[Schiendelman]

It's hard to read this because it's AI written, and at least three times as long as it needs to be.

TL;DR: A lot of credit crises happen because of a decrease in acceleration (the second derivative) rather than a decrease in point in time slope (first derivative), or absolute value. The author (or at least the person who prompted AI for the article) says they think AI capex has to continue to accelerate in order for the frontier model companies (and associated) to continue to pay their debts.

I think the theory is sound, but I'm also bullish on LLM/LBM market size being very undervalued today.

[jdlshore]

Agreed. The article is very sloppy, but that doesn’t mean the central thesis is wrong. I’m not conversant enough with financial theory to say one way or another. Anybody care to critique this?

[Schiendelman]

I don't agree it's going to break these companies; they are all capable of paying their debt service. The reason this was an issue in mortgages was that the mortgages were adjustable rate - the borrowers were defaulting when the rates adjusted because they weren't able to pay the new rates. I'm not aware of any financial instruments for these AI companies that would balloon like this.

[combilabs]

The article addresses the mechanism in the "How It Breaks" section. The problem lies in the interconnected contracts, not debt financing. If OpenAI cannot get new financing, it will have to cut costs, and one of its big costs are forward compute commitments it will have to break. Those commitments are a significant share of revenue for other companies like CoreWeave and Oracle, so if OpenAI has to cut significantly, it will be crippling to those companies. They in turn have commitments to suppliers like Nvidia, and on down the line the losses propagate, with corresponding drops in equity values in the sector which are extremely growth sensitive. Just as the same dollar showed up in the correlated revenue on the way up, so will the same dollar show up in the correlated losses on the way down. That's the balloon effect here.

[Schiendelman]

I'm extremely unconvinced. The article didn't at all address what OpenAI would really do if they couldn't get new financing. They'd sell the contracts for some of that compute to the fifty other companies desperate for it. With supply constrained, they'd likely make money on those sales, and there'd be little or zero downstream impact.

[combilabs]

>They'd sell the contracts for some of that compute to the fifty other companies desperate for it.

And if OpenAI were an isolated case, that might be all there is to it. But the whole point of the article is that it isn't isolated. So if OpenAI can't raise the money to execute on those contracts, it's unlikely that another AI company will have the spare cash to buy them at anything like full value. Remember, everyone in that market is riding the same second derivative. OpenAI is just a bellwether.

[aka-rider]

What will be the price? GPUs costs are x3, memory and SSD x5.

When OpenAI will be looking to sell and everyone knows it, the price would be x0,7 for depreciated used hardware or go bankrupt yourself.

[Schiendelman]

The scenario we're discussing is "if OpenAI can't find sufficient additional financing."

It looks like you want to discuss "if prices and demand drop", which would be a different scenario, which I do not believe is likely in the same timeframe barring some separate economic meltdown.

In today's world, the selling price would likely be an increase over what OpenAI paid. This isn't selling hardware, it's selling contracts for future hardware. It's basically futures trading.

[nothercastle]

Why do you think it’s ai written. I do not get that sense except that it’s overly wrong but perhaps you found some key ai areas?

[nojvek]

It has tell-tale signs. Not X, Y. Two X, One Y.

The way it does it headings and each paragraph neatly chunked. Fancy words, lengthy explanations that don’t add any meaning.

Someone wanted to write a long lesser and feel good.

A shorter human written one would be better.

[d4ng]

In light of the article, why are you bullish?

[nothercastle]

If you were bullish on housing in 08 you would have been right long term but the timeline would have been way off.

[d4ng]

It would have been a while until you could sell at profit and move, though.

[Schiendelman]

See my reply to the other person who replied to me :)

[ilovecake1984]

I think in the long term AI will be a massive market. I just don’t think it will be massively profitable. The hardware will get better, the models will be even more commoditised. It’s useful and it will get even more useful.. it’s just not going be a licence to print money.

[Schiendelman]

In the last year we've gone from "this is annoying but sometimes useful" to "I don't need to hire junior engineers". There are some 25 million software engineers worldwide today, and do not ignore Jevon's Paradox - there are far more people who would hire software engineers if the price of engineering was slashed 90% (which is likely an understatement).

If you don't think this will be profitable, you're saying you don't think software, the most profitable enterprise in human history, will be profitable. Looking at it that way, does it make sense why those of us on the business side of software disagree? I've produced more software (and quite good software) in the last six months than I did in the first ten years of my engineering career.

[ilovecake1984]

I’m not saying software will be unprofitable. I’m saying Ai models will be low margin businesses. I think AI is going to decrease the price of software and allow us to automate things we couldn’t. I just don’t see how this value accrues to AI companies.

[d4ng]

How much value in present terms was created by n hours of your time before LLMs, and how much value is created now by n hours of your time? If everyone is using LLMs, has that value calculation result changed?

People used to go to work by foot. The best runners would get to work first. Then the bicycle came along. The person who was the best runner before is now still getting to work first on the bicycle.

[Schiendelman]

Yes, that value calculation result has changed. The gap between building software and financial value is knowing which software to build and how to break into a market, not how much code you can write. There are not very many good product managers. Not that I'm necessarily good! But I try to be. :)

[d4ng]

Please see the bicycle analogy. Resources available at a given time are finite. If everyone is now churning out software at a higher rate with LLMs, what changes in terms of the value gained?

[d4ng]

I think I agree. The question is what happens before then. The S&P 500 went down about 40% following 2000 and 2008. The FTSE 100 showed similar losses.

My questions are: Will we still have jobs if there’s a crash? How can we start researching what the optimal hedge is against such a crash?