Microsoft will expand AMD’s presence in Azure with the deployment of Helios, a comprehensive data center platform that integrates Instinct MI455X accelerators, sixth-generation EPYC processors, Pensando networking, and ROCm software. The first systems are expected to start arriving in the second half of 2026 and will primarily be used to run advanced artificial intelligence models.
The key points of the Microsoft and AMD agreement in 20 seconds
- Azure will deploy Helios racks with 72 AMD Instinct MI455X accelerators.
- Microsoft will mainly use them for inference of advanced models.
- New virtual machines with EPYC Venice processors will also arrive.
- AMD Pensando will strengthen internal networks and various Azure services.
- The agreement broadens Microsoft’s options beyond NVIDIA dominance.
The collaboration goes beyond simply adding a new generation of graphics processing units (GPUs) to the Azure catalog. Microsoft and AMD are coordinating processors, accelerators, networking, systems, and software to deliver an infrastructure designed as a cohesive whole—an increasingly common approach given the size and power demands of new AI platforms.
Microsoft has confirmed three upcoming service families: ND MI455X v7 virtual machines for inference, HDv2 for AI agent-based workloads and data processing, and HXv2 for electronic design automation (EDA). The latter two will utilize AMD EPYC Venice processors based on Zen 6 architecture.
Helios turns the entire rack into an AI platform
Helios is AMD’s attempt to compete in the complete AI system market. Instead of merely selling accelerators, the company offers a rack design where computing, memory, networking, and software are all optimized to work together.
Each rack will integrate 72 Instinct MI455X accelerators, EPYC Venice processors, and Pensando Vulcano network adapters. Internal communication will use UALink, an open technology developed to connect accelerators and enable low-latency sharing of information.
AMD estimates each rack can deliver up to 2.9 exaflops of performance in FP4 calculations—a low-precision format used in certain AI workloads. This figure is not a direct comparison with supercomputers or conventional systems, as it depends on the type of operation and conditions, but it demonstrates the computation density AMD aims to concentrate in a single installation.
The architecture is designed for both training and inference, although Microsoft emphasizes its use for running advanced models, Azure AI services, and customer applications.
Inference is the phase where a trained model responds to queries, generates images, summarizes documents, or executes tasks. Its importance in data centers is growing as companies move AI from experimental stages to continuous-use applications.
This workload doesn’t always require the same infrastructure as training. Response times, query cost, user capacity, and energy consumption may be more critical than maximum raw performance over a few weeks.
Microsoft has not disclosed how many racks it will buy, which Azure regions will receive them first, or when customers can subscribe. AMD only states that Helios will begin shipping in the second half of 2026.
AMD gains ground in an Azure environment mixing silicon types
The agreement strengthens AMD’s position as a provider of infrastructure for Microsoft, but Helios is not the only AI platform in Azure.
Microsoft also uses NVIDIA accelerators and develops own chips, such as the Maia family. Its strategy involves combining multiple architectures to handle different workloads and reduce dependence on a single vendor.
For AMD, entering Azure with a full rack is more significant than just selling additional GPUs. Cloud providers now consider internal network performance, power consumption, cooling, memory availability, maintenance, and software maturity alongside hardware.
NVIDIA’s dominant position was largely built on this kind of integration: combining GPUs, NVLink interconnects, networking adapters, reference servers, and CUDA. AMD needs to demonstrate that ROCm and its open platforms can serve as a viable alternative in large-scale installations.
Microsoft’s deployment can help amplify this effort by testing Helios under real-world conditions, with large models and services for clients. It also provides a commercial reference that AMD can leverage with other companies and cloud providers.
Azure Foundry Managed Compute will offer access to part of this infrastructure so organizations can deploy and scale AI workloads without managing all physical components directly. Specific availability, pricing, and configurations have yet to be announced.
EPYC Venice and Pensando expand the partnership beyond GPUs
The collaboration includes two new lines of virtual machines based on AMD EPYC Venice. HDv2 targets data pipelines, information preparation, and AI agent workloads, while HXv2 will focus on semiconductor design and other engineering tasks.
Electronics design automation (EDA) demands large amounts of memory, computing capacity, and fast connections between servers. It’s a specialized market where cloud providers compete to attract chip manufacturers and companies needing to run simulations without building their own data centers.
Adding Venice also shows that AI growth isn’t driven solely by accelerators. CPUs prepare data, coordinate GPUs, manage storage and networking, run databases, and handle parts of applications unsuitable for accelerators.
Microsoft and AMD will further expand the use of Pensando data processing units (DPUs). These devices offload tasks such as packet processing, security, storage, and connection management from the CPU.
Azure already uses Pensando in some parts of its infrastructure. The next phase will integrate these technologies with Azure Boost, Microsoft’s system to migrate network and storage operations from virtual machines to specialized hardware.
The goal is to free CPU resources for applications and maintain stable performance when thousands of servers exchange data simultaneously. In an AI cluster, insufficient networking can cause expensive accelerators to wait for data instead of computing.
Another option for AI infrastructure
The announcement does not include revenue forecasts or the financial value of the contract. It also doesn’t guarantee AMD will quickly reduce NVIDIA’s lead, which still maintains broad adoption and a well-established software environment.
Helios confirms that the competition is shifting from individual chips to complete racks. Designing a fast GPU is no longer enough: vendors must figure out how to connect multiple accelerators, move data efficiently, cool the system, and provide tools to run models with minimal changes.
Microsoft gains another avenue to expand Azure and negotiate supply deals in a high-demand market. Meanwhile, AMD secures a large-scale customer to validate its GPUs, CPUs, networking, and software together.
The ultimate test will be when Azure publishes its commercial configurations and initial performance, availability, and cost data. Until then, many of the capabilities announced remain targets rather than proven results in live services.
Frequently Asked Questions
What is AMD Helios?
AMD Helios is a rack-scale AI platform that combines 72 Instinct MI455X accelerators, EPYC Venice processors, Pensando networks, and ROCm software environment.
When will AMD Helios be available on Microsoft Azure?
AMD will start delivering the first systems in the second half of 2026. Microsoft has not yet announced the commercial launch date or which Azure regions will initially offer them.
How will Microsoft use the MI455X accelerators?
Azure plans to mainly use them for inference of advanced models, Microsoft AI services, and customer applications. The platform is also designed for training workloads.
Will AMD Helios replace NVIDIA GPUs in Azure?
No substitution has been announced. Microsoft is expanding its hardware options with AMD while continuing to support other platforms, including NVIDIA accelerators and its own chips.
