AMD is partnering with data center infrastructure specialist Celestica to launch a complete rack-scale AI platform targeting large cloud and enterprise deployments by late 2026. The "Helios" system combines next-generation AMD Instinct MI450 Series GPUs with Celestica-designed networking switches in an open standards architecture.
Celestica will handle research, design, and manufacturing of scale-up networking switches for the platform, which follows Open Compute Project (OCP) Open-Rack-Wide specifications. These switches use advanced networking silicon to interconnect AMD's upcoming MI450 accelerators using Ultra Accelerator Link over Ethernet technology for high-speed communication within AI clusters.
The collaboration aims to address what both companies see as a growing need for standardized AI infrastructure that can be deployed quickly while improving supply chain reliability.
"Deploying AI at scale requires infrastructure that can be delivered quickly, consistently, and with the performance customers expect,"
said Steven Dorwart, senior vice president and general manager, Hyperscalers, Celestica.
AMD executive vice president and general manager of the Data Center Solutions Business Group Forrest Norrod described Helios as "a new blueprint for AI infrastructure" designed for the performance and flexibility demands of next-generation workloads. The platform targets cloud providers, enterprise data centers, and research institutions looking to deploy large-scale AI training clusters.
Celestica has been expanding its AI-focused hardware portfolio throughout 2025 and early 2026 with products including high-density storage systems and multiple AI-optimized switch families. The Helios collaboration represents their move into full rack-level platforms alongside GPU acceleration.
The open standards approach contrasts with proprietary rack-scale solutions currently dominating the market. By building on OCP specifications and UALoE connectivity, AMD and Celestica aim to offer customers more flexibility in deployment options while potentially reducing vendor lock-in concerns.
Availability is scheduled for late 2026, positioning Helios to compete directly against established rack-scale offerings from Nvidia and other AI infrastructure providers during what to be peak demand for large language model training capacity.
