AMD发布面向Agentic AI的Helion开放机架参考架构，合作伙伴涵盖鸿海、纬颖、广达云、戴尔、HPE、超微等，产品2026年11月才上市

AMD has introduced Helion, an open rack architecture designed for agentic AI workloads that could change how data centers allocate compute, power, and cooling. The system targets hyperscale and enterprise customers building AI agents that need continuous execution, tool use, and workflow coordination.

Open rack design targets shifting AI workloads
The launch underscores how quickly the infrastructure market is adapting to agentic AI, which moves beyond simple model inference toward persistent, multi-step processing. AMD said Helion is built around two configurations: a CPU-only rack and an integrated CPU-GPU system, giving operators more flexibility depending on workload demands.
The CPU-only version uses sixth-generation AMD EPYC processors and supports up to 32,768 CPU cores. AMD said that setup is aimed mainly at hyperscale data centers that want dedicated compute resources for AI agents while keeping model inference on separate infrastructure.
The integrated version combines EPYC CPU nodes, Instinct MI350P GPU nodes, and general-purpose computing resources in one platform. AMD said this option is intended for enterprises that want a more unified environment for deploying agentic AI applications without rebuilding their entire IT stack.
AMD Helion is based on Open Compute Project specifications and extends the company's open-platform approach used in its Helios rack-scale systems. The company said customers can adjust CPU density, GPU acceleration, power use, and cooling choices based on workload requirements. Both air-cooled and liquid-cooled versions are planned.
Dan McNamara, senior vice president and general manager of AMD's Compute and Enterprise AI, said agentic AI includes personal agents, enterprise applications, code execution, and complex workflows, each with different infrastructure needs. He said computing and infrastructure must therefore be adjustable to the workload, and that Helion combines EPYC performance, open software, and a flexible rack design to help partners build systems around customer requirements.
AMD also said it plans to submit Helion specifications to the Open Compute Project, which could help broaden industry adoption over time. AMD is developing the rack configurations with ecosystem partners and expects the configurations to become available in November 2026.
Partners signal broader competition in AI infrastructure
AMD's partner list suggests that competition in AI infrastructure is widening beyond chips to include rack integration, cooling, networking, and deployment services. The company is working with Foxconn, Wiwynn, Quanta Cloud Technology (QCT), Dell Technologies, Hewlett Packard Enterprise (HPE), and Supermicro on the architecture.
Foxconn said it will use its rack-scale integration and global manufacturing capabilities to support the open design. The company said it will integrate sixth-generation EPYC processors and flexible CPU-GPU configurations for hyperscale and enterprise deployments.
Wiwynn said it will focus on rack-scale integration and advanced cooling technologies, while QCT said the open blueprint will help it tailor system designs to customer requirements. Dell said it plans to add the Helion reference design to its portfolio and combine it with Dell PowerRack networking and storage equipment. HPE said it will continue work with AMD on open standards and rack-scale systems, including air cooling, direct liquid cooling, and networking architectures.
Supermicro said it will use its Data Center Building Block Solutions to support an AMD Helion rack-scale architecture compliant with Open Rack version 3 (ORv3) specifications. The company said the effort will combine rack design, system integration, liquid cooling, software, and deployment services to speed production use.
Cisco Systems and Lenovo Group have also indicated support for the architecture, with Cisco focusing on full-stack infrastructure and connectivity, and Lenovo on hybrid AI deployment flexibility.
AMD's approach reflects a broader market shift as AI systems move from isolated model calls to continuous, collaborative tasks that can run across multiple agents. In that environment, the design of the rack itself, along with cooling and power management, becomes as important as the processors inside it. Whether Helion gains traction will depend on how quickly partners ship products, how customers deploy them, and how well the systems perform on energy and workload efficiency.