Key points
- Helion supports up to 32,768 CPU cores in a single rack
- Partner systems are scheduled for release in November
- The design includes a token router for hybrid model calls
AMD has published an open rack blueprint called Helion, designed specifically for running AI agents rather than training large models. The architecture focuses on maximising the number of agents per watt, addressing growing power constraints for data centre operators.
What happened
AMD introduced the Helion blueprint at the Open Compute Project summit in San Jose, according to The Register. The design offers two configurations: a CPU-only version for hyperscale operators and a mixed CPU and GPU version for enterprises. The CPU-only rack utilises 128 sockets of AMD’s 256-core Venice chip on Zen 6, reaching a total of 32,768 cores. These blades operate at either 400W dense or 600W balanced power levels.
The mixed rack configuration contains fewer graphics components than typical training systems. It holds a maximum of 16 Instinct MI350P accelerators, distributed across two independent eight-GPU servers. AMD states these servers run without a scale-up fabric, distinguishing Helion from its Helios rack, which is designed for frontier training with 72 MI455X parts.
A key feature of the enterprise rack is a token router that directs agent model calls. This router sends requests either to frontier models in the cloud or to open-source models running locally within the rack. AMD’s release notes describe this component briefly, but it is central to the system’s operational flexibility.
Why it matters
The primary value proposition of Helion is power efficiency, particularly for sovereign AI operators in Europe. Matt Kimball at Moor Insights and Strategy noted that the goal is packing as many agents as possible into every available watt. This approach addresses significant grid limitations, such as Denmark’s pause on new connections amid a queue of 60GW against peak demand near 7GW.
European regulatory frameworks are also shaping this market dynamic. Brussels adopted a Cloud and AI Development Act in June to triple data centre capacity within seven years. However, a survey of 300 decision-makers in the UK, Ireland and the Nordics identified power availability as the leading obstacle to delivering sovereign capacity. Helion’s design directly targets this infrastructure bottleneck.
What happens next
AMD has not yet published pricing for the Helion systems. The company plans to contribute the specification to the Open Compute Project rather than having done so already. No European system builders have been named among the initial partners, though partner systems are due in November.
What to do and how to stay safe: AMD
- Review current data centre power capacity limits against projected AI agent workloads to identify potential bottlenecks.
- Evaluate whether existing GPU-heavy infrastructure can be repurposed for CPU-centric agent routing or if new racks are required.
- Monitor Open Compute Project updates for the final Helion specification and partner announcements before committing to hardware purchases.
- Assess the security implications of hybrid model routing, ensuring token routers maintain strict isolation between local and cloud-based model calls.
Step-by-step guide: Shadow AI usage: how to control unmanaged model access in your firm
General security guidance from the AppSec Scoop newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the core count of the CPU-only Helion rack?
The CPU-only version holds up to 32,768 cores using 128 sockets of AMD’s 256-core Venice chip.
When are partner systems expected to be available?
Partner systems built on the Helion blueprint are due in November.
How does Helion differ from AMD’s Helios rack?
Helion is designed for running AI agents with high CPU density, while Helios is a frontier training machine with more GPUs and HBM4 memory.
