AMD Helion: The Open Rack Blueprint for Agentic AI (2026)

Every AI agent you deploy needs a place to run — not just a GPU to think with, but a CPU to act with. When an agent spawns a code sandbox, calls an API, queries a database, coordinates with three other agents, and writes back to a system of record, that workload lands squarely on the processor, not the accelerator. SambaNova demonstrated this at Computex 2026: in a 45-second agentic data-analysis task, only 2 seconds ran on the GPU. The other 43 were pure CPU orchestration.

AMD’s answer to this structural shift arrived on October 12, 2026, at the OCP Global Summit in San Jose: Helion, an open rack reference design built from the ground up for agentic AI workloads. Where AMD’s Helios rack targets model training and inference with 72 MI455X GPUs and 31 TB of HBM4, Helion flips the ratio — up to 32,768 CPU cores in a single rack, purpose-built for the agent sandbox layer that sits between the model and the real world.

This guide covers the complete Helion architecture, its two configurations, the token router that distributes model calls, the EPYC Venice processors engineered for agent density, how it compares to NVIDIA Vera and Intel Xeon 6+, and what it means for enterprise teams planning agentic infrastructure in 2027 and beyond.

Table of Contents

• What AMD Helion Is (and What It Is Not)
• Two Configurations: CPU-Only Hyperscale and All-in-One Enterprise
• EPYC Venice: The 256-Core Agent Sandbox Processor
• The Token Router: Smart Model Distribution
• OCP Open Standard and the Partner Ecosystem
• The Three-Way Race: AMD Helion vs NVIDIA Vera vs Intel Xeon 6+
• AMD’s Own Numbers: 43% Token Cost Reduction
• What This Means for RPA and Agentic AI Architects
• Deployment Timeline and Availability
• FAQs
• Key Takeaways
• References

What AMD Helion Is (and What It Is Not)

Helion is not a product you buy from AMD. It is an open rack reference design — a blueprint that AMD is contributing to the Open Compute Project (OCP) so that OEMs and ODMs can build, customize, and sell their own systems based on it. Think of it the way you think of OCP’s earlier rack specs: AMD provides the architecture, cooling guidelines, and tray specifications; partners like Cisco, Dell, HPE, Lenovo, QCT, Supermicro, Foxconn (Hon Hai), and Wiwynn build the actual hardware.

The design is based on the OCP ORv4 specification — a 21-inch-wide rack format slightly wider and taller than a standard IT rack. Each rack holds up to 44 OpenU trays, each 48 mm tall, built to OCP’s DC-MHS (Data Center Modular Hardware System) standard for interoperability across data center, edge, and enterprise environments.

What makes Helion different from AMD’s existing Helios platform is its fundamental orientation:

The two platforms are designed to work together. A hyperscaler might deploy Helios racks for model inference and Helion racks alongside them for agent execution — the agents call models running on Helios (or in the cloud) while doing their CPU-heavy work on Helion. Enterprises with smaller footprints use Helion’s all-in-one configuration to run both the models and the agents in the same rack.

Two Configurations: CPU-Only Hyperscale and All-in-One Enterprise

CPU-Only Configuration: The Agent Density Play

The CPU-only configuration is Helion’s hyperscale variant. It packs up to 128 AMD EPYC 9996 processors — each a 256-core Venice chip — into a single rack, delivering 32,768 cores and 65,536 threads of agent execution capacity.

More Read

Unleashing the Power of Agno: Building Multi-Modal Agents with a Lightweight Python Library

Low-Code/No-Code Development Platforms: The Future of Software Development?

MCP 2026-07-28 Stateless Spec: What Every Agentic AI Architect Must Know

Blades come in two power profiles:

• 400W dense configuration — optimized for maximum agent density per rack unit, prioritizing throughput over single-thread speed
• 600W balanced configuration — higher clock speeds for workloads where individual agent latency matters more than total agent count

In this configuration, Helion racks connect to separate inference infrastructure — whether that is a Helios rack next door, a cloud endpoint, or a mix of both. Each agent gets a slice of the host system’s compute as a sandbox, running code, calling tools, querying databases, and managing state while the model calls happen elsewhere.

This mirrors the architecture already in production at scale. Meta’s Muse and OpenAI’s Dots both use a model where the agent’s code-execution environment is separated from the model-inference layer. The agents need CPUs; the models need GPUs. Helion supplies the CPU side at rack scale.

All-in-One Enterprise Configuration: Models and Agents in One Rack

The enterprise configuration adds up to 16 AMD Instinct MI350P GPUs across two eight-GPU servers, alongside EPYC CPU nodes and general-purpose compute — all in the same rack.

The MI350P is AMD’s air-cooled enterprise GPU, capable of running models up to 260 billion parameters on a single card. The two GPU servers in the enterprise Helion rack operate independently — there is no scale-up fabric connecting them — which means each server can run its own model or both can serve the same model behind a load balancer.

This configuration trades CPU density for self-sufficiency. An enterprise team deploying agentic workflows in a regulated environment — where sending prompts to a cloud endpoint is not an option — can run both the model and the agent fleet in a single rack, with the token router handling the distribution logic.

With up to 102 Tbps of networking bandwidth in the enterprise configuration, the rack can handle the data movement between agent processes and local model inference without becoming I/O-bound — a concern that Futurum Group analyst Daniel Newman has identified as the real binding constraint in agentic infrastructure.

EPYC Venice: The 256-Core Agent Sandbox Processor

Helion’s compute foundation is the 6th Gen AMD EPYC 9006 series, codenamed Venice. Built on TSMC’s 2nm process with 203 billion transistors, Venice delivers up to 1.8x the performance of the previous EPYC generation.

AMD has built the EPYC 9006 lineup around three tiers that map directly to agentic data center architecture:

The SP7 is the agent sandbox chip. With 256 cores and 512 threads per socket, a fully loaded Helion CPU-only rack delivers enough concurrency to run thousands of agent processes simultaneously — each with its own isolated slice of compute for code execution, tool calling, database queries, and I/O management.

AMD’s description of the three-tier model is worth internalizing because it maps directly to how production agentic systems are actually built:

• Agent sandbox tier (SP7 in Helion) — runs and coordinates agents, executes spawned code, manages memory and system state, handles I/O outside the GPU
• AI host node tier (LP in Helios or Helion enterprise) — feeds GPUs for model inference
• General-purpose tier (SP8) — databases, storage, application services, and AI-support tasks

Venice is in full production as of Q4 2026, with rollouts at major OEMs and cloud providers underway. This is not a future roadmap item — the silicon is shipping.

The Token Router: Smart Model Distribution

One of Helion’s most architecturally significant features is its integrated token router — a component that distributes agent model requests across different model endpoints.

In production agentic systems, not every agent call needs a frontier model. A code-review agent might need GPT-6 or Claude for complex reasoning, but a data-validation agent checking field formats can run on a smaller, faster, cheaper open-weight model locally. The token router makes this routing decision at the infrastructure level, sending each request to the optimal endpoint based on the task.

The router supports both proprietary model families (via cloud API endpoints) and open-source models (running locally on MI350P GPUs within the same enterprise Helion rack). This is the mechanism that enables the hybrid deployment model: agents running on Helion CPUs can call models running locally, in the cloud, or a mix of both — without the application layer needing to manage the routing logic.

AMD has not published detailed specifications on the token router’s decision methods, latency overhead, or configuration interface. The feature appears in AMD’s official materials but awaits independent benchmarking. For architects planning deployments, the router is best understood as a declared capability that will need validation once partner systems ship in November.

That said, AMD’s own internal IT team has already deployed a version of this routing approach. By routing between local open-weight models and frontier cloud models based on task complexity, AMD’s internal deployment achieved a 43% reduction in token costs and up to 3x faster response times for workloads that could run locally — numbers that provide a credible signal even while the Helion-specific implementation awaits public benchmarks.

OCP Open Standard and the Partner Ecosystem

AMD is contributing the Helion specification to the Open Compute Project, making it an open standard that any system builder can adopt and customize. This is a deliberate contrast to proprietary rack architectures — NVIDIA’s NVLink-based interconnects, for instance, create tighter integration at the cost of vendor lock-in.

Helion’s all-Ethernet networking approach (shared with Helios) means enterprises can build on existing network infrastructure rather than deploying proprietary interconnects. For RPA and automation teams already managing data center environments, this is a practical advantage: the rack plugs into what you have.

Confirmed Ecosystem Partners

The following companies have announced support for the Helion architecture (support for the design, not confirmed shipments as of October 2026):

• Cisco — integrating Cisco Cloud Control for fleet management of AMD local-AI devices, with agent observability and token-consumption controls
• Dell Technologies
• Foxconn (Hon Hai)
• Hewlett Packard Enterprise (HPE)
• Lenovo
• QCT (Quanta Cloud Technology)
• Supermicro
• Wiwynn

Each partner can tailor CPU and GPU mixes to customer workloads, which means the actual Helion systems that ship will vary by vendor. An HPE Helion rack might emphasize liquid cooling and high-density compute; a Lenovo variant might target edge-adjacent enterprise deployments with air cooling.

AMD’s broader ecosystem context is worth noting. At Advancing AI 2026, AMD announced major infrastructure agreements:

• OpenAI — a 6 GW agreement, with Helios infrastructure already running GPT-class workloads for three months
• Meta — MI350-series GPUs in production for ranking and recommendation, with MI450 planned under a 6 GW agreement
• Anthropic — up to 2 GW of MI450-series GPU capacity and up to $5 billion in AMD equity investment
• Cerebras — Helios racks paired with Cerebras wafer-scale engines for disaggregated inference, launching on Cerebras Cloud in H2 2026

These deals signal that the hyperscalers building the largest agent platforms are investing in AMD infrastructure at multi-gigawatt scale — and where Helios handles their model inference, Helion is positioned to handle their agent execution.

The Three-Way Race: AMD Helion vs NVIDIA Vera vs Intel Xeon 6+

The recognition that agentic AI is CPU-hungry has sparked a competitive race among the three major server processor vendors. Each is approaching the problem differently.

The strategic differences are revealing:

• AMD bets on open infrastructure and CPU density. Helion’s OCP contribution means enterprises avoid single-vendor lock-in. The Venice SP7’s 256-core design prioritizes running many agents concurrently.
• NVIDIA bets on vertical integration. Vera’s NVLink-C2C fabric creates a tight coupling between CPU orchestration and Rubin GPU inference. NVIDIA claims a 1.8x task-completion advantage over x86 in reinforcement learning sandbox loops specifically — a metric that matters for training-adjacent agent workloads but may not generalize to enterprise process automation.
• Intel bets on enterprise density. With 288 E-cores per socket and a liquid-cooled rack hitting 36,864 cores, Intel markets the highest core count and introduces “agents per rack” as the defining metric. Intel is already shipping, which is an advantage over both AMD and NVIDIA for teams deploying today.

For RPA and automation architects, the practical takeaway is that the CPU layer of the agentic stack is now a competitive market. A year ago, the conversation was entirely about GPU capacity for model inference. Today, three major vendors are competing specifically on agent execution density. That competition will drive prices down and capability up through 2027.

Futurum Group analyst Daniel Newman argues these chips are more complementary than substitutable: Intel serves the broad enterprise inference install base, Vera targets training-run economics, and AMD (via Helion) targets the open-standard infrastructure play. Most large enterprises will end up running a mix.

AMD’s Own Numbers: 43% Token Cost Reduction

AMD’s internal IT team has deployed a production agentic system using EPYC processors and MI350P GPUs with intelligent routing between local open-weight models and frontier cloud models. The reported results:

• 43% reduction in token costs compared to routing all requests to cloud endpoints
• Up to 3x faster response times for workloads that run locally on open-weight models

These are AMD’s self-reported numbers from their own deployment, not independently verified benchmarks. But they align with what other enterprises are finding when they implement model routing: a substantial portion of agentic workloads don’t need frontier-model capability, and running them locally on smaller models is both faster and cheaper.

Cisco, working with AMD, adds another data point. Cisco describes agentic AI workloads as “persistent, always-on demand” — “Humans click, but agents swarm” — and has built agent observability and token-consumption controls into Cisco Cloud Control specifically to manage the cost dynamics of large agent fleets. Cisco’s Cisco Cloud Control tool for managing fleets of AMD local-AI devices reached early availability in 2026, with U.S. general availability expected in early fall.

AT&T provides a third reference point. AT&T has processed over one trillion tokens on AMD Instinct GPUs in Microsoft Azure, and reports that its model-routing approach cut some AI costs by up to 80%. While AT&T’s deployment is on the inference side (not agent execution), it validates the routing-based cost optimization that Helion’s token router aims to bring to the agent layer.

What This Means for RPA and Agentic AI Architects

The Infrastructure Layer Is Now Purpose-Built for Agents

Until Helion (and the competing efforts from Intel and NVIDIA), enterprises deploying agents were retrofitting general-purpose compute for a workload that has a distinct profile: high concurrency, low per-agent compute, heavy I/O, tool-calling latency sensitivity, and a need for process isolation. Helion is the first open-standard rack design that treats agent execution as a first-class workload category with its own hardware optimization.

For automation architects building on protocols like A2A and MCP, this matters because the infrastructure no longer assumes your agents are just another container workload. A 256-core processor with PCIe Gen 6 and a dedicated token router means the physical layer is starting to match the software architecture.

CPU Capacity Planning Becomes a First-Order Concern

Most agentic architecture discussions in 2025 and early 2026 focused on model selection and GPU availability. Helion’s existence signals a shift: CPU capacity planning for agent execution is now a first-order infrastructure concern, not a secondary consideration you solve with “whatever servers you have.”

The three-tier model AMD describes — agent sandbox, AI host node, general-purpose — gives architects a concrete framework for capacity planning. If your agentic system has 500 concurrent agents, each running code, calling tools, and managing state, the question is no longer “do we have enough GPU?” but “do we have enough agent sandbox cores?”

Open Standards Lower the Switching Cost

By contributing Helion to OCP and using all-Ethernet networking, AMD is making a bet that enterprises will favor open, interoperable infrastructure over proprietary stacks. For RPA teams that have historically dealt with vendor lock-in at the software level (UiPath vs Automation Anywhere vs Blue Prism), this philosophy is familiar — and appealing.

An enterprise running agents on Helion infrastructure can swap CPU vendors, GPU vendors, or model providers without re-architecting the rack. That flexibility is worth real money over a 3-5 year infrastructure lifecycle.

On-Premises Agent Execution Becomes Viable at Scale

The enterprise Helion configuration — combining MI350P GPUs for local model inference with dense CPU compute for agent execution — makes a fully on-premises agentic deployment viable at rack scale for the first time. This matters for regulated industries where sending agent prompts to cloud endpoints creates compliance risk.

Financial services, healthcare, defense, and government organizations that have been cautious about agentic AI deployments due to data-sovereignty concerns now have a reference architecture for running the entire stack — model inference and agent execution — within their own data centers.

Power Efficiency Is the New Battleground

Matt Kimball of Moor Insights and Strategy describes Helion’s goal as “fitting as many agents as possible into each watt.” This is not an abstract concern. Europe’s data center capacity constraints — Denmark paused new data center connections in March 2026, and roughly 60 GW of projects are queued against peak demand near 7 GW — make power efficiency a binding constraint on agentic deployment scale.

The EU’s Cloud and AI Development Act, adopted in June 2026, aims to at least triple the bloc’s data center capacity within seven years, and the EU has opened bidding for up to seven AI gigafactories at about €30 billion total. But until that capacity comes online, enterprises will compete for power allocation — and the infrastructure that delivers the most agent throughput per watt wins.

Deployment Timeline and Availability

Note on the timeline: Tom’s Hardware reported conflicting dates — the article body initially cited H1 2027, while an update said AMD revised to November 2026. The most recent AMD materials and partner announcements point to November 2026 for initial availability. Architects should plan for Q1 2027 availability in practice, with early-access partner systems potentially available in November/December 2026.

Frequently Asked Questions

Can I buy AMD Helion directly from AMD?

No. Helion is a reference design, not a product. You purchase Helion-based systems from OEM/ODM partners including Cisco, Dell, HPE, Lenovo, QCT, Supermicro, Foxconn, and Wiwynn. Each partner will offer their own configurations and pricing.

Does Helion replace Helios?

No. Helion complements Helios. Helios is optimized for GPU-dense model training and inference (72 MI455X GPUs per rack). Helion is optimized for CPU-dense agent execution (up to 32,768 cores per rack). Many deployments will use both together — Helion for agent execution, Helios for model inference.

How does Helion compare to running agents on standard cloud VMs?

Cloud VMs are general-purpose compute. Helion’s EPYC SP7 processors are specifically designed for the concurrency profile of agent workloads — 256 cores per socket, PCIe Gen 6, and a token router for intelligent model-call distribution. For enterprises running large agent fleets (hundreds to thousands of concurrent agents), the density and efficiency advantages are significant. For smaller deployments (under 50 agents), standard cloud compute may remain more cost-effective.

Is the token router a hardware or software component?

AMD describes it as “integrated” but has not published detailed specifications on whether it runs as a dedicated hardware component, a firmware service, or a software layer. Independent validation will need to wait for partner systems shipping in November 2026.

What models can run on the MI350P GPUs in the enterprise configuration?

The MI350P supports models up to 260 billion parameters on a single card. In the enterprise Helion configuration with 16 MI350P GPUs across two servers, you could run multiple smaller models simultaneously or one large model with the parameters distributed across cards. AMD’s ROCm software stack supports major open-weight model families.

Key Takeaways

• AMD Helion is the first open-standard rack design purpose-built for agentic AI execution — up to 32,768 CPU cores per rack, announced October 12, 2026, with partner systems expected from November 2026.
• Two configurations serve different markets: CPU-only for hyperscale agent density (connecting to external inference infrastructure), and all-in-one enterprise for running models and agents in one rack.
• EPYC Venice SP7 delivers 256 cores / 512 threads per socket, built specifically for the agent sandbox tier where agents execute code, call tools, and manage state.
• The integrated token router distributes agent model calls between local and cloud endpoints — AMD’s internal deployment reports 43% token cost savings and up to 3x faster local responses.
• OCP open standard and all-Ethernet networking mean no vendor lock-in — eight major OEM/ODM partners are building systems.
• The CPU matters again: AMD, NVIDIA (Vera), and Intel (Xeon 6+) are now competing specifically on agent execution density — a category that did not exist in hardware marketing 12 months ago.
• For RPA and automation architects: CPU capacity planning for agent execution is now a first-order infrastructure concern. Plan for a three-tier model: agent sandbox, AI host node, and general-purpose compute.

References

1. Tom’s Hardware, “AMD debuts Helion reference design for agentic AI servers with up to 32,768 Zen 6 Venice cores” (October 2026) — tomshardware.com
2. Engineering.com, “AMD introduces open Helion rack blueprint for agentic AI” (October 2026) — engineering.com
3. The Next Web, “AMD puts 32768 CPU cores in a rack built for AI agents” (October 2026) — thenextweb.com
4. Igor’s Lab, “AMD Helion: Rack with up to 32,768 EPYC Cores” (October 2026) — igorslab.de
5. IT Europa, “AMD bundles chips, racks and partners into full-stack AI push” (2026) — iteuropa.com
6. Fierce Network, “AMD launches full-stack AI compute for agentic era” (2026) — fierce-network.com
7. StorageNewsletter, “AMD AAI 2026: 6th Gen AMD EPYC Server CPUs Power the Agentic Data Center” (2026) — storagenewsletter.com
8. Futurum Group, “COMPUTEX 2026: Are Agentic CPUs Rivals or Complements?” (2026) — futurumgroup.com
9. AMD Investor Relations / Press Release, “AMD Reports Second Quarter 2026 Financial Results” — ir.amd.com

Got a question about deploying agentic AI infrastructure? Drop a comment below or reach out on our contact page.