Are AI data centres a viable future?

There can be no doubt that AI is the flavour of the year, and over the past 12 months we’ve seen substantial investment announcements for AI data centres in the UK as a result of the visit by the US President Donald Trump. The total amount announced is some £150 billion, including £90 billion from Blackrock, £22 billion from Microsoft, £5 billion from Google and various other ancillaries.

Some of this funding will find its way into data centres; Blackrock have recently broken ground on a 720 MW data centre on a former coal fired power station site in the North East (Blyth), and Microsoft and Google for developments in Nottingham and North London respectively.
However, data centres are extremely complicated architectural machines with a lot of embodied carbon — both that used to obtain, process and manufacture the electrical and mechanical infrastructure and the IT equipment itself, and what’s used in the operation of that equipment (which in the UK will vary depending on the amount of renewables on the grid).

So, especially when we consider the energy and water needed to service AI data centres and their use of rare earth minerals, AI is not inherently sustainable — despite the protestations of its proponents.
Yet the market is going to get bigger — much, much bigger. Is this all really necessary, or are we building ‘white elephants’?

The current state of data centres in the UK

The UK data centre market consists of two main types of data centre: ‘enterprise’, or those owned and operated by the same entity (in some cases including hyperscale and AI factories); and ‘colocation’ datacentres, or professionally run facilities which cater for external clients. These clients can be retail (many customers) or wholesale (one main tenant). They usually sell just power, space and cooling. This legacy estate does not have the capabilities to service the new AI infrastructure, hence the rush to build new AI data centres.

The current data centre sector in the UK is estimated at between 1.5-1.6 GW of capacity, some of which is unused because of a low ‘rack density’ of only about 0.5 kW per rack (rack density is calculated by dividing the total amount of power consumed by the equipment by the number of server racks. Ascertaining just how much goes unused would require a far-reaching survey, or mandating the collection of such information via legislation.

There are also some 278 MW worth of projects under construction (which pre-date the current AI data centre boom and so are likely to be air cooled), 2 GW worth of projects ‘shovel ready’, and a further 3.5 GW worth of early stage projects which have not yet secured a power offer, planning permission or financing. This totals some 5-7 GW worth of projects in the planning cycle, all requiring power and water which is not currently available.

At the time of writing, the additional data centre projects that have been announced in the UK are big. Very big. 300 MW, 1.1 GW, 378 MW are just a few of the power capacity requirements mentioned. And whilst these systems are likely to be cooled via closed-loop systems, the amount of water required will be larger than traditional data centres.
According to technology companies, this quadrupling of the data centre fleet is due to the increased use of tech services, cloud and AI in particular — but the size of these data centres is due to the projected energy needs of AI infrastructure.

The dark side of data centre expansion

My concerns are based on five basic principles. The first is that the additional infrastructure required to deliver these AI data centres will be extremely expensive and involve the ripping up of acres of land for pylons, substations and new water distribution networks.

The second is that, assuming we can actually get them built in the first place, there is already an acute shortage of skilled personnel in data centre design, build and operations functions. Even that is nothing compared to the shortages in HV infrastructure and problems in the construction industry.

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

Thirdly, there are still supply chain problems: recent forecasts suggest that transformers are on three to five year lead times, with other components experiencing delays of between six months to two years.
Fourthly, community pushback. There are already citizens who are concerned about data centre builds in various UK locations. One in Wiltshire has over 1,000 letters of concern regarding noise, lighting and traffic movements (largely during construction) and this will only increase if government policy is to remove planning permission from local authorities and place them under central control.

Lastly, has everyone forgotten about the ongoing climate emergency? The principle cause of global warming has been the combustion of fossil fuels for energy, and unless data centres are powered by renewable energy they will cause even more damage to the environment — and the outlook is, already, not good.

White elephants

So why all this talk of white elephants? A white elephant is ‘a possession that its owner cannot dispose of without extreme difficulty, and whose cost, particularly that of maintenance, is out of proportion to its usefulness.

By focusing on providing for massive power and water consumption, the approach to designing new AI data centres is stuck in the 20th century. But product roadmaps show a very different future: for example, lower power chips which produce less heat and need less cooling, and very high-speed networking. These would result in less energy and resources being required to run the data centres.

Perhaps it would be better to adopt a more ‘mesh like’ edge data centre approach, similar to mobile phone technology: integrating many smaller data centres into the urban landscape using existing power distribution networks, requiring only minor local network upgrades rather than major grid transmission upgrade projects. There is also the potential to use the waste heat generated in local distribution networks, or for low heat industrial applications, in line with the DESNZ Heat Network approach.

We are at risk of allowing huge AI data centres to be built based on today’s technology and roadmap,  sucking precious resources away from other projects that are possibly more useful — for an outcome that is shrouded in mystery. Should we be thinking about a more integrated edge type approach?

We stand at a crossroads, and the direction of travel will impact us all.