Google bought DeepMind, a company specialised in Artificial Intelligence, in 2014. It is by using DeepMind that Google managed to win the Go competition against the world champion earlier in 2016.
Now Bloomberg is reporting that by leveraging on AI provided by DeepMind Google is able to cut its power consumption by 10%. It might not seem a lot, but when you consider that Google used 4,402,836 MWh in 2014 (the whole Telecom Italia network, by comparison, uses around half of that power) and that a single MWh may cost between 25$ and 40$ you start to appreciate that the figures are in the tens of millions $ in saving.
DeepMind works by analysing some 120 parameters in each data centre, including, of course, fans, cooling systems, windows. I assume there are plenty of sensors measuring temperature in racks and the trick is to find a way to maximise natural cooling and to let the electronics operate at a temperature that is not critical. At the same time, the differences in temperature in several points of the data centre can create air flow helping in the cooling.
The whole system is, as one can imagine, very complex and DeepMind has the capability to experiment and learn from it. This is what is really making a difference in today's AI system. They are not just "smart", they get "smarter" over time.
Interestingly, DeepMind seems to have the capability of understanding that it could do better, if there were just a few more data to analyse. And it prompts the engineers to deploy more sensors so that it can get those data. This is what actually surprised me most. AI is not just getting smarter, it is acquiring a sort of self awareness that it can further improve.