Four Mac Studios combined into a cluster with 1.5 TB of memory – launched an AI that doesn't fit into a single graphics card

The author built a cluster of four Mac Studio computers with 1.5 TB of unified memory provided by Apple, using Exo 1.0 and RDMA over Thunderbolt 5. The cluster achieved 3.7 teraflops in HPL and successfully ran large AI models like Kimi K2 Thinking (600+ GB) and DeepSeek V3.1, showcasing the benefits of combined memory for heavy AI workloads.

Apple NVIDIA Exo Moonshot AI DeepSeek Alibaba/Qwen

Huatai Securities: AI Agents Accelerate Demand for Computing Power and Memory, Speeding Up Supply Chain Localization

A research report from Huatai Securities suggests that the AI industry is shifting from large model pre-training to AI Agent commercialization in 2026, driving inference computing demand into an accelerated growth phase. The report highlights three main investment themes: computing power chain, power and passive components, and self-sufficiency. Additionally, South Korea's Ministry of Economy and Finance stated it will maintain close communication with the U.S. Treasury to ensure foreign exchange market stability.