Can AI also tackle climate change?

In the wake of AI, major US technology firms are investing heavily in the expansion of energy-intensive data centres – yet they remain committed to ambitious climate targets.

“You can see the computer age everywhere, except in the productivity statistics,” said economist and Nobel laureate Robert Solow in 1987, describing the “productivity paradox” of the computer age that was then just beginning. He had observed that, despite the increasing prevalence of computers and IT infrastructure in the 1980s, labour productivity barely rose – contrary to what tech enthusiasts had expected and promised.

It was not until the late 1990s, with the advent of the internet, better software and organisational adjustments, that productivity improved measurably as a result of this technology. That is why many economists today refer to the preceding phase as a “paradoxical period”, which could be explained by costly adjustments and transitions and could therefore be regarded as a transitional phase.

There is currently renewed debate over whether the strong presence of artificial intelligence (AI) will lead to productivity gains, as even now only minor positive changes in productivity are measurable in many industrialised nations.

Undeterred by this, major US tech firms continue to announce the expansion of their data centres and are spending vast sums of money (or raising funds on the capital markets) to do so. Faith in the significant social and economic benefits of AI therefore remains unshaken.

Paradoxically, the tech firms’ belief that they can achieve their ambitious climate targets despite the AI boom remains just as unshaken. Neither Alphabet, Meta, Microsoft nor Apple have taken advantage of the politically favourable climate under US President Donald Trump to drop their climate targets, which are not only ambitious but almost boastful.

Rising emissions despite reduction targets

In the early 2020s, all four set “net-zero” or even “carbon-negative” targets that they aim to achieve by the end of the decade. And that means they must reduce a large proportion of their emissions by then and offset the rest, so that the total emissions amount to net zero. “Carbon-negative” even requires removing more carbon dioxide from the atmosphere than is produced.

However, the current environmental balance sheets of these companies show: energy consumption rose rapidly after the targets were set, as did the greenhouse gas emissions actually generated on-site through electricity consumption; and emissions from the value chain are also climbing steadily. Or, to paraphrase Solow: “Net zero is visible everywhere – only not in the greenhouse gas balance sheets.”

The fact that the companies are publicly sticking to their promises may well be yet another expression of grandstanding and overconfidence. Perhaps, however, they see in their greenhouse gas balances merely a glimmer of hope that the paradox of climate pledges might resolve itself in a few years’ time, just as the productivity paradox observed by Solow did; there are indications that this might be the case.

Renewable energy

Almost all major cloud and software providers “source” their electricity 100 per cent from renewable energy. This works because, where no renewable energy flows through the grids, certificates can be purchased from other locations worldwide where it is available.

Furthermore, all major companies have invested in renewables, for example in the construction of wind farms, or have entered into long-term energy purchase agreements, known as “Power Purchase Agreements”. This ensures that such facilities are built on a sound economic footing. Both practices send market signals regarding the expansion of renewable energy.

As these capacities alone are not sufficient, Big Tech is also increasingly interested in other carbon-neutral technologies, primarily nuclear power. In addition to the annually increasing off-take agreements for renewable energy, several major technology groups are directly financing new projects or investing in them. When these capacities come online, greenhouse gas emissions from electricity generation will be reduced in the long term.

High-emission capital goods

Most Big Tech firms have their direct and electricity-related emissions well under control – at least after offsetting with carbon credits. Indirect emissions in the value chain (“Scope 3”) pose a significantly greater problem.

This is because, to achieve “net-zero”, emissions in the upstream and downstream value chain must also be reduced – for example, through the procurement of goods and services, the construction of buildings, or the transport of products to customers. For technology companies, the upstream category of capital goods currently accounts for the lion’s share of these emissions.

This includes, for example, emissions from cement production for newly built data centres, from construction vehicles, or from the manufacture of servers and chips. At Meta, for instance, these emissions rose from 2.5 million tonnes in 2021 to around 5.5 million tonnes in 2024. It is striking that the share of the capital goods category in total value-added emissions – which comprise 15 categories – has thus increased from 43 to 68 per cent. The other categories did not, therefore, develop to the same extent.

The category that is usually the strongest, “purchase of goods and services”, fell by around one million tonnes of CO₂ in absolute terms over the same period, or from 51 to 23 per cent in relative terms.

These emission figures are worrying in that data centres built today will only lead to higher energy consumption in the future – which in turn will have to be generated by renewable or at least carbon-free energy. Viewed positively, however, the figures also mean that once the investment boom in data centres subsides, value chain emissions are likely to normalise or even fall significantly if the increasing availability of renewable energy and efficiency gains – including those driven by AI itself – lead to lower emissions among suppliers and customers.

And so the hope remains that in a few years or decades, the climate target paradox will have to be commented on in a similar way to how Robert Solow once commented on the productivity paradox: “I did not predict the future, but described what was true at that time.” Solow, however, did not shy away from a detailed analysis of how the paradox was resolved.

Availability alone is not enough

The actual productivity turnaround required not only the availability of computers, but also appropriate processes and the right management. IT investments only translated into productivity gains when they were linked to far-reaching organisational changes, complementary investments, better indicators, training and, above all, process changes.

This describes the so-called “Walmart phenomenon”. For instance, at US retailer Walmart, productivity was significantly boosted by IT, for example, by using real-time data to transform warehouse management, purchasing practices and in-store processes.

It therefore stands to reason that not only are further “accompanying” changes needed to achieve productivity gains through AI, but also that more is required to meet climate targets than simply expanding renewable energy.

Firstly, we believe that establishing real-time energy and emissions monitoring would be important for planning operational processes according to energy availability and market conditions. Even today, non-time-critical computing workloads (for example, for AI training purposes) are shifted to times when more renewable energy is available on the grids and the price on the electricity exchange is therefore low.

Making grids more flexible, the use of increasingly cost-effective and larger storage facilities, and the integration of CO₂ signals into planning, reporting and remuneration – including, for example, internal CO₂ pricing – are important accompanying processes for achieving progress in the future. Neither AI nor renewable energy can be seen as panaceas to be applied across an economy. Both technologies must be effectively integrated. But that requires time and a willingness to adapt.