5 key takeaways from Data Centre World 2025

Download Listen

We were at Data Centre World 2025 in London last week to evaluate the key narratives among data centre operators, developers, suppliers and investors. Here are our five key takeaways from the event. 

The power squeeze is here

A dominant concern over the two days was the increasing strain on power infrastructure – both in terms of grid connections and managing power on site. As we’ve all heard enough times now, AI workloads are resulting in a significant demand uplift per rack for data centres, while causing power demands to fluctuate by up to 50% in a matter of seconds, overwhelming existing site and rack-level power systems.

In addition to the wider macro trend of power consumption increasing, attendees and presenters voiced concerns about the challenges of securing grid connections, with many grid applications failing to secure the requisite capacity or being pushed out of queues. In an adjacent talk, it was noted that the surge in Battery Energy Storage System (BESS) projects being rejected from the queue is leading to many applicants seeking speculative reallocation as data centre projects – although it was noted that many of these were likely to be unsuccessful due to the differing requirements and SLAs around reliability and redundancy.

In response to growing grid pressures, discussions emphasised that governments and operators must take urgent action – arguing that regulatory frameworks need to incentivise and enable on-site power generation through modular energy solutions collocated with data centre facilities. The UK’s expertise in small nuclear reactors (SMRs) was again referenced as a potential solution in the long-run (2035 and beyond), though this is not a feasible option in the near-term.

Some panellists also advocated for reducing the dependency on cooling solutions through hardware which can be run at temperatures higher than they do currently, potentially lowering average PUE from 1.4 to 1.2 and below. Another longer term mitigation was noted around dynamic power routing between data centres and the grid as a whole as a means to enhance efficiency and further reduce grid pressures.

Site selection parameters are changing

Access to power was the dominating factor in panels and talks on the topic of site selection and design. Many companies are seemingly turning to novel site selection requirements in an attempt to either differentiate the facility, optimise operations and/or costs, or generate new revenue streams from what are currently by-products:

• Underground data centres – can unlock benefits around cooling and security, as well as potential tax advantages on the cost optimisation side, while unlocking a differentiated value proposition which can be monetised for revenue uplift.
• Colocation with energy generation or storage – lots of discussion around SMRs and reducing the dependency on the grid, both for stability and resilience, as well reducing the impact on the national grid for social responsibility reasons (and the associated negative press).
• Colocation with heating customers – collocating with a site or building who could buy back the waste heat generated. Craig Walker, from CIONET, referenced the example of locating an edge data centre under an office building and selling back the heating waste to the local development. However, he noted complexity around aligning contract lengths with anchor customers of data centre services with the provision of heating services – could this be an area we see increasing partnerships?

 

Sovereignty rules are rewriting the data centre playbook

The role of regulation and geopolitics in shaping data centre strategy was a notable topic at DCW. Governments worldwide are implementing data sovereignty requirements, restricting where workloads can be hosted and creating new incentives to attract investment. One presenter highlighted a nascent sovereignty dynamic across developed and emerging markets, suggesting that developed markets could benefit by prioritising sovereign data hosting, allowing them to retain control over critical workloads. Meanwhile, non-sovereign workloads could be offloaded to emerging markets, which may offer cost advantages and scalability to operators, while providing new job and investment opportunities for the host country. For such emerging markets, however, industry leaders highlighted the limitations of the ‘chicken-and-egg problem’ – without reliable infrastructure, investors hesitate, yet infrastructure is dependent on investment.

In respect to sovereign data centre developments, many were also quick to highlight the importance of governmental backing for AI-infrastructure development, identifying the lack of current coordination both within the public sector, and between government, operators, and enterprises. Power constraints, permitting delays, and mismatched priorities were often criticised as slowing industry growth. Against this, attendees pointed to the Middle East as a leading example of rapid progress, where government-driven initiatives are accelerating data centre development – in tandem with strict data sovereignty requirements. While ambitions exist in other markets for innovative and mutually beneficial partnerships between data centres and enterprises, presenters stressed the need for the government to act as a bridge and driving force in enabling such collaborations. The message was clear: without structured public-private collaboration, driven by alignment across different public sector bodies, data centre expansion will continue to face bottlenecks, and countries that fail to align their policies with customer requirements risk falling behind in the global data centre race.

AI requirements are redesigning data centres—and sustainability is still on the agenda

A common theme across the event, particularly from vendor presentations, was how AI workloads are driving a fundamental rethink of data centre design, both within the rack and across the site. The legacy industry standard power density of 5kW per rack is changing, with the most dense training workloads necessitating 100kW+ racks and rapid adoption of liquid cooling and high-density architectures, and inferencing workloads still significantly above the 5kW per rack status quo. However, even with liquid cooling, heat rejection was marked as a critical challenge. Discussions also suggested that many sites may need to transition to medium-voltage power distribution to cope with rising energy demands.

It was also clear from presentations that sustainability is still very much on the agenda for data centres, in particular in Europe. A few of the newer discussion topics included innovations in low-carbon concrete and a push towards a standardised IT carbon disclosure label to increase visibility and simplify reporting, which could become a key procurement consideration.

Nobody knows what data centres will host in 15 years

If you asked someone topping out a data centre in 2010 what was going to be hosted there today, they wouldn’t have been close to the truth. Even excluding the more recent developments in AI, 2010 was the year in which Amazon finalised migrating its own retail business to AWS, and cloud computing was yet to take off at scale. Now, AWS is firmly a hyperscaler, part of a group who hold significant monopsony power over many data centre developers.

And the opacity of medium to long-term demand for data centre services still rings true today, and was a topic of conversation at DCW this year. Uncertainty around genAI adoption curves across regions, use cases and verticals, and the evolving sovereign AI and regulatory landscape will have lasting impacts on the increasingly global hosting market, and this is before we consider the implications of multi-modal applications of machine learning and LLMs, Nvidia CEO Jensen Huang’s CES 2025 keynote is a good place to start as a primer on medium-term AI innovation.

For many years data centre operators have operated agnostic of the workloads being run within their data centres – and for those operating or developing for a single-tenant, they can continue to do this at a relatively risk-free investment, guided in site selection and specifications by their anchor customer(s). However, it is worth considering why hyperscalers are so quick to engage in 10-15 year contracts – with energy and land availability constraining development in many key data centre markets, and demand forecast to increase dramatically, will market rates for rack fees in core markets increase dramatically?

In addition, for those serving multiple tenants, there is an increasing need to understand the changing requirements of a site – both in terms of site location and key specifications such as footprint versus power consumption. Do your sites enable future flexibility to the point where you will be able to compete on price and PUE with facilities completed in 2035, even if rack densities have diversified and increased on average? How does rack density (and therefore site-wide footprint versus power consumption) interplay with your commercial strategy and site location – are you targeting non-real time AI training workloads in a remote location with cheap and plentiful power, or more local inference and traditional compute workloads at an accessible location for enterprise customers?

These were some of the key questions being asked at DCW, with data centre operators seemingly ploughing ahead with inorganic and organic capacity expansion, led by today’s customer demands, as opposed to truly analysing the IT market of 2035, and proactively building a strategy to serve both today’s customers while reserving the flexibility to serve the expected markets of 2035 and beyond.