Microsoft and AMD just made their AI infrastructure relationship a lot more “locked in”, kinda. On July 20, 2026, AMD announced through its newsroom and an official post on X that an expanded strategic partnership, that essentially puts AMD’s newest GPUs, CPUs, DPUs, and that whole software stack right in the middle of Microsoft Azure’s AI infrastructure buildout. If you’ve been watching the AI hardware race, this is one of the clearest signals yet that AMD Helios isn’t just some roadmap doodle anymore. It’s getting dragged into the real cloud, like, infrastructure-level stuff.

AMD Helios is the company’s rack-scale AI platform, and it’s the headline part of this Microsoft Azure AI infrastructure expansion. Instead of moving “just chips”, AMD is shipping a whole integrated rack: AMD Instinct MI455X GPUs, EPYC “Venice” CPUs, Pensando networking, and the ROCm open software stack, all made to play together for big training plus inference, and that too, at scale.

Based on hardware outlets that have dug into the platform, one Helios rack includes 18 compute trays, each one with four MI455X accelerators and one Venice CPU. That totals to  72 GPUs, which is more than 4,600 EPYC cores (roughly 31 TB of HBM4 memory), and around 1.4 PB/s of aggregate memory bandwidth per rack. That, in turn, is said to translate to close to 2.9 FP4 exaflops of AI inference throughput. The MI455X chip itself reportedly carries 432 GB of HBM4 across four dies, with about 19.6 TB/s bandwidth per chip.

Also pretty important is how the rack talks to itself. Helios uses UALink-over-Ethernet, an open interconnect standard with support from a consortium that includes AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft. So, it’s being positioned as a direct alternative to Nvidia’s more proprietary NVLink (Tom’s Hardware). AMD’s whole angle here is simple-ish: give hyperscalers flexibility, instead of tightening the screws into vendor lock-in.

The CPU side is the other half of this Microsoft Azure AI infrastructure story. AMD EPYC “Venice” is the next-generation server processor from AMD, built on Zen 6 and made on TSMC’s N2 process node. It’s the first EPYC generation using the Gate-All-Around nanosheet transistors, not FinFET designs that have been around since about 2011. Venice shifts to eight Core Complex Dies, 32 cores each, pushing the flagship to 256 cores (which, for reference, is approximately a 33% jump versus the prior Zen 5-based EPYC “Turin”).

It also adopts a newer SP7 socket with 16 memory channels, up to 1.6 TB/s of memory bandwidth, plus PCIe Gen 6 support for quicker CPU-to-GPU data movement (Tech Times). 

AMD CEO Dr. Lisa Su has framed Venice like the “keep the GPUs fed” CPU at rack scale. She notes the company doubled memory and GPU bandwidth from the previous generation, so Venice can feed the MI455X accelerators without turning into the bottleneck. AMD expects first Venice-based systems to start shipping to customers in Q3 2026.

Past the Helios racks, Microsoft is rolling EPYC Venice into broader Azure usage through two new virtual machine series: Azure HDv2, which is targeted at agentic AI and data pipeline workloads.

Azure HXv2, which is built for semiconductor design and electronic design automation (EDA), reportedly combining Venice’s core density with 800 Gb InfiniBand networking. The pitch here is support for large-scale, MPI-based simulation workloads (QSOL IT).

Microsoft is also expected to bring out dedicated MI455X-based inference VMs, which would give enterprise customers a more direct lane to run production AI workloads on Azure Foundry Managed Compute, using AMD’s newest accelerators.

But AI infrastructure is actually how fast data moves between chips and racks and services. Here’s where AMD Pensando DPUs enter the stage. Microsoft has already deployed Pensando hardware across parts of Azure, and now this expanded partnership goes further by folding Pensando tech with Azure Boost, which is Microsoft’s system for offloading networking and storage processing away from host CPUs.

The goal is pretty straightforward, at least in intent. It is to improve networking performance, make it more efficient, and handle connection processing better across Azure’s fleet as AI traffic grows.

This announcement comes at a time when the AI infrastructure market is basically a two-horse situation between Nvidia and AMD. Hyperscalers are hungry for a credible second source of high-performance AI silicon, and not to have just one name on the shortlist. So, this is good news!

Microsoft’s increased commitment to AMD Helios builds on AMD’s earlier momentum, including wins with Meta and AMD’s MI400-series accelerators. Analysts at S&P Global Market Intelligence have projected around $7.2 billion in MI400-series revenue for AMD in 2026 alone.

There’s also a supply-side snag to watch out for. The HBM4 memory Helios racks depend on are reportedly fully allocated to hyperscale customers well into the future. That could affect how fast AMD scales Helios shipments beyond anchor customers like Microsoft.

Dr. Lisa Su, AMD’s Chair and CEO, said the companies are stretching their partnership across the full stack of AMD’s AI solutions on Azure. Meanwhile, Microsoft Chairman and CEO Satya Nadella pointed to customer demand for AI infrastructure that stretches across training, inference, data preparation, and reinforcement learning. In other words, the drive is bringing Helios into the Azure lineup because customers want the whole spectrum.

AMD says it will start shipping Helios systems to customers, including Microsoft, in the second half of 2026. EPYC Venice-based systems are expected to ship starting in Q3 2026. For a deeper technical dive, AMD’s own product pages cover the Instinct accelerator lineup, EPYC server processors, and Pensando networking solutions in more detail.

For now, the takeaway is pretty simple. That being, Microsoft Azure’s AI infrastructure is about to get a lot more AMD inside it, and that’s a meaningful data point for how the AI hardware market could keep shifting through the rest of 2026.