AMD has introduced AMD Helion, an open blueprint for rack-scale infrastructure built to support agentic artificial intelligence (AI) workloads. The platform gives customers flexibility to configure computing resources for AI agents, from large-scale agent execution to model inference in enterprise environments.
Based on a specification design from the Open Compute Project (OCP), AMD Helion combines hardware configurations with open software. It also includes an integrated token router that directs agents’ model calls across frontier and open-source models.
The announcement reflects the different computing demands of agentic AI. Unlike workloads limited to a single model interaction, AI agents can execute code, perform tasks, interact with models, and manage multistep workflows. These activities can require different balances of CPU performance, core density, GPU acceleration, and power consumption.
“Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements,” said Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD.
He added that AMD Helion brings AMD EPYC performance, open software, and flexible rack designs together to help partners build systems around customer requirements and make efficient use of available power.
Two rack configurations address different workloads
AMD Helion supports multiple configurations, including liquid-cooled and air-cooled systems. The company outlined two examples aimed at different deployment requirements.
CPU-only racks for hyperscale deployments
The first configuration uses sixth-generation AMD EPYC processors in a CPU-only rack that can support up to 32,768 CPU cores.
The system is intended for hyperscale customers that need dedicated computing capacity to execute AI agents. It connects to separate infrastructure responsible for model inference, allowing agent execution and model processing to run on distinct systems.
This approach gives large-scale deployments the option to allocate CPU resources specifically to agent workloads rather than combining every computing function in a single rack.
Integrated CPU and GPU racks for enterprises
The second configuration combines AMD EPYC CPU nodes, AMD Instinct MI350P GPU nodes, and general-purpose computing resources on one platform.
The all-in-one design targets enterprise deployments that require both agent execution and GPU-accelerated model inference. Bringing these resources together gives organisations another option for deploying agentic AI within their computing environments.
The two configurations illustrate AMD’s approach to infrastructure planning: workloads can use different combinations of processors and accelerators instead of relying on a single rack design for every deployment.
Open specifications and existing infrastructure
AMD Helion is intended to give customers flexibility in configuring CPU and GPU resources according to workload performance, power, and cost requirements. AMD also says the platform can be deployed within existing information technology infrastructure, allowing organisations to build AI capabilities around investments they have already made.
The company plans to contribute the Helion specification to the OCP community for further development and adoption. The contribution is intended to support a wider ecosystem of partners working on rack-scale infrastructure for agentic AI.
However, the announcement does not establish that all configurations are commercially available today. AMD said Helion rack configurations are under development with ecosystem partners, with availability planned for November 2026.
Partner development will shape the next phase
AMD Helion extends the company’s open-platform approach to infrastructure built for AI agents. Its proposed configurations separate large-scale agent execution from model inference or combine CPU and GPU resources in an integrated system.
The next milestone is the planned availability of partner-developed configurations in November 2026. The broader direction will depend on how partners adopt the specification and how customers apply the available designs to their AI workloads.
For now, AMD Helion outlines an infrastructure framework rather than a single fixed system. Its central proposition is flexibility: giving hyperscale operators and enterprises different ways to allocate computing resources as their agentic AI deployments grow.