Cause and effect: AI’s environmental costs should be counted now

Kate Crawford in the introduction to her book Atlas of AI argued, “…artificial intelligence is both embodied and material, made from natural resources, fuel, human labour, infrastructures, logistics, histories, and classifications.”

Quantifying human labour, especially of people employed in lower middle income countries, is a herculean task, with even governments failing to represent the numbers and the contribution of their people in blue collar jobs.

So perhaps a look at how technology works?

As with any digital technology, the life cycle of AI can broadly be divided into hardware and software.

The first begins from manufacturing that includes raw material extraction, component production (processors, memory, storage), assembly of AI hardware, installation in data centres, power and cooling infrastructure setup.

Software lifecycle, meanwhile, starts with data collection, model development, initial training and testing and validation before it can be deployed.

The next obvious question is: how is this harmful for the environment?

Information on this is scarce with limited information. A query fed into Anthropic’s large language model (LLM) Claude gave the following answer:

“I aim to have no direct physical impact on the climate since I’m a software system running on computers that are controlled by Anthropic.”

On further prodding, the chatbot responded as follows:

“Training Phase: High energy consumption for model training (can range from hundreds to millions of kWh); Computing hardware production impacts (rare earth metals, manufacturing emissions); Data centre cooling requirements;

Deployment/Inference Phase: Infrastructure maintenance and replacement, Network transmission energy costs;

End-of-Life: E-waste from retired hardware;

Notable factors affecting footprint: Model size and architecture, Power grid carbon intensity, Hardware efficiency and deployment scale.”

The response ended with a post-script that rather intuitively highlighted the difference in AI models of the company, suggesting a look at their own reports.

This begs the question: how much energy is actually utilised?

To be fair, from the production of devices to data centres that process each query on the cloud, the life cycle of digital technology is as energy intensive as any other human action.

The UNEP in 2024 estimated that building a 2kg computer needs about 800kg of raw materials; microchips need rare earth metals that are mined in destructive ways; data centres create electronic waste like mercury and lead; and the construction of data centres and cooling electrical components once operational requires water.

In fact, according to an estimate by Lotfi Belkhir and Ahmed Elmeligi, the tech sector is expected to contribute 14% of global greenhouse emissions by 2040, while a team in Sweden predicts that the electricity demands of data centres alone will increase about 15-fold by 2030.

Several independent estimates, however, have pegged the climate costs at varying levels.

“One query to ChatGPT uses approximately as much electricity as could light one light bulb for about 20 minutes,” Jesse Dodge, a senior research analyst at the Allen Institute for AI, told NPR in July 2024.

“So, you can imagine with millions of people using something like that every day, that adds up to a really large amount of electricity,” he said.

These statistics are not to establish AI as the only villain in this tech story: even a Google Search consumes energy.

But, for some perspective, OpenAI’s generative chatbot Chat GPT uses 10 times the energy of a Google search, a report by Goldman Sachs said.

A follow up question then is: how do different AI models differ in their energy utilisation and subsequent impact?

Experts insist that calculating the exact effect of AI on the climate crisis is an uphill task, especially in the absence of transparency from companies that own these models.

A key reason for this is that different types of AI — whether it’s a machine learning model that spots trends in research data or a large language model (LLM) that enables a chatbot to converse — all require different quantities of computing power to train and run.

For example, training GPT-3 led to carbon emissions equivalent to 552 tonnes of CO2 and consumed 1287 MWh of energy, the paper, Carbon Emissions and Large Neural Network Training, found.

This is the same amount of emissions that over 4500 direct round trips between Delhi and Chennai.

Google Flights estimate for the emissions of a direct round trip of a whole passenger jet between the two cities is 0.117 tCO2e.

Further, the water demand from data centres is expected to be six times greater than the consumption of Denmark.

Data centres use fresh water rather than surface water, so that the pipes, pumps and heat exchangers used to cool racks of servers do not get clogged up with contaminants.

Dr Venkatesh Uddameri, a Texas-based expert in water resources management, says that a typical data centre can use between 11 million and 19 million litres of water per day, roughly the same as a town of 30,000 to 50,000 people.

Microsoft’s global water use soared by 34% while it was developing its initial AI tools, and a data centre cluster in Iowa used 6% of the district’s water supply in one month during the training of OpenAI’s GPT-4.

Thus, experts suggest viewing AI’s impact on climate through a multifaceted lens.

In a vacuum created by lack of reliable information on environmental impacts of AI, a social media storm with several news reports decrying the emerging technology for spurring the crisis gained steam on social media. But, the same vacuum is also a reflection of a distrust of big tech in general.

As all may still not be lost. Experts point to the untapped potential to make AI greener: data centres can run on renewable electricity, chip designs are becoming increasingly energy efficient, and AI algorithms have the potential to be run smarter and faster.

This development, however, is constantly outpacing by the AI revolution.

For instance, Microsoft, a company with a firm 2030 carbon neutrality pledge, reported an increase in emissions of almost 30% in 2023 due to the growth of their data centres. Then there are significant regional variations as well.

An article by the United Nations University said that Google was able to run its Northern EU-based data centres more than 90% of the time on renewable energies in 2022, whereas for Asia-based centres, this number drops to less than 20%. Local level industrial regulations contribute to this variation.

While most companies working on AI don’t disclose their emissions, Google in its sustainability report released last year said that its greenhouse gas emissions rose last year by 48% since 2019, attributing this surge to its data centre energy consumption and supply chain emissions.

“As we further integrate AI into our products, reducing emissions may be challenging,” the report read.

Microsoft in its own report said its emissions grew by 29% since 2020 due to the construction of more data centres that are “designed and optimized to support AI workloads”.

To put things into perspective, there are over 8,000 data centres worldwide, a number that has nearly doubled since 2015. Collectively, these centres now consume as much electricity as the entire country of Italy. And it doesn’t stop there.

As AI becomes more widespread and AI tools grow more sophisticated, energy demand will only increase.

According to a Bloomberg analysis, Sweden could see power demand from data centres roughly double over the course of this decade — and then double again by 2040.

These data centres have the capacity to consume a combined 508 terawatt hours of electricity per year if they were to run constantly. That’s greater than the total annual electricity production for Italy or Australia. By 2034, global energy consumption by data centres is expected to top 1,580 TWh, about as much as is used by all of India.

In the UK, AI is expected to suck up 500% more energy over the next decade. And in the US, data centres are projected to use 8% of total power by 2030, up from 3% in 2022, according to Goldman Sachs, which described it as “the kind of electricity growth that hasn’t been seen in a generation.”

The preliminary data are there. As is the expert concern. But, if experience is anything, tech companies will only act in hindsight. With temperatures soaring, the window for hindsight may also have passed us.