AS ARTIFICIAL intelligence’s (AI) need for computing power increases with more use, data centres will need to scale up while processing moves towards users’ devices.
Based on a report by consultancy Bain, the cost of building large data centres could surge five years from now – from between US$1 billion and US$4 billion to between US$10 billion and US$25 billion. Power requirements will also rise in tandem with costs, with forecasts for data centres to hit 1 gigawatt (GW) or higher in the future.
The increase in power consumption will likely stress the electrical grids that have seen demand grow from electric vehicles and reshoring of manufacturing. With growth in electricity demand flat for the last 15 to 20 years, more investment will be needed to expand and strengthen the power grid. This includes adding new power sources, such as renewable power.
The large data centres of 1 GW or more will likely push innovation to build efficient hardware, cooling and software to support AI’s growth. The construction efforts will strain the labour pool, requiring 6,000 to 7,000 general and specialised workers at peak levels.
With a labour shortage in electrical and cooling sectors, many projects occurring at once will stress the entire supply chain.
Companies are also weighing the trade-offs in moving inference closer to the user, such as their own smartphones or computers rather than in the cloud. This will allow for lower latency, but requires new AI models and specialised equipment.
BT in your inbox
Start and end each day with the latest news stories and analyses delivered straight to your inbox.
A lot less chips
The accelerating adoption of AI will likely bring about a supply crunch for chips, specifically graphics processing units (GPU). Data centres will need to compete for supply with hardware makers that are proliferating AI-enabled devices, such as smartphones and computers.
Bain estimates that a demand increase of 20 per cent or more has a high likelihood of causing a chip shortage. Demand from large markets could easily pass the threshold, causing chokepoints in the chip supply chain.
Should data centre demand for GPUs double by 2026, suppliers of key components need to increase output by 30 per cent or more based on Bain’s forecasts. There will be a concentration of demand in advanced packaging and memory, with chipmakers needing to almost triple production capacity by 2026.
Similarly, demand from AI-enabled devices will strain advanced wafer fabs. Bain is forecasting that wafer fab output will need to be increased by 25 to 35 per cent, just to accommodate 31 per cent PC sales growth and 15 per cent smartphone growth between 2023 and 2026.
This requires building four or five bleeding-edge wafer fabs at an estimated US$40 billion to US$75 billion, justifying the fabs that chip foundries are already building.
This is all not accounting for supply chain risks such as extreme weather, natural disasters and geopolitical strife. Most of the pressure on GPU supply in the last 18 months was caused by disruptions to parts of the supply chain, such as advanced packaging capabilities.
Continued geopolitical tensions, trade restrictions and decoupling from China supply chains will pose risks to semiconductor supply. Bain expects larger supply chain risks to come from demand for high-bandwidth memory components and wafer fabs.
Companies will have to navigate the intricacies of the semiconductor supply chain, beginning with where components are being sourced. Computer and smartphone refresh cycles will need to be monitored, with a surge in demand likely to cascade across the supply chain, according to Bain.
The “just-in-time” inventory approach will likely give way to a “just-in-case” approach, with a move away from application-specific chips to industry-standard chips.
