Do we need an AI data centre, or do we just want one?

This is the second article about the Digital Infrastructure North conference. The first DIN article is linked here.

The second part of the day was about data centres. The speaker from the company that sells very expensive equipment to put into data centres thought it was very important that we built more data centres. Maps were shown, highlighting the AI Growth Zones across the country and the very big gap where the Liverpool – Manchester – Leeds – Sheffield corridor is. Lots of talk about the need for us to fill in that space with data centres, and pretty specifically AI data centres.

Now, there was some pushback on this from the attendees, but the overall view seemed to be that having AI data centres in the city region (to be clear, while billed as “Digital Infrastructure North”, it seemed very much focussed on Greater Manchester – after all, one of the AI Growth Zones is in the North East, but that seemed to only be seen as a threat) was A Good Thing, without much explanation of why. It was also explicitly mentioned a few times that to de-risk the building of these data centres, the public sector should act as a foundational user, a lynchpin with baseline demand.

Now, if I was a cynic, which obviously I am not, I would point out that de-risking in this case is only de-risking on one side. One side just takes on the risk for the other, with, frankly, limited incentives to do so. Leaving aside where all the AI data centre usage would come from in the public sector (where the majority of usage is going to be, let’s face it, through Copilot on M365, and thus unlikely to touch local data centres) what are the benefits that would accrue to the local area thanks to the commitment of public money to an AI data centre?

I’m not talking about the need for data centres in general. Even if AI is over-hyped, there is going to be value to the economy in increasing the total amount of compute available. I am talking about the need for data centres in a specific locality. Why Greater Manchester? What do the people in that area get out of them?

The first argument will always be that it will provide jobs. What jobs? At what salaries? Compared to what else? Data centres themselves do not take a lot of people to run. Small 5MW data centres may need 15 – 30 people, of which maybe half of them actually need to be on-site. 50MW – maybe 80 – 150. 500MW, the hyperscale ambition of the UK government’s AI plans, 500 – 900.

Let’s not forget what we are paying to get this. Very roughly, excluding the IT hardware, a data centre build in the UK would cost between £8 – £15 million per MW. Let’s be optimistic and say £10m per MW. A 5MW data centre would cost you £50 million to build – and provide 8 – 15 jobs locally, and the same number again which could be based anywhere (this is digital work, after all). At the 500MW scale, that’s £5 billion (with a b) to provide 250 – 450 local jobs, and the same number again potentially somewhere else.

(Note these job figures are a lot less than those from UK government press releases about AI Growth Zones – the government figures include short term construction roles, which presumably would feature in the building of any physical asset, not just data centres.)

So that’s a lot of money for not a lot of jobs. Now, it’s not as if the public sector would be paying all of this money – the argument is providing a base load of guaranteed compute capacity purchase. But the numbers involved give you an idea of how much the public sector would need to be committing to make a meaningful contribution to the running costs and loan paybacks needed.

And those costs are just for the physical building – the box to put the compute hardware in. The compute hardware itself will, for Blackwell chips, cost between 2 to 4 times as much as the building – and will depreciate a lot quicker. So your total construction and IT hardware costs are going to range from around £150 million for a 5MW data centre, to £15 – 25 billion for a 500MW facility. Now you see why there aren’t really any 500MW data centres – there are campuses that, across multiple buildings, reach 500MW of compute capacity, but no purpose built locations.

To put that in even more stark terms, the current UK data centre capacity – all of it – is about 1.6GW. Put in three 500MW sites, and you have doubled the UK’s capacity.

The first question, therefore, is whether any contribution the public sector could make in terms of base demand would even move the needle. If the public sector organisations in the area are technically competent and happy with running their own sovereign data centres, there may be opportunity. But if we are talking AI systems, then small, local inference focussed facilities could potentially support the public sector organisations, but the majority of, for example, local authority demand is likely to be via intermediaries like Microsoft’s Copilot. The benefit for Microsoft is reduced latency, but it is not clear this is a current business issue for them – will their customers notice a few millisecond improvement in latency?

The next argument would be that it would support more businesses and jobs around the data centre. This is the classic argument used for supporting or incentivising investment in factories – I grew up in the north east of England, and saw this play out with the support for Nissan to open a factory in the area. Smaller manufacturing companies and material suppliers sprang up in the area to support the main site – there was a known big purchaser available, so demand was there, and so companies were created or grew.

This doesn’t make sense for data centres, though. They aren’t looking for small, local firms to provide the GPU infrastructure. They aren’t buying massive data centre scale water cooling systems for server racks from a company based in someone’s garage. These are precision, high scale systems that are bought from specialist suppliers globally. My gut feeling is that there is no halo effect from building an AI data centre – the cause and effect is swapped, in that high technology companies tend to want more compute, and so data centres spring up near them, not the other way round.

(I note that the big AI companies have begun referring to AI data centres as AI factories – I assume because when people think of factories, they think of the manufacturing kind that provide many jobs, not the virtually depopulated data centres of reality.)

The only way I can see there being a benefit for skills, and potentially more companies and jobs, is if higher education institutes build an AI data centre. And even then, I’m not convinced – universities can produce highly trained and highly experienced AI professionals without them needing to be physically near a data centre.

All of this ignores the emotional desire from some for data centres in a locality. Regional pride is a real thing. But I’d argue that in this case the pride doesn’t require a physical data centre. Highly skilled people can work from anywhere, so the trick for the public sector in economic development is to ensure those people want to be in this area, not where the data centres are. Supporting cultural and social improvements is more likely to make people want to be here than putting big data centres in a neighbourhood.

The placing of data centres in the UK needs to take account of one thing – the competitive advantage of being in that location. There are loads of data centres in London, despite expensive real estate, electricity constraints, and so on, and that’s because being there delivers less latency, and customers in London are willing to pay for that. AI Growth Zones are being put in areas with other competitive advantages – cheap electricity from wind farms, or helping support the development of small modular reactors (because who doesn’t want more nuclear reactors knocking around?). Silicon Sands up in Blackpool has easy access to high bandwidth connectivity, and cheap wind farm electricity, both competitive advantages. Greater Manchester doesn’t have any of these.

So, yeah. Data centres in Greater Manchester – fill your boots if it’s a purely private sector build, with no sweeteners from the public sector (whether that is committing to buy compute, or getting government to waive business rates, or any of the other various ways of helping out) – that’s capitalism, baby, and if you think you can make money, then go for it.

But with my public sector lens, nah, it just doesn’t make sense. Invest in skills, invest in making the place somewhere people want to live. But don’t invest in a big air-conditioned cube at the end of the road.