How AI is radically changing the datacentre landscape

The surge of artificial intelligence (AI) applications has contributed to unprecedented demand on datacentre infrastructure.

Existing facilities are no longer fit for purpose and AI-ready capacity is in short supply, exacerbated by the existing demand from hyperscalers and cloud providers.

Hyperscaler providers, such as Amazon Web Services (AWS) and Microsoft, already consume the vast majority of datacentre capacity in Europe, and are looking to secure even more space to support the expansion of their digital services and outpace competitors. This is increasing the pressure on datacentre developers to increase supply across Europe.

AI technologies, particularly generative AI, require significantly more power than traditional datacentre workloads. For example, the development of an AI training model that looks at thousands of faces to learn what a nose should look like, requires more computational power and energy than a traditional computing environment. In addition, the rising uptake of AI across countries, industries, and functions, is rapidly increasing demand for datacentre space.

The International Energy Agency (IEA) believes AI-powered web searches will boost electricity demand tenfold. By 2026, the IEA expects total datacentre electricity demand to increase from approximately 460 TWh in 2022 to over 1,000 TWh.

Designing datacentres for AI

Datacentre design is also evolving to accommodate AI workloads, due to the need for more processing power compared to generic forms of computing. This generates more heat, which requires a radically different approach to cooling.

Datacentre operators are having to redesign their infrastructure to include liquid cooling, to ensure they can handle the higher power densities required for AI technology. This includes installing direct-to-chip or immersed solutions.

Datacentres like this will need larger dedicated areas to house the advanced cooling equipment. Given this complexity, many operators are choosing to build completely new liquid-cooled datacentres from scratch.

In Europe’s metro markets, where most colocation datacentres are found, there is a lack of power available. This is due to limited grid capacity, sustainability, and political pressure, as datacentre operators compete with residential developers and other commercial users for power.

Additionally, the availability of land that’s within easy reach of high-speed network connectivity and within reasonable proximity of datacentres where they have let space is limited. Consequently, where land with power and network is available, it is at a premium and datacentre operators are also looking further afield to find suitable locations.

Providers are struggling to keep pace with the demand for datacentre capacity from hyperscalers and cloud providers, and this trend is being exacerbated by the demand from next generation AI. As a result, available datacentre space has plummeted in the largest cities of Europe.

Can the surge in AI demand be accommodated?

Existing colocation facilities can to an extent support AI workloads if they can be retrofitted with specialised hardware and cooling equipment. But if, as expected, the use of AI continues to grow, new datacentre capacity will be required to fulfil the demand. This is unlikely to happen at scale in the traditional markets of Frankfurt, London, Amsterdam, Paris, and Dublin, because power and land availability is increasingly hard to find.

To accommodate the requirements created by AI, the industry’s development strategy will have to change. Datacentre operators will need to look outside the European metro markets when seeking locations to develop new capacity. This will lead to the development of smaller, secondary markets in countries such as the UK or France (e.g. Marseille or Lyon), where there may be more power and land available for datacentre purposes.

In the UK, this development trend is accelerating, which is encouraging investors, hyperscalers, and datacentre providers to purchase land for AI-ready data centre development. We estimate that at present 56% of the country’s colocation datacentres are located within 30 miles of London, although operators are shifting their focus outside the capital. For example, the datacentre operator Virtus has announced that it has purchased land in Saunderton, north west of London, where they plan to deliver 75MW of capacity for AI applications.

Lower latency connectivity will become more important for datacentre providers as inference AI is rolled out given the need to deliver services to users. In the meantime, equipment that powers AI training models are being implemented in datacentres; inference AI is expected to follow.

The outlook for AI-ready datacentres

There is no doubt that the AI boom has significantly impacted the datacentre market. Not only is there not enough capacity, but it is complex to create new AI-ready datacentres, as European grids are struggling to supply the power required for this new technology.

Alternative energy sources are being explored including Small Modular Reactors (SMRs) and renewable sources including wind and solar as a primary source of power, but these are not ready to be deployed at scale.

The need for new datacentre sites to not only have power at scale, but also access to high-speed networks, is making it difficult to find new locations.

To fulfil the demands created by AI, there is no doubt that operators are going to need to look outside the traditional datacentre markets when building new capacity.

Leave a Comment