谷歌CEO「劈柴」亲自下场分芯片，930亿美元填不饱「算力饥荒」

Core Viewpoint - Google is facing a "compute famine" despite its vast resources and plans to invest $91 to $93 billion in capital expenditures this year, leading to internal conflicts over chip allocation among various departments [1][25][26]. Group 1: Internal Dynamics and Resource Allocation - A new executive committee has been formed to decide how to allocate limited compute resources among Google Cloud, search, and DeepMind [2][13]. - The committee includes key figures such as Google Cloud CEO Thomas Kurian and DeepMind CEO Demis Hassabis, indicating a significant power restructuring within the company [4][8][13]. - The committee aims to simplify decision-making and ensure equitable distribution of resources among departments, which previously struggled to reach consensus [32][34]. Group 2: Strategic Importance of Compute Resources - The three core lifelines for Google are future AI development, growth through Google Cloud, and sustaining its extensive product matrix [15][20][22]. - Google Cloud is seen as a major growth engine that requires substantial compute power to serve clients and maintain expansion [20]. - The need for top-tier AI models necessitates vast compute resources for continuous iteration, making it critical for Google's competitive positioning [17][18]. Group 3: Financial Implications and Challenges - Despite plans for significant capital expenditure, the long lead times for building data centers and manufacturing chips mean that immediate relief from the compute shortage is unlikely [27][28]. - Google’s capital expenditure in 2023 was only $32 billion, which is considered conservative given the AI boom, contributing to the current compute challenges [29][30]. - CFO Anat Ashkenazi acknowledged that supply-demand imbalances are expected to persist into 2026, indicating ongoing challenges for the company [31]. Group 4: Execution-Level Dynamics - At the execution level, the focus shifts to immediate revenue generation, leading to prioritization of departments that can deliver the most profit [48]. - In DeepMind, resource allocation is complex, with some researchers enjoying privileges that allow them to access multiple compute pools, while others must navigate a more competitive environment [50][54]. - The grassroots level sees a culture of resource sharing and negotiation among researchers, turning compute power into a "hard currency" that relies on personal relationships and exchanges [57][59].