AMD unveils Helios rack-scale AI system to challenge Nvidia

AMD unveils its Helios rack-scale AI system, directly challenging Nvidia's market dominance. The system, developed in partnership with Cerebras, utilizes a "disaggregated inference" architecture and claims up to 30% better performance per dollar than Nvidia's offering. Major AI labs including Microsoft, OpenAI, Meta, and Anthropic have already signed on to deploy the new hardware, signaling a significant competitive push by AMD.

AMD-Cerebras partnership

• ▪In the joint system, AMD's Helios processes requests while Cerebras' wafer-sized chip generates responses
• ▪The AMD-Cerebras partnership will bring the Helios system into Cerebras' data centers later in 2026
• ▪AMD is partnering with chip startup Cerebras on a new approach to AI inference

Disaggregated inference architecture

• ▪The collaboration focuses on "disaggregated inference," which splits AI workloads across different specialized hardware
• ▪Nvidia and Amazon Web Services are also pursuing disaggregated inference setups to improve efficiency and lower costs
• ▪Analysts from UBS noted in June that limitations of current architectures are driving a shift toward disaggregated inference

Helios rack-scale system

• ▪Customers planning to deploy the Helios system include OpenAI, Meta, Microsoft, Oracle, and Anthropic
• ▪On July 23, 2026, AMD announced new products for data centers, including the Helios rack-scale AI system
• ▪AMD and AI lab Anthropic announced a partnership to deploy up to two gigawatts of GPUs via the Helios system

Competition with Nvidia

• ▪AMD also introduced its Venice-X CPU for data centers, which is expected to launch in 2027
• ▪AMD claims Helios delivers up to 30% more inference tokens per dollar than Nvidia's Vera Rubin NVL72 rack
• ▪AMD is attempting to capture market share from Nvidia in the data center chip sector, especially for inference computing
• ▪AMD CEO Lisa Su projected that the AI accelerator market will reach about $1.4 trillion by 2030
• ▪AMD's Helios system is designed to compete with Nvidia's Vera Rubin and Grace Blackwell rack-scale systems