The data centre industry is witnessing a seismic shift in scale, moving from a landscape of 20MW facilities to 1GW-scale developments that demand huge increases in power. Michel Grabon, technical director at Carrier, says this transition marks an era of high-stakes execution driven by the rapid solidification of AI.
Michel Grabon, technical director at Carrier
‘The scale of development is accelerating dramatically. The capacity of the incoming hyperscale facilities is unprecedented,’ he adds.
The scale of change
Just a couple of years ago, a large-scale data centre in Europe typically operated at a capacity of 20-80MW. Today, the industry is talking in gigawatts. While 800MW projects seemed ‘crazy’ only a few years ago, Grabon notes, 1,000MW (1GW) projects are now common in the United States, and Europe is rapidly following suit, with new projects ranging from 150-300MW.
This scale presents a big challenge for construction timelines, warns Grabon. ‘Traditionally, data centre projects have followed a rough rule of thumb: around one month of build time per megawatt of capacity,’ he explains. ‘If you go to 100MW, you can’t afford 100 months. Technology will change long before you finish building.’
The rapid pace of technological development – particularly around AI hardware – means operators must design facilities that can adapt to change. ‘Every six months you have new AI components,’ Grabon adds. ‘For data centre customers, it becomes very difficult to keep up.’
As a result, cooling infrastructure is being redesigned for speed and simplicity of deployment. Rather than assembling systems piece by piece on site, manufacturers are integrating more functionality into equipment before it arrives.
‘In the past, we were providing components,’ Grabon says. ‘Now, customers need whole solutions. They expect you to bring equipment to site that can be connected in one day.’
That shift is particularly important at hyperscale, he adds. When hundreds of megawatts of cooling are needed, even small delays multiply quickly. ‘You cannot have three days of commissioning per unit. Everything needs to arrive ready.’
To achieve this, Carrier is pushing the physical limits of size, designing chillers as large as a truck can transport – aiming for 3.2MW of capacity in a single movable piece – to reduce the number of units required on site.
Assessing thermodynamics
The heat produced by these increasingly powerful processors must be removed more efficiently than ever.
Rising power densities are driving the transition from air cooling to liquid cooling. However, removing heat from the chip is only part of the challenge. Ultimately, heat must still be rejected into the environment. ‘Today, we consider the air like an infinite sink,’ Grabon explains. ‘But when you have one gigawatt of heating in a space, the atmosphere around you becomes hotter and hotter.’
This is pushing the industry to explore alternative approaches, such as district heating or water-based heat-rejection systems. The technologies are not new, Grabon notes, but adoption has been slow.
Regulation may accelerate that shift, particularly in Europe. Germany, for example, already requires that 20% of data centre waste heat be reused. ‘Sometimes, regulation is the only way to get industry to do the right thing,’ says Grabon.
As data centres scale towards gigawatt capacities, their impact on the global energy grid is becoming impossible to ignore. Grabon believes this shift requires dense, decarbonised power sources, stating: ‘The only decarbonised energy that is available, dense and cheap is nuclear; it looks to me that this is the direction people will go.’
Manufacturers are increasingly developing integrated control strategies that coordinate chillers, pumps and fans across the whole cooling system. Rather than treating these elements as discrete components, such approaches seek to optimise performance at a system level, using supervisory controls to respond to variations in IT load and ambient conditions. Grabon highlights the inefficiency of the traditional, fragmented approach: ‘
‘You can consider those elements independently and achieve a certain level of effectiveness, but if you consider all those systems not as silos, but as one body working together with smart controls and optimisation layers, you can really reduce energy consumption with the same output.’
Carrier’s QuantumLeap is one example of this type of system-level approach, bringing together plant and controls within a unified framework. Reported benefits include reduction in overall energy consumption, although realised performance will depend on system configuration, control strategy and operating conditions.