What’s funny about the comparison to the subprime mortgage crisis is that there are, in all honesty, multiple different versions of The Stripper With Five Houses from The Big Short:
The AI companies that only have customers because they spend $3 to $10 for every dollar of revenue.
The venture capitalists that are ultra-rich on paper, heavily leveraging their firms in companies like Harvey (worth “$11 billion”) and Cursor (worth “$29.3 billion”) that burn hundreds of millions or billions of dollars and are now both too large to sell to another company and too shitty a company to take public.
The AI labs that have built massive businesses on selling heavily-subsidized subscriptions to customers who don’t want to pay for them and API calls to AI startups that can only pay them if infinite resources exist.
The AI data center companies that, thanks to readily-available debt, have started 200GW of projects (and only started building 5GW of them) for AI demand that doesn’t exist, entirely based on the theoretical sense that maybe it will in the future.
Oracle, who is building hundreds of billions of dollars of data centers for OpenAI (which needs infinite resources to be able to pay its compute costs), is taking on equally-large amounts of debt, all because it assumes that nothing bad will ever happen.
The customers of AI startups that are building lifestyles, identities and workflows around them believing that we’re “just at the beginning” on top of unsustainable AI subscriptions.
All of these entities are acting based on a misplaced belief that the world will cater to them, and that nothing will ever change. While there might be different levels of cynicism — people that know there’re subsidies but assume they’ll be fine once they arrive, or people like Sam Altman that are already rich and don’t give a shit — I think everybody in the AI industry has deluded themselves into believing they have the mandate of Heaven.
So let's for a moment assume that this is all there is to it. What's the hedge to enable oneself to keep capabilities? Ordered by cost estimate.
4 M5 Max MBPs clustered, with a 4-bit GLM-5 quant? (~$40k)
Gamble that between OpenAI and Anthropic, one winner takes it all and then we have no choice but to pay 20x current cost for inference? (~ $5k a month)
Buy 8x H200 so you can hot-ish swap between GLM-5 and Kimi K2.5? (~$400k)
Group-buy a Vera Rubin / Groq 3 rack with 12 people so everyone can run their own finetune (not sure how you're going to do the tune fwiw) (~$750k (*12))
Isn't it amazing how the "pleb solution" costs $40k?
Gamble that between OpenAI and Anthropic, one winner takes it all and then we have no choice but to pay 20x current cost for inference? (~ $5k a month)
Wild! $5k a month with wage growth being flat since the 1970s
Was listening to a podcast about the Radium Girls and they made $3.25 a day for making about 250 watches which adjusted for inflation equals about $95!
I think people would pay for AI if they could afford it but the mass majority can’t.
Even with massive government contracts the revenues must continue to grow to justify all of this spend.
Most importantly the losses need to shrink. There will be one big suck; their IPO. But Nvidia needs to get some ROI too, per #1447738:
Nvidia CEO Jensen Huang said the company's recent $30 billion investment in OpenAI "might be the last time" it invests in the artificial intelligence startup before it could go public toward the end of the year.
Maybe I'm reading too much into it, but to me it screams: burning more money would be reckless. IPO quick. Give us returns.
The utilization numbers are the part that should terrify investors. Epoch AI estimated that the average data center GPU sits idle 50-70% of the time outside of training runs. Microsoft reportedly hit just 30% average utilization across its Azure AI fleet in Q4 2025 before they started subleasing capacity.
Compare that to cloud computing, which took a decade to reach 60%+ utilization and only got there because fungible workloads (web hosting, databases, CI/CD) could fill gaps. AI accelerators are purpose-built. You can't just throw a Postgres database on an H100 cluster when inference demand dips.
The subprime comparison actually undersells it in one way: at least mortgage-backed securities had an underlying asset that existed. Half these AI capex bets are against demand curves that are purely theoretical.
That $3-to-$10 spend per dollar of revenue ratio mirrors the dot-com telecom overbuild almost exactly. WorldCom alone spent $11 billion on fiber optic capacity. At bankruptcy they were using 2.7% of it. Global Crossing, 360networks, same story. Hundreds of billions burned.
But here is the part nobody talks about: that "wasted" dark fiber became the physical backbone of AWS and cloud computing a decade later. The infrastructure outlived every company that built it. Pennies on the dollar at liquidation.
If AI data center buildout collapses the same way, the interesting question is who becomes the liquidation buyer. Bitcoin miners already specialize in acquiring stranded energy assets and excess compute at distressed prices. Riot Platforms bought Whinstone for $651 million specifically because cheap power infrastructure was available after a previous bust cycle. A wave of abandoned 200GW AI data center projects would be the largest stranded energy opportunity in history, and the only buyers with a business model that works at those margins are running SHA-256.
The grim conclusion:
So let's for a moment assume that this is all there is to it. What's the hedge to enable oneself to keep capabilities? Ordered by cost estimate.
Isn't it amazing how the "pleb solution" costs $40k?
Wild! $5k a month with wage growth being flat since the 1970s
Was listening to a podcast about the Radium Girls and they made $3.25 a day for making about 250 watches which adjusted for inflation equals about $95!
I think people would pay for AI if they could afford it but the mass majority can’t.
Even with massive government contracts the revenues must continue to grow to justify all of this spend.
Most importantly the losses need to shrink. There will be one big suck; their IPO. But Nvidia needs to get some ROI too, per #1447738:
Maybe I'm reading too much into it, but to me it screams: burning more money would be reckless. IPO quick. Give us returns.
The utilization numbers are the part that should terrify investors. Epoch AI estimated that the average data center GPU sits idle 50-70% of the time outside of training runs. Microsoft reportedly hit just 30% average utilization across its Azure AI fleet in Q4 2025 before they started subleasing capacity.
Compare that to cloud computing, which took a decade to reach 60%+ utilization and only got there because fungible workloads (web hosting, databases, CI/CD) could fill gaps. AI accelerators are purpose-built. You can't just throw a Postgres database on an H100 cluster when inference demand dips.
The subprime comparison actually undersells it in one way: at least mortgage-backed securities had an underlying asset that existed. Half these AI capex bets are against demand curves that are purely theoretical.
That $3-to-$10 spend per dollar of revenue ratio mirrors the dot-com telecom overbuild almost exactly. WorldCom alone spent $11 billion on fiber optic capacity. At bankruptcy they were using 2.7% of it. Global Crossing, 360networks, same story. Hundreds of billions burned.
But here is the part nobody talks about: that "wasted" dark fiber became the physical backbone of AWS and cloud computing a decade later. The infrastructure outlived every company that built it. Pennies on the dollar at liquidation.
If AI data center buildout collapses the same way, the interesting question is who becomes the liquidation buyer. Bitcoin miners already specialize in acquiring stranded energy assets and excess compute at distressed prices. Riot Platforms bought Whinstone for $651 million specifically because cheap power infrastructure was available after a previous bust cycle. A wave of abandoned 200GW AI data center projects would be the largest stranded energy opportunity in history, and the only buyers with a business model that works at those margins are running SHA-256.