Thanks, I took a look, couple things - the inlet temp on the VR is 45 C, but that’s not the radiator operating temp, you can probably run those chips closer to 90-100 C. And they’re building custom silicon for this, presumably that’ll be one of their design targets. Also, most bit flips should be fine when you’re running inference, you’re presumably running with some randomness anyway. If a node fails/becomes too unreliable, it can be detected and shut off.

Idk, building in the ocean seems a lot harder to me than space. Salt water is ridiculously corrosive, extreme pressure, etc. And one of the main justifications for this is massively increasing the output of solar and making it consistently output its nameplate capacity, which space is great for, and ocean is terrible for. The only benefit for that one I can see is some power savings on cooling, and a whole boatload of drawbacks, whereas we might not be able to keep up with demand with terrestrial power. So I can totally see why they never bothered to complete their subsea datacenter.

They’re definitely not aiming to put niche applications up, they want to run models by the bucketful.

I don’t see how the economics make it impossible? To be clear, I’m not saying that it’s something that’s going to end up happening, I have no idea if it will, but I don’t see how it’s structurally impossible, and I can see some things to commend it if token usage volumes grow like I think they will.