By Eamonn Ryan
With energy costs, environmental targets and uptime all under the microscope, the way we cool data centres is no longer just a technical concern – it is a strategic imperative.
Hitachi supports 800-volt architecture for next-generation data centres. Image supplied by Hitachi
According to McKinsey’s report, ‘The Cost of Compute Power: A USD7-trillion Race’1, global AI data centre capacity could reach as much as 125 gigawatts between 2025 and 2030 – roughly equivalent to Spain’s entire installed power generation capacity. Achieving this scale will demand extraordinary levels of investment and rapid innovation across both computing and energy infrastructure.
Artificial intelligence isn’t just transforming industries; it’s reshaping the very infrastructure that keeps the digital world running. Behind every algorithm, search query and cloud-based service lies a data centre humming with servers that demand constant, precise cooling. It’s the cooling systems that often dominate the energy equation. Depending on efficiency and design, cooling alone can account for anywhere from 7% of total consumption in advanced hyperscale centres to more than 30% in smaller or older operations.
Is it engineering… or science fiction?
As the quest ramps up to keep our ever-growing digital world cool, Lenovo is hinting that the future of data centres might blur the line between engineering and science fiction. Working with AKT II and Mamou-Mani, they’ve imagined facilities that could make your average server room with direct-to-chip liquid cooling feel positively quaint. Think floating cloud data centres drifting 20km above the Earth, villages of servers cozying up to rivers, or secret subterranean bunkers quietly humming under our feet. It’s enough to make you wonder: is this cutting-edge tech, or the opening act of the next sci-fi blockbuster?
The designs include:
- The Floating Cloud: A novel concept involving the suspension of the data centre in the air, at an altitude of 20–30km (safely away from commercial aircraft), with 24/7 energy from solar power and using pressurised closed liquid cooling loops to prevent air pollution. Smaller modules make the design feasible for floating in the air.
- The Data Village: Located close to water sources such as rivers or canals, the Data Village involves a modular, stackable brick or pod system of data centres linked to city needs. This location benefits from enhanced liquid cooling and the ability to transfer waste heat to power or heat local amenities like schools, or homes, as well as reducing latency thanks to its proximity to key locations. This model also extends into a Data Spa. Powered by geothermal energy, this concept blends into natural landscapes (such as a valley, lagoon, or geothermal pools), creating a low-visual-impact and biophilic data centre.
- The Data Centre Bunker: Utilising disused tunnels, bunkers or transport systems, the Data Centre Bunker places the technology underground to minimise the need for new spaces to increase capacity. This reduces land use while allowing for facilities to be set up in central locations with a lower impact. Benefitting from enhanced security resilience, the subterranean location creates a naturally efficient heat management system.
Image supplied by Lenovo Infrastructure
Data centres face a cooling crisis as ai demand surges
By Dean Wolson, general manager: Africa – Lenovo Infrastructure
As AI workloads drive unprecedented power densities, with racks drawing 20–100kW or more, cooling shifts from a supporting role to a frontline constraint. For South Africa and much of Africa, this is magnified by high ambient temperatures, a historically unstable grid and rapid growth in cloud and AI infrastructure. While the continent’s per-capita data centre electricity use remains below 1kWh per person today, it is projected to double by the end of the decade, with South Africa leading the surge in demand as the hub for hyperscalers and sovereign AI projects.
This is the thermal tipping point: air cooling, once good enough, can no longer scale to handle the energy intensity of AI’s future, particularly in Africa’s climate and grid context.
Air cooling has been the industry default for decades because it’s simple and cheap at low power densities. But when rack power draw jumps into the tens or hundreds of kilowatts, as
AI-optimised racks do, air becomes inefficient: the temperature delta required gets larger, fans and CRAC units consume more power, and thermal throttling risks increase. Research and market analysis across 2023–2025 show an industry pivot toward liquid cooling, direct-to-chip or immersion, because liquid moves heat far more efficiently than air and reduces total facility PUE (power usage effectiveness).
Air-cooled systems can support up to approximately 70kW per rack in the data centre, but hit a physical limit when it comes to removing heat, known as ‘specific heat capacity’. Beyond this, operators need to look at alternative cooling methods. Adding to the complexity is the rise in demand for AI workloads across every industry, requiring five times the amount of power. The industry is now seeing a shift from chips operating at 120 watts to 600 watts or more.
Africa and SA’s footprint today
Per-capita data centre electricity consumption in Africa remains low today, the IEA notes. Africa had less than 1kWh per person of data centre electricity consumption in 2024, but is expected to roughly double toward the end of the decade as cloud adoption, hyperscaler investments and sovereign AI projects scale up.
Market research firms place Africa data centre power and cooling markets as high-growth segments in 2024–25, signalling rapid investment but also rising on-site power demand.
South Africa is the regional leader in capacity and the obvious hub for hyperscaler and local large-scale deployments.Yet its grid has long operational constraints: coal still supplies the majority of electricity, and historically unstable supply (load-shedding) and rising peak demand have been persistent issues, though there were signals of improvement in 2025; the structural challenges remain. This combination of rising local demand from data centres, heat-intensified cooling loads and a constrained, carbon- intensive grid is why the cooling question is particularly acute in South Africa.
Higher ambient temperatures mean air cooling must work harder. In warm climates common across much of sub-Saharan Africa, the capacity margin available to dissipate high-density AI heat loads using air is much smaller than in temperate zones.
That raises two problems. Higher operational costs, which means more fan and compressor energy, and a higher probability of hitting physical thermal limits that force performance throttling or prevent deployments altogether. Ember’s 2024–25 electricity reviews show heatwaves already boosting residential A/C demand, meaning data centre cooling will compete more directly with other priority electricity uses.
Image supplied by UL Solutions ‘Safety Certification of IT Equipment Cooling in Data Centers White Paper
What scalable alternatives and mitigations look like
To avoid hitting the thermal tipping point, Africa’s data centre industry will need a mix of innovations and policy support.
The first is liquid cooling adoption, whether direct-to-chip or immersion, which has been proven to reduce cooling energy and enable far higher rack densities, essential for AI-heavy facilities. Second, operators must embrace on-site generation and co-located renewables, especially solar plus battery systems, which can ease peak grid loads and provide resilience against outages, a strategy that aligns with Africa’s accelerating solar boom.
A third opportunity is waste-heat reuse and district heating, which captures and exports server heat to improve overall efficiency, though practical deployment depends on local infrastructure. Fourth, governments should prioritise regulatory and power-market reforms: predictable tariffs, streamlined permits for on-site generation, and demand-response mechanisms can lower risk and unlock investment, as highlighted in OECD policy reviews.
Finally, Africa may need a hybrid edge vs centralised model, distributing smaller edge sites for latency-sensitive services while building a few high-capacity, liquid-cooled hubs in locations with abundant low-carbon power, stronger grid capacity or naturally cooler climates.
Reference:
- The cost of compute power: A $7 trillion race | McKinsey
