Interest in sustainable IT is growing. Andrew Nind MMaths MSc, and Siani Pearson PhD, from Just Algorithms Action Group (JAAG), explain how the way that digital technologies are assessed from a climate perspective can give a false impression, and encourage consideration of smaller systems.

Since BCS last reported on the carbon footprint of high performance computing, there has been encouraging progress internationally. Greenhouse gas (GHG) emissions from digitalisation appear to have been relatively stable in recent years: globally, the atmospheric release is comparable with the aviation sector, but the exponential growth warnings from some analysts — associated with video streaming for instance — have not materialised. This is partly explained by larger than expected efficiency gains in data centres, and partly by carbon offsetting, for example by IT companies sponsoring green projects elsewhere. It is also the result of clean energy sourcing; the electricity supplied to Amazon, Google, Microsoft and Apple is now almost 100% renewable. Additionally, amongst software developers, green awareness is improving: for instance the Green Software Foundation, established in 2021, now boasts 64 member organisations. Overall, the IT sector appears to be on track to reach net zero emissions by 2050.

Or is it?

Mismeasurement of emissions

Only by measuring and putting a value on the real carbon footprint of computation can the sector make the right decisions. So when a company says that its operations are carbon-free, we need to ask, ‘what about the embodied emissions in hardware, power networks, maintenance, storage and disposal?’ When supply chains and infrastructure are accounted for, calculated emissions roughly double — that is, they are closer to 4% than 2% of global GHG.

There is, however, a bigger problem than failure to account for supply chain impacts. A problem that deserves more attention than it gets.


In the early days of carbon markets, the European Union (EU) spent much time considering the thorny matter of what it called ‘additionality’; if an EU company seeks to offset its emissions via an overseas project, how could one be sure the project would not happen anyway? This was an easy question to ask, but a hard one to answer.

The relevance to IT emissions becomes apparent if we ask, say, what is the overall annual increase in GHG if a data centre is built in Switzerland? If we note that Swiss electricity has a carbon content of 12 gCO₂e/kWh, this means that for each kWh of electricity generated, the global warming is equivalent to that from the release of 12 grammes of CO₂ (the word ‘equivalent’ refers to the fact that there are different GHGs; CO₂ is used as the benchmark). Let’s say the data centre consumes 10 million kWh per annum. So is the answer 120 million gCO₂e? Unfortunately, as explained below, a realistic increase is actually… 3,500 million gCO₂e.

atNorth’s ICE03 data centre in Akureyri
atNorth’s ICE03 data centre in Akureyri. Source: atNorth

The key point is that the Swiss hydro assets would generate anyway, even if the data centre plans were scrapped. Assuming it’s built, electricity flows would be affected across a wide regional network, leading — in all probability — to a gas fired power station located in a different country (such as the Netherlands) increasing output. Gas is the swing fuel in the European power sector. Once a renewable station is commissioned, it generates whenever the sun shines, wind blows, or water flows, as long as it is technically operative: combined cycle gas turbines (CCGTs), unlike renewables, respond to fluctuations in electricity demand. The carbon content of natural gas is roughly 180 gCO₂e/kWh, and a typical CCGT outturn efficiency is slightly over 50%: hence, CCGTs emit about 350 gCO₂e/kWh: around 30 times the Swiss national average.

When an ICT company signs a renewable power purchase agreement (PPA), the renewable electricity should only be regarded as additional if it would not be produced in a counterfactual world without the PPA. Now it might be that, by providing a ‘route to market’, the PPA facilitates a higher level of renewable generation than would occur otherwise, but this is hard to demonstrate and easy to overstate. In most cases, we would expect electricity from existing renewables — and any new renewables specified in the contract — to be produced anyway, because PPAs are generally signed at competitive market prices.

This doesn’t mean it’s wrong to build the data centre. It does mean, though, that the benefits of building in a ‘low carbon country’ — see box below article — are debatable. This point is underappreciated. The Montreal AI Ethics Institute, for instance, stresses the importance of access to low carbon grids for training and deploying artificial intelligence (AI) systems. But when additionality is properly measured, location may be less important than site power consumption. ‘Proper measurement’ sounds complicated, but while a data centre’s impact over an extended geography may be difficult to quantify precisely, it is usually possible to estimate the broad changes to generation patterns and GHG emissions.

In short, when a corporation ‘proves’ to the outside world that it is carbon free, it may be attributing its impact on the planet to someone else, somewhere else.

Indirect benefits

Aside from direct effects like energy use and emissions embodied in hardware, IT technologies — AI in particular — are often credited with indirect benefits that make assumptions about human behaviour. This is best illustrated with an example.

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

Last November, a Harvard-based research group published analysis on the sustainability of machine learning on low power microcontrollers, or ‘tiny machine-learning’ (TML). TML was found to have a negative carbon footprint from emission reductions achieved when smart devices are inserted in domestic buildings that automatically turn off heating and lighting when they aren’t needed. The difficulty with this finding is that people don’t necessarily need smart devices to reduce energy consumption: rising prices or a publicity campaign, for instance, might lead to behavioural changes with the same outcome.

As with carbon offsetting, additionality — in this case, the indirect benefits — needed to be demonstrated and wasn’t. TML may be extremely promising, but its environmental boon was perhaps overstated by the Harvard analysis.

‘Small is Beautiful’?

To recap, the environmental impact of IT is measured correctly only when the following are considered: full supply chain impact, including embodied emissions in hardware, and additionality, a rigorous test that applies variously to carbon offsetting, clean energy — which is never unconstrained — and indirect benefits associated with human behaviour. Measured thus, the impact is probably very different to what is claimed, with implications for the best course of action, which may be lowering demand.

In the programming world, there is a well established correlation between the size of computer models, on the one hand, and related CO₂ emissions on the other. Packages like Green Algorithms quantify the impact. This suggests that we want — at least from the climate perspective — software developers to focus their efforts on creating relatively small and data-light algorithms where at all possible.

Case study

Iceland is a low carbon country. Its electricity is generated from hydro and geothermal, and much of the geothermal potential is undeveloped. Low ambient temperatures help to prevent overheating. So isn’t it a good idea to build data centres there? Well, not necessarily if they preempt green industrial developments — electric arc furnaces for steel manufacture, say, that could tap into those same geothermal resources. If these resources were infinite, there would be no issue, but they’re not; they’ll be developed at some point, and from a global environmental perspective, we should consider carefully how best to use them.

A range of techniques can be used to constrain computer consumption, including decreasing the number of features used in computation while preserving a high enough percentage of information. Economic incentives often align with reduced computing time and efficiency, but the trend for people trying to apply big data and machine learning in many different situations leads to proliferation of computationally intensive approaches when lower energy approaches might be sufficient or even better.

So Small may be Green as well as Beautiful. 50 years after its eponymous publication, and in a guise he wouldn’t have anticipated, Schumacher’s dictum is as relevant as ever.