Unveiled at the Open Compute Project summit in San Jose, Helion represents AMD's answer to a different workload than traditional AI infrastructure. Rather than focusing on model training, this rack blueprint emphasizes efficiency in deploying and running inference agents, with partner systems expected to launch in November.

The architecture comes in two variants. The CPU-only configuration, designed for hyperscale operators, accommodates up to 32,768 cores—achieved through 128 sockets of AMD's 256-core Venice chip built on Zen 6 architecture, available in either 400W dense or 600W balanced blade configurations. A second variant combines CPUs with GPUs for enterprise deployments, supporting a maximum of 16 Instinct MI350P accelerators distributed across two eight-GPU servers that operate independently without scale-up fabric.

This differs markedly from AMD's Helios rack, which remains the company's flagship training platform, equipped with 72 MI455X accelerators and 31 terabytes of HBM4 memory. Helion occupies a different tier in AMD's portfolio.

A particularly noteworthy feature of the enterprise configuration involves a token router that intelligently directs model inference requests either toward frontier models hosted in cloud environments or to open-source models running locally on the rack. According to Dan McNamara, who leads compute and enterprise AI efforts at AMD, this capability enables partners to optimize their use of available power resources.

Power efficiency forms the central narrative around Helion's design philosophy. Matt Kimball at Moor Insights and Strategy characterized the fundamental objective as maximizing the number of agents deployable per watt consumed. This focus proves particularly relevant for sovereign AI operators, positioning Europe squarely within the target market.

European power constraints drive timing

The launch arrives amid mounting electricity pressures across European infrastructure. Denmark's grid operator halted new connection approvals in March, with approximately 60GW worth of projects awaiting connection against peak demand hovering near 7GW. Similar constraints extend beyond Scandinavia.

Brussels responded by adopting a Cloud and AI Development Act in June, establishing an objective to expand the bloc's data centre capacity by at least threefold within a seven-year window. Data centre operators across the region echo consistent messaging: in a survey encompassing 300 decision-makers from the UK, Ireland and the Nordic countries, power availability emerged as the primary barrier to establishing sovereign capacity.

Specification and partnership status

Helion remains in relatively early stages despite its public unveiling. AMD has not yet announced pricing for systems built on the specification. The company intends to contribute the design to the Open Compute Project rather than having already completed that process. Notably, no European system builder appears among the named partners at this stage.

European investment in AI infrastructure continues regardless. The European Union has initiated a bidding process for up to seven gigafactories, allocating approximately €30B in total funding, with each facility designed to manufacture 100,000 chips annually from suppliers including AMD.

Source: The Next Web