Why are tech giants rapidly scaling their infrastructure?

The advent of artificial intelligence has led to many locations, one of which, maybe surprisingly to some, has been in construction. Amazon Web Services (AWS) has recently launched its Project Houdini, which is a novel method of construction to accelerate the process of building AI data centres. Instead of the traditional stick and built form of construction, which takes time, AWS is creating these data centres in skids that already contain racks, power distribution, and lighting, among other things. Each skid is built in a factory in a controlled setting. Why would AWS do this? If the wiring and power integration happen early on, Amazon can start installing the servers in 2-3 weeks as opposed to the usual 15 weeks. If you take a step back, this is happening because the industry is literally trying to catch up to the speed at which AI technology is moving. If data servers are built quickly, the technological work can begin early. 

AWS is not the only company doing this. Google, for instance, is building a large AI hub in Vishakhapatnam that would house a data centre with a gigawatt-scale compute capacity. Meta has built its latest data centre platform, Fairwater, which is built around GPU-dense clusters, like the NVIDIA GB200 and GB300, clubbed into a singularly integrated fabric. The motive of Fairwater is to provide an output for AI far higher than its traditional AI counterparts. In total, the amount of money put by companies for such growth is supposed to be over $600 billion.

Basically, no one could have predicted the amount of energy AI can consume, and this has put all giant tech companies in a race to train their respective proprietary AI models. But the race is not just to build the AI data centre, it is also to get the required chips and the cooling systems. Inevitably, this means that AI’s excessive power consumption is weighing heavily on the environment, largely through electricity use, water use, and pollution. Some next-gen AI campuses are projected to use city-wide electricity levels, which is putting a lot of load on the supply grids. 

However, things are not as black and white. If the data centres are carefully built in locations where there is an abundance of water and clean energy. With smarter siting, faster grid decarbonization, and operational efficiency, we can cut carbon impact by 73% and water impact by 86%, according to studies. There is an approach possible where AI can grow without completely eradicating our environment, and walking that line, through the less-is-more approach, is the healthiest bet.