GCRI · 人工智能与自主系统

算力与能源

Training and running AI requires specialised chips, data centres, electricity, cooling water and network capacity. Demand is growing quickly and concentrating in particular places, with effects on power grids, water resources and emissions. This page sets out how AI compute uses energy and water, how those uses are measured, and what evidence serves the grid operators, planners and policymakers who must accommodate them.

01画布

工作如何开展。

系统 · 围绕本主题示意

模型注册库

Compute and energy · Model registry · Release gate

一个模型在关口被拦截GCRI · 人工智能与自主系统

模型

M-904 triage assistant公共服务高19 of 24 suites已阻止
M-890 flood nowcast科学中10 of 12已发布
M-955 document search内部低8 中的 7已发布
M-242 agent planner智能体式高13 of 16评估中

模型卡

预期用途
分诊辅助,由人决定
不在范围内
资格决定
已知局限
在罕见案例上的准确率较低
负责人
模型所有者

事件日志

  1. D-28INC摘要中出现虚构引用
  2. D-6MIT已增加检索依据
  3. D-4INC通过上传文件的提示注入
  4. D-1MIT工具权限已收窄

Evaluation scores

vs release threshold

Task accuracy71 of 85 required稳健性71 of 75 requiredFairness gap (lower is better)3 of 5 requiredRefusal on misuse91 of 95 required
Model registry: the screen’s structure, terms and states for Compute and energy.示例名称和示意数值;是设计方案,并非记录或实时系统。

图Model registry: example names and illustrative values, not a record or a live system.

02问题

Compute and energy: how the risk develops.

Compute concentration, efficiency and demand. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.

机制

Training and inference both use energy

Training a large model uses a large amount of electricity once; inference uses less per query but runs continuously at scale, and can dominate over a model’s life.

机制

Data centres strain local grids

Large data centres need connections comparable to industrial plants. Clusters in one region can outpace grid and generation build-out.

机制

Cooling uses water

Evaporative cooling and electricity generation consume water. Where water is scarce, data centre siting competes with other uses.

机制

Efficiency and demand move together

Hardware and software efficiency improve quickly, but lower costs can raise total use. Net energy demand depends on both.

03地理

Energy now in East Asia.

Night lights, population density and land cover over East Asia, from NASA GIBS. Each layer carries its own date.

交互式地图需要JavaScript;图层列表和来源见下文。

图层 3

本页当前的能源状况

关于本图层
证据
观测
来源
VIIRS Black Marble(VNP46A2)(NASA LAADS DAAC)
提供方
NASA LAADS DAAC
许可协议
NASA 开放数据
年份版本
每日;卫星辐射亮度,已校正月光和大气影响
构建时的最新内容
2026-10-12
处理结果
约 500 米网格
不确定性
所示日期的卫星夜间辐射亮度;灯光是用电和停电的替代指标,并非对供应或获得电力情况的测量。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问
关于本图层
证据
估计值
来源
世界网格化人口 v4.11(NASA SEDAC(哥伦比亚大学 CIESIN))
提供方
NASA SEDAC(哥伦比亚大学 CIESIN)
许可协议
CC BY 4.0
年份版本
null;基于人口普查的估算,静态(2020)
处理结果
约 1 公里网格
不确定性
基于人口普查的 2020 年每平方公里人口估算,作为风险暴露背景展示;不是受任何危害影响的人数。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问
关于本图层
证据
模型生成(GIBS 发布的最新月份)
来源
MERRA-2(NASA GMAO / GES DISC)
提供方
NASA GMAO / GES DISC
许可协议
NASA 开放数据
年份版本
每月;再分析模型;GIBS 发布的最新月份
构建时的最新内容
2026-07-01
处理结果
约 2 公里网格
不确定性
再分析模型对陆地月平均入射太阳辐射的估算;是太阳能资源的指标,并非场址发电量评估。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问
关于本图层
证据
观测
来源
VIIRS NOAA-20 校正反射率(NASA LANCE)
提供方
NASA LANCE
许可协议
NASA 开放数据
年份版本
每日;卫星图像
构建时的最新内容
2026-10-14
处理结果
约 250 米网格
不确定性
卫星图像。卫星或模型栅格数据,按提供方发布的原样呈现。云层、条带缺口和反演误差会造成缺口和噪声;数值是指标,而非某地的测量值。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问
关于本图层
证据
模型生成(GIBS 发布的最新年份)
来源
MODIS Terra 和 Aqua 土地覆盖(MCD12Q1)(NASA LP DAAC)
提供方
NASA LP DAAC
许可协议
NASA 开放数据
年份版本
每年;卫星分类;GIBS 发布的最新年份
构建时的最新内容
2024-01-01
处理结果
约 500 米网格
不确定性
所示年份的卫星土地覆盖类别,含农田;该分类存在已知误差,并非农场登记册。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问
关于本图层
证据
观测(云层造成缺口)
来源
VIIRS NOAA-20(NASA LANCE)
提供方
NASA LANCE
许可协议
NASA 开放数据
年份版本
每日;卫星反演;云层造成缺口
构建时的最新内容
2026-10-14
处理结果
约 1 公里网格
不确定性
所示日期地表的卫星皮肤温度,是制冷需求的指标;不是气温,也不是高温预警。
经由以下途径到达
gibs.earthdata.nasa.gov,仅在您允许实时地图后访问

最新可用

NASA GIBS / EOSDIS · ASTER GDEM (METI, NASA) · Blue Marble (NASA) · © OpenStreetMap贡献者 (ODbL) · Natural Earth

正在加载最新的NASA影像。

拖动以移动。按住 Ctrl 或 ⌘ 并滚动、双指缩放或使用按钮缩放;悬停可查看指针下的数值;重置可返回该区域。

Energy now, framed on East Asia.NASA 发布的卫星和模型产品,各带自己的日期。它们是指标,而非对任何地点的评估。

数据:NASA GIBS · ASTER GDEM · Blue Marble · © OpenStreetMap 贡献者 · Natural Earth · NASA EONET · NASA/NOAA VIIRS · NASA JPL OPERA DIST-ALERT来源与许可

  • We acknowledge the use of imagery provided by services from NASA’s Global Imagery Browse Services (GIBS), part of NASA’s Earth Science Data and Information System (ESDIS). Colours redrawn from the published colour map; values averaged into hexagonal cells where shown (modified). NASA 开放数据
  • We acknowledge the use of imagery provided by services from NASA’s Global Imagery Browse Services (GIBS), part of NASA’s Earth Science Data and Information System (ESDIS). Colours redrawn from the published colour map; values averaged into hexagonal cells where shown (modified). NASA 开放数据
  • VIIRS Black Marble 夜间灯光,NASA 戈达德航天中心;由 NASA Level-1 与大气档案和分发系统(LAADS DAAC)分发;影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 NASA 开放数据
  • Center for International Earth Science Information Network (CIESIN), Columbia University. 2018. Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11. NASA Socioeconomic Data and Applications Center (SEDAC). 采用 CC BY 4.0 许可。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 doi:10.7927/H49C6VHW. CC BY 4.0
  • MERRA-2 再分析,NASA 全球建模与同化办公室;由 NASA GES DISC 分发;影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 NASA 开放数据
  • VIIRS 近实时产品(洪水、地表温度、积雪覆盖、校正反射率),NASA LANCE;影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 NASA 开放数据
  • MODIS 陆地产品(植被指数、总初级生产力、土地覆盖),NASA LP DAAC 和 LANCE;影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 NASA 开放数据
  • ASTER GDEM 是 METI 和 NASA 的产品(ASTER Global Digital Elevation Model V003,NASA LP DAAC);影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。引用格式:NASA/METI/AIST/Japan Spacesystems and U.S./Japan ASTER Science Team (2019). ASTER Global Digital Elevation Model V003. doi:10.5067/ASTER/ASTGTM.003. NASA 开放数据
  • 使用 Natural Earth 制作。免费矢量和栅格地图数据,见 naturalearthdata.com。 公有领域
  • © OpenStreetMap 贡献者(ODbL),经由 NASA GIBS(海岸线、边界、道路和地名)。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 ODbL
  • Blue Marble,NASA Earth Observatory;影像来自 NASA GIBS。颜色依据已发布的色标重绘;显示处的数值已按六边形网格平均(经修改)。 NASA 开放数据
  • 地球仪和平面地图上的背景信息:海岸线、边界、道路和地名来自 NASA GIBS 参考图层(OpenStreetMap 数据,ODbL,© OpenStreetMap contributors);Blue Marble 影像可作为可选底图;3D 地形和高程读数来自 ASTER GDEM 高程色彩指数(5 m 至 200 m 分级,为便于辨识而作了夸张处理)。边界按已发布的样式绘制,不代表对任何领土的立场。
  • 实时事件:NASA Earth Observatory Natural Event Tracker(EONET);事件仅供参考,并保留其原始来源链接。
  • Observed (EO): NASA/NOAA VIIRS (Suomi NPP). We acknowledge the use of imagery provided by services from NASA’s Global Imagery Browse Services (GIBS), part of NASA’s Earth Science Data and Information System (ESDIS).
  • Observed (EO): NASA JPL OPERA DIST-ALERT (contains modified Copernicus Sentinel data 2026). We acknowledge the use of imagery provided by services from NASA’s Global Imagery Browse Services (GIBS), part of NASA’s Earth Science Data and Information System (ESDIS).
  • NASA GIBS: We acknowledge the use of imagery provided by services from NASA’s Global Imagery Browse Services (GIBS), part of NASA’s Earth Science Data and Information System (ESDIS).
  • 过去一年的每日气温(ERA5,经由 Open-Meteo):包含经修改的 Copernicus 气候变化服务 2026 年信息。
  • 过去一年的每日气温和日照(NASA POWER):这些数据来自 NASA 兰利研究中心(LaRC)POWER 项目,该项目由 NASA 地球科学/应用科学计划资助。
  • 天气数据来自 Open-Meteo.com (https://open-meteo.com/), 依据 CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). 授权。
  • 可选图层(图层列表中共 10 个来源:实时、预报、观测和背景)启用后会在地图上和此处显示其提供方署名;预报为模型输出,衍生指数为估计值。

图层与日期

  • 夜间灯光,每日:VIIRS Black Marble(VNP46A2),NASA LAADS DAAC · 2026年10月12日
  • 人口密度,2020 年:人口:哥伦比亚大学 CIESIN,世界网格化人口 v4.11(2018),NASA SEDAC, CC BY 4.0 · 静态
  • 陆地地表入射太阳辐射,月均值:MERRA-2,NASA GMAO / GES DISC · 2026年7月
  • 真彩色:VIIRS NOAA-20 校正反射率,NASA LANCE · 2026年10月13日
  • 土地覆盖类型(IGBP),年度:MODIS Terra 和 Aqua 土地覆盖(MCD12Q1),NASA LP DAAC · 2024
  • 白天地表温度:VIIRS NOAA-20,NASA LANCE · 2026年10月13日

04跨领域

AI compute couples to energy and water.

Compute and energy analyses are designed for the 前瞻展望系列. The IEA’s Energy and AI report (2025) gives scenario ranges for data-centre electricity demand.

  • Measured energy and water use of computing facilities.
  • Grid connection studies.
05模拟器

在相互依存的网络中的一次级联。

A 危害足迹 takes out co-located assets in every layer; fragmentation, dependency between layers and overload of the survivors then spread the failure. The 3D view separates the layers by depth. Teaching model with stated illustrative values, not a prediction. The map on this page shows East Asia; the model itself is not tied to a place.

控制措施

每个图层的资产
18
危害足迹
占面积的 8%
依赖其他图层的资产
60%
备用容量(容差)
30%

读数

受危害冲击的资产
6%
仍在运行的资产
20%
初始损害的放大
14.3 times
结算步数
4
数字、方法与来源
步骤运行中的数据中心(%)运行中的网络(%)电力运行(%)
08994100
14444100
262278
362239
462233

方法

  1. 相互依存网络 Buldyrev 等(2010),Catastrophic cascade of failures in interdependent networks,Nature 464
  2. 过载级联 Motter and Lai (2002), Physical Review E 66
  3. 介数负载 Brandes (2001), Journal of Mathematical Sociology 25
  4. 多层网络 Kivela and others (2014), Journal of Complex Networks 2
  5. 基础设施相互依赖 Rinaldi, Peerenboom and Kelly (2001), IEEE Control Systems 21(6)
  6. 不确定性扇形图(参数蒙特卡罗) Spiegelhalter, Pearson and Short (2011), Visualising uncertainty about the future, Science 333
  7. 单因素敏感性(龙卷风图) Saltelli and others (2008), Global Sensitivity Analysis: The Primer, Wiley

数据

The model runs on the teaching values set by the controls. To calibrate it for a place, read it against these admitted layers and series, which this page’s map can draw:

  • 主要道路 Natural Earth
  • 铁路 Natural Earth
  • 海港 Natural Earth
  • 机场 Natural Earth
  • 城市范围 Natural Earth
  • 今晚的夜间灯光、停电(VIIRS 昼夜波段) NASA/NOAA VIIRS(Suomi NPP)· 登记 nasa.gibs.context-live-layers
  • 自然事件(NASA EONET) NASA EONET · 登记 nasa.eonet.events
级联在相互依存的网络中传播。示意性教学模型:它依据控件设定的数值运行,而非实测数据。方法:Interdependent networks (Buldyrev and others (2010), Catastrophic cascade of failures in interdependent networks, Nature 464);Overload cascades (Motter and Lai (2002), Physical Review E 66);Betweenness load (Brandes (2001), Journal of Mathematical Sociology 25)。
06指标

衡量什么,以及衡量得如何。

每个指标都注明其单位、频率、背后的来源类型以及数值将附带的证据标签。此处不显示数值:数值由计划依据其自身权限提供。

指标单位节奏来源类型标签
Data centre electricity consumptionterawatt hours per year年度Energy agency estimates估算
电能利用效率ratio of facility to IT energy年度Operator disclosures已声明
Water usage effectivenesslitres per kilowatt hour年度Operator disclosures已声明
Grid connection requests from data centresmegawatts requested每季度Grid operator queues已发布统计
Training compute of notable modelsfloating point operationsper modelCompiled model databases估算
Energy per inferencewatt hours per queryper measurementMeasurement studies估算
07孪生与级联模拟器

Model compute and energy as a system. Run the cascade.

Assembled from the Nexus Standards area profile 人工智能与自主系统风险标准: the hazards it names, the 系统 and sectors that carry the loss, the controls that dampen it and the data that observe it. Choose the components, pick an initiating hazard and step through the cascade. The Map view places the run on an example geography with open data. Every coupling weight is a stated assumption you can change. A teaching model, not a forecast, an assessment or a certification. To model a place of your own choosing, open the full System Builder.

灾害

  • 溃坝
  • Sea-ice loss
  • Coastal erosion (在下游引发)
  • 河流洪水(在下游引发)
  • Flash flood (set off downstream)
  • 大规模流离失所(在下游引发)
  • Biodiversity loss (set off downstream)
  • 大范围停电(在下游引发)
  • 供水短缺(在下游引发)
  • Arctic winter sea-ice collapse (set off downstream)
  • Polar vortex disruption (set off downstream)

控制措施

  • NS-0 章程与原则
  • NS-10 政策和条款(机器可读的义务:定义、约束、评估和执行)
  • NS-11 验证与监测(证据、证明、零信任符合性观测)
  • NS-12 软件、模型和 AI/ML 制品(供应链、模型卡、数据集说明)
  • NRS-H100 危害类别:机器人、自主化与自主武器(NHS-1 tech.autonomy-weapons)
  • NRS-H53 危害类别:信息物理系统中断(NHS-1 cyber.cyber-physical)
  • NRS-H61 危害族:级联失效(NHS-1 sys.cascade)
  • NRS-H113 危害类别:金融体系脆弱性(NHS-1 soc.financial-system)
  • NRS-A8 面向人工智能的风险数据与模型打包
  • NRS-H54 危害类别:信号与平台中断(NHS-1 cyber.platform)

观测

  • OSV open source vulnerability database (listed for reference)
  • Our World in Data datasets (listed for reference)
  • Research and development expenditure as a proportion of GDP

示例地理范围

  • North America: 243 hexagonal cells
  • 今夜夜间灯光与停电(VIIRS 昼夜波段),NASA GIBS
  • 穿透云层的地表水,30 m(OPERA DSWx-S1 雷达),NASA GIBS
  • 过去半小时降雨率(IMERG),NASA GIBS
  • Vegetation disturbance alerts, 30 m (OPERA DIST-ALERT), NASA GIBS
08来源

证据来自何处。

具名公共数据集与机构记录,各自注明许可类别。依协议获得的来源仅按其条款使用,绝不再次发布。

来源它提供什么许可类别
IEA Energy and AI analysis and data centre electricity estimatesGlobal and regional data centre electricity demandLicence per product (to check)
Epoch AI data on notable models and computeTraining compute and hardware trends开放,CC BY 4.0
TOP500 and Green500 listsPerformance and energy efficiency of supercomputersPublic, TOP500 terms
Grid operator interconnection queues and planning reportsRequested connections and load forecastsPublic records, terms per operator
Operator sustainability reportsEnergy, water and emissions disclosuresPublic documents, terms per operator
09部门特征

The sector signature of compute and energy.

The sectors compute and energy touches and how strongly they depend on each other: what a study has to observe and model together.

选择角点 · 悬停连线6 / 15 项关系

签名

能源及其合作伙伴

能源为基础设施提供交通和电信用电(0.8),为水提供泵送和处理用电(0.5),为粮食提供化肥、机械和冷链(0.5)。

W · E · I · F · H · B

部门
4 共 6
关系
6 共 15
耦合
0.46 占总量份额

当前最强的关系

  • 能源 → 基础设施 · 为交通和电信供电0.8
  • 水 → 能源 · 冷却水与水电0.6
  • 水 → 粮食 · 灌溉0.8
  • 基础设施 → 水 · 管道、水坝和处理设施0.6
典型城市区域的教学数值,不是任何地点的测量值。耦合是所涉关系强度的总和,以全部十五种关系的占比表示。
10决定

由谁决定,以及证据服务于什么。

人工智能与自主 系统 prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

谁来决定问题本页面服务的证据
Grid operators and energy regulatorsHow to plan generation and connections for data centre loadLoad analyses; connection queues
Water authorities and plannersWhere data centre water use is acceptableWater footprint evidence
Organisations deploying AIHow to reduce the footprint of their AI useEstimation methods; efficiency evidence

该主题产生的产出

  • Energy and water footprint estimation methods for AI workloads
  • Regional data centre load analyses
  • Efficiency trend summaries
  • Evidence notes for grid and water planning
11方法与证据

How compute and energy is modelled, measured and tested.

模型、其假设、数据和不确定性,以科学家或工程师会核查的形式呈现。结构遵循已发表的方法。所取数值仅为示意,不是研究发现。

系统模型

The couplings between the systems in compute and energy: signs of influence, a reinforcing and a balancing loop, and cross-links where stress jumps.

++−+人工智能数据社区经济R 增强B 平衡
Causal loop diagram from the page’s systems; signs show the direction of influence. Illustrative structure for a system dynamics or network model, not estimated couplings.
FAIR 指导原则 · GO FAIR 和 FORCE11W3C PROV · 万维网联盟关于一致处理不确定性的指南 · IPCCCEOS 分析就绪数据 · 地球观测卫星委员会
12核心问题

What researchers ask about compute.

  • How concentrated are chip manufacturing, cloud capacity and frontier model training, and what follows from that?
  • How much energy and water do training and inference use, and how is that changing?
  • How much do efficiency gains offset growth in demand?
  • Who has access to compute for research and public-interest uses?
13机制

Concentration, efficiency and demand.

Advanced chip fabrication is concentrated in very few facilities, and a small number of cloud providers supply much of the compute for AI. That concentration creates shared exposure to disruption, from natural hazards to export controls, and shapes who can build large models.

Hardware and algorithms become more efficient over time, but cheaper computation tends to increase total use, an effect known as the rebound or Jevons effect. Data centres draw large, steady electricity loads that can strain local grids and, where they use evaporative cooling, water supplies. Inference, the everyday use of models, can come to outweigh training in total energy use.

14方法与数据

Measuring compute and its footprint.

  1. 01

    Energy accounting

    Estimating energy per training run and per query, using 高性能、云与边缘计算. Providers disclose little, so estimates carry wide uncertainty.

  2. 02

    Supply-chain mapping

    Tracing chips, materials and facilities to identify concentration and single points of failure.

  3. 03

    情景分析

    Projecting demand under different assumptions about efficiency and adoption, with 概率与统计. Projections are scenarios, not forecasts.

  4. 04

    Access studies

    Assessing which researchers and institutions can obtain compute, and on what terms.

15方法

如何研究,以及方法无法说明的内容。

  1. 高性能计算、云计算与边缘计算公共研究的算力获取 · 工作负载的能耗测量在此应用
  2. 人工智能与机器学习分布偏移分析 · 针对特定用途的微调 · 可解释性方法在此应用
  3. 概率与统计基准分数的不确定性 · 校准评估 · 审计抽样在此应用
16Where this page’s references

Where this page’s references lead.

阅读本页时所对照的机构和文件。状态和版本会变化;每个链接都指向发布机构。

17问题

问题与解答。

How much energy does AI use?
Estimates vary widely. Data centres use a growing share of electricity, and AI is a fast-growing part of that, according to the IEA, but precise figures depend on assumptions and disclosure.
What is power usage effectiveness?
The ratio of a data centre’s total energy use to the energy used by its IT equipment; values closer to one mean less overhead for cooling and power conversion.
Why does AI use water?
Mainly for cooling data centres, especially with evaporative systems, and indirectly through the water used to generate electricity.
Will efficiency gains reduce AI’s energy use?
Not necessarily. Efficiency lowers the cost per task, which can increase total demand. The net effect depends on adoption and use.
Why does water matter for data centres?
Many use evaporative cooling, which consumes water. In water-stressed areas, this competes with other uses, linking AI to local water risk.
Does the platform allocate compute?
No. Compute allocations for Consortium projects are a separate matter of Foundry and the annual programme, and open only when their terms exist.
Can energy figures be trusted?
Published estimates vary with assumptions about hardware, utilisation and location. The platform reports ranges with their sources and assumptions.
报告贡献

从本页贡献证据。

报告 is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

GCRI人工智能与自主系统指数级技术Labs 证据
人工智能与自主系统

带来一种方法、一个数据集或一个问题。

信息和咨询现已开放。每项咨询都会获得一个编号,并告知下一步;申请、付款、算力分配和注册只有在相应团队、条款和审查能力就绪时才会开放。

The Global Centre for Risk and Innovation

Science · evidence · measurement

Not for profit

Our articles, maps and twins are free to read and free to republish. Readers and partners keep it that way.

Partner with us

Who funds us

GCRI is a not-for-profit institution. To ask who funds a piece of work, use the funding questions form.

Independence

Editors and the expert reviewers named on each article decide what we publish. Funders and partners do not review or approve articles before publication.

Licence

Articles: CC BY 4.0. Figures, maps, twins and data: CC BY 4.0. Other uses: ask through the permissions form.