El impacto de la inteligencia artificial en la sostenibilidad y su monitoreo.

Dialogic, on behalf of the Ministry of Economic Affairs, researched the impact on sustainability of developing and using AI, and how this can best be monitored. We conducted this research in collaboration with Pb7 Research and the University of Utrecht. Much has been written in the media and academic literature about the impact of AI on sustainability. Statements have been made regarding the amount of energy consumption that, for example, querying ChatGPT would generate, and compared to regular search queries. The estimates vary widely, often due to the boundaries set, calculation methods, and assumptions made. The goal of this research is not to confirm these estimates or to come up with a new estimate. The main objective is to provide insights into the underlying factors and how they relate, for the purpose of policymaking. To achieve this, this research delves into more detail on how AI operates and the chain between the use of AI and its ultimate impact on sustainability. Additionally, we offer insights into the expected future developments.

Offer of AI applications

The supply of AI services is rapidly evolving. Various market parties are trying to dominate the market and offer cheap or free AI services in this context. The main technological trend is the scalability of AI, particularly in terms of the size of AI models (number of parameters), the computing power used for training, and the computing power in inference (through inference-time scaling and the rise of agentic AI). Finally, we are seeing increasing differentiation (various model variants and corresponding prices, as well as dynamic pricing). The supply is highly dependent on the available (computing) capacity in data centers. Much of the AI used by Dutch users is delivered from abroad. The availability of computing capacity for AI and other purposes in the Netherlands is limited at least until 2030.

Use of AI applications

The use of AI involves the number of users (adoption), the intensity of use, and the quality of use. Adoption of AI is increasing among both consumers and businesses. It is still uncertain how the intensity of use will develop – which AI functions will users continue to use consistently? What is clear is that the development of use is strongly driven by supply. AI service providers are in fierce competition with each other, offering increasingly better AI services for free or at low cost. A key limitation here is the availability of computing capacity. Particularly, the use of AI for image and video generation requires a lot of computing power. In the future, we expect the development of AI use to be more demand-driven.

Energy consumption of AI

There are three relevant phases in which AI consumes energy:

• The collection and processing of the data needed to train AI is the first phase, requiring a limited (but substantial) amount of energy.
• The subsequent training and evaluation phase is highly energy-intensive. The amount of computing power (and therefore indirectly the amount of available energy) in this phase determines the size of the model and the size of the training dataset. With more efficient hardware, a 'better' AI can be trained given the same amount of energy. This energy consumption is essentially one-time and can be depreciated over the subsequent useful use of the AI.
• The inference phase: energy consumption here is primarily dependent on usage (number of requests and the size/quality of the generated response). AI for image generation consumes significantly more energy than AI that generates text, and video generation requires even more energy. Developments like inference-time scaling increase energy consumption in the inference phase. The rise of agents means proactive AI tasks are performed, which can lead to higher energy consumption. With more efficient hardware, batching, and improvements in the architecture of AI models, energy consumption per request can be reduced.

In the Netherlands, the energy consumption of AI is limited until at least 2030 by the availability of data center capacity. We estimate the current energy consumption of AI in the Netherlands for 2023 to be between 41 GWh and 107 GWh per year. Depending on the chosen future scenario, by 2030, we estimate a consumption of 2.9 TWh (mid-scenario) to 4.7 TWh (high scenario) per year.

Emissions from AI

The main form of emissions resulting from AI is CO2 emissions from the production of the required energy (scope 2 of the digital sector) in the training and inference phase. Currently, these emissions primarily revolve around data centers. These emissions may not necessarily occur in the Netherlands for Dutch use (both data centers and energy production can take place abroad), and vice versa (the Netherlands can produce for other countries). Due to current higher demand than supply, it is important to consider the emissions of data centers in the Netherlands used for AI. The emission per amount of consumed electrical energy (emission factor) is decreasing in the Netherlands. Additionally, data centers largely use green (emission-free produced) electricity, although there are nuances regarding the notion that this entails no emissions at all. Apart from these scope 2 emissions, there are limited scope 1 emissions (mainly from data centers' own generators). In scope 3, there are (upstream) emissions from the production of the necessary hardware and from suppliers, and (downstream) particularly from the recycling of such hardware. Reports like the EED and ETS data from the NEa provide valuable information for monitoring this. However, there is insufficient data available regarding (1) what portion of a data center's capacity involves AI (according to a consistent definition), (2) the extent to which locally available green energy and/or cooling options are utilized, and (3) on-site emissions (scope 1) and the energy mix (scope 2). Moreover, scope 3 emissions from external sources are challenging to assess.

Water consumption of AI

Data centers where AI training and inference take place consume water for cooling, with WUE being the standard measure of efficiency (water usage per unit cooling capacity). Only a portion of this water consumption involves drinking water – the rest is industrial water. In the Dutch context, it is particularly relevant to focus on drinking water usage. To measure this, we introduce the metric WUEp (drinking water consumption per unit cooling capacity). It's important to note that data centers, like other businesses, are the first to be cut off from drinking water during water scarcity. The total drinking water consumption of the sector currently does not exert significant pressure on the Dutch drinking water supply, nor does it endanger it. However, due to the risk of unavailability of drinking water, data centers in the Netherlands must seek alternative cooling methods that use less or no drinking water.