A new open-rack design makes it practical to build data-center racks specifically for agentic AI.
It packages CPU, GPU and general-purpose compute into a single blueprint that follows the Open Compute Project rack specifications. That means different vendors can build compatible racks instead of each team reinventing the integration work.
AMD calls the design Helion. It ships as two reference configurations: a CPU-only option built around sixth-generation EPYC processors, and a mixed option that combines EPYC CPU nodes with Instinct MI350P GPU nodes. AMD says Helion was developed with ecosystem partners and is planned to be available in November 2026.
Why this matters: agentic AI workloads are easiest to run when compute, networking and cooling are designed together. Helion treats the rack as a product rather than a custom engineering project, which shortens procurement and deployment cycles for large-scale services.
Think of Helion as a shipping container built for AI: everything inside is arranged, connected and cooled to run many processors as one coordinated system. One Helion configuration can hold about 32,000 CPU cores, and the mixed version pairs those cores with GPU nodes so large agentic services can both reason and act.
What changes now: hyperscalers get a repeatable rack design that lowers time and cost to deploy agentic workloads. System integrators and enterprises gain a standard they can buy or build to support those same workloads.
Practical limit: Helion is a data-center product. It does not make frontier agentic AI something a small team runs on a laptop, but it does remove a major barrier for organizations that already operate at rack or pod scale.
What to watch next: will customers standardize on Helion, and will independent benchmarks prove the design delivers better performance per watt or lower total cost? The answer will determine whether Helion becomes the new baseline for agentic AI infrastructure.
