AI’s water problem, and what NVIDIA actually announced
You commented “water,” so here it is. The article itself, the short version if you don’t want to read the whole thing, and what I honestly make of it.
Read the NVIDIA postThe short version
Data centers drink a lot of water
NVIDIA puts the old figure at roughly 2.6 million gallons per megawatt, per year. Their new design brings that to near zero. That gap is the whole story.
The trick is heat tolerance, not new plumbing
The coolant is allowed to run at up to 45°C, which is hotter than a hot tub sits at. Once your coolant can be that warm, you stop needing to chill it, and chilling is where most of the water and a lot of the electricity was going.
It’s a closed loop, so it gets filled once
The liquid is about 75% water and 25% propylene glycol. It runs through cold plates sitting directly on the chips, picks up the heat, gets cooled outside, and comes back around. NVIDIA says that in the right climate it can be cooled by outdoor dry coolers with no mechanical chillers at all.
This is the first generation where everything is liquid cooled
Older setups put liquid on the GPUs and CPUs and left everything else to fans and moving air. NVIDIA says its Rubin generation has no fans at all, and the servers now fit in two rack units instead of six.
The electricity side may matter more than the water
Cooling is about 40% of what a data center spends on electricity. NVIDIA puts the saving for a 50 megawatt site at over $4 million a year across energy and water together.
What it doesn’t solve
This is NVIDIA’s blog, about NVIDIA’s product, written by NVIDIA. Every number on this page is theirs, not mine, and I’d read it the same way you’d read any company talking about its own work.
Their own post is honest that it depends on geography. A site in the Scottish Highlands and a site in Phoenix are not the same problem, and warmer places will still need chillers some days of the year.
It’s also a design for new hardware. It isn’t something you go apply to the data centers that already exist, so none of this happens quickly.
Near zero is not zero. None of it touches the bigger question of how much electricity all of this uses in the first place.
So why do I still think it’s good news
The environmental cost is the objection to AI I’ve had the hardest time answering. I’m not going to pretend otherwise, and I’m not going to hand you a version of this where it isn’t a real problem.
What this article tells me is that the people building it are working on the part that worries me, rather than waiting for someone to make them. That’s worth something, even at the announcement stage.
You can think something is genuinely great and still want it to be better. That’s where I am with all of this, and it’s most of the reason this page exists.
You can think something is genuinely great and still want it to be better.
Guides go out to the newsletter before they go anywhere else. No hype, no daily emails, just the things that held up when I actually used them.