Diversify! Historically, the average length of a recession has been 12-24 months. So set up a system whereby you won’t screw’s yourself over by selling when things are low, but instead you can weather the storm. Build a rainy day fund. Determine how much cash you will need if you are out of a job and how long you think that will last, allocate some portion of that amount into low risk bonds. Russ way if you need cash…
But diversify into what? If we assume this takes down the US economy and bonds, what then? International bonds/stocks? Won't those also be too entangled? Precious metals?
Those gold guys have been decrying the collapse the US economy for 25 years now, so you'll be in good company.
The current commitment by hyperscalers is around 1.7T USD, reported liabilities 1.3T and this year global debt related to AI is 570B. So that’s around 3T total. For this to make sense AI must generate 2T in new revenue per year by the end of the decade. And that would be only a 10% ROIC. For context ROIC for big tech is around 35% so at 10% they will be barely breaking even. The SP500 gives 10-12%. With 10% ROIC from…
> GPUs become obsolete in 5 years.
Not only that, but they're typically amortized over 5 years, where the actual lifespan usually falls far shorter (1-3 years), adding to the artificial subsidy conditions we see today. So they're gaming the lenders into deferring interest payments as much as possible today so that new competitors don't have the same cheap financing advantage.[0]
That would only happen if they need to invest like this forever, otherwise it's just a short-term dent in their margins while they re-calibrate.
Why? GPUs are replaced every 3 to 5 years. This is going to be an ongoing operational cost forever. It will probably increase more if larger models require bigger VRAM sizes.
That cost has always been there and allowed for their lucrative margins. It's the upfront cost of building/populating their datacenters (many more than before) that is eating those margins.
> Everyone is in too deep to now admit that there’s a problem I'm not sure how to square this with the dramatic improvement in LLM capabilities in the last 8-9 months. If anything, it makes the earlier investments look prescient?
The problem is that the dramatic improvement in capabilities is not translating to a dramatic increase in revenue.
Source? Has Anthropic's annualized revenue not quadrupled in the last 7 months? And OpenAI's annualized revenue quadrupled since January 2025? Which is only unimpressive by comparison to Anthropic's meteoric revenue growth
I'd be with you if you claimed that the revenue hasn't translated into substantial profits. Being able to spend a lot of money to get less money back is not that impressive. But revenue by itself is on a dramatic rise as capabilities improve
> not translating to a dramatic increase in revenue. Completely false. AI and AI related revenues are growing exponentially .
I know it's easy to forget, and InsideOutSanta kind of anchored the conversation on "revenue", but profit is ultimately what matters. Back when Silicon Valley was merely insane rather than bat-guano crazy insane, it was commonly observed that it's not hard to build a business around selling a dollar for 95 cents. The point being that it doesn't necessarily mean much when you have a business doing that, because of cou…
I very much agree with this. Even the top tier models today, without the unit tests, without integration tests, and domain experts reviewing the code would flounder for 50% of the work they do. Sure they can write the unit tests and integration tests themselves, but at that point you aren't in need of a specific system being built, but rather an out of the box solution would probably fit your needs. It does speed up the grunt boilerplate work of development quite a bit, it does help with gnarly bugs and the like, but expertise is still needed. And we as engineers/programmers have systems in place that make using AI easier, we have the human context windows to be able to parse the technical jargon the AI spits out. Will AI for the masses be akin to slightly better automation?
If the margins aren't as high then there will be a repricing for all the massive cloud companies, which means several trillions worth of valuations to be cut from the companies. AWS/Azure/GCP/Oracle/SpaceX/etc neoclouds... are worth a combined 10+Trillion. That going down by 50-70% is going to be insane.
That would only happen if they need to invest like this forever, otherwise it's just a short-term dent in their margins while they re-calibrate.
This is a good chart that shows historical CAPEX spending. Hyperscalers have been through a couple CAPEX cycles like this, they all know what they are doing.
The current commitment by hyperscalers is around 1.7T USD, reported liabilities 1.3T and this year global debt related to AI is 570B. So that’s around 3T total. For this to make sense AI must generate 2T in new revenue per year by the end of the decade. And that would be only a 10% ROIC. For context ROIC for big tech is around 35% so at 10% they will be barely breaking even. The SP500 gives 10-12%. With 10% ROIC from…
That sounds reasonable, it's "just" $1k/yr for 2B workers (there are about 1.2B total "knowledge workers" in the world including gig drivers), or $10k/yr for 200M workers (there are 70M office and technical workers in the US). /s
In 4 years it better be 10x more important to have than a cell phone is today, or 10x more important than having internet/monitor/pc/printer is for an office worker today.
These alarms have been going off for a long time now. Everyone is already in too deep to admit that there’s a problem.
Cannot hear what you’re saying with all those alarms blaring non stop since a year. Someone should do something about them, maybe turn them off, I don’t know
Sink rate! Sink rate! Pull up! Pull up! Too low; terrain. Too low; terrain. Wind shear! Stall! Stall!
That would only happen if they need to invest like this forever, otherwise it's just a short-term dent in their margins while they re-calibrate.
Why? GPUs are replaced every 3 to 5 years. This is going to be an ongoing operational cost forever. It will probably increase more if larger models require bigger VRAM sizes.
We have probably hit a limit to scaling LLMs through raw parameter count alone, at least we're not seeing the exponential pace. I personally think we'll end up with a nice sigmoid curve plateauing in the sub 10T parameter regime. The amount of tokens processed (in inference) is increasing exponentially though (I've been following open router usage stats for years and it's always been exponential). We will of course make technological advances in hardware efficiency, and model parameter efficiency, but I think a much more plausible future is that VRAM needed for loading and serving individual models will slow down or even stop. We will need more chips, and more power, as demand continues to grow of course, but the operational lifetime of GPUs today will be a lot longer than the SoTA cards from 5 years ago.