Key points
- Helion offers CPU-only and mixed CPU/GPU rack designs for AI agents.
- Configurations support up to 32,768 Zen 6 cores and 102 Tbps networking.
- Partners are expected to deliver systems based on the design in November.
The short version
- AMD revealed the Helion rack reference design at the Open Compute Project Summit 2026.
- The design supports CPU-only hyperscale setups or enterprise mixes of CPUs and GPUs.
- Systems based on this blueprint are scheduled to arrive from partners in November.
The details
AMD introduced the Helion rack reference design at the Open Compute Project Summit 2026 in San Jose, California, according to The Register. This new blueprint is intended for agentic AI infrastructure, distinct from the company’s earlier Helios system, which focused on frontier-scale training and large-scale inference. The Helion design allows enterprises to combine agent execution and inference within a single rack, providing a consolidated approach to deploying AI agents.
The architecture supports two primary deployment paths: a CPU-only hyperscale configuration and an enterprise configuration that mixes CPU and GPU blades. The CPU-only option provides dedicated capacity for running agents while connecting to separate model inference infrastructure. In contrast, the enterprise setup integrates agent execution, GPU inference, and supporting workloads into one physical rack, simplifying the hardware footprint for organizations that need both capabilities close together.
AMD advertises that the Helion design can deliver up to 32,768 Zen 6 CPU cores. These cores are provided through compute nodes equipped with 256-core versions of the Venice EPYC processor. The system can also include up to 16 Instinct MI350P GPU accelerators. Networking bandwidth within the rack reaches up to 102 Tbps, ensuring high-speed data transfer between the numerous processing units and external resources.
The design accommodates various blade types to suit different performance needs. Dense deployments utilize EPYC 9006 chips rated at 400 watts, while performance blades feature a 96-core version of the chip clocked at 5 GHz. Balanced blades use 256-core processors rated at 600 watts. Additionally, the rack includes GPU inference blades with eight Instinct MI350P GPUs in a 5U server, as well as general-purpose blades for database, storage, or technical compute tasks.
The context
The Helion design complements AMD’s existing Helios system, which was launched earlier this year for different AI workloads. While Helios targets massive training jobs, Helion focuses on the execution of AI agents, a growing area of interest for businesses automating complex tasks. The naming similarity between Helion and Helios may cause some confusion, but the functional differences are clear: one is for training, the other for agent execution and inference.
Partners are expected to begin delivering systems based on the Helion reference design in November. These systems will offer both liquid-cooled and air-cooled versions, allowing organizations to choose the thermal management solution that best fits their data center infrastructure. The availability of these configurations in November marks a significant step in AMD’s strategy to expand its rack-scale AI lineup beyond just training hardware.
What to watch
Organizations should monitor the release of partner-built systems in November to evaluate hardware options. IT leaders should assess whether their AI agent workloads require the integrated enterprise configuration or the separated CPU-only setup. Network bandwidth requirements should be reviewed against the 102 Tbps capacity offered by the design. Thermal management strategies should also be updated to accommodate either liquid or air cooling solutions.
Background: Agentic AI
Agentic AI refers to systems that plan and execute complex sequences of actions to achieve a goal, rather than just answering questions. Unlike standard chatbots, these agents can interact with other software, browse the web, and run commands. This autonomy increases efficiency but also expands the potential for security breaches if the agent’s permissions are not strictly limited.
Read the full guide: Agentic AI Explained: How Autonomous Agents Change Security
What to do and how to stay safe: AMD
- Review current AI workload requirements to determine if agent execution and inference can be combined in a single rack.
- Evaluate data center cooling infrastructure to ensure compatibility with liquid or air-cooled server designs.
- Assess network bandwidth needs to support high-speed communication between CPU and GPU blades.
- Plan for hardware procurement in November when partner systems become available.
Step-by-step guide: Shadow AI Mistakes That Put Corporate Data at Risk
General security guidance from the TechSec Headlines newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the difference between Helion and Helios?
Helios is for frontier-scale AI training and large-scale inference, while Helion is designed for agentic AI infrastructure, combining agent execution and inference.
When will systems based on the Helion design be available?
Partners are expected to deliver CPU-only and mixed CPU/GPU configurations in November.
What processors are used in the Helion rack design?
The design uses 256-core versions of the Venice EPYC processor and can include Instinct MI350P GPU accelerators.
