Artificial Intelligence and Autonomous Systems
The Global Centre for Risk and Innovation (GCRI)
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.
Compute and energy · Model registry · Release gate
| M-904 triage assistant | Public service | High | 19 of 24 suites | Blocked |
| M-890 flood nowcast | Science | Medium | 10 of 12 | Released |
| M-955 document search | Internal | Low | 7 of 8 | Released |
| M-242 agent planner | Agentic | High | 13 of 16 | Evaluating |
vs release threshold
Fig.Model registry. Design wireframe with placeholder names and illustrative values, not a record or a live system.
AI energy use is the electricity consumed to train and run AI models, mostly in data centres that also use water for cooling. Demand is growing and concentrating regionally, straining grids and water resources, while efficiency improves. Disclosures are inconsistent, so estimates rest on stated assumptions.
Night-time lights, fires and gas flares, sunshine and heat, the signals of energy use, supply and demand, framed on East Asia; each layer is a dated NASA product.
The interactive map needs JavaScript; the layer list and sources are below.
Layers 2
Energy now
Keyboard, with the map focused
Hold Ctrl or ⌘ and scroll to zoom
Latest available
NASA GIBS / EOSDIS · ASTER GDEM (METI, NASA) · Natural Earth
Loading the latest NASA imagery.
Drag to move. Ctrl or ⌘ and scroll, pinch or the buttons to zoom; hover for the values under the pointer; reset returns to the region.
Night-time lights, fires and gas flares, sunshine and heat, the signals of energy use, supply and demand, framed on East Asia.Satellite and model products as NASA publishes them, each with its own date; indicators, not assessments of any place; claim ceiling INTEGRATED.
Layers and dates
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.
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.
Large data centres need connections comparable to industrial plants. Clusters in one region can outpace grid and generation build-out.
Evaporative cooling and electricity generation consume water. Where water is scarce, data centre siting competes with other uses.
Hardware and software efficiency improve quickly, but lower costs can raise total use. Net energy demand depends on both.
Each term as this page uses it. Where a standard-setting body defines the term, the wording follows that body.
| Term | Definition |
|---|---|
| Inference | Running a trained model to produce outputs for new inputs. |
| Power usage effectiveness | The ratio of total data centre energy to IT equipment energy. |
| Training compute | The total computation used to train a model, often measured in floating point operations. |
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Data centre electricity consumption | terawatt hours per year | annual | Energy agency estimates | ESTIMATED |
| Power usage effectiveness | ratio of facility to IT energy | annual | Operator disclosures | DECLARED |
| Water usage effectiveness | litres per kilowatt hour | annual | Operator disclosures | DECLARED |
| Grid connection requests from data centres | megawatts requested | quarterly | Grid operator queues | OFFICIAL |
| Training compute of notable models | floating point operations | per model | Compiled model databases | ESTIMATED |
| Energy per inference | watt hours per query | per measurement | Measurement studies | ESTIMATED |
The topic as a computational graph: named sources feed the methods, an evidence review gates the outputs, and the decisions stay with the authorities that hold them.
Fig. 02Nodes carry their type, two metric rows and a state; connectors show what feeds what; dashed frames name the owners; the timeline is the cadence. Design of the pipeline, not a record of a run.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| IEA Energy and AI analysis and data centre electricity estimates | Global and regional data centre electricity demand | Licence per product (to check) |
| Epoch AI data on notable models and compute | Training compute and hardware trends | Open, CC BY 4.0 |
| TOP500 and Green500 lists | Performance and energy efficiency of supercomputers | Public, TOP500 terms |
| Grid operator interconnection queues and planning reports | Requested connections and load forecasts | Public records, terms per operator |
| Operator sustainability reports | Energy, water and emissions disclosures | Public documents, terms per operator |
Artificial Intelligence and Autonomous Systems prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Grid operators and energy regulators | How to plan generation and connections for data centre load | Load analyses; connection queues |
| Water authorities and planners | Where data centre water use is acceptable | Water footprint evidence |
| Organizations deploying AI | How to reduce the footprint of their AI use | Estimation methods; efficiency evidence |
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.
Estimating energy per training run and per query, using high-performance, cloud and edge computing. Providers disclose little, so estimates carry wide uncertainty.
Tracing chips, materials and facilities to identify concentration and single points of failure.
Projecting demand under different assumptions about efficiency and adoption, with probability and statistics. Projections are scenarios, not forecasts.
Assessing which researchers and institutions can obtain compute, and on what terms.
The current service level is R0: information and enquiries. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.