Remix.run Logo
▲ TheAlchemist a day ago

IF the bubble is finally popping, just be careful guys - because no matter how cheap things look, they can always get a lot cheaper !

If what many people think is right, ie, that the whole business of AI is unprofitable and pretty much only exist for now because Nvidia & co is directly or indirectly financing their own demand, this whole house of card will collapse really quickly. As a reminder - before the 'boom' Nvidia was making a 3B profit, in a good year.

▲oefrha a day ago | parent | next [-]

The whole business of AI can’t be unprofitable, open models like DeepSeek are clearly being served profitably and if it turns out GPU, HBM, etc. should be cheaper than they currently are, serving them will only become more profitable. Given that, the frontier closed models with a pretty big price premium must be very profitable to serve as well, even if you don’t take their word for it. Training can stop today and the existing models (open ones at least) are still very useful and aren’t going anywhere. A valuation pop may drag down some companies, but the .net pop didn’t mean the whole business of internet is unprofitable.

▲zamalek 21 hours ago | parent | next [-]

> Training can stop today and the existing models (open ones at least) are still very useful and aren’t going anywhere.

Unless your competitor keeps training, then you'll be starved for customers. The idea that R&D can stop is absurd.

▲oefrha 21 hours ago | parent [-]

There are already plenty of inference-only providers in the business that are not training at all.

▲bayarearefugee 21 hours ago | parent | prev [-]

> The whole business of AI can’t be unprofitable

Offering compute is likely to remain profitable.

But unless someone suddenly pulls a moat of out of their butt, the layer of AI on top of that compute could very possibly end up being a minimally profitable business, like say VPS hosting.

Sure, you can make a few bucks doing that, but not what you want to try to build companies with trillion dollar valuations off of.

▲oefrha 16 hours ago | parent [-]

I believe inference as an industry will be as big or bigger than pre-AI AWS, just with current model capabilities, bubble popping or not. In fact if you add up this year’s AI revenue from all the big names, that should already be close to or surpass AWS’s 2024 revenue of $108B? Will it support multiple trillion dollar companies? No idea. But it’s not a few bucks either.

▲aurareturn 19 hours ago | parent | prev | next [-]

When Nvidia was making 3B in profit, we thought we might never solve Riemann Hypothesis and all of us were writing code word by word like cavemen.

▲topspin a day ago | parent | prev [-]

Rather nice not having to wade through all the bubble denier mental gymnastics today.