衡量谷歌级规模AI服务交付的环境影响

Cooper Elsworth, Keguo Huang, David Patterson, Ian Schneider, Robert Sedivy, Savannah Goodman, Ben Townsend,Parthasarathy Ranganathan, JeffDean, Amin Vahdat, Ben Gomes, and James Manyika Google, Mountain View, CA, USA Abstract The transformative power of AI is undeniable—but as user adoption accelerates, so does the need to understand andmitigate the environmental impact of AI serving. However, no studies have measured AI serving environmental metricsin a production environment. This paper addresses this gap by proposing and executing a comprehensive methodologyfor measuring the energy usage, carbon emissions, and water consumption of AI inference workloads in a large-scale,AI production environment. Our approach accounts for the full stack of AI serving infrastructure—including activeAI accelerator power, host system energy, idle machine capacity, and data center energy overhead. Through detailedinstrumentation of Google’s AI infrastructure for serving the Gemini AI assistant, we find the median Gemini Appstext prompt consumes 0.24 Wh of energy—a figure substantially lower than many public estimates. We also show thatGoogle’s software efficiency efforts and clean energy procurement have driven a 33x reduction in energy consumptionand a 44x reduction in carbon footprint for the median Gemini Apps text prompt over one year. We identify that themedian Gemini Apps text prompt uses less energy than watching nine seconds of television (0.24 Wh) and consumesthe equivalent of five drops of water (0.26 mL). While these impacts are low compared to other daily activities,reducing the environmental impact of AI serving continues to warrant important attention. Towards this objective, wepropose that a comprehensive measurement of AI serving environmental metrics is critical for accurately comparingmodels, and to properly incentivize efficiency gains across the full AI serving stack. 1Introduction consumed by idle systems provisioned for reliability and lowlatency, and the full data center overhead as captured by thePower Usage Effectiveness(PUE) metric [9]. The missing con-sensus on the energy-consuming activities to include in themeasurement—known as themeasurement boundary—has ledto published estimates for similar AI tasks varying by an or-der of magnitude. A lack of agreed upon methodologies mayhave contributed to a lack of first-party data when it is neededmost [10]. Artificial intelligence(AI) is reshaping industries and dailylife, driven largely by the accelerating capabilities ofLargeLanguage Models(LLMs). While much of the initial focuson AI’s environmental impact rightly centered on the energy-intensive process of model training [1,2,3], the surge in publicadoption of generative AI applications has shifted attentiontoward the footprint of AI model inference and serving. Withthese AI models now serving billions of user prompts globally,the energy, carbon emissions, and water impacts associated withgenerating responses at scale represents a significant and rapidlygrowing component of AI’s overall environmental cost [4, 5]. ContributionsThis paper presents the energy, emissions, andwater impacts for a production AI product by establishing a com-prehensive framework to measure critical aspects of serving AIat Google’s scale. First, we propose a full-stack measurementapproach that accounts for all material energy sources. Second,we apply this methodology to Google’s Gemini Apps productto provide the first analysis of three AI serving environmentalmetrics: In response, several research efforts and disclosures haveemerged to quantify the per-prompt energy consumption of in-ference (Wh/prompt). Early work provided coarse estimates ofenergy consumption per prompt, relying on high-level assump-tions about hardware specifications and model parameters [6].More recently, initiatives like the the AI Energy Score [7] andthe ML.ENERGY [8] benchmarks have advanced the field byemploying direct empirical measurements. These frameworksaim to standardize energy transparency by benchmarking mod-els on specific tasks using consistent hardware. In addition,other studies have expanded the aperture to consider the carbonemissions and water consumption associated with serving AImodels. •Energy/prompt:the energy consumption required toserve an AI assistant text prompt•Emissions/prompt:themarket-based(MB) emissionsgenerated by grid electricity generation (including renew-able energy procurement), and the embodied emissionsof the AI accelerator hardware•Water consumption/prompt:the water consumed forcooling machines and associated infrastructure in datacenters Despite this progress, the field lacks first-party data from thelargest AI model providers. Based on decades of deployingsoftware at scale, Google has a unique perspective on the opera-tional realities of maintaining a large-scale, globally-distributedAI production fleet, and serving software products at scale—such as web search. Characterizing and optimizing the environ-me