1 day ago
11
AI is now a driver of data centre expansion everywhere in the world—across Europe, the Middle East, and Africa. In Europe, the supplied projection puts data centre capacity growth at a compound rate of 25% to 2030, ahead of the effect produced by the shift to public cloud infrastructure over the past decade.
AI workloads change data centre design, and large language models and other AI systems draw more power and produce more heat than many enterprise workloads. Facilities planned around lower rack densities now face requirements that can exceed earlier assumptions. For colocation providers, cloud companies, and data centre operators, this creates engineering, delivery, and cost problems. AI-ready capacity has to be deployed inside limits in grid availability, fibre infrastructure, permitting, regulation, and sustainability reporting. Operators must also protect uptime and site economics.
The response should cover the power chain from the grid connection to the processor. A grid-to-chip model connects power conversion, distribution and cooling in one design, rather than treating each layer as a separate system.
Capacity demand and infrastructure limits
The constraints often begin before construction. In many regions, grid connections and fibre networks need upgrades before a data centre can operate at the required level. Those works can be slowed by planning and permitting processes, while local rules can restrict where facilities are built.
The pressure inside data centres is also changing. Racks are often operated at densities of about 5kW to 10kW, but AI workloads are already pushing some rack densities beyond 100kW, with projections reaching up to 1.2MW by 2028. At those levels, power distribution and heat removal become design issues.
A facility built for lower-density workloads may not cope with higher current, greater heat output and the closer relationship between IT equipment and cooling. Operators therefore have to consider power distribution, thermal management, and energy efficiency as one system.
Grid-to-chip design
The grid-to-chip approach starts from the premise that losses occur at each stage of the power path. In a high-density AI environment, small inefficiencies in conversion can create larger energy losses and added heat. That heat then raises cooling demand, which adds load to the facility.
An efficient model focuses on reducing losses between the grid and the processors, combining higher-voltage distribution, power conversion, and cooling systems designed for dense compute. Higher-voltage distribution can reduce current and resistive losses, while fewer conversion steps improve efficiency.
The same logic can be applied to operations. Embedded AI and machine learning systems are used to adjust cooling, monitor uninterruptible power supplies and batteries, and support energy orchestration. If the stated aim is lower energy consumption, longer equipment life and better uptime, large deployments following simple rules could save several million dollars a year in power, although, of course, the result would depend on site size, energy prices, load profile, and the nature of the system being replaced.
The change in design parameters is one of a move away from optimisation in silos. Power, cooling, and IT systems are often specified by separate teams or vendors, but in AI facilities, that separation can leave efficiency gains unused and make heat harder to manage. A more integrated design seeks to deliver power closer to the rack and align cooling capacity with the thermal profile of GPU clusters.
Modular build-out
Modular data centres are gaining relevance for AI projects, ranging from single-rack systems to containerised units. Modular allows capacity to be added in phases to reduce the risk of building more capacity than demand requires, while giving operators a way to deploy infrastructure before larger facilities or grid works are complete.
The main advantage is speed. Modular units, prefabricated and tested before arriving on site, reduce construction work at the data centre location. For AI services, where demand can change quickly, phased deployment may be more practical than a single large build.
A European telecom operator that used prefabricated modular data centres to expand a 5G edge network was expected to take about 2.5 years to build out, while a modular deployment could have been operational inside 16 months. Lower operating costs through energy efficiency, with improved uptime and resilience the gains.
Modularity does not remove every constraint. Some sites still face planning or regulatory limits even on containerised infrastructure. In those cases, modular systems may need external cladding or other adaptation to meet local requirements.
AI is increasing rack power, heat output, and presenting a need for closer coordination between power and cooling. Data centre operators will have to decide whether existing designs can be adapted or whether new projects should be planned around integrated, high-density infrastructure from the start. As GPU systems evolve, grid access, energy efficiency, and deployment speed are likely to remain limits on AI capacity growth.
Want to learn more about Cloud Computing from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Cloud Computing News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.