A data center operator is backing tougher UK grid-connection rules as land speculators flood the power queue with “zombie projects” that may never be built.
Many applicants seeking grid power for new data centers are not data center operators. Ofgem, the UK energy regulator, is now changing its grid-application criteria to separate these projects from real ones that would be developed and used as digital infrastructure.
“So many people are applying for power to build a data center who aren’t data center operators. They’re developers and speculators who want to increase their land value,” Séamus Dunne, managing director for the UK and Ireland at Digital Realty, told TechJournal.uk in an interview in London. “They have confused the whole development of data centers in the UK.”
He said land worth £2 million per acre can be worth between £10 million and £20 million per acre once its owner gets planning permission for a data center and a grid connection.
Demand connection applications surged from 41 gigawatts (GW) to 125GW between November 2024 and June 2025, mostly from data centers. In response, Ofgem launched a consultation on July 29 proposing a “Data Centre Commitment Fee” of £237,500 to £712,500 per megawatt.
Ofgem would refund the fee once a project is energized but forfeit it if it leaves the queue early. The consultation closed on September 16, and Ofgem plans to make decisions later in 2026 but has not set an implementation date.
“Developers like us would have to put upfront capital down as deposits to show seriousness of intent, which we think is a good approach, so long as it’s not too aggressive,” Dunne said. “That’s one way of only allowing real projects to be developed, and that’s probably the biggest constraint we have right now.”
The other constraint is building costs and the supply chain. Because so many data centers are being built, generators, chillers, coolers and other equipment needed to build them are in short supply.
“We’re dealing with all of those things one by one, and we’re still committed to developing our capacity,” he said. “We really want to build our platform with connectivity, to have data center capacity in the right place for enterprises and businesses to deploy their AI (artificial intelligence) and their IT.”
He said that rules out big projects to build super data centers in the Nordics to train large language models (LLMs). Digital Realty’s target customers include financial services, life sciences, retail, manufacturing and public sector organizations.
Digital Realty, headquartered in Austin, Texas, operates more than 350 data centers across six continents. At one time, it had about $10 billion to $11 billion in new data center construction underway. It employs a little under 500 people in the UK, or about 1,000 including indirect staff.
Europe’s priciest power
The interview, which covered AI infrastructure, power, and cooling, took place on October 14 at the launch of ORCA Computing’s PT-3 quantum computer, installed in a Digital Realty data center in London.
“The United States is by far the most advanced IT-based and AI-based economy in the world. Generally, what we see is Europe follows a little later, and mostly we see the UK as a first adopter of technology, far more innovative and progressive than the rest of Europe,” Dunne said.
He said this wave of innovation differs from previous ones in three ways:
UK power prices are the highest in Europe;
China is making progress in IT infrastructure, particularly AI, and is choosing to deploy in Europe; and
Most European economies, particularly Germany and France, have recognized that they need to catch up quicker, with Paris and Frankfurt adopting policies to accelerate AI compute deployment.
“The price of power in the UK is the highest in Europe, and that’s a problem,” he said. “The UK needs to be concerned about its power pricing and the cost of doing business.”
Chinese hyperscalers are expanding across the continent at the same time.
“Chinese hyperscalers like Alibaba and Tencent are investing hugely across Europe. The question is, what are they investing for?” Dunne said. “Are they investing for Chinese companies, following other Chinese companies like the car manufacturers, perhaps pharma, or are they trying to grow their European-based business?”
Cooling is another pressure point. Graphics processing units (GPUs) make up roughly 10% of the compute deployed in Digital Realty’s data centers worldwide, he said, adding that this is not a hard figure.
“To get to the rack power densities that the Nvidia roadmap is showing, the only cooling solution is direct liquid-to-chip cooling,” he said. “There’s no question that direct liquid-to-chip cooling will grow significantly, and we’ll have to increase the density of power in our data centers. But the hybrid mix of CPU (central processing unit) to GPU is still not understood yet.”
High-density GPUs such as Nvidia’s Vera Rubin will need that cooling. But building the capacity is getting harder as shortages stretch beyond equipment to the people who build and deploy data centers.
“The supply chain for everything, including cooling equipment, generators, chillers, the plumbing, the people who are able to build and deploy it, the general contractors, everything is constrained,” Dunne said. “Timelines have just got longer and more expensive, and it’s a big problem.”
Inference over training
Digital Realty is betting its growth on enterprise AI use rather than on how models are built. Training is mainly done in the US, while in Europe it is probably growing most in the Nordics.
“We want to only grow with AI inference, not with AI language model training. AI language model training is where GPUs are being consumed right now, but we’re more interested in the applications that will come with AI inference deployments,” Dunne said.
He said GPU servers could make up 30% to 40% of data center servers across the industry within a few years, though nobody knows for sure. Growth is accelerating at hyperscalers and neoclouds, but it is unclear how much will reach enterprise applications.
Businesses can buy compute as GPU-as-a-service from neoclouds such as CoreWeave and Nebius, from hyperscalers such as Amazon, Microsoft and Google, or from frontier model companies such as Anthropic, OpenAI and xAI. They can also own their infrastructure as private AI.
“More and more, what we see is that trials and experiments are done in a cloud model, but the deployment for production is done privately,” he said. “Companies are not sending their proprietary data elsewhere. That’s too secure and precious. For example, a bank is not sending data to the Nordics to train on its own data.”
Private deployments mostly sit in colocation facilities. Hyperscalers and frontier model companies build LLMs using publicly available internet data, but Dunne said companies will train smaller language models on their own data closer to home, a shift he sees as a big opportunity.
Location matters as much as the type of compute. The biggest demand for data center capacity is in south-east England, where connectivity and latency matter. West London is the biggest area of demand, followed by the Docklands, because of latency and networks.
“It depends on the workload and the application. Some applications need very low latency, and some applications don’t need it, so that’s why we think about hybrid IT,” Dunne said.
He cited an enterprise that runs 500 applications to show how needs differ:
Back-office applications and email can be consumed through cloud services, accessed via ExpressRoute or Direct Connect nodes in a data center;
Trading applications need very low latency and sit near other trading firms in West London, the City of London or the Docklands; and
Customer relationship management (CRM) and enterprise resource planning (ERP) systems need some latency to branch offices, but not low latency.
The UK government has designated AI growth zones with more power and land available, and has invited applications to build in them. But Dunne said they will mainly attract neoclouds and AI training, and that few enterprises, such as financial services firms, will build data centers there.
“Globally, to run businesses, you want to build around population centers, where something called data gravity exists,” he said. “That’s generally population centers and cities.”
Skills are the final squeeze. Dunne said Digital Realty has raised the shortage of engineers with industry people and officials, and that the main solution is apprenticeships. The company takes on 20 apprentices a year, a long-term bet as each takes four years to qualify.
With power costly and equipment scarce, he said the company remains committed to building UK capacity for enterprises running AI rather than training models.