I'm saying that environmental externalities are hugely overblown[0] and to a large extent are part of the cost model. Datacenters are not like the legacy industries, they have to pay for the land and energy they use.
> Your point about efficiency is particularly obtuse: efficiency doesn't mean burning less fuel, and getting more compute per Joule. Not reducing the overall carbon footprint, but accelerating the burn rate as fast as we can build.
Efficiency does mean getting more compute per Joule. They are doing that. They are also building out as fast as they can, because both of those address the same problem: demand for compute outstripping supply for it. They can't just pursue the build-out strategy, because hardware production is now supply-constrained too - that's where the "RAMpocalypse" came from.
Yes, this is a huge build-out, and little to none of that is 100% carbon-neutral, so ecological footprint adds up. But that's normal and expected and a kind of tradeoff humanity has been making forever: building new stuff, be it hospitals or airports or data centers, has an environmental footprint, and we only hope what we get from it is more important to us on the margin, and that we can offset the environmental costs some other way.
> I'm not even sure I agree that demand is outstripping compute -- nvidia's circular demand-inflating investment/buildout/loan situation certainly muddies the waters on that, see my point 3.
I'm not talking about financial "demand". I'm talking about real demand, for compute. There's nothing to doubt here, it's pretty clear that all major AI players are constantly running at capacity - it's obvious in the very structure and limits they place on even the highest tiers. Unless you believe half the data centers are just spinning busy-loops and burning energy to inflate stock prices and generate a fake reason to build out more compute capacity - if yes, then I don't know what to tell you.
--
[0] - Like the water story, where the reported datacenter usage numbers look big in isolation, but once you relate them to "how much water is there" and "how much is used by other industries", it turns out to be a nothingburger. The only numbers that survive are those showing the story is really about those other industries having a competitor for cheap water now, and having to pay a bit more than they're used to.
> reported datacenter usage numbers look big in isolation, but once you relate them to "how much water is there" and "how much is used by other industries", it turns out to be a nothingburger.
Water scarcity is a problem around the world; even my rainy hometown of Vancouver had steep water restrictions this year. The question "how much water is there" is incredibly disingenuous, to the point of feeling like a motte and bailey: yes, the universe, or the earth contains vast oceans of the stuff but datacenters are using drinking water which is quite scarce. Just because "other industries" use a lot of water doesn't make it okay -- I level exactly the same criticism at other industries which waste water at such an egregious degree.
To be clear: if data centers only charged up a pool of water to use as coolant, that's fine. But when the weather is warm, they run millions of gallons of tapwater down the drain just to cool off. And heat spikes are happening with increasing frequency, severity and duration.
We as a species need to conserve water and emit less CO₂, full stop. The tech industry doesn't get a pass just because other industries are dirty.