[scellus]

He writes as if only datacenters and network equipment remain after the AI bubble bursts. Like there won't be any AI models anymore, nothing left after the big training runs and trillion-dollar R&D, and no inference served.

[rjh29]

Who's going to pay to run those models? They are currently running at a huge loss.

[WalterSear]

Anthropic said their inference is cash positive. I would be very surprised if this isn't the norm.

[harvey9]

I can run quite useful models on my PC. Might not change the world but I got a usable transcript of an old foreign language TV show and then machine translated to English. It is not as good as professional subtitles but i wasn't willing to pay the cost of that option.

[mike_hearn]

There's a gazillion use cases for these things in business that aren't even beginning to be tapped yet. Demand for tokens should be practically unlimited for many years to come. Some of those ideas won't be financially viable but a lot will.

Consider how much software is out there that can now be translated into every (human) language continuously, opening up new customers and markets that were previously being ignored due to the logistical complexity and cost of hiring human translation teams. Inferencing that stuff is a no brainer but there's a lot of workflow and integration needed first which takes time.

[quesera]

Running the models is cheap. That will be worthwhile even if the bubble pops hard. Not for all of the silly stuff we do today, but for some of it.

Creating new LLMs might be out of reach for all but very well-capitalized organizations with clear intentions, and governments.

There might be a viable market for SLMs though. Why does my model need to know about the Boer wars to generate usable code?

[logicchains]

They're not running at a loss. Training runs at a loss, but the models are profitable to serve if you don't need to continuously train new models.

[dcre]

They are obviously running free users at a loss. Can you point to evidence of negative margins on subscriptions and enterprise contracts?

[qgin]

The models get more efficient every year and consumer chips get more capable every year. A GPT-5 level model will be on every phone running locally in 5 years.

[antonvs]

I run models for coding on my own machines. They’re a trivial expense compared to what I earn from the work I do.

The “at a loss” scenario comes from (1) training costs and (2) companies selling tokens below market to get market share. Neither of those imply that people won’t run models in future. Training new frontier-class models could potentially become an issue, but even that seems unlikely given what these models are capable of.

[myhf]

Is there a genuine use case for today's models, other than for identifying suckers? You can't even systematically apply an LLM to a list of text transformation tasks, because the ability to produce consistent results would make them less effective sycophants.

[Juliate]

The point is, after the bubble burst, will there be enough funds, cash flow and... a viable market, to make these still run?

[muldvarp]

Inference is not that expensive. I'd argue that most models are already useful enough that people will pay to run them.