[YetAnotherNick]

Loss to who? Now all of a sudden, we are caring about investors and sovereign funds?

And I think we passed the threshold for crash down for AI, even if AI companies wont be that profitable. Nvidia/cloud providers will be profitable as long as there is demand for AI.

[dhosek]

Their loss, big deal. Let them suffer. The problem is that when they crash they bring a lot of other stuff down along with them. The people who lost money in the 2008 crash were not the ones who suffered the aftermath.

[YetAnotherNick]

Because in 2008 ordinary everyday people invested in overvalued things like house.

[dhosek]

It went beyond that. I didn’t own a house, but my employer, facing an economic downturn, laid off about half its employees.

[YetAnotherNick]

Because house is part of economy. Crypto went from $4T market cap to $2T and no wide impact was observed. Same will be true for $4T AI market cap.

[Mistletoe]

Almost every single person’s retirement has exposure to this unless they have some sort of Bitcoin/gold/small cap value type portfolio.

[rsoto2]

Uhh I think a lot of people and their families likely have investment exposure to nvidia/hyperscalers. if places like Amazon spent unrealistically on ai or their stock goes down massively that could mean major job losses too.

If AI companies aren't that profitable...then they're going to stop spending so much money on GPUs to train AI models. A gigantic amount of Nvidia's profits would go bust overnight.

[YetAnotherNick]

But inference is increasing dramatically. Google says they now do inference of 3.2 quadrillion tokens per month, 7x increase in a year.

Claude code and others are here to grow even if they don't do any further training.

[rsoto2]

.. so what?

The cost(and size) to train models is also increasing and is still 60% of the cards that Nvidia is selling. Losing 60% of your most profitable revenue stream I think would do bad for a company regardless of how much inference is increasing "dramatically"(all this means is the GPUs are dead sooner and the cost to do this massive inference increases too)