Unveiling its AI Opportunities Action Plan at the start of 2025, the UK government announced it has secured £14bn from private sector partners to help build brand new AI data centres across the country.  

The infrastructure – which could see several such facilities built around the UK – is required, say ministers, to take full advantage of the opportunities presented by the development and use of AI across industries around the UK. 

Coupled with a further £25bn the government says had already been pledged for new data centre facilities prior to the latest announcement, and promising to create more than 13,000 AI-based jobs, the Action Plan seems highly impressive on the surface. A media statement which accompanied the plan’s launch describes it as “game changing”. 

But an instant leap into building AI data centre facilities without first understanding the real – and different – computational requirements of AI could lead to building an infrastructure which is greater than demand. We’ve already seen that Microsoft has cancelled data centre leases in Europe they say due to an “an oversupply relative to its current demand forecast”. Perhaps also as a result of its decision not to support additional training workloads for OpenAI?

The distinct phases of AI development 

AI development can be split into two distinct phases. 

The first being training. This is the stage at which the AI models are being trained, where the algorithm is being fed data to enable it to learn patterns and make accurate predictions or decisions. Whilst this phase can be short, it is intense and uses a significant amount of computational power and therefore infrastructure. 

The second phase, is inference. In this phase, the knowledge gained from the training phase is used to make predictions or decisions. This phase needs to be fast and efficient, especially in real time applications but it uses significantly less computational power which is used over a much shorter timeframe.  

In short, training a new AI model is resource intensive, but over a relatively short period, typically somewhere between 3 and 12 months. Compare this short-term use of data centre infrastructure to the average colocation contract of 8 years and its difficult to see how hosting such projects in purpose built infrastructure could be commercially viable. 

When it comes to deploying AI enabled applications, this can be achieved using a more typical data centre infrastructure, this time, location becomes more important as locating such AI applications closer to the end user to limit latency is a key consideration. 

The biggest challenge to the data centre industry in the UK is power 

The average data centre needs around 5mW to 10mW of power. For exclusively AI, that rockets to more like 100mW. When you consider that the UK’s National Grid is currently under massive strain due to the age of our electricity grid, coupled with the long wait time for new connections and the fact that the UK currently has the highest electricity charges in the world, investing in an AI data centre is not without its challenges. 

In addition to power, there’s the question of the UK’s fibre network. Is it in good enough shape to support the vast connectivity requirements which would underpin a world-leading AI industry? According to research the country ranks 51st globally for internet speed – slow compared to many developed nations. 

Is AI in danger of becoming the data centre industry’s dotcom bubble? 

By the end of 2024 US tech giants Microsoft, Google, Amazon, Meta and The Stargate Project – a joint venture between OpenAI, Oracle, Softbank and MGX – had collectively allocated around $395bn to AI infrastructure for 2025 alone (a figure published on LinkedIn by data centre and AI expert Rich Miller). 

But just a few weeks into 2025, the unheralded arrival of DeepSeek threatened the status quo. Its launch showed that such huge resources, and the latest technology, might not be needed for AI after all. As previously mentioned, by late February 2025, reports suggested Microsoft had cancelled lease agreements with data centres operators, totaling around 200mW, partly related to its decision not to support additional training workloads for OpenAI. 

The growth in AI infused applications must and will continue, and resources are obviously required to ensure its success, just perhaps not on the scale being touted. More thought needs to be given to how and where training AI takes place, and where the final deployment of the AI enabled application will be made. 

Investment in AI is to be welcomed. But thinking of making investments in centralized AI infrastructure could be risky. Instead, we need to think about how we can have AI enabled data centres evenly distributed across the UK, which have sufficient access to power and connectivity to handle AI training and deployment. 

As a leader in the industry, Asanti is already in discussion with government officials to advise on AI strategy and the resources required to underpin it. We look forward to bringing our influence and experience to ongoing conversations with policymakers and the wider data centre industry.