‘AI demand is becoming exponentially more data intensive and energy intensive’, noted Howard, of Aurora Energy Research. He sees the new model of AI doubling roughly every five months, with power demand climbing at a slightly lower (though still challenging rate) – doubling every year.

To put AI in context, he compared its computing power to the human brain. While the brain operates at 20 petaflops (equal to 20 x 10¹⁵ floating-point operations per second), AI operates a billion times faster, at 10²⁴ flops. Developing an AI training model takes about four months at a constant load in excess of 100 MW. That totals more than 300 GWh with energy costs in the tens of millions of dollars. Development takes deep pockets, as an AI training model can cost billions of dollars.

Although data demand for AI is high and continuing to surge globally, Howard recognised that there is ‘a huge amount of uncertainty’ due to poor public data with different methodologies and other factors at play. The estimates of current power demand by AI vary from 500–700 TWh, which is roughly the power consumed by Germany, Canada and Brazil together.

By 2030, this figure is forecast to rise up to 2,000 TWh, due to faster AI adoption and digital integration in the workplace driving increased demand. However, there may well be power savings due to improvements in hardware efficiency and software optimisation, counterbalanced by local regulatory hurdles and grid delays.

There are also challenges around AI monetisation given that ‘first mover advantage’ could give way to disadvantage (as seen in the dot.com boom and bust), as cheaper but effective AI models come on the market – like DeepSeek, China’s inexpensive AI open-source model.

Howard also mentioned supply chain pressures with competing demand for materials (like copper and aluminium) which could slow construction and strain energy supplies. As well as geopolitics – like the tariff hikes by the Trump administration.

Nevertheless, the demand for data centres is rising by about 200 TWh per year. That’s roughly the power demand of the Netherlands and Poland combined.

Location, location, location
Location really matters when it comes to siting data centres. Data centres need to be where they can get fast grid access; reliable, robust supply; and, ideally, at cheap cost. By 2030 data centres could account for 2–6% of global power demand, according to Aurora. They will be concentrated in specific localities, such as Virginia and Texas in the US, London in the UK, Dublin in Ireland, Frankfurt in Germany, and Madrid in Spain.

Typically, data centre operators want to achieve timescales of two to three years for planning and construction, whereas the energy sector is mostly working with timescales of five to 10+ years, anticipating grid connections in the 2030s. ‘That doesn’t really wash when you are trying to build a data centre,’ remarked Howard. And that’s not all; they are choosy customers about the quality of the power, too. ‘Data centres want firm access to the power grid, with a guarantee of 99.9999% uptime – which means only minutes of downtime per year.’ He also claimed: ‘The greenness of that power is at best a secondary consideration.’

Between 2024 and 2030, Aurora predicts a 92% increase in data centre demand in Great Britain – although there are long queues for fast grid access; a 73% increase in Germany; and an 80% increase in Northern Virginia, US. Data centre demand is also growing in Dublin, driven by tax incentives, but with current data centre power demand exceeding 20% there is a moratorium on new construction until 2028.

There are numerous data centres around London, but as capacity is becoming a bit constrained, they are spreading across the country. The biggest, a £10bn project led by Blackstone subsidiary QTS, is located in Cambois, near Blyth, Northumberland, and involves repurposing a former power station.

However, Spain is set to leap-frog other European data centre locations, with 339% growth forecast from 2024–2030, given access to strong wind and solar power resources that provide 95% green power matching. However, concerns have been raised by the recent Iberian power cut (despite the fact that investigations into the causes of that incident remain ongoing). ‘Though Spain comes out favourably as a European data centre location, it is not cost competitive in global terms,’ said Howard. Still, Spain shows limited cost differences between grid and green options, driven by low-cost solar generation that supports high renewable shares without significant premiums.

But the US outcompetes Europe on the cost of grid power.

Generally, ‘diversified portfolios across sites and technologies help smooth cost spreads,’ suggested Howard.

Data centres want firm access to the power grid, with a guarantee of 99.9999% uptime.

Green credentials
Some data centre operators are intent on demonstrating their green credentials. At the other end of the spectrum, some are moving directly to power on site with gas turbines despite the carbon emissions, while others express interest in the potential of small modular nuclear reactors (SMRs).

Howard considers that gas turbines will be an ‘interim option’ on the road to renewables, but favours a hybrid solution. This typically combines a number of different power generation solutions, including green gas and storage technologies. For system resilience, hybrid solutions have the potential to connect to the grid as well as running in ‘island mode’ during grid failures. ‘Most operators are looking at having a multitude of different sources on- and off-site, with access to renewables, natural gas and access to the grid – for a robust, hybrid solution with reliability,’ he said.

Today, the most common approach for data centres to procure power is by direct grid supply from a distribution or transmission grid via a retail tariff. This approach offers a relatively secure supply, and a carbon footprint in line with grid carbon capacity. Moreover, offtakers can buy green certificates to improve their green credentials.

Data centres are also entering power purchase agreements (PPAs) with generators. The PPAs can be ‘sleeved’ (where a licensed utility company acts as an intermediary between a power generator and an offtaker to deliver electricity) or virtual, depending whether a supplier or intermediary is involved. There is also potential to source a higher share of green power, or secure PPAs directly with utilities or aggregators to access tailored generation profiles.

Behind the meter, there are options for direct power supply using natural gas, biofuel, SMRs or batteries which are connected directly to the data centre, bypassing the grid. In fact, most data centres have behind-the-grid back-up systems.

Adding a battery energy storage system (BESS) onsite will not only smooth out the profile of bills but also provide another source of power to guarantee uptime. ‘Adding batteries effectively reduces the amount of renewable capacity needed to draw on, and can reduce the excess MWh required also, as a low-cost solution to hit the goal of 99.9999% by green power matching,’ explained Howard.

Data centre optimisation
Finally, Howard described a tool which Aurora has developed for optimising a data centre’s procured power mix.

The tool was built on its LUMAS valuation software for PPA simulation. It provides pricing across 15 markets, including Germany, the Netherlands, Great Britain, Poland, Spain, Finland, Sweden and Denmark.

The tool can profile the data centre, looking at load factors for a particular location, and put an ‘optimiser’ in the middle to establish the best mix of renewable capacities to hit a certain level of green power matching.

‘Some people talk about 24/7 green power, but I think matching at this level would be extremely difficult,’ said Howard. ‘The model allows us to try out different capacity mixes.’

To do that, the model examines half-hourly load factors, taking into account weather and other locational load factors, PPA prices, on-site BESS configuration, green power matching requirements and half-hourly demand profiles for flexibility of DC demand.

Nevertheless, Howard recognises there is tremendous uncertainty about the future power demands of data centres: ‘Even looking ahead to 2030 is a tough game,’ he admitted. Aurora sees rough doubling of power demand but acknowledges ‘there’s a really wide range of views out there, and many factors at play’. Nevertheless, data centres require quick and highly reliable access to the grid and are currently chasing locations round the world. Bearing in mind the need to be 99.9999% on… where ‘green power is nice to have’.

• Further reading: ‘Power hungry: How AI fuels data centre energy demand and calls for more sustainability’. Artificial intelligence is changing the game for data centres across the globe. However, its adoption is not simply about handling complex tasks faster, it also consumes far more power and creates a surge in greenhouse gas emissions. As a result, tech giants are under pressure to make their vast data centres more energy efficient and sustainable in concert with rapid AI growth.
• AI is poised to drive a significant increase in electricity consumption from data centres worldwide, according to new analysis from the International Energy Agency and BloombergNEF. However, it is also expected to provide opportunities to cut costs, enhance competitiveness and reduce emissions.