The environmental costs of artificial intelligence are commonly regarded as a key risk that may emerge from the widespread diffusion and general advancement of the technology. There has been an exponential increase in capabilities and resources allocated to frontier AI models in recent years – a trend that is likely to continue in the coming decades.

Of the key inputs in AI development, compute and data storage are the most resource-intensive. Evidence has suggested this poses material risks related to the unequal global distribution of the harms and benefits of AI. The global South – particularly developing countries – may miss out on many economic benefits due to significant capital costs, while simultaneously enduring the hazards of AI growth in the form of possible environmental degradation and labour disruption.

As a general rule, more compute entails better predictions and more robust machine learning systems. While algorithmic efficiency is clearly another crucial component of AI development, simply increasing compute has proven a remarkably reliable strategy for the upscaling of AI models, allowing them to leverage larger training datasets, and helping them become ‘smarter’.

Greater energy demands

The scale of compute used for model training has been growing exponentially. This compute investment has enabled the development and widespread deployment of large language models, and it assists researchers and financial institutions across the world in operationalising ever-larger multi-variable datasets. Financial firms use text-to-data AI to leverage qualitative inputs (such as social media posts) and create ‘structured’ data. This creates more parameters for financial decision-making.

While text-to-data machine learning has existed for decades, the scale of its use cases and flexibility has grown considerably along with compute. Many AI in finance firms are, above all, providers of text-to-data tools that help companies leverage unstructured data using machine learning.

Yet more compute does not necessarily entail greater emissions. Research has found that the 550% increase in data centre computing capacity between 2010 and 2018 was accompanied by only a 6% increase in energy consumption. This relatively modest increase in AI energy consumption is due to improved algorithmic and hardware efficiency, whereby AI developers create more highly optimised systems that do not require as much data or compute to perform tasks.

Despite this, however, AI energy consumption is still exponentially increasing. And with massive global investments in AI taking place across the world, there is reason to believe AI energy costs will continue to grow.

Moreover, research has suggested that the increase in AI energy consumption may be due to both the training process (training the model from data) and the subsequent inference process (using the trained model to answer users’ prompts), both of which are showing an exponential increase. Historically, training costs have received more attention, but recent estimates suggest that – depending on the AI model – the energy needed for the inference process may be higher. For example, Google reported in 2022 that 60% of its AI-related electricity use was linked to the inference process alone. This can partly explain why concerns about the technology’s impact on global emissions are increasingly commonplace in AI safety literature. De Vries found that implementation of an LLM assistant for Google browser searches would lead the company to consume as much energy per year as the country of Ireland.

Impacts on the global South

AI advancement is often justified by the argument that the risks are outweighed by the benefits for economic growth and standards of living, creating improved healthcare services, for instance. Yet the well-understood disproportionate impact of climate change on the global South portends more significant risks for already vulnerable populations and regions.

There are at least three good reasons to believe that the economic benefits of AI are less likely to accrue to the global South, relative to the global North.

First, AI development and industry are highly concentrated, globally and within countries. This means that most frontier models are created in major western cities. As AI developments progress, it appears likely that fissures in capacity and opportunity will continue to grow. In just the last two years, the US has substantially widened its position as the global leader in AI.

There is some evidence of AI-based technologies aimed at supporting communities in the global South. These include computer vision to identify banana or cassava disease, reducing antimicrobial resistance in Ghana, and systems for supporting education in Colombia, West Africa and Thailand. Yet AI progress is primarily driven in the global North, making the distribution of AI economic benefits more highly concentrated. Many of these investments are the result of Google, Meta and Microsoft projects, which may be helpful but do not fundamentally alter the structure of global AI power.

Second, advanced AI requires significant electrical and physical capital in order to operate. Many global South countries continue to have limited connectivity. The World Bank found that just 35% of people in developing countries have access to the internet. This number is 80% for developed economies. Sub-Saharan Africa, in particular, suffers from a low electricity capacity (80% of the population) and internet (36%).

This limited physical infrastructure can partly explain why it is more challenging to introduce AI technology in a diffuse manner. A consequence of this could be the emergence of worse global inequalities, as advanced countries pull further away from developing countries in terms of health outcomes, standards of living and economic dynamism.

Third, AI may result in the automation of labour that induces work reshoring, depriving digital sectors in global South countries of foreign capital and income. There is evidence that this may already be occurring in global South countries’ IT sectors. Specifically, many of the areas of comparative advantage global South countries have developed in IT services are those that are highly exposed to AI-enabled automation. Many countries, such as India, have begun investing heavily in IT skills, but it is unclear if this will be sufficient to stem the potential outflow of capital and employment to the global North.

As AI advances, it will become increasingly important for institutions and collaboration to help the global South capture more of the benefits of AI while minimising risks. At present, there is a danger that developing countries in particular will be overwhelmingly exposed to the negative potential effects of AI – including the results of potentially elevated global emissions – without access to the benefits that, in theory, will make AI the most transformative technology of our lifetimes.