AMD used the Open Compute Project (OCP) Global Summit in San Jose on October 12, 2026 to introduce AMD Helion, an open rack-scale reference design built specifically to run autonomous AI agents rather than just large language models. The announcement, styled officially as the AMD Helion™ Agentic AI Rack, marks AMD’s clearest attempt yet to define a standard for the next phase of AI infrastructure: systems that spend as much compute on orchestration, tool use, and multi-step reasoning as they do on raw model inference.
Unlike a single boxed server, Helion is a blueprint. AMD built it on an Open Compute Project specification design, meaning server makers and hyperscalers can assemble their own racks around AMD’s reference rather than buying a fixed appliance. That distinction matters for a market that has spent the past two years watching Nvidia set the pace with its own tightly integrated rack-scale systems. AMD’s pitch with Helion is openness and configurability over a single, vendor-defined box.
What AMD Helion Actually Is
AMD Helion is not one product but a family of two rack configurations, both built to the same open mechanical and electrical standards. The first is a CPU-only rack designed for hyperscale deployments that need enormous amounts of general-purpose compute to run the orchestration layer behind AI agents — the code that decides what an agent should do next, calls external tools, retrieves data, and manages long-running workflows. The second is an all-in-one configuration that mixes CPU and GPU compute in the same rack for enterprises that want a single, self-contained system rather than separately provisioned agent and inference clusters.
Both configurations ship in liquid-cooled and air-cooled variants, letting operators choose based on the thermal and power constraints of their facility rather than being locked into one cooling approach. AMD said it plans to contribute the Helion specification itself to the Open Compute Project, a move that would let other hardware vendors build compatible racks rather than treating the design as AMD-exclusive intellectual property.
Inside the CPU-Only Configuration: 32,768 Cores
The headline number from the announcement is 32,768 CPU cores. AMD’s CPU-only Helion configuration is powered by 6th generation AMD EPYC processors and scales up to that core count per rack, according to AMD’s own announcement. That is an enormous figure even by data center standards, and it reflects a specific bet: that the “thinking” layer of agentic AI — the part of the stack that routes requests, checks outputs, calls APIs, and coordinates multiple agents working on the same task — is fundamentally a CPU-bound problem, not a GPU-bound one.
That framing is the core of AMD’s argument for why agentic AI needs different infrastructure than the GPU-dense racks built for training and raw inference. A single user query to an AI agent might trigger dozens of smaller tool calls, retrieval steps, and verification passes before it produces a final answer. AMD’s position is that hyperscalers will want to decouple that orchestration workload from their GPU inference clusters, running it on dedicated CPU-only racks that can scale independently. The CPU-only Helion rack is aimed squarely at that use case.
The All-in-One Configuration: EPYC Meets Instinct MI350P
For organizations that don’t operate at hyperscaler scale, AMD’s all-in-one Helion configuration combines AMD EPYC CPU nodes, AMD Instinct MI350P GPU nodes, and general-purpose compute nodes inside a single rack. This is the version aimed at enterprises deploying agentic AI applications that need both orchestration compute and model inference in one footprint, without having to stitch together separate CPU and GPU clusters themselves.
AMD also described a software layer riding on top of the hardware: an integrated token router that directs an agent’s model calls across a mix of frontier and open-source models depending on the task. That is a meaningful detail, because it signals AMD isn’t just selling silicon — it’s trying to own part of the orchestration software stack that decides which model handles which step of an agentic workflow, a layer that is becoming as commercially important as the chips underneath it.
Why the Token Router Matters
Agentic AI workloads rarely rely on a single model for every step. A coding agent might use a large frontier model for planning and a smaller, cheaper open-source model for repetitive validation checks. By building model routing directly into the Helion software stack, AMD is positioning its hardware as model-agnostic infrastructure rather than a platform tied to one AI vendor’s ecosystem — a contrast with how some GPU-centric platforms have historically bundled hardware and software more tightly together.
Built on Open Compute Project Standards
The “open” in AMD Helion’s branding is not just marketing language. According to AMD’s own announcement, the rack design layers together several existing OCP standards rather than inventing a proprietary mechanical format from scratch. That matters for data center operators who have already standardized their facilities around OCP specifications and don’t want to retool for a single vendor’s rack geometry.
AMD Helion — Rack Configuration Summary
=========================================
1) CPU-only configuration (hyperscale agent execution)
- Processor: 6th Gen AMD EPYC
- Max CPU cores per rack: 32,768
- Cooling: liquid-cooled or air-cooled
- OCP standards used: ORv3 (rack/busbar), DC-MHS
(host processor modules), DC-SCM (management/
security), HPRv2 (power shelves), OCP NIC 3.0
(networking), Caliptra (silicon root of trust)
2) All-in-one configuration (enterprise agentic AI)
- CPU: AMD EPYC nodes
- GPU: AMD Instinct MI350P nodes
- Plus: general-purpose compute nodes
- Software: integrated token router across
frontier and open-source models
- Cooling: liquid-cooled or air-cooled
AMD’s plan to hand the Helion specification over to the Open Compute Project community is consistent with how it has approached data center hardware for much of the past decade: compete on open standards rather than closed ecosystems, and let scale and partner adoption do the rest. If AMD follows through on contributing Helion to OCP, other server manufacturers would be free to build their own Helion-compatible racks, which could accelerate adoption faster than a single-vendor product ever could.
Liquid and Air Cooling, By Design
Both Helion configurations support liquid-cooled and air-cooled deployments, according to AMD. That flexibility is not a minor footnote. Data centers vary enormously in how much liquid cooling infrastructure they already have in place, and a growing share of the industry’s capital spending in 2026 is going toward retrofitting facilities for the power density that modern AI racks demand. By supporting both cooling approaches natively in the reference design, AMD is lowering the barrier for operators who haven’t yet completed a liquid-cooling buildout but still want to deploy dense agentic AI compute.
Availability: November 2026, Not Already Shipping
AMD says Helion rack configurations are currently under development with ecosystem partners and are planned to become available in November 2026. That is an important detail to get right: Helion was announced today, but it is not a shipping product yet. Reports characterizing it as already available, or pushing its arrival into the first half of 2027, run against AMD’s own stated timeline. The gap between “unveiled” and “shipping” is just a few weeks on paper, which is an aggressive turnaround for a rack-scale reference design of this complexity, and it suggests AMD has been working with select partners well before today’s public reveal.
Industry Reaction: What AMD Is Saying
AMD has so far kept its public commentary focused on explaining why agentic AI needs a different infrastructure model rather than making head-to-head performance claims against competitors. Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD, framed the problem in terms of workload diversity rather than raw throughput: “It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements.”
That statement gets at the thesis underpinning the entire Helion launch: agentic AI isn’t one workload, it’s dozens of different ones running side by side, from a consumer chatbot calling a weather API to an enterprise system executing a multi-step financial workflow. AMD’s bet is that no single, fixed hardware configuration can serve all of those cases efficiently, which is why Helion ships as two distinct configurations rather than one. Independent analyst commentary and AMD’s stock reaction to the announcement were not yet available in the hours following the reveal, and this article will not speculate on market moves that have not been confirmed.
AMD Helion vs Nvidia’s Rack-Scale AI Systems
The competitive subtext of today’s announcement is impossible to miss. Nvidia has spent the past two years setting the template for rack-scale AI infrastructure with integrated systems like the GB200 NVL72 and its newer GB300 platform — tightly engineered, vendor-defined racks that bundle Nvidia GPUs, networking, and software into a single purchasing decision. AMD’s answer with Helion is deliberately different in structure: an open specification, built on OCP standards, that lets customers choose CPU-only or mixed CPU/GPU configurations rather than buying into one fixed GPU-centric design.
That is a strategic choice as much as a technical one. Nvidia’s integrated approach has won it the lion’s share of AI training and inference infrastructure spend industry-wide, in part because it reduces the integration work operators have to do themselves. AMD is betting that as agentic AI workloads diversify — some orchestration-heavy, some GPU-heavy, some evenly mixed — operators will increasingly want the flexibility to pick their own balance of CPU and GPU compute rather than accept a fixed ratio baked into someone else’s reference design. Whether that bet pays off will depend heavily on how quickly server OEMs and hyperscalers actually build and deploy Helion-based racks once the specification reaches partners this November.
Historical Context: AMD’s Long Game in the Data Center
Helion is the latest step in a strategy AMD has pursued since it re-entered the server CPU market with EPYC nearly a decade ago: compete on open standards and partner ecosystems rather than trying to out-spend rivals on proprietary vertical integration. AMD’s EPYC line has steadily taken server CPU market share from Intel over the past several years, and the Instinct GPU line has positioned AMD as the most credible alternative to Nvidia in AI accelerators, even if it remains well behind Nvidia’s overall AI data center revenue.
Helion extends that playbook up a layer, from individual chips to entire rack architectures. By contributing the specification to the Open Compute Project, AMD is effectively inviting the same dynamic that helped EPYC gain ground against Intel: let the broader ecosystem build on top of an open standard, rather than asking every customer to buy a complete AMD-only stack. It’s a slower path to dominance than Nvidia’s integrated approach, but it is consistent with how AMD has chipped away at entrenched incumbents in the past.
Why Agentic AI Changes the Infrastructure Math
The broader industry context here is the shift from “generative AI” to “agentic AI” as the dominant framing for enterprise AI investment in 2026. Generative AI workloads are largely GPU-bound: a prompt goes in, a model produces an output, and the GPU does almost all the work. Agentic AI workloads look different. An autonomous agent might make several model calls per task, interleaved with tool calls, database lookups, retries, and verification steps that run on conventional CPU infrastructure, not GPUs.
That shift is exactly what AMD’s CPU-only Helion configuration, with its 32,768-core ceiling, is built to address. If a meaningful share of future AI infrastructure spending goes toward orchestration and agent execution rather than pure model inference, a vendor that can offer massive CPU scale as a first-class product, not an afterthought bolted onto a GPU rack, stands to capture a different slice of the market than the one Nvidia has dominated so far.
AMD Helion Rack Configurations at a Glance
| Attribute | CPU-Only Configuration | All-in-One Configuration |
|---|---|---|
| Primary processor | 6th Gen AMD EPYC | AMD EPYC CPU nodes |
| GPU included | None | AMD Instinct MI350P |
| Max CPU cores per rack | 32,768 | Not disclosed by AMD |
| Target user | Hyperscalers running agent orchestration at scale | Enterprises deploying self-contained agentic AI |
| Cooling options | Liquid-cooled or air-cooled | Liquid-cooled or air-cooled |
| Software layer | AMD open software stack | Integrated token router across models |
| Design basis | Open Compute Project specification | Open Compute Project specification |
| Planned availability | November 2026 | November 2026 |
The OCP Standards Stack Behind Helion
| OCP Standard | What It Governs in Helion |
|---|---|
| ORv3 | Rack frame and busbar power distribution |
| DC-MHS | Host processor module design |
| DC-SCM | System management and security controllers |
| HPRv2 | Power shelf design |
| OCP NIC 3.0 | Networking interface cards |
| Caliptra | Silicon root of trust for firmware security |
Stacking six separate OCP standards into a single reference design is a statement in itself: AMD is signaling that Helion is meant to interoperate with existing OCP-based data center infrastructure rather than demand a bespoke facility build-out. That is a meaningfully different approach from vendors that ask customers to adopt proprietary rack geometries wholesale.
What This Means for Hyperscalers and Enterprises
For hyperscale cloud providers, the CPU-only Helion configuration offers a way to scale agent-orchestration capacity independently of GPU procurement, which has been one of the most constrained and expensive line items in any AI infrastructure budget over the past two years. Being able to add tens of thousands of CPU cores dedicated to agent execution without competing for scarce GPU allocation is a meaningful operational advantage, assuming the economics hold up once real pricing and performance benchmarks become public.
For enterprises, the all-in-one configuration is the more immediately relevant option. Most companies deploying agentic AI applications today are not provisioning racks at hyperscale; they need a self-contained system that can run both the orchestration layer and model inference without requiring a team of infrastructure engineers to stitch together separate CPU and GPU clusters. Whether AMD and its OEM partners can price that configuration competitively against existing GPU-server options will be one of the clearest early signals of whether Helion gains real traction outside of hyperscaler deployments.
Open Questions AMD Hasn’t Answered Yet
Several details remain unconfirmed as of this announcement. AMD has not published pricing for either Helion configuration. The company also has not disclosed exact GPU density figures for the all-in-one configuration in its own materials, leaving open questions about how it will stack up against integrated GPU-dense competitors on raw accelerator count per rack. And while AMD says it plans to contribute the specification to OCP, it has not given a firm timeline for when that formal contribution will happen, which matters for how quickly third-party server vendors could start building compatible hardware.
Predictions: Where AMD Helion Goes From Here
- Partner hardware announcements will follow fast. Expect server OEMs to announce Helion-compatible systems in the weeks around the November 2026 availability window, mirroring how OCP-based designs typically roll out through multiple manufacturers rather than a single vendor.
- Pricing will become the real battleground. Once AMD or its partners disclose Helion pricing, the comparison that matters most won’t be core counts or cooling options — it will be cost per agent-task throughput against Nvidia’s integrated rack systems.
- CPU-only agent infrastructure becomes its own market category. If AMD’s framing holds up, expect competitors to start marketing dedicated “agent orchestration” compute tiers separate from GPU inference tiers, following the split AMD has drawn with Helion’s two configurations.
- OCP contribution timing will shape adoption speed. The faster AMD formally hands the Helion spec to the Open Compute Project, the faster other vendors can build compatible racks — a slow handoff would blunt the openness advantage AMD is leaning on.
- Nvidia will respond with its own orchestration-focused messaging. Given how directly Helion’s CPU-only framing targets a gap in GPU-centric rack designs, expect Nvidia and its partners to sharpen their own messaging around CPU/orchestration capacity in upcoming rack-scale announcements.
The Bottom Line
AMD Helion is less a single product launch than a statement about where AMD believes AI infrastructure spending is heading next. By splitting its rack-scale reference design into a CPU-only configuration built for massive agent orchestration and an all-in-one configuration that blends EPYC CPUs with Instinct MI350P GPUs, AMD is betting that agentic AI’s compute needs are too varied for any single fixed hardware ratio to serve well. The open, OCP-based approach is a deliberate contrast to Nvidia’s more tightly integrated rack systems, and it echoes the same ecosystem-first strategy that helped AMD’s EPYC processors claw back server market share from Intel over the past decade.
None of that guarantees Helion succeeds. Pricing is unknown, GPU density in the all-in-one configuration hasn’t been disclosed, and the November 2026 availability window is still a few weeks away. But the direction of the bet is clear: AMD thinks the next phase of the AI buildout will be won not by whoever has the most GPUs in a rack, but by whoever can most flexibly match compute — CPU and GPU alike — to workloads that look nothing like the single-shot model inference that defined the first wave of the generative AI boom.
Frequently Asked Questions
What is AMD Helion?
AMD Helion is an open, rack-scale reference design for agentic AI infrastructure, unveiled by AMD on October 12, 2026 at the Open Compute Project Global Summit. It is based on an Open Compute Project specification design and comes in CPU-only and combined CPU/GPU configurations.
When will AMD Helion racks be available?
AMD says Helion rack configurations are currently under development with ecosystem partners and are planned to become available in November 2026.
How many CPU cores does the AMD Helion CPU-only rack support?
The CPU-only configuration, powered by 6th generation AMD EPYC processors, supports up to 32,768 CPU cores per rack, according to AMD’s announcement.
What GPU does the AMD Helion all-in-one configuration use?
The all-in-one configuration combines AMD EPYC CPU nodes with AMD Instinct MI350P GPU nodes and general-purpose compute nodes in a single rack.
Is AMD Helion a finished product or a reference design?
Helion is a reference design and open blueprint based on Open Compute Project specifications, not a single fixed appliance. AMD plans to contribute the specification to the OCP community so partners can build compatible systems.
How does AMD Helion compare to Nvidia’s rack-scale AI systems?
AMD is positioning Helion as a more open and configurable alternative to Nvidia’s integrated rack-scale systems, such as the GB200 NVL72 and GB300, emphasizing Open Compute Project standards, independent CPU or combined CPU/GPU configurability, and a choice of liquid or air cooling rather than one fixed vendor-defined design.
Does AMD Helion support liquid cooling?
Yes. Both the CPU-only and all-in-one Helion configurations support liquid-cooled and air-cooled deployment options, according to AMD.
What is the token router feature in AMD Helion?
AMD describes an integrated token router as part of the Helion software stack that directs an AI agent’s model calls across a mix of frontier and open-source models, depending on what a given task requires.