AMD launches Helion: an open AI rack with 32,768 CPU cores
AMD just launched Helion, an open rack blueprint packing 32,768 CPU cores to run AI agents without relying entirely on expensive GPUs.
Tools · Source: AMD Newsroom
What happened
AMD just unveiled Helion. It is a new open blueprint for agentic AI infrastructure based on an Open Compute Project specification. The system is specifically engineered to handle the complex workflows of AI agents. Rather than forcing all AI workloads onto graphics processors, Helion gives customers the flexibility to match compute performance, core density, and power consumption to the exact tasks their agents perform.
The standout configuration is a massive CPU-only rack. It packs up to 32,768 cores using 6th Gen AMD EPYC Venice processors, which are built on the Zen 6 core architecture. This specific setup targets hyperscale customers who need massive dedicated capacity for agent execution. It allows them to handle agent coordination and logic on processors while connecting to entirely separate infrastructure for model inference.
For enterprise deployments, AMD offers an all-in-one rack version. This configuration combines the EPYC CPU nodes with AMD Instinct MI350P GPU nodes and general-purpose compute on a single platform. Both setups feature an integrated token router. This software layer automatically directs agent model calls across different frontier and open-source models. The first configurations will be ready by November 2026, with broader availability planned for the first half of 2027.
Key facts
- 32,768 — Maximum number of CPU cores in the CPU-only Helion rack configuration
- MI350P — The AMD Instinct GPU nodes used in the all-in-one enterprise configuration
- November 2026 — When the first configurations of the Helion rack will be ready for deployment
Why it matters
AI agents do not just generate text. They execute code, call external tools, manage complex workflows, and wait for other systems to respond. Running these coordination tasks on expensive graphics processors is a massive waste of capital and power. AMD is providing a hardware architecture that matches the actual reality of agentic workflows. You can now scale agent execution on high-density processors while reserving your graphics cards strictly for heavy inference tasks. This separation of concerns is critical for scaling AI systems efficiently.
The second-order effect here is the disruption of closed hardware ecosystems. The open nature of Helion shifts power away from proprietary hardware stacks. By contributing this blueprint directly to the Open Compute Project, AMD allows the broader hardware community to build, modify, and deploy these racks. This open approach drives down infrastructure costs across the board. Cheaper compute for agent coordination means developers can build more complex autonomous systems without burning through their startup runway. It turns existing IT infrastructure investments into a scalable foundation for the next wave of intelligent workloads.
For builders
Cheaper agent execution at scale
Hyperscalers and large enterprises can now offload agent coordination entirely to CPU-only racks. This saves highly expensive GPU cycles for actual model inference, drastically lowering the total cost of running autonomous workflows at scale.
Built-in token routing software
Helion includes an integrated token router to manage complex model calls. Builders can seamlessly direct agent requests between massive frontier models and smaller open-source models based on the specific requirements of each task.
Leverage through open hardware
The OCP-based design means you are not locked into a single vendor proprietary rack architecture. Infrastructure buyers gain significant leverage to negotiate better pricing and customize their liquid or air cooling setups.
My take
Agents spend most of their time waiting on API calls and running basic scripts, not doing heavy matrix math. AMD clearly gets this. By cramming 32,768 CPU cores into an open rack design, they are attacking the exact bottleneck in agentic AI. If you are building AI agents, this is the exact type of infrastructure that will make your unit economics actually work at scale.
Original reporting: AMD Newsroom. This is my rewrite and opinion.