AMD has introduced AMD Helion, an open infrastructure architecture for agentic artificial intelligence (AI) that allows server racks to be configured according to the needs of agent execution, model inference and enterprise computing. The proposal includes systems with up to 32,768 CPU cores based on sixth-generation AMD EPYC processors, as well as configurations that combine CPUs and AMD Instinct MI350P GPUs. The company expects the systems developed with its partners to be available in November 2026.
AMD Helion in 30 seconds
- AMD Helion is an open rack design for agentic AI workloads, based on Open Compute Project (OCP) specifications.
- The CPU-focused configuration can reach up to 32,768 cores with sixth-generation AMD EPYC processors.
- The integrated alternative will combine EPYC nodes, AMD Instinct MI350P accelerators and general-purpose computing.
- The designs will support air or liquid cooling and can be adapted to each facility’s power, performance and density requirements.
- AMD plans to contribute the specification to OCP and expects the configurations developed with its partners to reach the market in November 2026.
The announcement reflects a shift in infrastructure needs as companies move from using AI models to answer questions to deploying agents capable of carrying out tasks, calling tools and coordinating workflows. These applications need capacity to run the agent itself, communicate with models and process auxiliary operations. Not all of them require the same mix of CPU and GPU.
AMD wants to offer a hardware foundation that allows these functions to be separated or combined within the same rack. The company presents Helion as an adaptable reference design, not as a single configuration that every organization must install. The availability announced corresponds to systems that are still in development with industry partners.
Two configurations for different AI workloads
AMD’s proposal includes two main approaches. The first concentrates compute capacity in processors, while the second combines CPUs and GPUs to cover more functions within the same platform.
CPU racks for running agents at scale
The configuration aimed at large cloud service providers uses sixth-generation AMD EPYC processors and is designed to reach up to 32,768 CPU cores per rack.
The goal is to dedicate compute capacity to running agents and connect it to a separate infrastructure that handles model inference. This separation allows for a design where processors manage agent logic, tools and execution flows, while other systems handle requests to the AI models.
The figure of 32,768 cores describes the maximum announced capacity for that reference configuration. AMD doesn’t detail in the announcement all the components needed to reach that maximum, nor does it publish comparable results for performance per agent, power consumption per task or operating cost.
That’s why the core count alone isn’t enough to determine how many agents a rack can run under real conditions. Actual capacity will depend on the workload, available memory, communications, concurrency and the resources each agent needs. AMD has already pitched a similar agentic AI framing for its EPYC Venice processors, so Helion builds on ground AMD has been preparing for months.
An integrated CPU and GPU rack
The second option combines AMD EPYC nodes, AMD Instinct MI350P accelerators and general-purpose computing resources. It’s designed for organizations that prefer to concentrate different AI and enterprise processing functions on the same platform.
In this scenario, CPUs can handle agent execution and auxiliary tasks, while GPUs are dedicated to workloads that require acceleration, such as inference for compatible models. The exact distribution will depend on the software and the requirements of each deployment.
Flexibility also extends to cooling. AMD includes air-cooled and liquid-cooled versions, a decision that affects equipment density, data center design and power and heat-dissipation needs.
The company says partners will be able to adjust the ratio of CPUs to GPUs for each customer. However, the announcement doesn’t provide pricing, per-rack power figures or performance comparisons that would allow calculating the total cost of ownership against other platforms.
AMD wants to bring Helion to the Open Compute Project
AMD Helion is based on a reference design tied to Open Compute Project (OCP) specifications, an initiative that promotes open designs for data center infrastructure. AMD says it plans to contribute the Helion specification to this community to support its development and adoption.
This approach allows manufacturers and integrators to adapt the architecture to their products and to each facility’s requirements. AMD has announced support from companies including Cisco, Dell Technologies, Hon Hai Technology Group, Hewlett Packard Enterprise (HPE), Lenovo, QCT, Supermicro and Wiwynn.
These partners’ involvement points to joint work on the design, integration and commercialization of systems. It doesn’t by itself mean that all of them will sell the same configurations or that all the products are already available.
Cisco highlights the importance of integrating compute, networking, management and security to run AI workloads. Dell says its collaboration aims to offer a validated path for deploying agent infrastructure at scale, while HPE focuses on open systems and cooling options.
Supermicro, for its part, says it will work to bring the design to its data center infrastructure solutions. Other manufacturers highlight the ability to adapt the equipment to the needs of cloud providers and enterprise customers — a path already visible in HPE’s own ProLiant Gen13 servers built around sixth-generation EPYC chips for enterprise AI.
AMD expects the Helion configurations developed with its partners to be available in November 2026. Until they reach the market, the final commercial models, pricing, specific configurations and real-world performance results remain unknown.
What changes with agentic AI
The architecture responds to a different need than the one driving large model-training systems. Agents can spend much of their time executing code, querying services, using tools or coordinating tasks — activities that don’t always require a high-capacity GPU.
For those workloads, an infrastructure with enough CPU cores can be useful for keeping numerous processes active and managing calls to models hosted on other systems. By contrast, deployments that combine agent execution and inference can benefit from combining CPUs and GPUs in the same rack.
AMD’s proposal tries to cover both situations through an adaptable architecture. Its usefulness will depend on how tasks are distributed, the cost of the infrastructure, and the equipment’s ability to maintain efficient operation under sustained loads.
The design’s openness can also make it easier for manufacturers and integrators who want to customize components or cooling systems. Even so, an open design doesn’t automatically guarantee full interoperability, a lower price or a performance advantage over closed alternatives. Those aspects will have to be evaluated with the final products.
Frequently asked questions
What is AMD Helion?
AMD Helion is an open rack architecture designed to run agentic AI workloads. It supports CPU-focused configurations as well as systems that combine EPYC processors with Instinct MI350P accelerators.
How many cores can an AMD Helion rack have?
AMD is announcing a configuration based on sixth-generation EPYC processors that can reach up to 32,768 CPU cores per rack. The figure corresponds to the planned design and isn’t a direct measure of how many agents it will be able to run.
When will AMD Helion be available?
AMD expects the configurations developed with its partners to be available in November 2026. The company has stated that the systems are still in development.
Which manufacturers are collaborating on AMD Helion?
The partners mentioned by AMD include Cisco, Dell Technologies, Hon Hai Technology Group, HPE, Lenovo, QCT, Supermicro and Wiwynn. Their involvement covers the development and adaptation of infrastructure for different workloads.
