[mlyle]

Parts of it (e.g. water consumption per query) are overblown.

But the degree of data center buildout and resource use, if exponential growth just continues a little longer, is going to end up being a big number. AI datacenters are already stretching electrical power grids and increasing peaker power plant use.

Data centers right now are about 5% of electricity use in the US. AI could easily double that share.

[cm2012]

Yes, any major manufacturing ends up as a big number. It is still usually worth doing!

[mlyle]

It's worth doing when the -true cost of resources- is represented in their price.

If it's not, then more of the thing ends up being done than is socially optimal.

E.g. artificially cheap agricultural water -> lots of water-heavy crops being produced like alfalfa and exported for less than the cost of the water.

I think we might disagree about the degree to this is true, but I think most of us can agree that the true cost of energy is not completely included in its price.

[hrideshmg]

I was thinking about this the other day. Surely, a datacenter, even one optimized for machine learning workloads could switch gears and do other kinds of computations.

Even if the bubble were to pop, i feel like the worst that could happen is that we would have a bunch of inactive datacenters that could be switched on to meet demands of the growing internet. Kind of like how nuclear plants operate.

cmiiw to think along these lines though.

[mlyle]

Well, stuff tends not to get completely wasted, but:

- AI datacenters are gold-rush rush jobs with interesting things like their own gas turbine generators etc.

- It's not clear that serving the internet needs us to double the amount of datacenter footprint. If anything, a lot of workloads are getting more power and space efficient.

- Most expensive thing is that we're filling them full of GPUs and with RAM tied up to the GPUs. That's infrastructure that we've paid the resource costs for and it's difficult to repurpose to something else.

I do think AI is going to grow a lot, so I'm not sure how much of the buildout will need repurposing. But I do think doubling our datacenter footprint and doing it in environmentally yucky ways will probably have some lasting effects and consume a lot of resources.

[zozbot234]

These are more like HPC supercomputers than garden variety datacenters. That's why there's so much concern re: water use for the electricity being supplied. (That's easy to address in principle, of course: wind and solar power use up negligible amounts of water compared to other sources.)

[mlyle]

Most of the water concern is evaporative cooling of the datacenter itself. But IMO not too much of a concern. The energy use and the resource use to make the chips, etc, is bigger.