AMD distinguishes Helios production from the planned Helion agent rack

Helios moves into customer deployments

AMD's OCP 2026 keynote report describes Helios production ramp-up with customer deliveries, deployments and workload validation underway. The rack integrates Instinct GPUs, EPYC processors, Pensando networking and the ROCm software ecosystem.

AMD says it co-engineered the system with Meta and the Open Compute Project community. It reports initial systems installed and running at OpenAI, Meta, Microsoft and Oracle. These are statements from AMD's announcement, rather than an independently measured performance comparison across those deployments.

Helion is a separate planned configuration

The report also describes Helion, an agent-focused rack being developed with ecosystem partners. AMD plans availability in November 2026. That timing distinguishes it from the Helios systems the company says are already running.

Helion includes CPU-only configurations for hyperscale deployments and integrated CPU/GPU configurations for enterprises. An integrated model router directs requests across frontier and open-source models. AMD says partners can tailor configurations to workload, power requirements and existing infrastructure, and that it plans to contribute the architecture to OCP.

The distinction reflects different demands inside an agent workflow. Model inference can be only one step alongside database queries, business applications, code execution and coordination. The announcement presents rack design and routing as ways to match resources to those demands.

Evaluate the system against a real workload

Franklin analysis: an infrastructure comparison should start with the work a rack must execute. Record model-serving demand separately from the CPU work and application interactions around it, then test the routing and deployment assumptions against that mix.

Customer installation, planned availability and measured workload performance are different evidence. Keep each visible when comparing options. AMD's report establishes its deployment account and roadmap; it does not supply a universal cost or throughput result for every agent workload.