Global electricity demand from Data Centres set to double

A recent report from the International Energy Agency (IEA) has outlined that global electricity demand from data centres and artificial intelligence is expected to double by 2026 to over 1,000 terawatt hours. The facilities consumed 460TWh of electricity in 2022, roughly 2% of the world’s total (although this share is much larger in particular countries). For context, the average UK household consumes about 2,800 Kilowatt hours (kWh) of electricity each year⁽¹⁾. One TWh is equivalent to one billion kWh.

Data and energy requirement growth

In an era dominated by technological advancements and data-driven decision-making, the demand for electricity, especially from data centres, has reached unprecedented levels. Data centres are the backbone to our digital lives, supporting everything from our photos on the cloud to patient data in hospitals to large research clusters. Their 24/7 operations and large cooling systems require vast amounts of power and the recent boom in large language models (LLM) and machine-learning (ML), generally known as artificial intelligence (AI), is driving the industry’s electricity consumption to unseen levels.

The LLM ChatGPT, currently uses around 10 times as much electricity than Google search, according to the IEA. A typical Google search consumes 0.3 Wh of electricity and a typical ChatGPT query consumes roughly 2.9 Wh of electricity. If Google alone incorporates generative AI at scale across all of Google Search, considering the nine billion daily Google searches that are conducted, the IEA estimated that such a move would require an additional 10 TWh of electricity per year. GPT-4 was trained on around 25,000 Nvidia A100 GPUs for 90–100 days. Assuming the GPUs were installed in Nvidia HGX servers, which host 8 GPUs each, would equate to 3,125 servers.

Thermal Design Power

To estimate the electricity consumption, we can take the thermal design power (TDP) unit of an Nvidia HGX server, which, denoted in watts, expresses the power consumption of a piece of hardware under maximum theoretical load. Although Nvidia does not directly disclose TDP information, we can take the unit of the similar Nvidia DGX server, which is 6.5 kW.

An Nvidia DGX server running at maximum power for 1 hour, will consume 6.5 KWh, according to the TDP. With up to 100 days model training for GPT-4, it would take about 2,600 hours per server to train the model. If we assume the servers ran at full power constantly, we can multiply the number of hours by 6.5 kW, then each server consumed up to 16,900 KWh of electricity. Multiply by the 3,125 number of servers, this equals to 52,812,500 KWh total demand.

In 2023, Nvidia shipped 100,000 AI servers that, according to the IEA figures, consume an average of 7.3 TWh per year. BSI have architected similar sized 20,000 GPU environments for financial trading companies, which have critically been deployed in Nordic data centres, which harness 100% renewable energy.

Resource consumption

Part of the issue with calculating not just energy usage, but carbon emissions, is that a company's given carbon footprint relies on many more factors than electricity consumption; the location of data centres, the type of grid powering those data centres, the load on that grid and the number of requests a model deals with on a given day all impact the way electricity consumption translate to carbon footprint.

The issue of electricity consumption is one component of the overall environmental impact. Depending on the location on the data centre, there are various other Scope 1-3⁽²⁾ resource usage considerations, such as water usage, which in many cases have yet to be quantified and the associated environmental impact fully realised.

In the past, data centres have relied on air cooling to maintain safe operating temperatures, when rack power requirements remained well below 20 kilowatts (kW). However, today's high-performing racks can easily exceed 20 kW, 30 kW or more.

Microsoft-backed OpenAI, which runs ChatGPT, requires water cooling to cool the supercomputer. The water is pulled from the Raccoon and Des Moines rivers in central Iowa. In its latest environmental report, Microsoft disclosed that its global water consumption spiked 34% from 2021 to 2022, a sharp increase compared to previous years, which analysts attribute to its increased AI demand.

Last year, researchers at the University of Oxford estimated ⁽³⁾ that US data centres collectively consumed around 1.7 billion litres of water per day. However, they note that measuring data centre water consumption is often challenged by a lack of transparency. Microsoft has agreed not to use evaporative cooling in future datacentres at its Phoenix campus in Goodyear, touted as a sustainability improvement because it will reduce water consumption in the drought-prone region. In its ESG report ⁽⁴⁾, Microsoft claims it's using water consumption metrics to guide water reclamation efforts as it seeks to be net "water positive" by 2030.

Energy grid capacity

There are around 600 known commercial data centres in the UK, split predominantly between colocation and hyperscale operators. Recent planned announcements from these operators to build, in addition to market trends, have produced estimates of an almost doubling in colocation facilities. This is accompanied by a potential more than ten-fold increase in hyperscale, would amount to approximately just under 6% of the UK’s total electricity consumption by 2030.

According to the National Grid, forecasting ⁽⁵⁾ beyond 2030 towards 2050 is more difficult to extrapolate, given the timeframe for un-foreseeable developments in data processing needs and solutions. The uncertainties around future data processing applications, the longevity of Moore’s law, the role of future disruptive technologies, lack of government progress ⁽⁶⁾ and agency tracking the UK datacentre landscape, will all play a part.

However, if we were to use existing market data and modelling projections, this could point to annual data centre electricity consumptions in the UK of between 3.6 TWh in 2020 to as much as 35 TWh by 2050.

In the US the energy consumption was at 200 TWh of electricity in 2022, which was equivalent to 4% of the country's electricity demand. The IEA expects that number to reach 260 TWh, about 6% of the country's total energy consumption, by 2026, powered largely by a growth in cloud-based services and increased adoption of 5G networks. Furthermore, the report indicated that an increased adoption of AI into large, existing applications could increase electricity consumption to between 620 and 1,050 TWh by 2026.

The risks, however, won’t be felt evenly. Approximately 33% of the world’s 8,000 data centres are located in the United States, estimates the IEA. Another 16% are in Europe. In Ireland, data centres will make up nearly one-third of the country’s total electricity demand by 2026.

Ireland, Germany, Singapore, China and the Netherlands have introduced restrictions ⁽⁷⁾ on new data centres in recent years, to comply with more stringent environmental policy requirements. Together with the electrification of transport vehicles, networks, electric steel furnacing ⁽⁸⁾ and industrial reshoring, data centre growth is putting some electricity networks at high risk of capacity outages, particularly during certain weather events.

Renewables

In the UK, the Climate Change Committee (CCC) has estimated that up to 85% - 90% of our energy requirements will be provided through electrification of our energy production and renewables. 15% of this energy mix is to be provided by Green Hydrogen to power hard-to-abate sectors.

The Committee’s Sixth Carbon Budget published recommended deployment of renewables at scale, including 40 GW of offshore wind by 2030 and sustaining that build rate to support deployment up to 140 GW of offshore wind by 2050, raising further opportunity for growth and job creation.

In the United States, the Inflation Reduction Act (IRA)⁽⁹⁾, a federal (public) nationwide initiative of $783 billion in provisions relating to energy security and climate change investment was made last year. Hawha Q-Cells ⁽¹⁰⁾ has received a $2.5bn investment, collectively employing around 4,000 people directly, across the state of Georgia. It's the largest investment in the history of the US solar industry of any kind. As a percentage of GDP, this figure stands at 3x the amount of the recently rescinded green investment pledge of £28 billion in the UK.

To reach net-zero emissions by 2050, annual clean energy investment worldwide will need to more than triple by 2030 to around $4 trillion. A team of Oxford University researchers⁽¹¹⁾, led by Professor Doyne Farmer, highlighted in a report ⁽¹²⁾, that a rapid transition to green energy sources such as wind and solar power could save anywhere between $5 to $15 trillion future expenditure, compared to taking no action, with public investments notably having a lower cost of capital.⁽¹³⁾

Pumped storage hydro (PSH) ⁽¹⁴⁾ has a central role within our future net-zero grids and generative circular economies ⁽¹⁵⁾. No single technology on its own can deliver everything we need from energy storage, however no other mature technology can fulfil the role that pumped hydro storage is capable of. It is a mature, cost-effective energy-storage technology capable of delivering storage durations in the critical 10–50 hour duration bracket, at scale and to cover fluctuations associated with decentralised renewable energy grids, such as wind and solar fleets. There are currently four operational in the UK, with an installed capacity of 135 GWh and a 600 GWh facility recently commissioned ⁽¹⁶⁾.

The Nordic solution

Green energy tariffs and carbon certificates, which are advertised as providing energy from renewable sources, do not actually guarantee the overall grid supply will be from clean energy. There are several options to choose from when considering how to decarbonise your AI systems and infrastructure. We are working with a number of data centre partners, which take advantage of a reliable mix of power from geothermal and hydroelectric sources.

Located in the various Nordic countries and owing to their unique geology, lie host to the highest number of hydroelectric dams in the world and Europe's largest renewable substation. Along with geothermal renewable energy sources also, stable power grids and year-round cool temperatures, these certain locations make for excellent future ready data centre hosting. Due to cool local temperatures and various available infrastructure cooling options, whether air or liquid-cooled, this further reduces energy use requirements.

Replacing the majority of the conventional mechanical cooling in a data centre, where over-cooling can take place, can also result in significant operational and capital cost savings.

With all year-round direct energy supply, it allows us to provide consistent and ultra low-cost power provisioning, dramatically lowering carbon emissions and TCO. All hosting facilities are committed to long-term sustainable growth and mitigating environmental impact of its operations.

Energy efficiencies

Machine-Learning technology ⁽¹⁷⁾ can be applied to data centres to optimise their energy usage and increase their efficiency. Hyperscaled data centres can run large-scale operations without significant increases in electricity consumption.

Where data centres don’t benefit from round year cool temperature locations, water cooling can be used. Arriving at a true total cost of ownership (TCO) can be a complex process. Liquid cooling is typically thought to have a much higher Capex, however with technological advancements of modular cooling from Carbon-Z and Submer, this is changing.

According to a cost study conducted by Schneider Electric, the Capex for chassis-based immersion cooling for a 10 kW rack is comparable to air cooling the rack using hot aisle containment. The greater efficiency that comes with liquid cooling can also translate to lower Opex, especially as densities grow.

Conclusion

There is a discernible shift towards renewables, where production now costs less ⁽¹⁸⁾ than from fossil fuels. Direct renewable energy sources, such as solar, wind and hydropower, presently contribute to approximately 25-30% of the global electricity supply. By 2030, it is projected that renewables will generate more than a third of the world's global energy supply.

As global electricity demand continues its upward trajectory, understanding the pivotal role of data centres and the evolving landscape of renewable energy is paramount. The convergence of technological innovation and sustainable energy practices is not just a necessity but a collective responsibility. The path forward involves embracing cleaner energy sources, optimising data centre efficiency, and harnessing the potential of AI and HPC in shaping a sustainable, circular and energy-efficient future.

At BSI, we supply AI technology solutions and offer options for hosting systems in sustainable 100% renewable powered data centres.