2—that significantly expand the availability of AMD silicon inside Microsoft’s cloud.
None of these VMs are available to provision today. Microsoft offered no pricing, regional rollout plans, or a specific launch date. AMD’s own roadmap points to broader Helios deployments starting in the second half of 2026, so enterprises should treat this as a signal of what’s coming, not an immediate resource.
The actual hardware: three new VM families, three distinct jobs
Microsoft named three forthcoming Azure instance types, each aimed at a different part of the enterprise infrastructure stack.
ND MI455X v7 is the headliner. It is a GPU-powered VM series built on AMD’s Helios platform, which pairs next-generation Instinct GPUs, EPYC “Venice” server processors, Pensando networking, and AMD’s ROCm software into a single rack-scale design. The company says these instances will run production AI inference—the kind of work that serves trained models at scale for reasoning, search, and agentic AI applications. Helios is AMD’s attempt to offer cloud operators an integrated alternative to Nvidia-centric systems, complete with compute, networking, and software, rather than requiring separate assembly.
HDv2 targets AI data systems. These VMs provision nearly 500 physical sixth-generation EPYC cores, 4 TB of RAM, 32 TB of local NVMe storage, and 400 Gbps Azure Boost networking. The configuration is clearly designed for the CPU- and storage-intensive steps that feed AI pipelines: data preparation, search indexing, reinforcement learning, and agent coordination. In many workloads, those stages become bottlenecks long before a GPU is ever involved, so a beefy CPU-based instance can accelerate the entire cycle.
HXv2 is aimed at electronic design automation (EDA), scientific simulation, and other technical-computing jobs. Each HXv2 instance will pack 176 sixth-generation EPYC cores running above 5 GHz, with 3D V-Cache, up to 4 TB of memory, and 800 Gbps InfiniBand for tightly coupled MPI workloads. Microsoft first introduced the HX series with AMD in 2023 for RTL simulation tasks, and this follow-up roughly doubles the core count while adding the clock speed and cache typically demanded by chip designers.
| VM Family | Primary Target | Key Specs | Expected Availability |
|---|---|---|---|
| ND MI455X v7 | AI inference (Helios) | Next-gen Instinct GPUs, EPYC Venice CPUs, Pensando networking, ROCm | H2 2026 |
| HDv2 | AI data prep, search, agentic AI | ~500 EPYC cores, 4 TB RAM, 32 TB NVMe, 400 Gbps Azure Boost | H2 2026 |
| HXv2 | EDA, HPC, simulation | 176 EPYC cores >5 GHz, 3D V-Cache, up to 4 TB RAM, 800 Gbps InfiniBand | H2 2026 |
What it means for you: impact by role
For most home and business Windows users, nothing changes at their desk. This is a cloud infrastructure play. But for anyone managing Azure workloads, the announcement adds concrete new options to the roadmap.
Windows admins and IT pros responsible for cloud capacity: you now have a timeline for when AMD’s most current server silicon will appear in Azure. If your organization runs AI inference, data engineering, EDA simulations, or other high-throughput compute on Azure, start evaluating whether these instance types fit your cost-performance profile. Because AMD lags Nvidia in GPU market share, pricing could be aggressive when these VMs reach general availability. Keep an eye on preview sign-ups.
Developers and data engineers building AI pipelines on Azure: the HDv2 series, with its massive core count and local NVMe storage, may let you collapse steps that currently require separate compute and caching clusters. That simplifies architecture and potentially lowers latency. The ND MI455X v7 series, meanwhile, gives inference-heavy applications a non-Nvidia option that runs on Microsoft’s own infrastructure with ROCm support—important if you’re trying to avoid vendor lock-in or optimize for different cost structures.
HPC and semiconductor teams: The HXv2 nearly doubles the cores of its predecessor, adds 3D V-Cache, and bumps clock speeds above 5 GHz. If you rely on Azure for electronic design automation (RTL simulation, place-and-route), computational fluid dynamics, or other MPI-bound workloads, this is a spec bump worth tracking. The move from 5th-gen to 6th-gen EPYC also means newer architecture, which typically brings IPC gains on top of the raw core increase.
No Azure customer can spin up these VMs today. Microsoft did not publish preview dates, and AMD’s own “H2 2026” window encompasses anything from July to December of that year. Plan accordingly.
How we got here: AMD’s deepening Azure footprint
This is not AMD’s first appearance in Azure’s high-end compute lineup. Microsoft has offered AMD EPYC-based VMs for general-purpose and memory-optimized workloads since 2019. The HX series, introduced in 2023 with 5th-gen EPYC, was built specifically for AMD’s strengths in high-frequency, cache-heavy EDA workloads. All three new families represent a deepening commitment, not a sudden turn.
The Helios platform itself is new. AMD announced Helios in 2025 as a rack-scale design that integrates GPUs, CPUs, networking, and software, directly challenging Nvidia’s DGX and HGX reference architectures. By landing a cloud provider as large as Azure for Helios, AMD gets a critical endorser that can validate the platform’s real-world efficiency and economics. For Microsoft, adding AMD-based inference capacity diversifies its GPU fleet, which has historically leaned heavily on Nvidia. Given the still-constrained supply of Nvidia H100 and B100 GPUs, alternative silicon helps Azure guarantee capacity for customers.
The jump to 6th-generation EPYC (“Venice”) is also timed to coincide with a broader industry shift. As AI models proliferate, inference is becoming the dominant operational cost—not training. Microsoft’s framing of ND MI455X v7 as an inference-first series aligns with that trend: companies are serving millions of queries against open-source and proprietary models, and they need cost-efficient compute that can handle volume. AMD’s pitch has long been that its Instinct GPUs offer competitive price/performance for inference, even if Nvidia still leads on raw training throughput.
What to do now
- For capacity planners: Watch the Azure updates page and AMD’s investor communications for specific availability windows. The “H2 2026” label gives you at least a year to factor these series into your next budget cycle. If you’re already running on Azure’s existing HBv3, HBv4, or HX instances, start modeling what an HXv2 migration could look like for your EDA jobs.
- For AI/ML teams: If inference cost is a growing line item, begin evaluating AMD’s ROCm software stack in your CI/CD pipelines. You can test AMD-based VMs in Azure today on existing NCas_T4_v3 or NVads_A10_v5 series, but the Helios-based instances will represent a significant generational leap. Prepare your container images and model-serving frameworks to work smoothly with ROCm.
- For data engineers: Keep an eye on how Azure positions HDv2 for data preparation. If the instance can replace multi-node Spark or Dask clusters with a single massive VM, it could simplify workload orchestration dramatically. Start designing experiments that leverage large memory and local NVMe for in-memory joins and pre-processing.
- No immediate action: There is no setting to toggle, no Windows update to apply, and no license to buy right now. This is a cloud roadmap announcement. Mark your calendar for mid-2026 and check Microsoft’s regional availability when previews open.
Outlook: what to watch next
The biggest open question is performance. AMD has released Horizon AI benchmarks for its Instinct GPUs, but real-world throughput, latency, and power efficiency inside Azure will determine whether ND MI455X v7 becomes a genuine Nvidia competitor or a niche option. Microsoft typically publishes VM-level benchmark data closer to preview, so watch for that.
Pricing will be the other decider. If Azure prices these AMD instances below comparable Nvidia GPU VMs, they could attract cost-sensitive inference workloads in droves. Conversely, if pricing parity persists, adoption may depend on developer familiarity with ROCm, which still trails CUDA in mindshare and ecosystem maturity.
Finally, the Helios platform’s integration of Pensando DPUs for networking hints at potential performance gains in multi-node inference or distributed model serving. If those gains materialize, expect Microsoft to highlight them when the VMs approach general availability. For now, the message is clear: AMD’s next-generation data-center hardware is coming to Azure, and Windows admins who plan cloud capacity should start penciling it into their roadmaps.