@robb Their advice and a lot of replies are missing that data on disk uses very, very little power. The water is to cool the DC because of electricity use and electricity use is ~all for CPU and TPU work. There's some cost to spinning platters but historical, cold data is a tiny fraction of recent data. There's some cost if AI is trained on all data including old but it's a tiny fraction of training cost and giant model training is surprisingly rare and much smaller than inference (*).

@robb (*) depends a bit on exact model. Datacenters are heavily optimised to reduce electricity (the main opex cost) and so cooling and so water costs. It's because there's so many that there's a problem. It's nearly always more efficient in electricity to do things in a DC, not at home. Of course cooling at home is usually passive and there are no water-eating cooling towers. And you hold your own data. But "delete old pictures" is the "turn off your TV at the socket". Every big counts.