Theres “big bets” and then there’s heavily leveraged bet the farm with leveraged debt. The first one hurts when it pops. The second one wipes people out.
In 2008 the USA managed to export the shit assets generated by the mortgage boom , mostly to Europe. The consequence for the European economy and political setting have been quite sharp. For example it's hard to imagine a Brexit vote if the economic conditions of 2006 had persisted in the UK.
If the AI boom unwinds it's quite hard to see how the USA will export the pain this time. The issue is that the boom is tied to physical infrastructure in the USA and these assets will be stranded. Corporate bonds used to finance it are mostly owned by US pensions and 401k. We are likely to see many US insurance companies liquidated... and that will kill the annuities for many retirees.
The grid expansions that are being created in the USA will mean significant debts on US utility providers as well, if the data centres profitably come on line then this will get paid back. If not then American households are going to have to pay. Chip producers (Tiawan SK) will also get burned.
The USA will forfeit tech leadership, the US bond market will not permit Keynesian expansion. Basically the USA will be looking at a great depression. US hackernews friends need to find a cover story that stops their neighbours realising that they were part of the shit storm, and they also need to sharpen their camping skills and general outdoor hardiness. Alternatively -> try and get a French passport?
In many places if people charge their car and run the AC on a hot day the grid struggles. So the grid investment won’t go unused if AI implodes.
But that infra in the middle of nowhere doesn't necessarily benefit consumers charging their cars.
I didn't know about this. How did it work?
The end of democracy, that's what all of this is about and why we've seen so many of the super rich build private bunkers or buy entire private islands, talking about shock collars for their staff to keep them from rising up. They're preparing.
You shall own nothing and rent everything, and the easiest way to achieve that is to wipe out everyone's retirement savings.
> Chip producers (Tiawan SK) will also get burned.
They'll be fine. They already got paid, and if the AI bubble crashes, there's years of pent up consumer demand just waiting for affordable slots at the chip fabs.
> Basically the USA will be looking at a great depression.
The US is looking at a return to medieval times. Rich elites lording over the poor masses.
Alternatively, we'll see World War 3. It's notable enough that most countries have slowly risen the temperature on disaster preparedness, even those not in direct line of fire of Russia.
I mean, the premise is "who's going to get wiped[...]?", but the harder question is why would they? And that answer is not trivial because these companies are actually generating money and are competitive. The people and investors running these companies and investing are also not just standing still.
I think one can assume these companies are all a massive scam, but having used LLMs in my day to day every day from the very same companies this article talks about, I don't feel so. I only see the integration of intelligence becoming deeper, and even if cheaper per unit(token), many more tokens will be required. Productivity unlocked is further unlocking business and social opportunities we are only now exploring and getting to know. In the very essence it is very easy to see how this will turn out badly: after all a collapse passes all the known floors on the way down, but we have no idea where we are going up with this technology. I don't buy the comparison with the dotcom bubble because it is very easily falsified by simple metrics as the profitability of these technologies in the present. In the dotcom companies were not profitable, and many other counter examples disprove the comparison applicability. I will not even comment on the comparison with the subprime loans which were related to real estate, which has very little unknown upside.
Anyway, have some faith people this is hacker news, and with LLMs I never was able to do so much cool tinkering in my whole life. I have custom modifications on my gnome shell, I have privately developed file servers, I have automations that make use of my rooted TV, I have a prototype web page to help my folk's small business. I am able to support and learn from the LLM about topics that would be too expensive to get input from. For example with a calibrated microphone the LLM generated the custom equalized bands for my type of amplituner. It looked at a picture of my living room and suggested moving certain furniture like the curtains and measure it. With LLMs I was able to take pictures of nice rugs in the internet and visualize them in my living room before actually buying to great effect. So many possibilities... Focusing on the known unknowns is the easy part, let us have our imaginations run wild and let the whimsical hacker spirit return.
I think that the first issue you raise is "why would they" and the answer is (IMHO) because there are a bunch of people who are trying to make money anyway that they can and as far as they are concerned it's just bad luck and social darwinism if they get the cash and other people are left holding the bag.
You are right that the products that have emerged from the AI boom have real value, but it's not clear that anyone can capture the value from consumers. LLMs are more or less commodity items. Listen, for sure, for serious, if you gave me $20m I could build a tier 3 LLM in six months. Fella, I am not boasting, I am a small fry old AI researcher and there are 100 like me in every bar in my town (London, near enough) but I know I could do that. I also know that the tricks to move from tier 3 to tier 1 are increasingly available *FROM THE FRONTIER MODELS THEMSELVES* . The number 1 use case at OpenAI for AI is... managing the clusters. And managing the clusters has been the hard bit since about 2017. This is not the case of a rare product that everyone wants, it's basically the same as Florida condos - yes a view over the ocean is great, yes a nice condo is good to live in, but also there is a lot of beach front in Florida and building condos is not rocket science (I am bearish on SpaceX too but for other reasons).
No one will give me $20M sadly, and if they did I'd probably abscond to somewhere I hoped was safe with it and spend the rest of my life raising litters of kittens. This is definitely a good hobby and is actually quite like managing big clusters, but with a better emotional balance.
Anyway, the next point is profitability.
These companies are not profitable. OpenAI is not profitable, Anthropic is not profitable. In fact they are burning cash and losing money hand over hand. They are subsidising subscribers and enterprises to use their models because if they don't, and they force market costs, then they won't have market share. Look to openrouter's stats if you don't believe me : https://openrouter.ai/rankings
The path for Amazon to be profitable was to use investor cash to build a modern last mile distribution and supply chain across the US and then the globe. The AI gamble was to create a capability to serve AI that could effectively monopolise the market and that they could raise the drawbridge like Google did with search. But Google created a market out of you and me for advertisers, where as OAI and Anthropic are asking you and me to buy their products. Brother, I will, but only if they are cheapest and best, which means that this will be a competition, and in a competition I do not care if you go bust and the data centre has to be auctioned to the lowest bidder because if that happens I will then see a cheap service come into use that asset to replace the leavered up investors who sadly lost their pants. If you have to sell off an Amazon distribution center you've got a big problem, an Amazon distribution center only has real value because of the last mile fulfillment and the global supply chain that's attached to the store. A data center for LLMs is not the same, tokens are tokens. I can buy them from Slough or Texas, the Texas ones might cost be less because gas is cheap there, but we're talking $1.40 per million vs $1.00.
That's it.
A $3Trn commodity service.
Also, I could be wrong, but I see two additional problems with the AI business model:
* AI accelerationists believe that generative AI will lead to compounding software complexity that opens paths to incredible economic growth. But there's a good chance it will mostly just expand breadth-wise, i.e. flood the world with mediocre code that follows well-trod paths and ships a lot faster instead of inventing new abstractions or reifying concepts that actually allow for compounding complexity. I don't think consumers want to be flooded with more software. I think they actively don't want this.
* Unless software development is forever "finished" in its current state, or unless actual general intelligence is somehow invented, the training work can never stop. If the money runs out, current models will still be useful for a while, but eventually the industry will lap the training data and model output will degrade.
There is a paradox here, I think that there is a lot of denial about it.
Also, it’s mostly not R&D either. It’s infrastructure buildout for a thing that doesn’t have a demonstrated business case that actually pays for all the infrastructure.
There remains a scary small amount of truly new money coming into this circular flow between VCs, chip makers, startups, datacenters and such. At some point there needs to be a $hit ton of new cash coming in from the outside to actually pay for all this and that’s not materializing.
While there might be new stuff, these business's are currently the ones funding AI - if they all get wiped or substantially reduced, then there will be less ai revenue then there is current tech revenue - potentially at reduced margins.
It’s been sort of funny to watch the tech industry realize that the main thing being disrupted by AI is tech itself.
I think a lot of SaaS tools have a great future. But they will require less engineers to build.
1 - People build more and more complex implementations, and your "working mess" software DOES still need time to maintain it, is unstable, and they will still want something that "just works" so they hand it off to another person / group. (This would be AI effectively creating "more software + jobs"
2 - People build more and more complex implementations, and AI no longer creates a "working mess" as it improves, it creates "working software" that does not need time to maintain it, is stable, and they do not need to hand it off to another person / group.
I think it's more likely that scenario 1 is going to be the case. I've seen the things people are building around my company. They are enthused, creative, and useful - but need a ton of work, and are continuing on the path of "usable mess", and are now getting fatigue because they need to maintain the tooling they built...
I'm finding it almost hard to feel much angst about this cycle, not due to a lack of compassion, but because of the learned reality that if I spend all my energy on today's Big Finance misadventure, well it seems there'll be another tomorrow anyway...