The Environmental Consequences of the AI Boom
The environmental consequences of rapidly growing AI workloads, including electricity demand, water consumption, carbon emissions, hardware manufacturing, and the efficiency-versus-scale paradox.
Key Takeaways
- AI may be digital from the user’s perspective, but it depends on physical infrastructure consuming electricity, water, materials, land, and hardware.
- Data-centre electricity demand is expected to grow substantially, with AI-focused workloads growing faster than overall data-centre demand.
- Energy efficiency per AI task is improving, but rising usage and increasingly demanding workloads can offset those gains.
- Water consumption varies dramatically by data-centre design, climate, electricity source, hardware efficiency, and utilisation; there is no universal “water per AI query” number.
- AI can also enable energy and emissions reductions in other sectors, so its environmental impact cannot be evaluated from data-centre consumption alone.
AI has a physical footprint
The typical user experience looks like:
1
2
3
4
5
User
↓
AI
↓
Answer
The physical infrastructure is considerably larger:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
User
↓
Network
↓
Data Centre
├── Accelerators
├── CPUs
├── Memory
├── Storage
├── Networking
├── Power systems
└── Cooling
↓
Model inference
↓
Response
Training large models adds another level of infrastructure demand involving large compute clusters, storage, networking, and cooling.
AI is therefore not only a software problem.
It is an infrastructure problem.
Electricity demand is rising
The International Energy Agency’s 2026 analysis projects global data-centre electricity consumption to approximately double from 485 TWh in 2025 to 950 TWh in 2030. It expects electricity consumption from AI-focused data centres to grow even faster, roughly tripling over the same period.
The significance is not simply the global percentage.
Data-centre demand is geographically concentrated.
A data centre can therefore represent a significant local load even when the global percentage appears relatively small.
That can create pressure on:
1
2
3
4
5
6
generation capacity
transmission
transformers
cooling infrastructure
grid planning
local utilities
The IEA explicitly notes that local impacts can be much more pronounced than the global share of electricity consumption suggests.
AI workloads are becoming more demanding
Not every AI request has the same computational footprint.
A simple text response is fundamentally different from:
1
2
3
4
5
6
long reasoning
large-context processing
image generation
video generation
multi-step agents
large tool workflows
The IEA notes that energy use per AI task has fallen substantially as hardware and software improve, while also highlighting rapidly increasing demand and more energy-intensive workloads.
This creates the central paradox:
1
2
3
4
5
6
7
Efficiency per task ↓
+
AI usage ↑
+
Workload complexity ↑
=
Total resource demand may still ↑
Efficiency is therefore necessary but not automatically sufficient.
The efficiency paradox
Imagine an AI service becomes:
1
10× more efficient per request
but usage becomes:
1
20× larger
The total workload still increases.
The relationship is:
1
2
3
4
5
Total consumption
=
Consumption per task
×
Number and intensity of tasks
This matters because AI adoption expands the number of workloads as well as the capabilities of each workload.
AI agents add another layer
A traditional request may be:
1
2
3
4
5
User
↓
1 model request
↓
Answer
An agentic workflow might be:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
User
↓
Planning
↓
Tool call
↓
Tool result
↓
Reasoning
↓
Another tool call
↓
Code execution
↓
Additional reasoning
↓
Final result
One human request can therefore produce many model operations.
This links AI architecture directly to environmental efficiency.
Consider the difference between:
1
Agent reads 50,000 tokens
and:
1
Agent retrieves 5,000 relevant tokens
The latter may reduce unnecessary computation.
Context engineering, compression, dynamic tool discovery, and selective model use can therefore have a resource-efficiency dimension in addition to a cost dimension.
The environmental impact of any individual optimisation still needs to be measured; the relationship is not automatically one-to-one.
Water is not a fixed number
Statements such as:
1
"One AI prompt uses X litres of water."
can be misleading because water consumption depends heavily on infrastructure.
A 2025 Lawrence Berkeley National Laboratory study found more than 10,000-fold variation in workload-level data-centre water use. The researchers identified server efficiency, grid water intensity, server utilisation, cooling system, infrastructure efficiency, climate, inactive servers, and server refresh cycle as major determinants.
Therefore:
1
2
3
Same workload
≠
Same water use everywhere
Location and infrastructure matter.
Direct and indirect water use
Water can enter the footprint through:
1
2
3
4
5
Direct:
Cooling at the data centre
Indirect:
Water associated with electricity generation
These two mechanisms can behave very differently.
A site may use:
1
less direct water
while relying on:
1
electricity with higher upstream water intensity
or the reverse.
There is consequently no universal cooling design that minimises every environmental metric simultaneously.
The LBNL research specifically concludes that water-efficient outcomes depend on combinations of multiple site-specific factors rather than one universal recipe.
Carbon emissions
Electricity consumption becomes a carbon problem depending on the generation mix.
The IEA estimates that data centres already represent a significant and growing source of indirect CO₂ emissions from electricity consumption and expects emissions to increase as electricity demand grows.
But an important distinction is required:
1
2
3
Data-centre emissions
≠
AI-only emissions
AI is part of the broader data-centre workload.
As AI becomes a larger proportion of that workload, its contribution becomes increasingly important, but it should not automatically be equated with the full data-centre footprint.
Hardware has an environmental footprint too
Electricity is only one input.
AI infrastructure depends on:
1
2
3
4
5
6
7
8
accelerators
HBM and memory
CPUs
network equipment
storage
power electronics
cooling hardware
servers
Manufacturing these components requires:
1
2
3
4
5
6
raw materials
energy
water
chemical processing
manufacturing facilities
transport
There is also a hardware lifecycle question.
Higher-performance hardware can improve energy efficiency per task, but frequent replacement can increase manufacturing demand.
The equation becomes:
1
2
3
4
5
6
7
8
9
New generation
↓
Higher performance
↓
Potentially better efficiency
↓
Faster refresh cycle
↓
Additional manufacturing footprint
The environmental outcome therefore cannot be determined from accelerator efficiency alone.
Local impacts matter
Global statistics can obscure regional constraints.
A data centre may have a modest global footprint while creating substantial local demand for:
1
2
3
4
5
6
electricity
water
land
construction
network infrastructure
cooling
The IEA notes that data-centre electricity demand is highly geographically concentrated and that this concentration makes local grid effects more significant than global averages would suggest.
This means site selection becomes an environmental and infrastructure decision.
AI can also reduce environmental impact
A complete analysis cannot look only at consumption.
AI can potentially improve:
1
2
3
4
5
6
7
power-grid optimisation
renewable forecasting
industrial efficiency
building energy management
transport optimisation
predictive maintenance
energy-demand forecasting
The IEA identifies multiple applications where AI could improve energy-system efficiency and enable emissions reductions. At the same time, it cautions that adoption, infrastructure constraints, and rebound effects influence the eventual outcome.
So the equation is not:
1
AI = environmental damage
nor:
1
AI = environmental solution
It is:
1
2
3
AI creates environmental costs
+
AI can enable environmental benefits
The balance depends on deployment.
The rebound effect
Efficiency improvements can create more usage.
Suppose a model becomes:
1
90% cheaper to operate
That may lead organisations to run:
1
2
3
4
more workloads
more agents
more experiments
more generated content
The result can be:
1
2
3
Efficiency ↑
Usage ↑↑
Total demand ↑
This is the same basic rebound effect seen in other technologies.
A highly efficient system does not automatically produce lower total resource consumption when usage expands rapidly.
Token optimisation has a second-order environmental effect
This is where AI architecture connects back to sustainability.
Consider:
1
2
3
4
5
Need:
5,000 relevant tokens
Actual:
50,000 tokens
The additional context is not simply a billing difference.
It represents additional information processing.
At very large scale:
1
2
3
4
5
6
7
unnecessary context
↓
additional inference
↓
additional compute
↓
additional electricity
That does not mean every token saved corresponds to a fixed quantity of electricity or water.
It means that eliminating unnecessary computation is directionally relevant to resource efficiency.
This is why techniques such as:
1
2
3
4
5
6
progressive disclosure
context compression
dynamic tool loading
persistent memory
smaller models for simple tasks
code-based data processing
are becoming interesting not only for cost and latency, but potentially for compute efficiency.
The real environmental equation
AI sustainability is multidimensional:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Model efficiency
+
Hardware efficiency
+
Data-centre efficiency
+
Cooling strategy
+
Electricity source
+
Water intensity
+
Utilisation
+
Hardware lifecycle
+
Workload volume
+
Workload complexity
Optimising only one variable is not enough.
A data centre can have excellent server efficiency and poor water conditions.
A site can have low direct water use and high-carbon electricity.
A model can become dramatically more efficient while workload volume explodes.
The future is an infrastructure optimisation problem
The useful question is not:
1
"Is AI environmentally friendly?"
A more useful engineering question is:
1
2
"What infrastructure and workload choices minimise environmental
impact for the required AI capability?"
That requires better measurement of:
1
2
3
4
5
6
energy
water
carbon
hardware lifecycle
utilisation
workload intensity
It also requires understanding where efficiency improvements are being cancelled by growth in demand.
AI is therefore becoming not only a software engineering challenge, but also an infrastructure and resource-management challenge.
References
- International Energy Agency — Key Questions on Energy and AI, 2026.
- International Energy Agency — Energy and AI.
- Lawrence Berkeley National Laboratory — The Water Use of Data Center Workloads, 2025.