© Supplied by Amazon AI is still new and will probably change the world in ways we cannot yet conceive, but it may save more energy than it consumes.
Staying at the cutting edge of technology while maintaining a competitive economy requires energy-hungry AI. But decarbonising the power system and simultaneously increasing the load makes an already tough challenge even tougher. How big a concern is the rising power demand required by AI?
The recent period of rapid data centre growth, in fact, tells a very different story to forecasts of a large and potentially unmanageable boom in power consumption. The UK currently has over 500 data centres, yet saw 275 TWh of electricity flow through its system in 2024 – substantially less than the 390 TWh of electricity generated in 2008.
This reflects increases in energy efficiency, deindustrialisation and price-related energy conservation, which have all reduced electricity demand.
Data centres, largely clustered around London, may cause local grid problems, and they are a source of energy demand growth, but they are not currently overloading the system. At present, the issue is grid access in places where there are also fibre optic cables, internet connection points and a high density of consumers.
Digital services vs. data centre power demand
Nor will data centres overload the grid anytime soon. Demand for digital data is growing exponentially, but that is different from data centre power demand. Power demand growth overall has been negative not just because of factors external to the growth of data centres, but also to some extent because of them.
The processing power of microchips shows constant gains; the ability to process the same amount of data using less power offsets the exponential rate of growth in demand for digital data. Data centres also consolidate computing power, which reduces the need for distributed computing power – i.e. a company housing its own servers – which is generally less efficient.
In addition, digital alternatives can replace more physical energy-intensive activities, which results in a net reduction in power demand across the economy.
All these factors concentrate power demand in data centres, which makes them and their power consumption much more visible. As a result, the spotlight has been turned on data centre capacity and new construction as metrics for assessing future demand.
However, again, this is misleading. Computing power does not equal power demand as most data centres are over provisioned – they must have sufficient capacity, plus a safety margin, to work reliably at full tilt on the hottest day in summer when cooling requirements are most extreme.
© Supplied by NAThey may also be over-provisioned to account for expected growth, as it is more expensive to retrofit increased power supply than to cater for growth in the construction phase. And, in any case, they may not always be working at full capacity.
Older, less efficient data centres, will also undergo refurbishment or replacement, reducing the amount of energy required to process a set amount of data.
As a result, power demand for data centres is increasing in a linear fashion, not at the exponential rate evident for digital services. At the same time, total power demand is not rising, in part, because data centres reduce power demand for computing elsewhere and replace other energy-intensive practices, and, in part, because of other factors in the economy reducing electricity consumption.
The Irish experience
Not convinced? As of April 2024, Ireland had 82 data centres with 14 more under construction. The country will soon host about one fifth the number of UK data centres, but within a national electricity system which is nine times smaller, generating just over 30 TWh in 2023.
In 2023, Irish data centre power demand accounted for 20.7% of the country’s total electricity demand, almost certainly the largest such share globally. As electricity prices remain high, this has caused concern for both government and other sectors keen to secure power. The clustering of demand has caused local congestion, resulting in some new data centre planning applications being refused.
However, even with the construction of 82 data centres in a small electricity system, Ireland’s electricity consumption has been steady since 2008; it has not grown.
Power savings elsewhere in the system – i.e. the decline of distributed computing and more generalised energy efficiency – are offsetting the more visible and measurable increase in demand from data centres.
© Supplied by NADoes AI change the equation?
These are conventional data centres, where the average rack requires somewhere between 8-12 kW. AI data centres consume a lot more energy – more than 60 kW per rack. As a result, the construction of AI data centres, alongside the continued construction of conventional data centres, implies a sharp increase in the amount of power required to feed massive new computing requirements.
In January, the UK government announced its AI Opportunities Action Plan, which aims to expand the country’s compute capacity by at least 20 times by 2030. If this were achieved, in theory, it would catapult UK data centre power consumption forward, giving rise to concern that they will absorb a huge amount of UK electricity generation at the same time that more power demand is required by other sectors such as heating and transport in the effort to decarbonise.
Such is the concern, that the action plan itself references the possibility of the need to build a series of Small Modular Reactors (SMRs) – an experimental, and financially unproven technology – to power the AI revolution.
The plan highlights the more than £25 billion announced in private sector investment in UK data centres since July 2024. These investment plans have been supported by the UK’s designation of data centres as critical national infrastructure, which in turn promises more streamlined permitting.
As part of the Action Plan, the government will create AI Growth Zones, which will have enhanced access to power and support for planning approvals. To drive this forward, the government is working on its Compute Strategy, which will be published this Spring.
DeepSeek upends power demand projections
However, with the ink still wet on the Action Plan, China’s DeepSeek has thrown current projections for AI data centre power use into extreme doubt.
DeepSeek has trained an AI chatbot at a cost of about $5.5 million, while US companies like Meta AI and Microsoft’s training costs ran, in 2024, into the tens of billions of dollars. DeepSeek says that its servers consume 50-70% less energy than its US competitors. It has created an AI model using a fraction of the energy on which current forecasts for AI data centres are based.
If DeepSeek, reportedly close to launching a second AI model, has opened the way to much less energy intensive and cheaper AI development, then forecasts for data centre energy demand need to be adjusted substantially downward.
© Supplied by NAAI – more unknowns than knowns
DeepSeek’s revelation simply underlines that there are more unknowns than knowns when it comes to AI, which means many things to many people and has many different branches of development. AI is likely to be both over and under-estimated in equal measure.
However, as with the internet and digitalisation, it is reasonable to assume that it will broadly make economies more efficient, even if it also invents new reasons for consuming energy, just as digitalisation has, run by its powerhouses, the data centres.
The question is whether the impact on power demand will be substantially different to the impact of data centres so far, which itself seems to have been mis-evaluated.
China’s Huawei provides a global digitalisation index, which it says shows a positive correlation with GDP. At the top sits the US, which has more data centres than any other country in the world – more than 5,000 compared with 450 in China and just over 500 in Germany and the UK respectively.
Yet, US power demand in 2023 was 4,494 TWh, a mere 3% above the 5,319 TWh generated a decade earlier, a very slow rate of increase during the period in which the growth of digital services has been fastest. Moreover, US GDP is 12.1% above its pre-pandemic level, compared with 4.7% in the eurozone and 3.2% in the UK.
Power demand will rise from sectors such as heating and transport as decarbonisation progresses. The pool of displaceable distributed computing will eventually be exhausted. The UK, like Europe, continues to develop energy transition related reindustrialisation policies. The external content in which AI develops will be different to the one experienced by digitalisation.
However, these processes will also be enabled by the growth of digitalisation and AI. In the longer run, rather than posing a major challenge for power demand, AI, in fact, could provide the net energy savings critical to total, or near total, economic decarbonisation.
Ross McCracken is a freelance energy analyst with more than 25 years experience, ranging from oil price assessment with S&P Global to coverage of the LNG market and the emergence of disruptive energy transition technologies.
