The impact of AI on sustainability and its monitoring

Dialogic, commissioned by the Ministry of Economic Affairs, researched the sustainability impact of developing and using AI, and how this impact can best be monitored. This research was conducted in collaboration with Pb7 Research and Utrecht University. There has been extensive discussion in the media and academic literature about the impact of AI on sustainability. Various statements have been made regarding the energy consumption involved in tasks such as querying ChatGPT, compared to regular search queries. The estimates vary widely, largely due to different delimitations, calculation methods, and assumptions. The main goal of this research is not to validate these estimates or to come up with new ones. The core aim is to provide insight into the underlying factors and their interrelationships, for the purpose of policy-making. To achieve this, the study delves into the workings of AI in more detail and explores the chain from AI usage to its ultimate sustainability impact. Additionally, it provides insights into expected future developments. **Supply of AI applications** The AI services market is rapidly evolving. Various market players are trying to dominate the market by offering cheap or free AI services. The primary technological trend is the scaling up of AI, particularly in terms of the size of AI models (number of parameters), the computational capacity used for training, and the computational capacity for inference (via inference-time scaling and the rise of agentic AI). Furthermore, there is an increasing level of differentiation (various model variants with associated prices, as well as dynamic pricing). The availability of AI services strongly depends on the (computational) capacity in data centres. Much of the AI used in the Netherlands is supplied from abroad. The availability of computational capacity for AI and other purposes in the Netherlands is limited at least until 2030.   **Usage of AI applications** The usage of AI involves the **number** of users (adoption), the **intensity** of usage, and the **quality** of usage. The adoption of AI is increasing among both consumers and businesses. It is difficult to predict how the intensity of usage will develop – which AI functions will users continue to use consistently? It is clear, however, that the development of usage is largely driven by supply. AI service providers are fiercely competing with each other and offering increasingly better AI services for free or at low costs. A critical limitation here is the availability of computational capacity. Particularly, the usage of AI for image and video generation requires significant computational capacity. In the future, we expect that the development of AI usage will be more demand-driven. **Energy consumption of AI** There are three relevant phases in which AI consumes energy: - The **collection** and processing of data needed for AI training is the first phase, requiring a limited (but substantial) amount of energy. - The subsequent **training and evaluation phase** is highly energy-intensive. The quantity of computational capacity (and thus indirectly the amount of available energy) in this phase determines the size of the model and the training dataset. More efficient hardware can train a 'better' AI given the same amount of energy. This energy consumption is essentially one-time and can be amortised over the subsequent useful use of AI. - The **inference phase**: the energy consumption here primarily depends on usage (number of requests and the size/quality of the generated response). AI for image generation consumes significantly more energy than AI generating text, and video generation requires even more energy. Developments like inference-time scaling increase energy consumption in the inference phase. The emergence of agents means proactive AI tasks are carried out, which can lead to higher energy consumption. With more efficient hardware, batching, and improvements in AI model architecture, the energy consumption per request can be reduced. In the Netherlands, the energy consumption of AI will be limited until at least 2030 due to data centre capacity availability. The current energy consumption of AI in the Netherlands is estimated to be between 41 GWh and 107 GWh per year for 2023. Depending on the chosen future scenario, projections for 2030 range from 2.9 TWh (mid-scenario) to 4.7 TWh (highest scenario) per year. **Emissions from AI** The primary form of emissions resulting from AI is CO2 emissions from the production of required energy (scope 2 of the digital sector) in the training and inference phases. Currently, these emissions primarily revolve around data centres. These emissions do not necessarily occur in the Netherlands for Dutch usage (both the data centres and energy production can be located abroad), and vice versa (the Netherlands can produce for other countries). As demand currently exceeds supply, it is important to examine the emissions from data centres in the Netherlands used for AI. The emission level per amount of electricity consumed (emission factor) is decreasing in the Netherlands. Additionally, data centres largely utilise green (emission-free produced) electricity, although there are nuances to the claim that it generates no emissions. Apart from these scope 2 emissions, there are limited scope 1 emissions (mainly from data centre's own generators). In scope 3, we observe (upstream) emissions from the production of required hardware and by suppliers, and (downstream) emissions mainly from recycling. Reports such as EED and ETS data from the NEa supply relevant data for monitoring. However, there is insufficient data available on (1) which part of a data centre's capacity involves AI (according to a consistent definition), (2) the extent to which locally available green energy and/or cooling methods are used, and (3) emissions on-site (scope 1) and the energy mix (scope 2). Additionally, assessing scope 3 externally is challenging. **Water consumption of AI** Data centres where AI training and inference occur use water for cooling, and **WUE** is the standard measure for their efficiency (water usage per unit of cooling capacity). Only a portion of this water usage concerns drinking water – the remainder is industrial water. In the Dutch context, particularly examining drinking water consumption is essential. Introducing the measure **WUEp** (drinking water usage per unit of cooling capacity) is crucial. It is noteworthy that, in times of water scarcity, data centres – like other enterprises – are the first to lose access to drinking water. Currently, the total drinking water consumption of the sector does not exert significant pressure on the Dutch drinking water supply nor poses a threat to it. However, due to the risk of unavailability of drinking water, data centres in the Netherlands need to explore alternative cooling methods that use less or no drinking water.