ENERGY + MATERIALSgeneration, grids, fuels, minerals, manufacturing, waste
HUMAN SYSTEMSmarkets, infrastructure, institutions, behavior, inequality
Optimizing one layer can shift cost into another.
The planetary context
AI enters an Earth system that is already under growing pressure.
1.43°Cestimated 2025 global temperature above the 1850-1900 averageWMO
11 years2015-2025 were the hottest eleven years on recordWMO
6 of 9planetary boundaries assessed as transgressed in a 2023 framework updateSCIENCE ADVANCES
~1Mspecies estimated to face extinction risk in the IPBES global assessmentIPBES
Keep a double ledger
Every AI idea has a benefit pathway and a footprint pathway.
ENABLING EFFECTS
better forecasts and earlier warnings
more efficient grids, buildings, transport, and industry
faster scientific discovery and experimentation
stronger monitoring, verification, and enforcement
ENVIRONMENTAL EFFECTS
electricity and associated emissions
direct and indirect water consumption
chips, servers, buildings, minerals, and land
e-waste, induced demand, rebound, and burden shifting
Net impact is the change in the whole system, not the sum of marketing claims.
This is not an LLM story
Planetary AI relies on several technical families working together.
TIME-SERIES FORECASTINGweather, load, renewable output, demand, prices, and hazards
COMPUTER VISIONsatellite imagery, wildlife, infrastructure, smoke, methane, and land change
GRAPH MODELSweather meshes, river networks, power grids, molecules, and materials
HYBRID PHYSICS + MLlearn hard-to-model processes while preserving known dynamics
CONTROL + OPTIMIZATIONdispatch, storage, cooling, microgrids, buildings, and experiments
FOUNDATION MODELSreusable representations for Earth observation, weather, and science
The control loop
Prediction changes the planet only when it closes a loop.
SENSEcollect weather, grid, satellite, laboratory, and field observations
ESTIMATEinfer hidden state, detect events, and quantify uncertainty
PREDICTsimulate possible futures under changing conditions
ACTdispatch, warn, repair, conserve, invest, or regulate
MEASURE AGAINobserve outcomes, drift, side effects, and changing behavior
Open-loop accuracy is not closed-loop success.
Six opportunity pathways
AI can help us see earlier, coordinate better, and discover faster.
01OBSERVEextract usable signals from satellites, sensors, meters, and scientific instruments
02FORECASTanticipate weather, hazards, demand, renewable output, and ecosystem change
03OPTIMIZEcoordinate grids, buildings, factories, transport, water, and logistics
04DISCOVERsearch materials, chemistry, batteries, catalysts, and fusion controls
05VERIFYdetect emissions, deforestation, waste, damage, and policy compliance
06ADAPThelp communities prepare infrastructure, services, and investments for changing risk
01
The physical system behind AI
Every model request travels through chips, servers, cooling, buildings, power systems, water systems, and supply chains.
A query reaches far beyond the screen
AI energy use is produced by a chain of technical and physical systems.
WORKLOADTokens, pixels, model size, precision, memory, and repeated callscomputational work
HARDWAREAccelerators, memory, storage, network, utilization, and idle powerIT electricity
FACILITYPower conversion, cooling, lighting, backup, and building systemssite overhead
GRIDGeneration mix, transmission, congestion, timing, and marginal supplycarbon and reliability
SUPPLY CHAINMining, fabrication, construction, replacement, transport, and retirementembodied impact
Training and inference are different workloads
One large training run and billions of small uses create different environmental profiles.
TRAIN + ADAPT
large, concentrated computation over a defined period
data preparation, pretraining, tuning, evaluation, and failed experiments
cost can be amortized across future uses
new model versions restart part of the cycle
RUN + SERVE
repeated requests across the product lifetime
latency, reliability, context length, media, agents, and tool calls matter
low utilization and peak capacity can waste resources
scale of use can dominate lifetime demand
Report the life of the service, not only the headline training run.
What drives the workload
Model size matters, but it is only one term in the energy equation.
COMPUTEOperationsarchitecture, parameters, tokens, image resolution, steps, and repeated reasoning
DATA MOVEMENTMemorymoving weights and activations can consume substantial time and energy
EFFICIENCYPrecision + hardwarequantization, accelerators, kernels, sparsity, and utilization alter energy per task
DEMANDHow often it runsusers, agents, retries, background tasks, and product design determine total volume
Global electricity demand
Data-center electricity is growing much faster than total electricity demand.
415 TWhglobal data-center electricity consumption in 20241.5% OF GLOBAL USE
485 TWhupdated IEA estimate for 202517% ANNUAL GROWTH
~950 TWhcentral IEA projection for 2030ABOUT 3% OF GLOBAL USE
3×projected growth in electricity use by AI-focused data centers from 2025 to 2030IEA 2026
Global averages hide local shocks
A small global share can be a very large regional load.
4.4%share of U.S. electricity used by data centers in 2023LBNL
11.8%central 2030 estimate in Berkeley Lab's 2025 updateU.S. TOTAL
9.5-15.3%scenario range for the U.S. share in 2030UNCERTAINTY
~50%share of U.S. electricity-demand growth through 2030 attributed to data centers in IEA's 2025 outlookGROWTH, NOT TOTAL
The new infrastructure economy
AI is now an energy, construction, finance, and industrial-policy story.
$500BApproximate global data-center investment in 2024, nearly double the 2022 level. IEA projects $4.2 trillion in cumulative data-center investment from 2025 through 2030 in its base case.
>$400B2025 capital expenditure by five large technology companies, driven by data-center investment
+75%IEA's expected increase in that group’s capital spending in 2026
2,500 GWgeneration, storage, and large-load projects stalled in grid connection queues worldwide
What will power the growth
Renewables supply much of the increase, but fossil generation remains part of the near-term mix.
RENEWABLESmeet nearly half of additional global data-center electricity demand through 2030 in IEA's base caseFASTEST EXPANSION
NATURAL GAS + COALremain important where clean generation, grids, storage, and flexibility cannot arrive quickly enoughNEAR-TERM EMISSIONS
NUCLEARplays a larger role toward the end of the decade and beyond, including new and existing plantsFIRM POWER
FLEXIBILITYshifting non-urgent work can reduce peak strain and better match clean supplyDESIGN LEVER
Water is a local constraint
Cooling water is only one part of AI's water footprint.
66B litersEstimated direct U.S. data-center water consumption in 2023. The same Berkeley Lab analysis estimated roughly 800 billion liters of indirect consumption through electricity generation.
DIRECTevaporative cooling, humidification, and facility operation
INDIRECTwater consumed in electricity generation and fuel supply
EMBODIEDsemiconductor fabrication, construction, and materials processing
CONTEXTwater source, season, watershed stress, temperature, and competing use
Hardware has a life before and after the data center
Chips turn minerals, manufacturing, and electricity into a rapidly aging asset.
62 Mtelectronic waste generated globally in 2022UNITAR + ITU
22.3%documented as formally collected and recycled2022
82 Mtprojected annual e-waste by 2030GLOBAL TREND
<1%of rare-earth demand met by e-waste recyclingCIRCULARITY GAP
AI hardware is part of the broader electronics economy, not a separate waste stream.
The full life cycle
Operational electricity is visible, but it is not the whole footprint.
01EXTRACTminerals, water, land, energy, and labor
02FABRICATEchips, memory, boards, servers, and cooling equipment
03CONSTRUCTbuildings, substations, transmission, backup, and networks
04DEVELOPdata, experiments, training, evaluation, and failed runs
05OPERATEinference, storage, cooling, redundancy, and updates
06RETIREreuse, refurbish, recycle, export, landfill, and contamination
Choose the system boundary before the metric
The answer changes as the boundary expands.
PLANETARY + SYSTEMIC EFFECTS
ENERGY, WATER + SUPPLY CHAIN
FACILITY + HARDWARE
WORKLOAD
Workload: What did one run consume?
Service: What did the product consume over its useful life?
Infrastructure: What generation, grid, water, and hardware did it require?
System: What behavior, demand, markets, and physical outcomes changed?
Planet: Were impacts reduced, shifted, delayed, or amplified elsewhere?
No single metric is enough
Measure resource intensity per useful outcome, then keep the components visible.
PUEtotal facility energy relative to IT energy; shows overhead, not workload value or grid carbonENERGY OVERHEAD
WUEwater consumption relative to IT energy; depends on cooling and local conditionsWATER INTENSITY
CARBONoperational emissions plus allocated hardware manufacturing emissionsLOCATION + TIME MATTER
FUNCTIONAL UNITper forecast, image, solved case, user outcome, or other useful unitWHAT VALUE WAS DELIVERED?
The efficiency toolbox
The cleanest computation is often the computation the product never needed.
AVOIDremove low-value generation, retries, polling, background agents, and duplicate work
RIGHT-SIZEuse rules, search, small models, specialists, or human judgment when sufficient
COMPRESSquantization, distillation, pruning, sparsity, and shorter contexts
REUSEcache outputs, batch requests, retrieve known answers, and share representations
UTILIZEmatch hardware, keep accelerators busy, and reduce idle reserved capacity
SCHEDULEshift flexible work to cleaner, cooler, less congested hours and locations
CO-DESIGNoptimize model, software, hardware, cooling, and power together
EXTEND LIFEmaintain, refurbish, redeploy, and recycle hardware responsibly
Carbon-aware is not automatically planet-aware
Shifting a workload can lower one impact and raise another.
TIMERun when the grid is cleaner, but respect deadlines, forecast error, and whether clean power is actually marginal.
PLACEMove to a lower-carbon region, but check network cost, data rules, local water stress, and transmission constraints.
COOLINGReduce water use with dry cooling, but measure any increase in electricity, heat, cost, or equipment stress.
Multi-objective optimization should keep carbon, water, reliability, cost, and justice visible.
The optimization paradox
Efficiency lowers impact per unit. Rebound can raise the number of units.
DIRECT REBOUNDCheaper and faster inference encourages more prompts, richer media, longer contexts, and always-on agents.
INDIRECT REBOUNDSavings are spent on other energy, travel, goods, or computation rather than becoming absolute reductions.
SYSTEMIC REBOUNDAI changes markets and behavior, such as shifting travel from public transport toward autonomous vehicles.
The outcome metric must track total physical change, not only efficiency.
02
AI for planetary systems
The strongest applications connect observation, prediction, physical action, and verification inside a real institution.
How AI weather forecasting works
Machine-learning forecasts learn how atmospheric state changes over time.
01OBSERVATIONSsatellites, stations, balloons, aircraft, ships, radar, and buoys
02DATA ASSIMILATIONcombine incomplete observations with a model to estimate the current state
03LEARN DYNAMICStrain on decades of reanalysis to predict the next atmospheric state
04ENSEMBLEgenerate multiple plausible futures to represent forecast uncertainty
05INTERPRET + WARNforecasters translate probabilities into products, warnings, and decisions
Operational weather AI in 2026
Machine-learning weather forecasting has moved from research papers into daily operations.
ECMWF AIFS2025deterministic system became operational in February; ensemble system followed in July
AIFS v2May 2026both systems upgraded with more variables, waves, land fields, and stratospheric levels
2025 PERFORMANCE5-15%typical medium-range error reductions for many upper-air and surface variables compared with physics-based controls
Performance is strong but uneven: cyclone tracks improved while intensity retained a weak-intensity bias.
Flood forecasting
AI can extend useful flood forecasts into river basins with few gauges.
Warn earlier where observations are scarce.
A global model can transfer hydrologic patterns across basins, combine weather and watershed features, and forecast river conditions where local gauge history is limited.
7-DAY RIVER FORECASTSoperational systems estimate near-term river conditions and significant flood risk
UNGauged BASINSshared learning can improve access where conventional local models are weak
OPEN FRAMEWORKGoogle released its hydrology modeling framework in June 2026
ACTION LAYERwarnings need local thresholds, communication, evacuation, and anticipatory support
Earth-observation foundation models
A reusable model can turn satellite archives into many environmental tools.
Pretrain once, adapt many times.
Learn common spatial and temporal patterns from large unlabeled satellite archives, then fine-tune with smaller task-specific datasets.
FLOOD + WATER EXTENTsegment inundation and changing surface water
FIRE + LAND CHANGEmap burn severity, deforestation, land use, and recovery
AGRICULTURE + BIOMASSclassify crops, estimate productivity, and measure above-ground biomass
LANDSLIDES + HABITATdetect terrain change, disturbance, and ecological conditions
Methane detection
Lightweight AI is helping turn satellite observations into verified mitigation.
1.3M+satellite observations analyzed by UNEP's MARS since 2023OBSERVATION
12-15×more data processed with AI-assisted workflows while retaining expert reviewCAPACITY
40+methane mitigation actions linked to MARS since becoming fully operationalACTION
1.2 Mtestimated methane from sources that were mitigated after AI-assisted detectionREPORTED OUTCOME
Every AI-flagged detection is independently reviewed before a notification is issued.
The grid is a real-time balancing system
Electricity supply and demand must stay synchronized while both keep changing.
FORECASTdemand, solar, wind, outages, weather, and prices across minutes to years
ESTIMATEinfer voltages, flows, equipment state, and hidden failures from incomplete sensors
DISPATCHcoordinate generation, storage, flexible load, and network constraints
PROTECTmaintain frequency, voltage, stability, cybersecurity, and recovery margins
The grid is a constrained control problem, not a marketplace that can ignore physics.
AI can unlock more value from existing energy assets
Better sensing and decisions can expand capacity before new infrastructure arrives.
175 GWadditional transmission capacity that IEA estimates AI-enabled methods could unlock in existing linesWIDESPREAD ADOPTION CASE
$110Bpotential annual power-plant operations and maintenance savings by 2035IEA ESTIMATE
PLANNINGscenario screening, climate-informed needs, siting, and interconnection studiesLONG HORIZON
OPERATIONSdynamic line ratings, predictive maintenance, anomaly detection, and dispatchREAL TIME
Buildings and industry
AI creates value when it controls physical systems, not when it only describes them.
HVAC CONTROLforecast occupancy, weather, thermal state, and comfort to reduce waste
INDUSTRIAL PROCESSoptimize temperature, pressure, flow, fuel mix, yield, and quality
PREDICTIVE MAINTENANCEdetect degradation before equipment fails or operates inefficiently
FLEXIBLE DEMANDshift non-urgent loads away from expensive, congested, or high-carbon hours
IEA estimates about 10% savings for optimized building HVAC and 8% for light industry by 2035 under widespread adoption.
Renewables, storage, and flexible demand
AI can coordinate variability across generation, storage, and use.
Make flexibility predictable enough to operate.
Forecast what the system will have, estimate what it can move, and choose actions that preserve reliability at lower cost and emissions.
WIND + SOLAR FORECASTSpredict output and uncertainty from weather, history, and asset state
BATTERY CONTROLschedule charge and discharge while managing degradation and reserves
EV + BUILDING LOADcoordinate flexible demand without surprising users
MICROGRID CONTROLbalance local generation, storage, critical load, and islanded operation
Materials discovery
AI can search chemical possibility faster than experiments alone.
381,000new stable crystal entries on the updated convex hull reported by GNoME, drawn from 2.2 million candidate structures stable relative to prior work. The paper found 736 matches to independently realized experimental structures.
GENERATEpropose candidate compositions and structures
PREDICTgraph networks estimate energy and stability
VERIFYhigher-fidelity quantum calculations test candidates
LEARNnew results improve the next search round
EXPERIMENTsynthesis and characterization determine what is real and useful
Self-driving laboratories
AI becomes more powerful when prediction and experiment form a closed loop.
SELECTchoose the next candidate with the highest expected information or value
PLANpropose a synthesis recipe from literature, models, and prior failures
3RIGHT-SIZEUse the smallest model, data pipeline, and serving architecture that meet the requirement.
4OPTIMIZEImprove algorithms, software, hardware utilization, cooling, and equipment lifetime.
5MATCH RESOURCESSchedule and locate flexible work where energy is cleaner and water stress is lower.
6CLOSE THE LOOPMeasure absolute outcomes after deployment, check rebound, and redesign or stop.
State of the field · August 2026
AI and planetary systems are moving from demonstrations toward infrastructure.
ENERGYData-center electricity use grew 17% in 2025.The IEA reports rapid growth alongside grid bottlenecks and a scramble for power.
WEATHERECMWF made AIFS v2 operational in May 2026.AI forecasts now sit inside a major public forecasting institution.
METHANEUNEP reported verified mitigation actions in July 2026.Detection becomes valuable when alerts connect to operators and follow-up.
FLOODSGoogle open-sourced its hydrology framework in June 2026.Open tooling can expand research, adaptation, and local capacity.
REPORTINGEU data-center performance reporting is expanding.Energy and water indicators are moving toward a common rating scheme.
MEASUREMENTSCI for AI was ratified in December 2025.Practitioners now have a more specific method for measuring AI carbon intensity.
Joint capability
The strongest planetary systems combine machine scale, human judgment, and institutional authority.
AI SYSTEMintegrates large sensor streamsfinds weak and nonlinear patternsruns forecasts, simulations, and searchacts quickly and consistently within constraints
SCIENTIST + OPERATORtests mechanism and physical plausibilityinterprets uncertainty and rare conditionsrecognizes unsafe or irrelevant outputsadapts action to the actual system state
INSTITUTION + COMMUNITYsets goals, rights, limits, and prioritiesowns physical action and accountabilitycontributes local and lived knowledgedecides who benefits and how harm is repaired
Planetary value emerges from the joint system, not from the model alone.
The near future · 2 to 7 years
The next wave will connect better models to operational systems and verified action.
WEATHER ENSEMBLESmultiple fast AI forecasts with calibrated uncertainty inside public warning systems
OPEN EARTH MODELSadaptable foundation models for regions, hazards, land, water, and ecosystems
GRID DIGITAL TWINSsimulation, forecasting, and constrained control for more clean power and resilience
FLEXIBLE DATA CENTERSworkloads that respond to grid carbon, congestion, water stress, and reliability
AUTONOMOUS LABSclosed-loop discovery for batteries, catalysts, refrigerants, cement, and carbon removal
VERIFIED POLLUTION RESPONSEsatellite detection linked to repair, enforcement, and transparent follow-up
LOCAL ECOLOGICAL SENSINGlow-cost networks operated with communities for biodiversity, forests, farms, and water
ADAPTATION COPILOTSrisk tools tied to budgets, infrastructure plans, public services, and accountable action
The farther horizon · plausible, not promised
The most ambitious future applications treat Earth as a system we can understand, not a machine we can fully control.
EARTH DIGITAL TWINSinteractive simulations that compare policy and infrastructure pathways before deployment
CLEAN-ENERGY DISCOVERYautonomous science that searches vast spaces of molecules, materials, devices, and processes
FUSION CONTROLlearned simulation and real-time control that accelerate progress toward viable fusion systems
BIODIVERSITY NERVOUS SYSTEMplanetary sensing that detects ecological change early while preserving local authority
CIRCULAR MATERIALSproducts and manufacturing designed for low impact, repair, reuse, separation, and recovery
ADAPTIVE INFRASTRUCTUREbuildings, grids, transport, and water systems that anticipate hazards and recover gracefully
ENERGY COMMONScoordinated microgrids and flexible resources that strengthen community reliability and ownership
NET-POSITIVE COMPUTEtransparent global accounting that makes every major workload answer for absolute planetary value
The closing test
When someone calls AI a planetary solution, ask what changes in the real world.
01What physical or human outcome improves, compared with what baseline?
02Where is the system boundary, and what has been left outside it?
03What electricity, water, hardware, minerals, land, and labor make it possible?
04What happens when it scales, becomes cheaper, or changes behavior?
05Who decides, who benefits, who bears risk, and who can stop it?
AI IS ANAMPLIFIERTHE SYSTEM SETSDIRECTIONMEASUREMENT KEEPS USHONEST