Why AI Demands a Greater Focus on Data Centre Sustainability

The business world is being changed every day by the impact of AI. Up to a third of European businesses have adopted AI, representing a 32% growth rate over the previous year, which if maintained, could contribute an additional €600bn (US$657bn) in gross value added to the European economy by 2030. 

These figures are further supported by IDC estimates that last year, enterprises worldwide spent US$166bn on AI solutions, including software, hardware and services, with expectations of 27% growth year on year to US$423bn by 2027. The immediate effect of this growth in demand is that forecasts are a more than five-fold growth in the market for European data centres by 2035, from less than 4GW capacity today to around 20GW in 10 years.

All of this comes at a cost, however. That growth in demand has put pressure on energy grids, supplies and other resources, such as land and water. We caught up with Thierry Chamayou, VP of Cloud and Service Providers at Schneider Electric, to discuss how data centre builders and operators can achieve their sustainability goals and maintain their obligations, while addressing demand for services.

Balancing AI's power hunger with sustainable data centre operations

There is no doubt that AI is power hungry. One estimate was that training a large language model such as GPT3 consumed in the region of 1,300 megawatt hours (MWh) of electricity, or about the same as 130 US domestic homes annually. That is not insignificant and is likely to grow as more deployments are made with ever greater adoption. 

“To support this level of consumption, data centres need to cool and condition the workloads,” says Thierry. “That, among other things, takes water. Not every data centre operator releases those water consumption figures, but from publicly available information, we can make good estimates. For example, one of the hyperscalers says that in 2021, its data centres consumed 4.3bn gallons of water. That’s about 16.27bn litres. That is a staggering amount that runs to about 450,000 gallons per day for one facility.”

Other hyperscalers prefer to talk about a measure of water usage effectiveness (WUE), similar to the power oriented PUE, where smaller is better. 

“WUE is a ratio of the consumption of water in litres and the consumption of energy in kilowatt-hours (kWh). By this measure, another hyperscaler says its data centres use 0.19 litres of water per kWh,” Thierry continues. “WUE makes it easier to understand efficient use of water in cooling, but it doesn’t give a full picture of actual amounts, which could vary from facility to facility. Yet another hyperscaler claims that its data centres globally use 0.49 l of water per kWh.”

What is evident from these figures is that data centres generally and even more so driven by AI demand, are consuming vast amounts of water and energy, without looking at the impact of land, materials, or other considerations such as embodied carbon or circular practices. 

“This leaves the challenge for vendors, data centre builders and operators of how to meet demand while reducing consumption, improving efficiency and being able to demonstrate it,” says Thierry. “The first step is to ensure that designs are as efficient as possible from day one. Digital design and modelling tools can ensure that facilities are properly configured to avoid issues such as hotspots, or stranded capacity.”

Secondly, a modular approach whether in a dedicated hall, or via container style modules, allows easy scaling and rapid deployment. 

“Preconfigured modules can be factory tested to ensure they are as efficient as possible from the day they are plugged in,” Thierry adds. 

Schneider Electric's vision for sustainable data centres and the green energy transition

With AI in mind, Schneider Electric has carried out a lot of research to understand the specific requirements of AI workloads in the data centre. Industry partnerships also offer an opportunity to combine strengths to achieve more. 

“Our partnership with Nvidia has allowed us to develop reference designs, including with the new generation of Blackwell processors, to ensure efficiency and reduced consumption with vastly improved performance,” he says. “This generation of processors now offers up to 25 times the processing power of previous generations of technology, while using up to 30 times less power and with a massively reduced physical footprint for high-density deployments.”

The next generation of data centre infrastructure management (DCIM) systems are also powerful tools to provide granularity in control and monitoring to give a full picture of performance. 

“The Intelligence gathered from DCIM allows for more accurate information to feed into digital twins to experiment with power management, for example, to model on-site renewable energy sources (RES), such as wind, solar, or in some cases, hydro,” Thierrycontinues. “This can support energy resilience and reduce reliance on the grid in times of peak demand or fluctuation. Combined with measures such as extended on-site energy storage such as battery energy storage systems (BESS), these facilities allow operators greater flexibility where there may be existing or expected energy constraints, or as workloads grow.”
The natural evolution of these measures is for data centres, or a campus of more than one facility, to be configured as a microgrid. 

“With RES and dispatchable power sources, such as generators running on hydrotreated vegetable oil (HVO), data centres can become self-sufficient in energy generation for extended periods, supported by BESS,” says Thierry. 

However, with digital management systems supporting two flows of energy, it means the data centres can also serve as RES balancing capacity for the public grid, or to supply surplus energy back to the grid during periods of high demand. 

“In that way, through what is termed the prosumer model, the data centre can support the accelerated integration of RES for the public grid, while also enjoying a monetary benefit from any surplus generated,” Thierryadds. “This represents a significant boost in energy resilience and security for facilities.”

The combined initiatives of better design, scalability, efficiency and operation and management, as well as new techniques for the specifics of AI workloads, will allow data centre owners and operators to meet AI demand, while being responsible and sustainable. 

“With new energy mix options that not only support energy resilience and security, but also contribute to RES adoption for the public grid, data centres can play a significant role in the green energy transition for the future,” says Thierry. “A green energy future that will be enabled by the AI that was initially a cause for concern.”

******

Make sure you check out the latest edition of Data Centre Magazine and also sign up to our global conference series - Tech & AI LIVE 2024

******

Data Centre Magazine is a BizClik brand