AI’s Infrastructure Challenge: Power, Cooling, and the Cost of Getting It Wrong
Sat 26 Sep 2026
As AI workloads place new demands on data centres, electricity availability and cooling capacity are becoming central to AI infrastructure planning. Insights from London, Frankfurt, and Paris have revealed how organisations are reassessing the infrastructure needed to support digital growth.
An enterprise may have approved its AI strategy, selected its hardware and secured funding, yet still face delays because the electricity connection is unavailable or the chosen data centre cannot accommodate the required computing density.
These constraints are changing how organisations plan AI infrastructure. Decisions about where workloads run, how facilities remove heat, and when capacity becomes available can determine whether a technology investment is delivered on schedule.
Research and industry discussions across Tech Show London, Tech Show Frankfurt, and Tech Show Paris have examined different aspects of the problem. Together, they suggest that infrastructure feasibility needs to be established earlier in AI investment planning, with greater attention to the operational and environmental consequences of those decisions.
AI Roadmaps Face Physical Infrastructure Constraints
The scale of the challenge is substantial.
The International Energy Agency projected that global data centre electricity consumption could more than double to approximately 945TWh by 2030, with AI expected to drive much of the increase.
However, the availability of electricity varies significantly by location, and additional generating capacity does not automatically translate into timely grid connections.
In the UK, the Government identified access to grid connections as the “single biggest blocker” to establishing AI Growth Zones.
A Tech Show London Insights analysis examined how these infrastructure constraints can expose weaknesses in enterprise planning.
AI leaders may forecast demand through models and expected usage, while infrastructure teams translate those requirements into computing capacity. Facilities teams must then determine whether the necessary power and cooling can be delivered.
Problems can emerge when these assumptions are tested after major investment decisions have been made.
A site may have sufficient overall electrical capacity but lack the ability to support the density of a proposed GPU deployment. Another location may meet the technical requirements but face a connection timetable that delays production.
The consequences of delayed power availability have already emerged in major UK projects. Microsoft recently withdrew from Nscale’s planned £2 billion AI data centre in Essex following delays to the site’s electricity connection, leaving the developer exploring alternative power sources.
The London analysis argued for a shared capacity plan connecting workload forecasts with infrastructure requirements and delivery dates.
Higher Rack Densities Are Changing Cooling Requirements
The pressure on infrastructure is felt beyond the availability of electricity.
At Data Centre World Frankfurt, Dimitri Wolf, Global Sales Manager at aquatherm GmbH, questioned whether conventional data centre designs remained appropriate for the thermal demands of high-density computing.
During his session, Are We Cooling the Wrong Thing?, Wolf referred to the long-established 19-inch rack design.
“It was never meant to handle 50 or even hundred kilowatts of pure thermal power inside the rack,” he said.
As processors generate more heat within smaller physical spaces, cooling requirements can affect facility design and the viability of existing infrastructure.
Equipment suppliers are also developing more coordinated approaches to infrastructure design. Trane Technologies and Eaton recently introduced a reference design combining electrical and thermal systems for AI data centres, with the companies claiming improvements in energy efficiency and installation costs compared with conventional approaches.
The Frankfurt Insights article also explored what happens to that heat after it has been removed.
Wolf discussed potential industrial uses for recovered heat, but noted that the availability of suitable recipients depended on location. Data centres developed outside established district heating networks may have limited immediate opportunities for reuse.
He described aquatherm’s thermal plug concept, which allows operators to prepare a cooling system for a future heat recovery connection.
The broader engineering consideration is whether facilities can accommodate changing workloads without requiring extensive reconstruction.
Preparing for future heat reuse may preserve options, although its eventual value depends on local demand and the infrastructure needed to transport the heat.
Audiences Want Evidence of Sustainable Infrastructure
A survey involving attendees of Data Centre World, Cloud & AI Infrastructure, the Cloud & AI Leaders Summit and Tech Show Paris found that infrastructure energy consumption was the leading sustainability concern, cited by 60.3% of respondents. Data centre cooling followed at 44.1%, while 36.8% identified responsible AI as a priority.
The research also found that 51.5% of respondents wanted more technical content on energy, cooling and infrastructure, while 50% wanted more real-world use cases and customer feedback.
These findings, detailed in the Tech Show Paris Insights analysis, suggested an appetite for evidence showing how infrastructure decisions affect environmental performance.
The research identified a further gap between interest in sustainable technologies and the ability to find relevant solutions.
Only 30.9% of respondents reported having clearly identified solutions capable of improving their infrastructure’s sustainability, while 47.1% had identified them only partially.
The figures reflect the surveyed event audience rather than the wider technology industry. Nevertheless, they provide an indication of the practical information these professionals were seeking.
For operators and technology suppliers, the findings suggest that claims about sustainability need to be supported by details of implementation, performance and operational limitations.
Infrastructure Planning Requires Shared Accountability
The three perspectives converge around the decisions organisations make before committing to additional computing capacity.
London’s analysis identified the need for shared ownership of infrastructure assumptions. Frankfurt demonstrated how changing hardware requirements affect cooling design and future heat reuse opportunities. Paris showed that surveyed technology professionals wanted more evidence of how these challenges were being addressed.
These considerations also influence cost.
A facility designed for the highest possible AI demand may leave expensive capacity unused if adoption develops more slowly than expected. A design that cannot accommodate future density requirements may require further investment.
Energy efficiency and heat recovery can improve the operating case for a facility, but their benefits depend on technical conditions and whether the proposed systems can be used effectively.
For enterprise leaders, the practical starting point is a capacity plan that connects expected AI workloads with compute requirements, power availability, and cooling capability. The plan also needs to account for delivery timing and the cost of adapting infrastructure as demand changes.
The IEA’s projections illustrated the potential scale of future electricity demand, while the experiences discussed across London and Frankfurt showed why individual projects need more detailed assessments than headline capacity figures alone.
The Paris research has added another expectation: organisations should be able to demonstrate what their infrastructure decisions achieve.
As AI investment continues, credible deployment plans will depend on whether the physical infrastructure can support the workloads being proposed, and whether the resulting costs and environmental impacts can be measured.
