AMD gains attention as Helios targets AI inference

Nvidia remains the clear leader in AI chips through its CUDA ecosystem, training position and broad data center stack, but AMD is positioning itself around the industry’s shift toward inference. Its Helios rack-scale AI system combines 72 Instinct MI455X GPUs, 18 sixth-generation EPYC “Venice” CPUs, Pensando networking equipment and ROCm software. AMD says Helios can deliver up to 30% more inference tokens per dollar than its main competing solution under specific internal test conditions.

AMD has lined up several large customers and partners for the platform. OpenAI plans to bring related systems online starting in the fourth quarter of 2026 and accelerate deployment in 2027, Anthropic plans to use up to 2 gigawatts of MI455X GPUs, and Meta is testing Helios and the next-generation EPYC platform. AMD is also working with Cerebras on disaggregated inference and announced the acquisition of Taalas in August to strengthen inference chip capabilities.

Stanley Druckenmiller’s Duquesne Family Office opened a new AMD position as of the end of the second quarter of 2026 after exiting Intel, Micron Technology and Broadcom positions established in the first quarter. The AMD stake accounts for roughly 0.8% of disclosed assets. Technically, AMD closed at $456.75 on August 24, below its 20-day moving average of $479.21 and 60-day moving average of $507.78, leaving support around $450 as a key level for any rebound attempt.