A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret.
Archived copies of the article:
A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret.
Archived copies of the article:
Physics puts some useful limits on things that can be applied here. All of the water used for cooling is considered wastewater, and it is generally treated chemically in a way that is difficult to process back to clean water. Water has a specific thermal mass, and we can see what hardware is being used in these data centers. From personal experience a flow rate of 1.5L/min is standard. Each rack is limited in power not by how much the supply can put out, but by what they can cool.
Even doing napkin math with the Blackwell systems that have been out for a couple of years it is as bad as they say and likely worse with the coolants and passivators going into the water.
I have advised a few water cooled systems, and did some work on Cheyenne in college. Not once was water reclamation even mentioned, it was all pumped right to sewage.
Wait so it isn’t a closed system that cycles? That’s the “thing” I’m always told.
In the context of these large data centers no, it usually isn’t recycled water. It gets filtered coming in and run through, then right down the drain. Closed loops do exist in the data center world, but they don’t use water, they use a dielectric coolant, and are orders of magnitude more expensive to set up and maintain. You don’t usually see systems like that in use at scale like this, the cooling towers would be immense.
If they reuse the water they also have to remove the heat. Down the drain it becomes someone else’s problem.
Damn. Thanks for sharing!
Far out that’s depressing. Thanks for sharing your insight!