AMD introduced its Helios rack-scale AI system at the Advancing AI 2026 conference on July 23. The solution integrates 72 Instinct MI455X GPUs, 18 6th Gen EPYC Venice CPUs, Pensando networking hardware, and the ROCm software stack into a single unit.
The system achieves a peak performance of up to 2.9 exaflops in FP4 and includes 31 TB of HBM4 memory. AMD states that Helios provides up to 30% more tokens per dollar compared to the Nvidia Rubin NVL72 rack. The hardware targets frontier model training and large-scale inference for agentic AI workloads.
Major partners have committed to the platform: Anthropic will deploy up to 2 GW of MI450-series GPUs, while OpenAI and Meta have each signed 6 GW multi-generational agreements. Microsoft and Oracle are also among the initial customers for the new rack solution.
Shipments are scheduled to commence in the second half of 2026, with volume increasing throughout 2027. Additionally, AMD open-sourced its Optima solution for internal AI process optimization and reported a fourfold increase in energy efficiency between 2024 and mid-2026.
AMD projects the server CPU market will exceed $200 billion by 2030, while Bank of America forecasts the sector will surpass $170 billion. The company aims for a 20-fold improvement in AI energy efficiency by 2030.
