A while ago I was thinking, “Gee, AI is so complicated, how can I keep up with the landscape?”
After reading these articles go by so often, it feels like what I actually can’t keep up with is the bond market. To paraphrase Trotsky, you may not be interested in the bond market, but the bond market is interested in you. I want to be able to read the signals at the bottom of this article, and divine some kind of prediction that can guide me… I don’t know, to choose whether I should buy a house or change the investment strategy in my retirement fund or something. But I’m just seeing all these signals go by, waiting for the story to be written, which only happens when the dust settles.
I guess I’ll go back to not understanding AI, instead of not understanding the bond market.
Don’t forget there is a constant pressure for everyone to have an opinion on how this is going to end while hoping it ends tomorrow so they can be vindicated. The dotcom bust took a decade to grow and collapse. I think it is too early to make predictions with AI. I mean the sentiment here is either it will dry up the world and kill us all or transcend humanity, there’s no gray area. I don’t want to fall into the emotional sieve that seems to drive everything.
That's the financial stakes here. That's why it's all or nothing. You're spending on a level that is only justified by the bonafide machine god being ushered into existence, not productivity or coding tools (and on relatively short time horizon). So if this doesn't change the near term trajectory of humanity to a parabolic move upward there is going to be a lot of economic pain. It's not just the spending, it's that the expectations for the returns to justify them are in a relatively short period of time.
> if this doesn't change the near term trajectory of humanity to a parabolic move upward there is going to be a lot of economic pain
"trajectory of humanity to a parabolic move upward" is poorly defined here. Whether we are headed to a machine god ruled scenario or "just" incredibly powerful productivity tools, there will be a lot of economic pain for some (most) and a lot of economic gain for a few.
I've yet to a see an LLM/agent-based business plan in where scaling massively with an order fewer workers is not a central part of the value proposition.
I think these large numbers are casually thrown about, but the real meaning is mind boggling. 1 trillion dollars is the entire US defense budget - aircraft carriers, nuclear submarines, health care, salaries, stealth fighters ect. The hidden AI debt alone is more than that https://asia.nikkei.com/business/technology/five-us-tech-gia... just for five tech giants (not to mention all the other smaller players like neoclouds)
1T is big in the absolute sense, but that's simply the scale these big tech companies operate at. Go back to 2024 or 2025 and you'll see as a group they are making a net income of $400B+. The scale at which these companies do anything is just staggering.
What’s weird is how emotional people get on this. I told publicly (because I was asked, not out of an obligation to have an opinion), that the prices we pay for LLMs are likely to go up because that’s what happens when the ratio of operational assets to foreign capital drops due to the capital having been turned into heat rather than operational assets. The grief I got from people, dear Lord…
I think that opinion is as reasonable as any. I feel compelled to argue against it (I even thought out the arguments in my head!) but my compulsion to have an opinion on HN is a disease, and you made a point of saying that you gave the opinion because asked.
Revolutionary technology + massive adoption ≠ good investment
Investors have poured money into a bottomless pit, attracted by the growth and glamour of the industry. The airline industry since its birth has had a collective net loss, in aggregate, despite moving hundreds of millions of people.
Commodity Product, no switching costs. Infinite competition
The airline industry since its birth has had a collective net loss, in aggregate, despite moving hundreds of millions of people.
The industrialisation essentially socializes the cost across a lot more people though, so even though it doesn't make a profit it does mean people can have air travel without it costing millions per flight for the few people who can afford it. Essentially the economies of scale from having lots of flights isn't enough to make it profitable but they are enough to make it affordable.
There's no spare money to extract from the airline industry but it's still very useful. The same could be true for AI in the long term.
Sometimes the goal of an industry is to exist rather than to make a profit, because the benefit to society is more important than profit. People don't like that though so they do a bit of creative accounting or head-in-the-sand denial around it.
in the case of airplanes the only thing thats the private market is the planes and the ticket, the entire system of airports, safety, navigation is state subsidized and when the market fails it gets bailed out. the oil is subsidized by constant warfare. it's just an illusion for reganomics so a few rich ppl can make a buck off of a public utility.
It isn't a commodity product in my opinion. Far from it. I think it will ultimately be a monopoly or duopoly for SOTA. The mid to low end is commodity, yes. But SOTA models are not commodities.
The number of competitors for SOTA drops by a few every year. The winners make more money, get more revenue, buy more compute, train better model with compute, buy best talent, and the cycle goes.
I think it's easier to fall behind and never catch back up than people think. One disastrous training run can leave a lab months to a year behind. For example, Meta's disastrous LLAMA 4 models. Meta is lucky to have their ads business as a funding source. However, Anthropic's revenue is growing so fast, that ability to use ads as a funding source to stay in the race may not last much longer for Meta.
To me, SOTA LLM training is very much like new chip fab nodes. One disastrous node can put you behind for many years or forever. The cost to build the next chip node doubles every every 4 years (Rock's law). The cost to train the next SOTA model likely has some similar power law which means over time, it's too costly for losers to keep up. The only reason TSMC isn't a defacto monopoly for advanced chip nodes is strictly due to geopolitics.
All things equal, let's say your SaaS startup uses GPT 5.0 (release 10 months ago) and my business uses Fable 5. We have the same business goals, same talent level, same strategies. I think the chance of my business winning against yours is higher.
you won't get debt if you don't have assets that can be repossessed, so having debt means these AI companies have assets: that's a strong thing, not a weak thing. interest rates are what they are, and they go up and down for reasons exogenous to your industry; debt regardless of interest is always "cheaper" than equity, and the shareholders expect to make their money from equity, paying interest on debt as a type of impedance matching and cost of keeping more equity.
so everything is going according to plan, and nobody knows the future, and predicting collpses has never been a profitable business.
I didn't have to read past the first few confusing contorted and convoluted paragraps of this article to decide to come over here and explain it, this is all straightforward corporate finance 102 and the article is fluff
I agree with the general sentiment, but I feel like it is also a bit reductive. Assets in this space are near impossible to evaluate and can fluctuate in value greatly based on other actors. In a hypothetical scenario where, say, google releases a new frontier model that somehow leapfrogs the competition by 5 months all of a sudden the value of the Asset of Fable 5 and GPT 5.6 might completely crater.
I would imagine Anthropic et al. are largely leasing land/buildings, so as the other commenter said… must be the server racks that are acting as collateral (if anything). Generally enterprise hardware depreciates very harshly. I’m used to paying $10 for Intel Xeons that once retailed for over $5,000. I expect to pick up some NVIDIA Blackwell 6000s for $100 each someday.
Yep, a friend recently told me that he remembers working somewhere that gave away old empty server racks - they were unnecessary, and expensive to store, so why keep them?
GPUs have a five year lifespan before they become obsolete and start experiencing reliability issues. We're already 1-2 years into that five year lifespan.
That wasn't what it seemed like at the time. Amazon didn't post profits, sure, but they sure as hell weren't a giant money suck either, they didn't need billions in financing to run their business. There were a lot of Amazon bears, but they were concerned about the high valuation, not about them going broke (since even the most pessimistic bear can read a cashflow statement).
- Gen AI: Too Much Spend, Too Little Benefit?: https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too... (Goldman Sachs)
- AI’s $600 Billion Question: https://sequoiacap.com/article/ais-600b-question/ (Sequoia Capital)
- The Simple Macroeconomics of AI: https://www.nber.org/system/files/working_papers/w32487/w324... (MIT / Daron Acemoglu)
After reading these articles go by so often, it feels like what I actually can’t keep up with is the bond market. To paraphrase Trotsky, you may not be interested in the bond market, but the bond market is interested in you. I want to be able to read the signals at the bottom of this article, and divine some kind of prediction that can guide me… I don’t know, to choose whether I should buy a house or change the investment strategy in my retirement fund or something. But I’m just seeing all these signals go by, waiting for the story to be written, which only happens when the dust settles.
I guess I’ll go back to not understanding AI, instead of not understanding the bond market.
"trajectory of humanity to a parabolic move upward" is poorly defined here. Whether we are headed to a machine god ruled scenario or "just" incredibly powerful productivity tools, there will be a lot of economic pain for some (most) and a lot of economic gain for a few.
I've yet to a see an LLM/agent-based business plan in where scaling massively with an order fewer workers is not a central part of the value proposition.
You’re talking about about an amount that is a 13% of the total US government spending, of which is 20% of the entire US GDP.
I’m not saying it’s insignificant but it’s only a few percent of the US GDP and it represents spending over several years.
In what bubble does this pressure exist?
Revolutionary technology + massive adoption ≠ good investment
Investors have poured money into a bottomless pit, attracted by the growth and glamour of the industry. The airline industry since its birth has had a collective net loss, in aggregate, despite moving hundreds of millions of people.
Commodity Product, no switching costs. Infinite competition
The industrialisation essentially socializes the cost across a lot more people though, so even though it doesn't make a profit it does mean people can have air travel without it costing millions per flight for the few people who can afford it. Essentially the economies of scale from having lots of flights isn't enough to make it profitable but they are enough to make it affordable.
There's no spare money to extract from the airline industry but it's still very useful. The same could be true for AI in the long term.
Sometimes the goal of an industry is to exist rather than to make a profit, because the benefit to society is more important than profit. People don't like that though so they do a bit of creative accounting or head-in-the-sand denial around it.
The number of competitors for SOTA drops by a few every year. The winners make more money, get more revenue, buy more compute, train better model with compute, buy best talent, and the cycle goes.
I think it's easier to fall behind and never catch back up than people think. One disastrous training run can leave a lab months to a year behind. For example, Meta's disastrous LLAMA 4 models. Meta is lucky to have their ads business as a funding source. However, Anthropic's revenue is growing so fast, that ability to use ads as a funding source to stay in the race may not last much longer for Meta.
To me, SOTA LLM training is very much like new chip fab nodes. One disastrous node can put you behind for many years or forever. The cost to build the next chip node doubles every every 4 years (Rock's law). The cost to train the next SOTA model likely has some similar power law which means over time, it's too costly for losers to keep up. The only reason TSMC isn't a defacto monopoly for advanced chip nodes is strictly due to geopolitics.
I can't prove it. It's just my opinion.
so everything is going according to plan, and nobody knows the future, and predicting collpses has never been a profitable business.
I didn't have to read past the first few confusing contorted and convoluted paragraps of this article to decide to come over here and explain it, this is all straightforward corporate finance 102 and the article is fluff
Question is: is that worth enough to cover the debt after the market crashed?
Until they didn't.
Not sure what your point is.
https://youtu.be/LxJW7hl8oqM?is=IjdyHwZchaiMHk4C