Material impacts of AI

Estimating the environmental impacts of AI and computationally involved tasks is a challenge.

Let’s start with a simple metric: 

Question: What percentage of global energy consumption is allocated directly to computers and their infrastructure (Information and Communication Technologies)?

Answer: About 4-6% in 2020 according to the UK government. Though there are ranges in estimates, they are mostly in the low-mid single digits.

TASK: Compare this 4-6% value with how much energy is expended in other sectors. Think heating, transport, manufacturing, etc. The 2024 report on energy consumption in the UK over time may be a good starting point.

Most of our global energy consumption is attributed to moving, or changing the temperature of things. The energy costs associated with a computer’s operation is simply not of the same scale. Think about the typical tenement in Glasgow and its level of insulation. How many computers would you need to have running AI processes to make up for the lack of loft insulation in winter? 

Note: For a computer to operate, it needs to dissipate heat into the environment. The amount of heat dissipated proportional to the amount of ‘computation’ can not be made infinitely efficient.

But only considering the direct energy consumption of individual computer systems isn’t the whole story. We also need to explore other, broader ways of considering the material impact of AI systems. 

TASK: Perform a (qualitative) Lifecycle Assessment of some aspect of an AI system (this could be a full data-centre, a smartphone, or a single computer chip). What are we missing if we only look at energy consumption as our metric?

Image by Erin Gilliard

Understanding AI impacts on the environment:

An accessible and comprehensive evaluation of the environmental impacts of AI systems can be found in the 2025 National Engineering Policy Centre report, “Engineering Responsible AI: foundations for environmentally sustainable AI“.

In summary, across their lifecycles, AI systems and services require energy, water, and critical materials.

Image by Adrien Book

• These requirements are resources that are directly consumed during production, use, and end-of-lifecycle management of the compute and infrastructure hardware underpinning AI systems and services. 
• They are also indirectly consumed during activities such as energy generation and distribution. 
• Consuming these resources can cause air pollution, water pollution and thermal pollution, as well as creating solid waste. These impacts, in turn, negatively affect human and environmental health and can result in heightened risk of social inequity or exclusion, especially as the environmental impacts of AI are typically unevenly distributed.
• The software lifecycle includes data collection, system development, training, validation, delivery, maintenance and retirement.
• Decisions made at each stage of the software lifecycle impact resource consumption by compute and infrastructure hardware. 
• In recent years, software development at the frontier of AI has tended to prioritise scale, size and capability, leading to increasingly resource intensive software lifecycles, particularly impacting the training of models and their inference. 
• This tendency has been compounded by the rapid expansion of both the development and uptake of AI.
• AI systems can take many forms, with different levels and profiles of resource intensity depending on their design and use. 

Resource consumption for an AI system or service has 2 stages:

1. Training Stage
2. Inference (use)

Depending on a system’s design and the way it is used, the two stages may have different impacts on resource consumption. For example, an agent-based system may consist of multiple agents that collectively require fewer training resources than a single Large Language Model. However, if these agents are designed to query one another via inference regularly, the system’s total resource consumption may still increase. 

Energy Generation and the AI Arms Race

A very significant portion of AI development is taking place in the US. Look at the figure below, from IOT Analytics, published 2025, how many of these leading AI companies are based outside the US?

Image by IOT Analytics

TASK: Investigate how much of the development of AI is grounded in the US today? Illustratively, the US government allocated 500 Billion USD towards funding the development of AI through Project Stargate. That is over 10 percent of UK annual GDP.

The way the US approaches AI is, at present, the dominant approach to AI, and this has significant implications for all aspects of AI development. In this section, we will focus on the environment, but you are encouraged to also consider the extent of global cooperation we should exercise when developing AI (see also AI & Governance, Week 2, for more on this theme).

Case Study: US Big Tech, Fossil Fuels and Nuclear Power

Start with the following articles, and consider how environmental considerations may be cast aside in the AI arms race.

Elon Musk’s xAI powering its facility in Memphis with ‘illegal’ generators

Data Centers Fuel Energy Concerns – and a Transatlantic Split

AI boom is driving a surprise resurgence of U.S. gas-fired power

TASK: Consider whether the development of AI should be decentralised from Silicon Valley? If so, what strategies could help move towards this goal?

*Hint: we can actually all do something about this…European Alternatives and British Alternatives.