Data centre energy consumption: a risk to AI development?

The expansion of artificial intelligence (and especially generative AI) has transformed the way in which companies and institutions use digital infrastructures.  Consequently, a serious issue has emerged regarding the growing energy consumption of data centres.

With increased demand for cloud services, distributed processing and the daily use of LLM models, electricity management has become one of the most critical challenges facing the entire tech sector.

In such a scenario, the quest to develop data centres which are able to offer sustainability as well as ensuring efficiency and reliability has become essential in order to avoid innovation being held back by energy issues and environmental limitations.

Why is data centre energy consumption growing so rapidly?

Data centre energy consumption has rocketed mainly do to wider use of high performance GPUs, necessary for the training and inference of AI models.  These processes require massive calculations which generate very high loads and the heat produced by the hardware components requires advanced cooling systems, adding to the amount of electricity consumed.

In light of this, it is safe to say that the growth in generative AI is influencing three critical areas: calculation potential, cluster density and cooling systems.  Each improvement in the hardware increases the thermal complexity that the data centres must manage.  The main issues can be summarised as follows:

• the use of GPUs and specialised accelerators with higher energy consumption compared to traditional servers;
• the need for ever larger models requiring energy intense calculation infrastructures;
• the growing environmental impact of data centres linked to indirect CO₂

The IT sector is now facing a paradoxical situation as technological innovation advances rapidly but the electricity networks available in many countries are unable to match its pace.  This is bringing significant consequences, such as:

• a slowdown in the building of new computational hubs;
• increased costs for cloud providers and companies that use AI;
• difficulties in the expansion of AI infrastructures caused by the limited availability of electricity;
• the need to rethink national energy planning in order to support the digital transition.

Many regions have already imposed limits on the creation of new data centres due to the risk to network stability, reflecting how sustainability has become a strategic priority.

This issue does not only affect technology; data centre energy requirements have wide-reaching political and social implications.  Local communities often oppose new data centres, because they fear that they will divert energy from homes and traditional businesses.

At the same time, awareness of the environmental impact of data centres is rising, with calls for greater transparency regarding energy consumption and more tangible measures to aid the energy transition.   Tech companies are therefore obliged to operate in a scenario in which sustainability, governance and social responsability have become competitive factors.

How the geography of data centres is changing and possible alternatives

The location of data centres is not only a question of connectivity logics or IT security. The top priority is now energy availability.  Companies look for areas with:

• an abundance of renewable energy sources, such as hydroelectric, solar and wind;
• robust elecricity networks with capacity for expansion;
• favourable climate conditions for natural cooling to reduce the energy consumption of cooling systems.

This new development model is creating an authentic ‘global digital energy map’, where colder regions, Nordic territories and areas close to power plants are becoming strategic hubs for AI infrastructure.

As well as identifying advantageous locations, the sector is exploring several alternative technologies, which could one day play a significant role in reducing data centre energy consumption.  Some of the most promising are:

• small modular nuclear reactors (SMRs), considered a possible solution in the medium to long term to ensure a continuous, stable energy supply;
• autonomous micro-grids powered by renewable sources with integrated storage systems;
• sustainable data centres based on optimised energy use architectures;
• hybrid generators, or in some cases, a return to gas power stations, albeit in contrast with sustainability targets.

In the short term, however, it is interesting to note how many operators are investing in data centre energy efficiency via innovative cooling technologies and high performance hardware architectures.  These improvements are not however sufficient to keep up with the increase in AI related energy requirements.

Reducing data centre energy consumption also means rethinking the development of artificial intelligence models.  The most efficient techniques in this sense include:

• distillation of models to reduce complexity without affecting performance;
• quantising to reduce numerical precision and lighten computational load;
• pruning to eliminate non-essential parameters;
• hybrid inference between edge computing and cloud to distribute the load.

The above techniques enable not only the creation of less energy-intense models, but also the reduce the impact on cloud infrastructure and computational resources, thereby supporting the transition towards a more sustainable AI.

The future of data centres: a balance between innovation and sustainbility

Considering these factors, it is logical to expect a constant increase in data centre energy consumption over the coming years, due mainly to the need to power ever more complex and widespread models.

This scenario should not be seen only as a threat however, but also as an opportunity for innovation, starting with the cooling systems and architecture of the digital systems.

Emerging technologies, the use of renewable energy sources by data centres and the evolution of AI-friendly hardware will contribute to creating a new balance between computational capacity and sustainability.  The challenge will be to reconcile the growth of AI with its environmental impact and energy availability, while ensuring a high efficiency future for global digital infrastructures.

The road towards achieving these objectives remains rocky and full of challenges which will require full collaboration between organisations, manufacturers and governments.  It will require a joint effort so that energy requirements are no longer an obstacle to neural network development, but rather a starting point for changing the way humanity makes use of the planet’s precious resources.