AMD has published an open rack blueprint called Helion, designed specifically to run autonomous software agents rather than train massive foundational models.
The 32,768-Core Architecture and Blade Configurations
The scale of the CPU-only Helion rack relies on raw socket density. It accommodates 128 sockets of AMD’s 256-core Venice chip built on the Zen 6 microarchitecture. These are housed across blades configured for either 400W dense operations or 600W balanced performance.
For workloads requiring hardware acceleration, the mixed CPU and GPU rack takes a different structural approach. Contrary to expectations of a dense graphics deployment, this version is relatively sparse. It holds a maximum of 16 Instinct MI350P accelerators distributed across two eight-GPU servers.
According to AMD’s system documentation, these two eight-GPU servers operate independently without a traditional scale-up fabric. This engineering choice separates the Helion platform entirely from frontier training clusters. By comparison, AMD’s Helios rack carries 72 of the newer MI455X parts and 31 terabytes of HBM4 memory, retaining its role as the company’s dedicated training machine.
Token Routing and Power Efficiency Constraints
The most compelling mechanism within the Helion enterprise rack is its token router. Detailed exclusively in AMD’s release materials, this hardware-software feature dynamically directs an AI agent’s model calls. Depending on the complexity of the task, the router sends requests either to massive frontier models hosted in the cloud or to smaller open-source models running locally inside the rack.
Dan McNamara, who manages compute and enterprise AI at AMD, stated that the point was to let partners make efficient use of available power. Power availability dictates the architectural envelope for modern data center operators.
Matt Kimball at Moor Insights and Strategy noted in a field note that the job is packing as many agents as possible into every watt. Kimball highlighted sovereign AI operators as a primary market, bringing Europe directly into the operational frame.
Infrastructure Bottlenecks Across European Markets
Grid capacity constraints continue to throttle data center expansions across Europe. Denmark’s primary grid operator paused new electricity connections in March, facing roughly 60 gigawatts of pending projects against a peak local demand of approximately 7 gigawatts.
Legislative pressure adds another layer of urgency. Brussels adopted the Cloud and AI Development Act in June, setting a target to at least triple the European Union’s data center capacity within seven years.
Data center decision-makers confirm these structural limitations. In a survey of 300 professionals across the UK, Ireland, and the Nordics, power availability emerged as the leading obstacle to delivering sovereign compute capacity.
Ecosystem Status and Gigafactory Bidding
Despite the release of the Helion specification, the platform remains in an early stage of commercialization. AMD has not yet published pricing details, intends to contribute the specification to the Open Compute Project rather than having done so already, and has named no European system builders among its initial partners.
Meanwhile, European infrastructure acquisition moves forward through separate channels. The bloc has opened formal bidding for up to seven AI gigafactories valued at approximately €30 billion in total. Each facility is intended to house 100,000 chips sourced from major suppliers, including AMD, leaving room for rack-level agent architectures once the grid infrastructure catches up.
