To me the main question is what happens after that..
We all move to frozen open source models running on 2nd hand Oracle/Coreweave GPUs? 10 years before someone dares make another training run?
Culturally, do we all collectively sober up once money dries and hallucination are still here? Pendulum swing, AI consideredharmful moment? How to promote healthy use when cognitive surrender is so engrained in us?
What happens if there's a new GPT2 scale (i.e. not astroturf/mass histeria marketing) breakthrough?
I mean the companies are delaying the inevitable. We have companies putting limit on the monthly token spend. Just 1% (you see what I did) of the companies can spend fearlessly on AI. In time most of the companies will be burnt out of their funding. I hope AI can make a case to be affordable.
Income by implemented features, losses by bugs and savings by personal. Sticks and stones may break even my bones, but ledger words will never hurt me.
Zitron is literally the worst person to raise alarms about the financials of the AI ecosystem because he's so hyberbolic and pollutes his own arguments with nonsense.
Take:
> When somebody decides to build an AI data center, they form a special purpose vehicle (much like a CDO), which then raises debt, in some cases slices it into tranches and, in most cases, sells them to institutional investors, asset managers or banks.
This is just such a weird and wrong comparison. A CDO's assets are other people's debt claims. The same mortgage bond could be split among many CDOs at once, those CDOs could be re-tranched into further CDOs, and thanks to credit default swaps, synthetic CDOs could reference bonds nobody in the deal actually owned. So basically exposure to a fixed pool of mortgages could be manufactured without limit.
A data center SPV's assets are the building, the power interconnect, the GPUs, and the customer contract. If the SPV fails, the loss is limited to what those things are actually worth. There are no multipliers as there are with CDOs.
Later in the post, Zitron even concedes this:
> What differs this from the subprime mortgage crisis is that the systemic risks aren’t driven by derivatives or complex financials but by the sheer scale of costs to build an AI data center, a catastrophic misunderstanding of the AI industry itself and the dangerous lending standards of private credit.
He claims this isn't important:
> When every single debt deal is over $500 million and usually numbering in the billions, we don’t need a vast web of different contracts to create a systemic risk, just clusters of projects that either fail to keep up with their SPVs’ debt or bonds that go unpaid by destitute or defunct data center developers.
But here's the thing: systemic risk isn't a function of how big the losses are. Instead, it's a function of who takes the losses and whether they propagate.
Equity holder losses just get absorbed by equity holders. What happened in 2008, on the other hand, was that the losses hit leveraged intermediaries funding long assets with overnight money, so one firm's distress became another firm's funding withdrawal.
Big deal sizes don't create that type of situation. A $10 billion SPV default is a $10 billion loss distributed across whoever bought the debt.
He brings up Lehman but that's literally the worst example for his argument. Lehman's losses were trivial against its $600 billion balance sheet. It failed because of a funding run. Repo counterparties refused to roll, the clearing banks demanded more collateral and prime brokerage clients pulled their balances. This doesn't happen in an SPV because SPV debt is term debt. It's sized and dated to match the asset. There are no runs on a term loan. When an SPV breaches its DSCR defaults, the lenders take the assets. It's not pretty, but it's contained. It can't spread beyond its own confines and multiply because there is no maturity mismatch, which is what killed Lehman.
1. LLMs are remarkable and they've uncapped a supply/demand loop for software that has previously been much more tightly constrained than anyone realized. It turns out that if software is much cheaper and faster to make, people find ways to use a lot more software, so much so that the world is temporarily completely out of all the parts you need to make the machines that turn electricity into software.
2. The leading tech companies have spent wildly on something that seems likely to turn out to be a commodity that sells for a few points over what it costs to provide it on the expectation that the gains in software developer productivity would also apply to every other industry in a reasonable time, replacing human workers and increasing productivity.
Recklessly taking a trillion+ in debt to corner the market, only to find you can't actually corner it and can't built a real moat with this technology as long as everybody knows how it works (and everybody that wants to know, knows how it works), seems precarious, to me.
They have to make a lot more than $100/month off of everyone using their services for this to work out for them, and nobody wants to spend a lot more than $100/month for these services. People start looking around for alternatives the moment Anthropic says, "Well, first one's free, but we're going to take away the best model on the subscription plans pretty soon, of course."
Ed may be wrong on some points. But, it's hard for me to look at how much money the big guys have spent and not wonder, "Who's going to buy the services at the prices they need to charge?" It isn't going to be me.
Nobody with a minimum of credibility says that. Everyone says this is clearly a bubble and yet it could go on another couple of years.
Nobody is timing it. But if you want to deny the obvious by numbers go ahead
Zitron's screeds reminds me of all the analysts who predicted the collapse of TSLA based on the fundamentals, and recommended shorting it. How does that saying go? "markets can remain irrational a lot longer than you can remain solvent".
There is an immense, coordinated, desperate effort to keep the biggest party in the history of the financial system going, to pour enough fuel into the engine that escape velocity may be achieved against the forces of mathematics.
To me the main question is what happens after that..
We all move to frozen open source models running on 2nd hand Oracle/Coreweave GPUs? 10 years before someone dares make another training run?
Culturally, do we all collectively sober up once money dries and hallucination are still here? Pendulum swing, AI consideredharmful moment? How to promote healthy use when cognitive surrender is so engrained in us?
What happens if there's a new GPT2 scale (i.e. not astroturf/mass histeria marketing) breakthrough?
I mean the companies are delaying the inevitable. We have companies putting limit on the monthly token spend. Just 1% (you see what I did) of the companies can spend fearlessly on AI. In time most of the companies will be burnt out of their funding. I hope AI can make a case to be affordable.
This year a lot of companies got the green light to buy AI subscriptions for 1 year.
The fun begins the next financial year when leaders need to take stock and justify the costs.
Income by implemented features, losses by bugs and savings by personal. Sticks and stones may break even my bones, but ledger words will never hurt me.
Zitron is literally the worst person to raise alarms about the financials of the AI ecosystem because he's so hyberbolic and pollutes his own arguments with nonsense.
Take:
> When somebody decides to build an AI data center, they form a special purpose vehicle (much like a CDO), which then raises debt, in some cases slices it into tranches and, in most cases, sells them to institutional investors, asset managers or banks.
This is just such a weird and wrong comparison. A CDO's assets are other people's debt claims. The same mortgage bond could be split among many CDOs at once, those CDOs could be re-tranched into further CDOs, and thanks to credit default swaps, synthetic CDOs could reference bonds nobody in the deal actually owned. So basically exposure to a fixed pool of mortgages could be manufactured without limit.
A data center SPV's assets are the building, the power interconnect, the GPUs, and the customer contract. If the SPV fails, the loss is limited to what those things are actually worth. There are no multipliers as there are with CDOs.
Later in the post, Zitron even concedes this:
> What differs this from the subprime mortgage crisis is that the systemic risks aren’t driven by derivatives or complex financials but by the sheer scale of costs to build an AI data center, a catastrophic misunderstanding of the AI industry itself and the dangerous lending standards of private credit.
He claims this isn't important:
> When every single debt deal is over $500 million and usually numbering in the billions, we don’t need a vast web of different contracts to create a systemic risk, just clusters of projects that either fail to keep up with their SPVs’ debt or bonds that go unpaid by destitute or defunct data center developers.
But here's the thing: systemic risk isn't a function of how big the losses are. Instead, it's a function of who takes the losses and whether they propagate.
Equity holder losses just get absorbed by equity holders. What happened in 2008, on the other hand, was that the losses hit leveraged intermediaries funding long assets with overnight money, so one firm's distress became another firm's funding withdrawal.
Big deal sizes don't create that type of situation. A $10 billion SPV default is a $10 billion loss distributed across whoever bought the debt.
He brings up Lehman but that's literally the worst example for his argument. Lehman's losses were trivial against its $600 billion balance sheet. It failed because of a funding run. Repo counterparties refused to roll, the clearing banks demanded more collateral and prime brokerage clients pulled their balances. This doesn't happen in an SPV because SPV debt is term debt. It's sized and dated to match the asset. There are no runs on a term loan. When an SPV breaches its DSCR defaults, the lenders take the assets. It's not pretty, but it's contained. It can't spread beyond its own confines and multiply because there is no maturity mismatch, which is what killed Lehman.
I don't think Zitron will ever admit he's been wrong about LLMs
Two things can be true:
1. LLMs are remarkable and they've uncapped a supply/demand loop for software that has previously been much more tightly constrained than anyone realized. It turns out that if software is much cheaper and faster to make, people find ways to use a lot more software, so much so that the world is temporarily completely out of all the parts you need to make the machines that turn electricity into software.
2. The leading tech companies have spent wildly on something that seems likely to turn out to be a commodity that sells for a few points over what it costs to provide it on the expectation that the gains in software developer productivity would also apply to every other industry in a reasonable time, replacing human workers and increasing productivity.
Recklessly taking a trillion+ in debt to corner the market, only to find you can't actually corner it and can't built a real moat with this technology as long as everybody knows how it works (and everybody that wants to know, knows how it works), seems precarious, to me.
They have to make a lot more than $100/month off of everyone using their services for this to work out for them, and nobody wants to spend a lot more than $100/month for these services. People start looking around for alternatives the moment Anthropic says, "Well, first one's free, but we're going to take away the best model on the subscription plans pretty soon, of course."
Ed may be wrong on some points. But, it's hard for me to look at how much money the big guys have spent and not wonder, "Who's going to buy the services at the prices they need to charge?" It isn't going to be me.
Because he‘s not?
I thought the AI bubble was supposed to pop in two weeks a month ago?
the market can remain irrational longer than I can stay solvent
Nah, it must have already happened, I saw a video on YouTube that said it was imminent a year ago.
Nobody with a minimum of credibility says that. Everyone says this is clearly a bubble and yet it could go on another couple of years. Nobody is timing it. But if you want to deny the obvious by numbers go ahead
Zitron's screeds reminds me of all the analysts who predicted the collapse of TSLA based on the fundamentals, and recommended shorting it. How does that saying go? "markets can remain irrational a lot longer than you can remain solvent".
There is an immense, coordinated, desperate effort to keep the biggest party in the history of the financial system going, to pour enough fuel into the engine that escape velocity may be achieved against the forces of mathematics.