On July 20, 2026, Microsoft announced an expansion of Azure AI and HPC infrastructure that brings AMD’s Helios AI platform and next-generation EPYC data-center processors to three upcoming Azure offerings. The central signal is that agent workloads are becoming a full infrastructure problem, not only a GPU or model problem.
The first offering is Azure HDv2 for data processing, search, reinforcement learning, and large-scale agent coordination. Microsoft describes configurations with nearly 500 physical sixth-generation AMD EPYC CPU cores, 4TB of RAM, 32TB of local NVMe storage, and 400Gb Azure Boost networking. These are vendor-published specifications; availability, regional rollout, and pricing still need to be checked against Azure’s live catalog.
The second is Azure HXv2 for electronic design automation and other technical-computing workloads. Microsoft says HXv2 will provide 176 sixth-generation EPYC CPU cores, clock speeds above 5GHz, 50% more addressable cache per core, and VM sizes with nearly 2TB or 4TB of RAM. It also includes 800Gb InfiniBand for large MPI simulations and distributed-memory applications.
The third is ND MI455X v7, powered by AMD’s Helios rack-scale solution and aimed at the reasoning, search, and agentic workloads behind modern AI services. Taken together, the offerings show Azure separating data preparation, chip design, and production inference into distinct infrastructure paths rather than treating every AI workload as the same job.
That separation matters for agents. An agent needs more than an accelerator: CPUs prepare data and context, manage tools, schedule retries, move results, and coordinate multiple steps. As an agent becomes a long-running service rather than a single response, data movement, networking, memory, and scheduling can become bottlenecks alongside model inference.
Microsoft is therefore emphasizing an open, heterogeneous platform. Organizations can choose different CPU, GPU, memory, and networking configurations for different workflows instead of putting every task on the same accelerator. The cost model becomes more detailed as a result: data preparation, search, coordination, and inference need to be measured separately to show where a task actually consumes resources.
The announcement remains a vendor description of upcoming offerings, not an independent performance test or proof that every region has launched. Buyers still need to validate end-to-end latency, retry rates, network and storage costs, and the deployment and governance overhead of moving an agent across different hardware paths.
