Hacker News Reader

Comment by harpiaharpyja | Hacker News Reader

Do these companies that are buying up all the supply actually need all that RAM right now? Or are they buying it all up in anticipation of future need? If the latter, honestly this might be a case where some kind of regulation really ought to step in.

View on HN

To the extent that most of this is going into AI and people are having their ChatGPT and Gemini requests throttled because of lack of capacity, they need it now.

AI is dramatically more compute- and memory-hungry than past computing models, so if that's what people are using, it's going to require a large build-out of computing capacity to support the requests that are being made right now.

Some of the spike is speculation, and the overshoot seems to be correcting itself now. But the deal that sparked it was a contract promising to buy future capacity, not just doing a big block order for a bunch of stock 'in case' (which isn't unusual: if you're a big buyer, you will almost certainly buy most things this way).

It seems like it wasn’t really a binding contract? At least the OpenAI one, some are saying it was more a letter of intent kind of thing?

Sounds like market manipulation to me... But I suppose that is not punished...

That's because you're not paying attention to the wider world. Every inference provider has been maxed out for years now. Nvidia switched from mainly selling training servers to inference servers two years ago?

People complain about the llm's hallucination, but humans making up conspiracies to shore up their own ignorance is a much worse problem

They're not directly buying up RAM sticks per se, but rather placing orders for say a GPU resulting in wafer capacity getting redirected

And since enterprise GPUs have enormous leads times right now yeah it is in anticipation of future needs

Depends what you mean be 'need', but it is mostly going into powered on systems not being stockpiled.

Reply on news.ycombinator.com