AMD debuts Helion reference design for agentic AI servers with up to 32,768 Zen 6 ‘Venice’ cores

Helion is AMD’s blueprint for a large data-center rack. It defines how processors, GPUs, trays, cooling, and networking can fit together. The design follows OCP’s ORv4 specification and supports 44 OpenU trays. This matters because customers can receive a common design from several manufacturers instead of relying on one proprietary rack.

AMD will not sell Helion racks directly. Instead, AMD contributes the reference design to the Open Compute Project. Original equipment manufacturers and original design manufacturers are expected to turn that blueprint into products. AMD had not named supporting partners during the press pre-briefing, although Lisa Su was expected to announce partners at the OCP Global Summit.

Helion complements AMD’s Helios rack, which targets AI training and inference. Helion is more flexible and mainly CPU-focused, but it can also support GPUs and networking. The updated article says racks should become available in November 2026, giving manufacturers time to develop their own implementations.

In its densest CPU-only configuration, Helion can hold 128 EPYC 9996 processors. Each processor uses a 256-core Zen 6 design called Venice. Multiplying those figures gives 32,768 Zen 6 cores in one rack. That makes the configuration especially suitable for workloads that can divide work among many users or tasks.

The rack is designed around 44 OpenU trays, each 21 inches wide and 48mm tall. AMD says Helion supports both air cooling and direct liquid cooling. The CPU-only arrangement sacrifices space that could otherwise hold accelerators, but it maximizes general-purpose processing capacity.

The figure describes available hardware, not a guarantee that every application runs faster by the same amount. Software must be able to use many cores efficiently. Helion’s intended use includes agentic AI sandboxes, where many separate activities can run at once. AMD also offers a hybrid configuration with up to 16 MI350P GPUs and 102 Tbps of networking.

A CPU-only Helion rack could handle software development tasks and other work assigned to individual agent sandboxes. The article describes these servers as a possible foundation for agentic AI systems. Their large core count lets a host divide computing capacity among many users, agents, or background jobs.

For example, an agent might receive a slice of the host system while building software. The rack would provide the general-purpose CPU capacity for that sandbox. A separate inference server would act as the backbone for heavy AI processing. In the article’s examples, this resembles systems associated with Meta’s Muse and Axiom Dots.

Heavy inference usually happens elsewhere because the CPU-only rack is not optimized as the complete AI accelerator system. Helion can instead be configured with MI350P GPUs, but adding them reduces CPU capacity. This split lets hyperscalers provide many CPU sandboxes while sending demanding model inference to systems designed for that workload.

Agentic AI refers here to systems in which personal AI agents perform tasks for users rather than only returning one-off answers. The article links this model to Meta’s Muse and Axiom Dots. Each user gets a sandbox, meaning an allocated slice of computing capacity where the agent can work.

A sandbox could let an agent build software or perform another CPU-based task without taking over the entire server. The host divides its resources among many users. At the same time, an inference server provides the backbone for generating the AI responses and handling heavier model processing. This creates a two-part system: local task execution and shared inference.

Large servers may need many sandboxes because many people can use agents simultaneously. That demand helps explain Helion’s emphasis on dense CPU capacity. The article presents this as a model hyperscalers may pursue as agentic AI grows, while enterprises could choose Helion’s hybrid form for data management, networking, and AI in one rack.

AMD is offering two configurations because customers have different infrastructure needs. A CPU-only rack maximizes EPYC capacity for agent sandboxes and other general-purpose workloads. A hybrid rack trades away some CPU capacity for GPUs and networking. This gives buyers a choice instead of forcing every deployment into one balance of resources.

The hybrid design can include up to 16 MI350P GPUs and 102 Tbps of networking. AMD targets this version at enterprises seeking an all-in-one rack for data management, networking, and AI usage. The CPU-only version is more closely associated with hyperscalers building capacity for agentic AI.

The distinction also reflects how AI systems are organized. Some agent tasks, such as building software, can run on CPUs. Heavy inference may need a separate accelerator-oriented system. A hybrid Helion rack can combine more of those functions, while a CPU-only deployment can concentrate on serving many independent sandboxes.

OCP’s ORv4 specification provides the physical foundation for Helion. AMD describes the design as standardized, with a 21-inch rack that is slightly wider and taller than a standard IT rack. Using that specification gives manufacturers a common target for rack dimensions, trays, cooling, and system layout.

DC-MHS adds a modular interoperability framework. The standard is designed to bring interoperability among key elements of data-center, edge, and enterprise infrastructure. In practical terms, compatible components can be designed around shared expectations rather than one vendor’s unique format. Helion supports 44 OpenU trays and can use air or direct liquid cooling.

These standards do not mean every rack will be identical. AMD is contributing a reference design, while OEMs and ODMs will create the products. The shared specifications should make it easier for those manufacturers to build systems that fit the same ecosystem. AMD did not name partners in the pre-briefing, but partners were expected to be announced at the summit.

A core count measures available parallel processing capacity, not an automatic speed multiplier. A single task may contain steps that depend on earlier results, or it may use only a limited number of threads. Memory access, communication, and software overhead can also limit gains. Those constraints mean 32,768 cores cannot make every program 32,768 times faster.

Consider building software. Some files or tests can be processed at the same time, but other steps must wait for shared results. Helion could assign separate workspaces or jobs to different cores, while each individual job uses only the parallelism its software supports. The rack’s total capacity is therefore larger than the speed available to one task.

Serving many simultaneous users is different. Each user’s agent sandbox can receive its own slice of compute, allowing many independent activities to progress concurrently. In that situation, extra cores increase overall throughput and user capacity. That is why Helion’s large CPU count fits agentic AI platforms even when one agent’s task does not scale across every core.