AMD announces Helion rack design for AI agent hardware

What happened

AMD has unveiled Helion, a new reference design for a server rack intended to run AI agents. An AI agent is a type of software that can perform tasks autonomously. The Helion design follows the Open Compute Project (OCP) ORv4 specification, which means it is slightly taller and wider than a standard IT rack [2].

AMD is offering two main ways to set up the Helion rack:

• A CPU-only configuration: This version is aimed at hyperscalers (large-scale cloud providers) and can hold up to 128 EPYC 9996 CPUs. This setup provides up to 32,768 Zen 6 ‘Venice’ cores [2]. This configuration is designed for agent execution and connects to separate systems for model inference [3].
• An enterprise configuration: This all-in-one version combines agent execution, GPU inference, and other workloads in a single rack [3]. This hybrid setup can include up to 16 Instinct MI350P GPU accelerators and up to 102 Tbps of networking bandwidth, though adding GPUs reduces the total CPU capacity [2].

The hardware can be deployed using either air-cooling or direct liquid-cooling [2]. The design supports 44 OpenU trays, which measure 21mm in width and 48mm in height [2].

AMD provides different types of blades (individual server units) for these racks [3]:

• Dense blades: Use EPYC 9006 chips rated at 400 W [3].
• Balanced blades: Use 256-core processors rated at 600 W [3].
• High-performance blades: Use a 96-core version of the chip clocked at 5 GHz [3].
• GPU inference blades: Feature eight Instinct MI350P GPUs in a 5U server [3].
• General-purpose blades: Intended for storage, databases, or technical compute workloads [3].

While the Helios rack was built for large-scale AI training, Helion is described as a more flexible solution [2]. AMD plans to contribute the Helion specifications to the OCP community [3].

Who is affected

• Hyperscalers: These large providers can use the CPU-only configuration for agent execution [2].
• Enterprises: Businesses can use the hybrid configuration for integrated data management, networking, and AI usage [2].

What to do next

• Partner-built Helion configurations are scheduled for November 2026 [3].
• The Helion rack is expected to be available starting in the first half of 2027 [2].