Workflow
GPU告急!亚马逊自建“调度帝国”
AmazonAmazon(US:AMZN) 半导体芯闻·2025-04-22 10:39

Core Viewpoint - Amazon is facing significant challenges in acquiring sufficient AI chips, particularly GPUs, which has led to delays in its retail projects. In response, the company has initiated a comprehensive internal reform to manage and allocate its GPU resources more effectively [1][2]. Group 1: GPU Resource Management - Amazon launched "Project Greenland" in July 2024 to create a centralized GPU capacity pool for better management and allocation of limited GPU resources [1][3]. - The company has implemented strict approval processes for GPU requests, requiring detailed data and ROI justification for each project [3][4]. - A formal set of principles, or "creed," has been established for GPU allocation, emphasizing high ROI, selective approval, and efficiency in decision-making [4][5]. Group 2: Investment and Financial Impact - Amazon's retail business has invested approximately $1 billion in AI projects, which have generated an estimated $2.5 billion in operating profit and saved around $670 million in variable costs [7][8]. - The retail department plans to increase its total spending on AWS cloud infrastructure to about $5.7 billion in 2025, up from $4.5 billion in 2024 [8]. Group 3: AI Project Implementation - Over 160 AI-driven projects are currently being developed within Amazon, including shopping assistants and automated fraud investigation services [7]. - The company is focusing on enhancing its AI capabilities to optimize logistics and customer service, which are critical for maintaining competitive advantage [7][8]. Group 4: Future Projections - Internal forecasts suggest that GPU supply constraints will ease by 2025, allowing for increased resource allocation to high-priority projects [2][10]. - Amazon's self-developed AI chip, Trainium, is expected to meet the retail department's GPU needs by the end of 2025 [10].