Hacker News Reader

Comment by vb-8448 | Hacker News Reader

The projected ai capex for this and next years is above 1000b/year. 673b/year doesn't sound so weird, if they manage to spent so much, obviously.

View on HN

Capex covers a lot of non-gpu stuff, but yeah it doesn't sound crazy to me.

I couldn’t find the 1000b number, but if that is “ai buildout capex” that means you think that 67% of every dollar spent on building data centers, training models, and running inference, is going directly to nvidia.

Which is just silly on the face of it. Data centers need concrete, copper, DRAM, SSDs, and labour. That alone will cost more than 33%.

> Data centers need concrete, copper, DRAM, SSDs, and labour. That alone will cost more than 33%.

Surely the land is a double digit percentage of their budget? I know they build in the middle of nowhere, but even then.

Yeah but remember 1) stuff gets more expensive the more it's "processed"/further up the chain. It's hard to think of something further up the value chain than a modern gpu server. Just look at one input, asml machines. 2) a lot of dram, labor etc is also baked into the gpus.

Sure. Obviously GPUs are expensive. My computer is more expensive and more processed than a 2x4, but that doesn't mean that my computer is a significant percentage of the value of my house.

And it's pretty unclear what is included in the "capital expenditure" numbers. E.g. does it include training costs? does it include research costs? etc.

I'm not sure. I think they count training as R&D? Hopefully we learn just they're doing their accounting with anthropic ipo

Maybe 67% is too much, but the GPUs are the most expensive thing in all those datacentres.

Anyway, according to the latest articles[0], 1000B is the lower bound.

[0] https://sg.finance.yahoo.com/news/ai-infrastructure-investme...

The most expensive thing in my house isn't even 1% of the value of my house and everything in it.

Imagine a house full of racks of gpus!

I can imagine that. I could probably fit more GPUs in my closet than the cost of my house.

Now imagine adding a complex liquid cooling system to my house, upgrading the wiring to support the extra power draw, and some full time staff to maintain the space and keep the GPUs running. Wouldn't be hard to switch the balance back away from GPUs toward the infrastructure around them.

And let's also imagine I include my massive GPU usage bill in the house column. And while we're at it let's also include my full time staff of researchers in the house column.

No way staff, complex liquid cooling and everything else is going to cost more than the hardware itself.

All articles that I found on "ai datacenters cost breakdown" say that >60% is for the HW inside.

Well then we're left with the assumption that 100% of the new data center buildout will be done with Nvidia GPUs, with absolutely ZERO being AMD GPUs or google TPUs.

I suggest looking into Nvidia's DGX and DSX systems. Basically, everything of a data center except for RAM+SSD comes now from Nvidia as a vendor.

Nvidia's margin is so high because they deliver way more than just a GPU.

Reply on news.ycombinator.com