The news
On July 20, 2026, Advanced Micro Devices (AMD) and Microsoft (MSFT) said Microsoft will deploy AMD Helios on Azure, putting AMD's rack-scale AI system to work serving frontier AI models for Microsoft's own services and its cloud customers.
Helios packages AMD Instinct MI455X graphics processing units (GPUs), 6th Gen AMD EPYC "Venice" central processing units (CPUs), AMD Pensando networking and AMD's ROCm software into one integrated rack, according to AMD. The company said it will begin shipping Helios to customers, including Microsoft, in the second half of 2026.
In a blog post the same day, Scott Guthrie, Microsoft's executive vice president of Cloud + AI, said Helios will power a new Azure instance, ND MI455X v7, which Microsoft described as designed for reasoning, search and agentic AI workloads. AMD said enterprise customers can run production AI workloads on AMD-powered infrastructure through Azure Foundry Managed Compute.
The agreement also covers CPUs. Azure is adding two virtual machine (VM) series on EPYC Venice: HDv2 for agentic AI and data pipelines, and HXv2 for semiconductor design. Microsoft said HDv2 offers nearly 500 physical CPU cores, 4 terabytes of RAM, 32 terabytes of local NVMe storage and 400 Gb Azure Boost networking, while HXv2 offers 176 cores running above 5 GHz and 800 Gb InfiniBand.
On networking, AMD said Microsoft is broadening its use of AMD Pensando data processing units (DPUs), chips that take networking work off a server's main processors, and that the companies are integrating AMD silicon with Azure Boost to lift networking performance across Azure's fleet. Microsoft CEO Satya Nadella said the collaboration gives customers "the performance, scale and choice they need."
Neither company disclosed how many Helios racks Microsoft will deploy or what the agreement is worth, a gap WinBuzzer also noted in its July 23 coverage.
The numbers
- Helios shipments to customers begin
- Second half of 2026 (AMD)
- New Azure VM series on EPYC Venice
- 2 (HDv2, HXv2)
- HDv2 physical CPU cores
- Nearly 500
- HDv2 memory and local storage
- 4 TB RAM, 32 TB NVMe
- HXv2 CPU cores
- 176, above 5 GHz
- Deployment scale and deal value
- Not disclosed
Why CEOs should care
For technology buyers, the practical change is another large pool of AI inference capacity inside a cloud many enterprises already use. Teams serving models on Azure should ask their Microsoft account team when ND MI455X v7 capacity will be offered, in which regions, and at what price compared with Nvidia-based instances. They should also test whether their models and serving software run on ROCm without major rework, because software portability decides whether a second chip supplier is usable in practice.
For CFOs negotiating multiyear Azure commitments, competing accelerator families inside one cloud can create pricing leverage, but only if workloads can move. Ask whether committed spend can shift between GPU families as new instances launch, and avoid locking long reservations to hardware whose pricing Microsoft has not yet published. Chip and hardware design firms should look at HXv2, which Microsoft aimed at register-transfer level (RTL) simulation and other engineering workloads.
For CISOs and boards, new silicon in the network path deserves the same scrutiny as any new cloud layer. Security teams can ask Microsoft how Pensando DPUs and the Azure Boost integration affect tenant isolation and firmware update practices. Boards should treat the announcement as a signal of direction rather than a capacity guarantee, since no deployment volumes were disclosed.
The bigger picture
The announcement fits a broader effort by cloud providers and AI developers to spread AI compute across more than one chip supplier. WinBuzzer reported that a Futurum Group estimate would put Nvidia's (NVDA) GPU market share at more than 95% and AMD's at about 4.5%, and quoted Counterpoint Research analyst Neil Shah, who attributes Nvidia's ecosystem lead to its CUDA software. AMD pitches Helios as an integrated rack of compute, networking and software, which lets it compete for whole-system AI deployments rather than individual chip sales.
What happened next
On July 23, 2026, at its Advancing AI 2026 event, AMD said Helios was in production and named Microsoft among the companies deploying it at scale. AMD said each Helios rack holds 72 MI455X GPUs and 18 EPYC CPUs. On August 4, 2026, AMD's second-quarter results listed the Azure Helios deployment, the new EPYC VM series and the Pensando expansion among its highlights, and CEO Lisa Su said Helios was beginning to ramp.
Microsoft's July 20 post did not give availability dates, regions or pricing for ND MI455X v7, HDv2 or HXv2. Those are the details to watch as AMD's second-half 2026 shipments proceed.
