Power, cooling and connectivity requirements for AI result in significant bottlenecks emphasised by scarcity of microprocessors. The outcome is a new playing field for information and communication technology (ICT) infrastructure.

With skyrocketing power and water consumption, current AI-powered data centres are putting huge pressure on grids and on civil infrastructure, with a direct impact on the way AI data centres are built. Resulting operating costs become deterrents for many players.

In this initial phase, small data centres have been favoured (from 1MWh to 4MWh). However, plans shared by Big Tech forecast gigawatt hour (GWh) or more data centres to be operated soon. This exponential growth in terms of capacity requirements results from computer power hungry large language model (LLM) training and fine tuning, but also from inference workloads, each category requesting different needs in terms of connectivity, latency and more generally service-level agreements (SLAs).

Impact on the value chain will be significant. All players will have to revisit their business model, in a race for capacity and efficiency, with a growing number of partnership pre-empting access to resources.