To be fair to banks this is what they do with almost all debt their underwrite.
They hold it to help stabilize the market for it while its being sold and then offload it once the market has been established.
Nothing out of hte ordinary here, but to be fair the original article isnt' a financial site so they would have no real idea about how this work behind the scenes.
This doesn't seem like a story at all given how robust this debt is trading.
Isn’t this a perfectly normal thing banks do with all large debt? Any bank underwriting any $40b loan was always going to syndicate it.
Not to do so would be reckless.
What is unusual about this is that syndications usually clear in a few weeks but this has been stuck in distribution for six months and still not closed.
> Matthew Moniot, co-head of credit risk sharing at Man Group, told the Financial Times the deal sizes involved are "out of scale to anything we've thought about, ever."
Which of course is a valid point to raise, but you should just say "corporate bond issuance is at a all time high", rather than some clickbait headline about how "Banks are now trying to offload the AI data center debt" and it's "out of scale to anything we've thought about, ever."
What is unusual about this is not that it is a single debt of that size but that it is Oracle that is the poor quality indicator.
And it's just one of them: they know that there is about three trillion dollars of this risky shit out there either in the hyperscaler/FAANG balance sheets or off the books, so they are going to struggle to sell this, and Oracle's position gets worse every day.
We are getting a lot closer to the Margin Call crisis meeting moment where one bank decides it cannot assess the risk of their share of it, are we not?
But this time it is happening in a world where retail investors, lawyers, government officials, elected representatives are much more likely to have seen Margin Call and The Big Short.
Everyone knows these loans are dogshit but if anyone moves first to say so then everyone else’s loans basically are screwed too. So everyone is incentivized to pretend like there’s no problem even though everyone knows there’s a problem.
Listening to Ed Zitron on various podcasts, there are several issues with Oracle stocks:
1. they are one place above junk;
2. they are being treated as effectively junk;
3. if they get classified as junk, any pension portfolios holding Oracle stock must sell as they can't hold junk stocks.
If that happens then there will be a cascade of effects. Not just in the AI/datacentre sector but elsewhere, as Larry Ellison holds a stake in Paramount/helped finance the Paramount/WB merger.
But if Oracle crashes and burns, is that going to tank the whole industry, or is it more likely that the bigger players just eat it up and take its customer?
Oracle is a really serious problem, because they actually are too big to fail. Nobody can take Oracle's core business that quickly. So multiple governments, internationally, are going to have to step in to preserve the various outposts of Oracle that keep them and major organisations running.
Those bits could be sold off to different global consulting firms, I suspect.
(But aren't those global consulting firms also going to be freaked out about their own AI risks? And what does the whole AI-assisted coding thing do to a sober assessment of what all of Oracle's global business is worth? Lots of unanswered questions)
Also that is not the AI data centre bits; those will have to be chopped off and dumped.
If that doesn't seriously spook the markets — a company as well-known and formerly blue-chip as Oracle, needing an international bailout — then it would be kind of surprising.
Massively not an expert, but at that level, things are not a binary "good/bad investment". Financial institutions have policies regarding finely sliced debt types, ratios, risks, lengths, sectors, etc.
I personally think it's all hugely overcomplicating things but that's the way modern world works.
(until it every now and then, on regular intervals, doesn't:)
I saw an article a couple of days ago on one of the popular sites (ft or Bloomberg or something), about how now is a great opportunity to purchase private company credit.
Guess this is the submarine https://paulgraham.com/submarine.html
Private credit is far different than the debt of a public ally traded company. It is far riskier as those companies can't easily tap the credit or equity markets for new money if they get into trouble the way a public company can.
But the debt we are talking about is not Private credit debt, its completely different animal with far different risk characteristics
Had temporary acronym overload, unpacking for anyone else:
The giant Republican wishlist bill from the beginning of Trump's term has relaxed the banks' minimum Enhanced Supplementary Leverage Ratio so they can make riskier borrowing/bets.
You might as well say that banks giving mortgages caused that. Sure it happened as part of the process but that wasn't the thing that caused the problem.
> In the early 2000s, the debt underpinning CDOs was generally diversified, but by 2006–2007—when the CDO market grew to hundreds of billions of dollars—this had changed. CDO collateral became dominated by high risk (BBB or A) tranches recycled from other asset-backed securities, whose assets were usually subprime mortgages. These CDOs have been called "the engine that powered the mortgage supply chain" for subprime mortgages, and are credited with giving lenders greater incentive to make subprime loans, leading to the 2007–2009 subprime mortgage crisis.
No, to be more specific, banks selling the collateralized mortgages that were subprime in risk as if they were prime and these risk ratings being accepted at their word and via collusion with the ratings agencies is what caused this.
There is no issue with collateralized debt as long as the risk is properly assessed, the whole point of securitization is that you can take on higher risk for higher reward of an asset that was previously just one single piece of risk.
The nuance here, though, was that the CDOs and similar instruments were sold in a tranched fashion, and the super-senior tranche was made to look as if it was very low risk, as normal for tranches instruments.
But when your entire series of tranches all consists of trash, then which tranche you were in didn't really matter.
CDO and CDO2 instruments hid this distinction really well, and unfortunately the ratings agencies bought into it and sold crap as AAA securities.
> Cloud computing giant Oracle Corp is on the cusp of a junk-grade credit rating and its shares have tanked as the artificial intelligence hyperscaler becomes the fall guy of investors' concerns over the sector's debt-fuelled spending binge.
Honestly, that would be more like the railway mania in the 1840s or, really, insert any "Panic of X" here. Speculative mania has been a thing for centuries and will continue to be a thing for a centuries.
The specific problem of the 2008 financial crisis was that the banks were going out of their way to obfuscate the source of their CDOs so that, even when it was clear in 2007 that the subprime mortgage market was unraveling, you couldn't look at your instruments and figure out whether or not you were affected. The problem wasn't the CDO, it was the obfuscation. And as far as I'm aware, the current debt offerings aren't being obfuscated--if you want to avoid the data center debt because you think it's junk, just avoid the stuff being touted as data center debt.
It’s also hard to ignore differences. Bubbles gonna bubble, but the consumer is quite healthy this time around. Burry being short doesn’t make it 2008.
> Wage growth has remained relatively muted over the past year. That, combined with higher inflation, means Americans are earning less for an hour of work. Average hourly earnings, adjusted for inflation, fell 0.5 percent in September, and were down 0.7 percent from a year earlier. It was the sixth straight month of year-over-year declines in real earnings.
> The share of seriously delinquent auto loan balances has hit a record high this year, while nearly 13% of credit card balances are seriously past due — close to their 2010 peak in the aftermath of the housing crisis — according to recent data from the Federal Reserve Bank of New York.
> The ability of U.S. families to stay current on their debts worsened over the past three years, hitting levels not seen since the aftermath of the global financial crisis, the Federal Reserve reported Friday.
> Wage growth has remained relatively muted over the past year. That, combined with higher inflation, means Americans are earning less for an hour of work. Average hourly earnings, adjusted for inflation, fell 0.5 percent in September, and were down 0.7 percent from a year earlier. It was the sixth straight month of year-over-year declines in real earnings.
Where are they getting this data? This directly contradicts their claim:
Another factor in that case was that the government regulated, in some cases, the ratings of bonds that entities were allowed to buy (they couldn't be as highly levered with lower rated bonds) and so there was an incentive for bond ratings to be falsely inflated, happily carried out by the rating agencies. Of course collateralized junk is still junk, and many people at the time knew that, but there was money to be made in looking the other way.
And for this one when insurance companies are forming syndicates because their underwriting becomes greater than any one company, it just sounds frightening, the sound this thing is going to make when it finally pops.
yeah, either they discount it enough to look worthwhile to someone, or they keep it and take the risk. Risk assessments are inherently subjective, so both parties can think they're making a deal.
>We lent £1,000 to John Sozzleby at 10%. We quantify the risk level as “iffy”, which is 42.6 qualm points above "dodgy". and are willing to sell you the debt for £700. its a great deal!
This is common. Selling debt is normal. If there is money to be made by selling the debt and going back to issuing more loans, they’ll try to find a buyer.
Everyone is so primed to believe that all things related to data centers are bad that they’ll assume this is a special event. The person writing the headline knows that.
I remember in Summer 2007 feeling that there was an impending implosion coming based on the shady practice of using 2nd mortgages as downpayments - it had become a common practice. Started doing a bit of shorting (not too much), made a little, lost a little and then sort of lost interest in doing it just about the time things fell apart in '08. All that to say that these things can take a lot longer than you'd expect. You can look at the situation (like the circular funding as well as this article) and decide that yes it's not sustainable. But it can sustain longer than you expect.
Oh, I am well aware it's hard to time it. But still getting a general sense of the impending crash can help preparing in different ways. You can park part of your assets, not completely withdraw. Or pay extra for piece of mind by hedging them. Sure, it will cost, maybe won't be needed at all, or still misfired, but I'll sleep better with less anxiety, so that's still a win in my book...
With these things it's not about spotting the signs but rather predicting the timing. Everyone and their dog can see the signs but that's not enough to be actionable
The big two AI labs are still not listed on the stock market, and SPCX is still above it's listing price. Unless you can remain liquid more than the market remains irrational better find some other use for your investment funds. No?
Not sure I agree. The credit default swaps have been extremely expensive lately, meaning that someone is already thinking about an AI bubble popping that will take private credit down which will make banks take a serious hit (11% of their outstanding loans are to linked to AI). On top of that we are in a time where the economy is propped up basically by AI spending and delinquencies on all sort of debt are at historically high rates.
Sorry if this appears obtuse, but what exactly is an “AI Data Center”?
Is it a data center with nothing but GPUs, that does nothing but inference all day and all night?
Is it a data center custom-built for only AI companies to use?
Do these “AI Data Centers” have data storage? VPS? …
Or, as I suspect, is “AI Data Center” simply a fake media buzzword that really means “Normal Data Centers” but with political baggage and stigma built-in.
Who is building the infrastructure to resell the gpus from failed data centers if / when this collapses. It would be a waste for everything to get wasted
The GPUs used in data centers cant really be used outside the same context. They cant connect to standard PCs or servers. They are highly specialized. The resale would need to be at the server or rack level for the same use case.
Indeed, also even the air cooled ones are designed for high static pressure fans (ie small fans that spin at really high RPMs). Those fans are really really loud.
You can just mod them to run on consumer fans that are designed to spin much slower but have higher diameter
Yes, they're "just" pcie, though units like H100 and newer (or did it start with pascal? i really don't know) have variants that use a proprietary connector called SXM, but it's always just pcie lanes underneath. In some versions there are interposer and mezzanine boards, but again, pcie with associated bits and bobs like retimers and pcie switches. Not impossible to integrate, just annoying.
I mean it's not that hard. It's just hard if your frame of reference is "can I download it within one click on my $6000 Macbook".
This whole subthread is just sad. "Oh no, it's not the same form factor. It is useless".
Where's the "Hacker" in "Hacker News" there.
It's a shame that we need to defer to "China" as a shorthand for "actual engineers" just because our own engineers lack imagination and drive that would validate being given that title in the first place.
> It's a shame that we need to defer to "China" as a shorthand for "actual engineers" just because our own engineers lack imagination and drive that would validate being given that title in the first place.
There are a lot of reasons other than skill why China is the way it is. Culture and locality are big ones. I don't know a whole lot but I disagree that any "actual engineer" could easily come up with a sustainable business for this stuff in the US -- for example, harvesting and processing used parts for their components is a hugely laborious process, and China has a huge culture of recycling and reworking that simply does not exist in the US (and would largely be socially unacceptable to introduce!). Take for example the pre-modded 4090s, or the BGA desktop motherboards I mentioned, or any number of other products that are based not merely on sheer engineering. It's not exactly engineering on its own that causes these things to happen, but more the fact that nearly anybody can just throw together an operation like that and have tons of workers available to make it sustainable.
I do believe that your assessment that it will come from China is right.
Though, at least the "coming up with the idea" is something that could be done in the West I would argue. We too could read datasheets and probe stuff. We just chose not to.
Please don't misunderstand me as attacking you. I was attacking that other guy further up the thread with the empty GitHub account and the equally empty ambitions, trying to drag all of us crabs back into this bucket we find ourselves in.
There have been some examples of folks ripping off the SMX interface and rewiring the card for PCI-E but I'm not sure how efficient that would really be or what kind of market there could be
Why? It's hard to see what they're good for outside of AI data center workloads. Your average member of the general public arguably doesn't have much use for a "graphics processing unit" that isn't actually capable of processing graphics
I don't see how. I'm not saying that consumer-grade hardware isn't useful because enterprise-grade hardware exists. I'm saying that enterprise-grade hardware isn't particularly practical for most consumers. Personal computers were many things, but what they generally weren't was people installing IBM S/390s in their living rooms. Because some of the same design features that make them great for centralized, high volume business workloads also makes them kind of a PITA if that isn't what you need.
For my part, I do work on machine learning projects at home. But I'm nowhere near operating at a scale where datacenter GPU features like being optimized for 4- and 8-bit calculations would be a net benefit. Nor do I have any particular desire to have one of those things making my home office sound like I live next to a steel mill.
It's not too far off from why I don't have a box truck parked in my garage even though it's technically more capable than my Kia. And on the occasions when that is what I need, the local hardware store will gladly rent one to me for $20 per day.
They can have a pretty good LLM running in there, or cad design. For the average joe, having the free tier of AI without the breach of privacy is pretty good already.
Do they normally spend six months trying to offload a single tens-of-billions-of-dollars debt at a discount that is associated with one of the USA's longest standing blue chip tech stocks (whose debt rating has fallen to one level above junk)?
The sirens might not be wailing yet but some of the lights are flashing.
This tracks with recent reporting that AI debt issuance fell 50% month over month in September[0]
> Tech companies raised $23 billion in AI-related debt globally in September, less than half the amount raised in August, according to Morgan Stanley data covering public bonds and private placement transactions.
> The decline marks a continued drop from June, when companies raised a record $113 billion for AI infrastructure including data centers and chips. In the US investment-grade market, where blue-chip tech groups borrowed around $306 billion between January and August, AI-related bond issuance stopped completely last month.
The CDOs on some loans are getting to the point where it’s too expensive to even bet that the loans will fail as the CDO market seems to have concluded the loans will fail.
> The CDOs on some loans are getting to the point where it’s too expensive to even bet that the loans will fail as the CDO market seems to have concluded the loans will fail.
Did you mean CDS? As your sentence doens't make sense as its written.
To be fair to banks this is what they do with almost all debt their underwrite.
They hold it to help stabilize the market for it while its being sold and then offload it once the market has been established.
Nothing out of hte ordinary here, but to be fair the original article isnt' a financial site so they would have no real idea about how this work behind the scenes.
This doesn't seem like a story at all given how robust this debt is trading.
My mortgage has been sold almost a half dozen times already.
Isn’t this a perfectly normal thing banks do with all large debt? Any bank underwriting any $40b loan was always going to syndicate it.
Not to do so would be reckless.
What is unusual about this is that syndications usually clear in a few weeks but this has been stuck in distribution for six months and still not closed.
> Matthew Moniot, co-head of credit risk sharing at Man Group, told the Financial Times the deal sizes involved are "out of scale to anything we've thought about, ever."
Isn't that just because corporate bond issuance is at a all time high?
https://www.economist.com/content-assets/images/20260711_EPC...
Which of course is a valid point to raise, but you should just say "corporate bond issuance is at a all time high", rather than some clickbait headline about how "Banks are now trying to offload the AI data center debt" and it's "out of scale to anything we've thought about, ever."
What is unusual about this is not that it is a single debt of that size but that it is Oracle that is the poor quality indicator.
And it's just one of them: they know that there is about three trillion dollars of this risky shit out there either in the hyperscaler/FAANG balance sheets or off the books, so they are going to struggle to sell this, and Oracle's position gets worse every day.
We are getting a lot closer to the Margin Call crisis meeting moment where one bank decides it cannot assess the risk of their share of it, are we not?
But this time it is happening in a world where retail investors, lawyers, government officials, elected representatives are much more likely to have seen Margin Call and The Big Short.
Yes this is basically the Margin Call situation.
Everyone knows these loans are dogshit but if anyone moves first to say so then everyone else’s loans basically are screwed too. So everyone is incentivized to pretend like there’s no problem even though everyone knows there’s a problem.
Listening to Ed Zitron on various podcasts, there are several issues with Oracle stocks:
1. they are one place above junk;
2. they are being treated as effectively junk;
3. if they get classified as junk, any pension portfolios holding Oracle stock must sell as they can't hold junk stocks.
If that happens then there will be a cascade of effects. Not just in the AI/datacentre sector but elsewhere, as Larry Ellison holds a stake in Paramount/helped finance the Paramount/WB merger.
[1] https://www.youtube.com/watch?v=rHaPfdnRoZ4 -- Mostly Economics Podcast #55: Who pays when the AI bubble bursts with Ed Zitron (12:36 - 23:53)
Oracle's force majeure notice on a data centre build-out is not helping, I am sure.
(Can't remember, is it this Texas one?)
But if Oracle crashes and burns, is that going to tank the whole industry, or is it more likely that the bigger players just eat it up and take its customer?
Oracle is a really serious problem, because they actually are too big to fail. Nobody can take Oracle's core business that quickly. So multiple governments, internationally, are going to have to step in to preserve the various outposts of Oracle that keep them and major organisations running.
Those bits could be sold off to different global consulting firms, I suspect.
(But aren't those global consulting firms also going to be freaked out about their own AI risks? And what does the whole AI-assisted coding thing do to a sober assessment of what all of Oracle's global business is worth? Lots of unanswered questions)
Also that is not the AI data centre bits; those will have to be chopped off and dumped.
If that doesn't seriously spook the markets — a company as well-known and formerly blue-chip as Oracle, needing an international bailout — then it would be kind of surprising.
I'm confused. Why would you offload it if you thought it was a good investment?
Massively not an expert, but at that level, things are not a binary "good/bad investment". Financial institutions have policies regarding finely sliced debt types, ratios, risks, lengths, sectors, etc.
I personally think it's all hugely overcomplicating things but that's the way modern world works.
(until it every now and then, on regular intervals, doesn't:)
They make money as middlemen, not the ultimate holders of debt.
I saw an article a couple of days ago on one of the popular sites (ft or Bloomberg or something), about how now is a great opportunity to purchase private company credit. Guess this is the submarine https://paulgraham.com/submarine.html
Private credit is far different than the debt of a public ally traded company. It is far riskier as those companies can't easily tap the credit or equity markets for new money if they get into trouble the way a public company can.
But the debt we are talking about is not Private credit debt, its completely different animal with far different risk characteristics
Found it: head of JP Morgan asset management calling "once in a lifetime opportunity to invest in high quality company debt" https://www.cnbc.com/2026/10/10/jpmorgan-sees-once-in-a-gene...
Hah, this is remarkably short turnaround. The OBBB only lifted the ESLR on April 1!
Had temporary acronym overload, unpacking for anyone else:
The giant Republican wishlist bill from the beginning of Trump's term has relaxed the banks' minimum Enhanced Supplementary Leverage Ratio so they can make riskier borrowing/bets.
thank you for this link, all these years on the internet and I had not come across it
This is just what banks do O_o underwrite loans, then collateralize (i.e. bundle) to people who want the risk.
> This is just what banks do…
Sure. Sometimes they cause global financial crises as part of that process.
https://en.wikipedia.org/wiki/2008_financial_crisis
You might as well say that banks giving mortgages caused that. Sure it happened as part of the process but that wasn't the thing that caused the problem.
No, banks collateralizing mortgages caused it.
https://en.wikipedia.org/wiki/Collateralized_debt_obligation
> In the early 2000s, the debt underpinning CDOs was generally diversified, but by 2006–2007—when the CDO market grew to hundreds of billions of dollars—this had changed. CDO collateral became dominated by high risk (BBB or A) tranches recycled from other asset-backed securities, whose assets were usually subprime mortgages. These CDOs have been called "the engine that powered the mortgage supply chain" for subprime mortgages, and are credited with giving lenders greater incentive to make subprime loans, leading to the 2007–2009 subprime mortgage crisis.
It's hard to ignore the similarities.
No, to be more specific, banks selling the collateralized mortgages that were subprime in risk as if they were prime and these risk ratings being accepted at their word and via collusion with the ratings agencies is what caused this.
There is no issue with collateralized debt as long as the risk is properly assessed, the whole point of securitization is that you can take on higher risk for higher reward of an asset that was previously just one single piece of risk.
This is correct.
The nuance here, though, was that the CDOs and similar instruments were sold in a tranched fashion, and the super-senior tranche was made to look as if it was very low risk, as normal for tranches instruments.
But when your entire series of tranches all consists of trash, then which tranche you were in didn't really matter.
CDO and CDO2 instruments hid this distinction really well, and unfortunately the ratings agencies bought into it and sold crap as AAA securities.
> unfortunately the ratings agencies bought into it and sold crap as AAA securities.
It's very hard to understand something when your salary depends on you not understanding something.
They just kept the music going, more ratings issued was more money for them. Paid by the bank creating the security, what could go wrong
> subprime in risk as if they were prime
What if this winds up being the case for the trillions dumped into AI in hopes of future revenues?
https://www.reuters.com/business/oracle-corp-goes-high-stake...
> Cloud computing giant Oracle Corp is on the cusp of a junk-grade credit rating and its shares have tanked as the artificial intelligence hyperscaler becomes the fall guy of investors' concerns over the sector's debt-fuelled spending binge.
Honestly, that would be more like the railway mania in the 1840s or, really, insert any "Panic of X" here. Speculative mania has been a thing for centuries and will continue to be a thing for a centuries.
The specific problem of the 2008 financial crisis was that the banks were going out of their way to obfuscate the source of their CDOs so that, even when it was clear in 2007 that the subprime mortgage market was unraveling, you couldn't look at your instruments and figure out whether or not you were affected. The problem wasn't the CDO, it was the obfuscation. And as far as I'm aware, the current debt offerings aren't being obfuscated--if you want to avoid the data center debt because you think it's junk, just avoid the stuff being touted as data center debt.
Is the risk on AI debt properly assessed?
Are these companies worth their incredibly high valuations?
It’s also hard to ignore differences. Bubbles gonna bubble, but the consumer is quite healthy this time around. Burry being short doesn’t make it 2008.
> the consumer is quite healthy this time around
Inflation's way up. Real wages are down.
https://www.nytimes.com/2026/10/14/business/economy/inflatio...
> Wage growth has remained relatively muted over the past year. That, combined with higher inflation, means Americans are earning less for an hour of work. Average hourly earnings, adjusted for inflation, fell 0.5 percent in September, and were down 0.7 percent from a year earlier. It was the sixth straight month of year-over-year declines in real earnings.
https://www.cbsnews.com/news/household-debt-delinquencies-ri...
> The share of seriously delinquent auto loan balances has hit a record high this year, while nearly 13% of credit card balances are seriously past due — close to their 2010 peak in the aftermath of the housing crisis — according to recent data from the Federal Reserve Bank of New York.
https://www.cnbc.com/2026/10/09/american-debt-delinquency.ht...
> The ability of U.S. families to stay current on their debts worsened over the past three years, hitting levels not seen since the aftermath of the global financial crisis, the Federal Reserve reported Friday.
> Wage growth has remained relatively muted over the past year. That, combined with higher inflation, means Americans are earning less for an hour of work. Average hourly earnings, adjusted for inflation, fell 0.5 percent in September, and were down 0.7 percent from a year earlier. It was the sixth straight month of year-over-year declines in real earnings.
Where are they getting this data? This directly contradicts their claim:
https://fred.stlouisfed.org/series/LES1252881600Q
Another factor in that case was that the government regulated, in some cases, the ratings of bonds that entities were allowed to buy (they couldn't be as highly levered with lower rated bonds) and so there was an incentive for bond ratings to be falsely inflated, happily carried out by the rating agencies. Of course collateralized junk is still junk, and many people at the time knew that, but there was money to be made in looking the other way.
The cause is always a failure of underwriting. That was true of the Savings and Loan Crisis, the Great Financial Crisis.
It feels like each bubble gets bigger.
> It feels like each bubble gets bigger.
Well you would expect that given that the market keeps getting bigger and the money in circulation keeps getting bigger.
Inflation alone means each bubble should be bigger than the last all other things being equal.
And for this one when insurance companies are forming syndicates because their underwriting becomes greater than any one company, it just sounds frightening, the sound this thing is going to make when it finally pops.
I reacted the same. What could be a bad sign though is if they try to sell it and they dont get what they want.
yeah, either they discount it enough to look worthwhile to someone, or they keep it and take the risk. Risk assessments are inherently subjective, so both parties can think they're making a deal.
>We lent £1,000 to John Sozzleby at 10%. We quantify the risk level as “iffy”, which is 42.6 qualm points above "dodgy". and are willing to sell you the debt for £700. its a great deal!
|This is just what banks do O_o underwrite loans, then collateralize (i.e. bundle) to people who want the risk.
... and from the linked piece ...
| Banks do not sell down debt they are confident in
... I have no claim to expertise but both of these can't be true.
I wouldn't worry about it anyway, you can be sure when the day comes "they'll be to big to fail".
This is what a lot of, if not all, IPOs are for too.
Lots of names to obfuscate the same financial yada yada.
They tend to be quite, you know, disapproving of needing to sell it at a discount.
yes just like mortgage backed securities in 2007
Next thing you know, they're going to take the money in your savings account lend it out to anyone who asks for a loan!
/s
Major base rate fallacy.
This is common. Selling debt is normal. If there is money to be made by selling the debt and going back to issuing more loans, they’ll try to find a buyer.
Everyone is so primed to believe that all things related to data centers are bad that they’ll assume this is a special event. The person writing the headline knows that.
Your understanding of finance is limited.
Stop trying to be someone you ain’t.
It's normal to sell debt. It's not normal to underwrite bad debt.
So, is this the type of signs we should be on the lookout anticipating an impending implosion?
I remember in Summer 2007 feeling that there was an impending implosion coming based on the shady practice of using 2nd mortgages as downpayments - it had become a common practice. Started doing a bit of shorting (not too much), made a little, lost a little and then sort of lost interest in doing it just about the time things fell apart in '08. All that to say that these things can take a lot longer than you'd expect. You can look at the situation (like the circular funding as well as this article) and decide that yes it's not sustainable. But it can sustain longer than you expect.
Oh, I am well aware it's hard to time it. But still getting a general sense of the impending crash can help preparing in different ways. You can park part of your assets, not completely withdraw. Or pay extra for piece of mind by hedging them. Sure, it will cost, maybe won't be needed at all, or still misfired, but I'll sleep better with less anxiety, so that's still a win in my book...
This is the thing about being right too early has the same effect of being wrong.
Which is why I hate when people criticize some obviously overvalued shit, and the response is "why don't you short it".
Sometimes the best way to be safe of something about to explode is to just get away from it.
Yep, market can remain irrational for longer than you can stay solvent.
With these things it's not about spotting the signs but rather predicting the timing. Everyone and their dog can see the signs but that's not enough to be actionable
If I knew the answer to that I’d go short some stock…and maybe not post about it.
The big two AI labs are still not listed on the stock market, and SPCX is still above it's listing price. Unless you can remain liquid more than the market remains irrational better find some other use for your investment funds. No?
Just short Oracle and Coreweave.
My chop block list has those, plus micron, amd, nvidia, sandisk, western digital, skhynix, corning and marvell.
Those should really eat a lot of shit if any of the two ai companies stops buying everything no matter the price.
Why Corning?
They make the fiber optic cables and such that connect everything.
Depends on what price the banks offload at.
When the debt in 2007 sucked, it crashed the economy. But when the price was good in 2006, everyone celebrated.
Note: prices are getting worse for bonds today but nothing panic worthy yet.
Not sure I agree. The credit default swaps have been extremely expensive lately, meaning that someone is already thinking about an AI bubble popping that will take private credit down which will make banks take a serious hit (11% of their outstanding loans are to linked to AI). On top of that we are in a time where the economy is propped up basically by AI spending and delinquencies on all sort of debt are at historically high rates.
I am panicking a bit TBH
oh ffs please
When are we getting CDOs with Nvidia GPUs as collateral as Nvidia envisions?
That's some very nice collateral you have there. But can it run crysis?
It can run a financial crysis
"bank does what banks normally do" is not news
Sorry if this appears obtuse, but what exactly is an “AI Data Center”?
Is it a data center with nothing but GPUs, that does nothing but inference all day and all night?
Is it a data center custom-built for only AI companies to use?
Do these “AI Data Centers” have data storage? VPS? …
Or, as I suspect, is “AI Data Center” simply a fake media buzzword that really means “Normal Data Centers” but with political baggage and stigma built-in.
Who is building the infrastructure to resell the gpus from failed data centers if / when this collapses. It would be a waste for everything to get wasted
Maybe as ship's ballast.
The GPUs used in data centers cant really be used outside the same context. They cant connect to standard PCs or servers. They are highly specialized. The resale would need to be at the server or rack level for the same use case.
Aren't they PCIe like everything else or am I missing something? I know they have some fast interconnects. If not, is an adapter to PCIe possible?
Cooling is usually what kills you with server gear at home. I don't have a cold aisle, or a case that flows that much air. Can I water cool it enough?
No video outputs would probably be my biggest annoyance. It should be fine at headlessly rendering games for streaming though, as well as AI.
If there's a will, there's usually a way.
Lots of the newer stuff is watercooled
Yes they use pcie electrically, but its not the same slot shape and there is usually carrier boards
You wouldn’t repurpose these for home gaming, it would still be for llms
Indeed, also even the air cooled ones are designed for high static pressure fans (ie small fans that spin at really high RPMs). Those fans are really really loud.
You can just mod them to run on consumer fans that are designed to spin much slower but have higher diameter
Yes, they're "just" pcie, though units like H100 and newer (or did it start with pascal? i really don't know) have variants that use a proprietary connector called SXM, but it's always just pcie lanes underneath. In some versions there are interposer and mezzanine boards, but again, pcie with associated bits and bobs like retimers and pcie switches. Not impossible to integrate, just annoying.
So, sort of specialized? but not really.
> If there's a will, there's usually a way.
That's the spirit!
I bet China will find a way. They have stuff like desktop-size motherboards with BGA soldered mobile chips, just cause they can.
They already have PCI-e SXM risers and external multi-SXM boards for V100's.
I mean it's not that hard. It's just hard if your frame of reference is "can I download it within one click on my $6000 Macbook".
This whole subthread is just sad. "Oh no, it's not the same form factor. It is useless".
Where's the "Hacker" in "Hacker News" there.
It's a shame that we need to defer to "China" as a shorthand for "actual engineers" just because our own engineers lack imagination and drive that would validate being given that title in the first place.
> It's a shame that we need to defer to "China" as a shorthand for "actual engineers" just because our own engineers lack imagination and drive that would validate being given that title in the first place.
There are a lot of reasons other than skill why China is the way it is. Culture and locality are big ones. I don't know a whole lot but I disagree that any "actual engineer" could easily come up with a sustainable business for this stuff in the US -- for example, harvesting and processing used parts for their components is a hugely laborious process, and China has a huge culture of recycling and reworking that simply does not exist in the US (and would largely be socially unacceptable to introduce!). Take for example the pre-modded 4090s, or the BGA desktop motherboards I mentioned, or any number of other products that are based not merely on sheer engineering. It's not exactly engineering on its own that causes these things to happen, but more the fact that nearly anybody can just throw together an operation like that and have tons of workers available to make it sustainable.
I do believe that your assessment that it will come from China is right. Though, at least the "coming up with the idea" is something that could be done in the West I would argue. We too could read datasheets and probe stuff. We just chose not to.
Please don't misunderstand me as attacking you. I was attacking that other guy further up the thread with the empty GitHub account and the equally empty ambitions, trying to drag all of us crabs back into this bucket we find ourselves in.
There have been some examples of folks ripping off the SMX interface and rewiring the card for PCI-E but I'm not sure how efficient that would really be or what kind of market there could be
Here is an example of an SXM2 adapter board to use a regular pcie slot
https://ebay.io/m/imDgiL
I volunteer as tribute!
I'd much rather be in the recovery/asset remarketing market rather than the primary/reseller one when this thing pops, for sure.
JOHN TULD: And Mr. Sullivan that challenge is?
PETER SULLIVAN: We have to hold these assets on our books longer than we might ideally like to.
JOHN TULD: Yes.
JOHN: So. Now that we know the music has stopped. What do we do about it?
JARED: Sell it all. Today.
JOHN: Is that even possible, Sam?
SAM: Yes. But at what cost?
Same thing they did with bad mortgages just before the 2008 crash.
They should offload the overprovisioned, hyper concentrated GPUs back to the people
Why? It's hard to see what they're good for outside of AI data center workloads. Your average member of the general public arguably doesn't have much use for a "graphics processing unit" that isn't actually capable of processing graphics
I run free local models on my A4500. I wouldn't mind a little upgrade... if I could afford it.
You sound exactly like the people who didn't understand personal computers because mainframes existed.
I don't see how. I'm not saying that consumer-grade hardware isn't useful because enterprise-grade hardware exists. I'm saying that enterprise-grade hardware isn't particularly practical for most consumers. Personal computers were many things, but what they generally weren't was people installing IBM S/390s in their living rooms. Because some of the same design features that make them great for centralized, high volume business workloads also makes them kind of a PITA if that isn't what you need.
For my part, I do work on machine learning projects at home. But I'm nowhere near operating at a scale where datacenter GPU features like being optimized for 4- and 8-bit calculations would be a net benefit. Nor do I have any particular desire to have one of those things making my home office sound like I live next to a steel mill.
It's not too far off from why I don't have a box truck parked in my garage even though it's technically more capable than my Kia. And on the occasions when that is what I need, the local hardware store will gladly rent one to me for $20 per day.
They're buying out RAM and consumer GPUs too.
What exactly is the average consumer supposed to do with an NVIDIA H100 or GB200?
They can have a pretty good LLM running in there, or cad design. For the average joe, having the free tier of AI without the breach of privacy is pretty good already.
For the average joe, spending $10k on the server-grade hardware needed to run the GPU, plus $100-$200 per month for power is not a great investment.
This is completely normal. Banks underwrite large loans and then package a bunch of them in different ways to sell to different investors.
Do they normally spend six months trying to offload a single tens-of-billions-of-dollars debt at a discount that is associated with one of the USA's longest standing blue chip tech stocks (whose debt rating has fallen to one level above junk)?
The sirens might not be wailing yet but some of the lights are flashing.
Oracle data enter debt.
'nuff said.
This tracks with recent reporting that AI debt issuance fell 50% month over month in September[0]
> Tech companies raised $23 billion in AI-related debt globally in September, less than half the amount raised in August, according to Morgan Stanley data covering public bonds and private placement transactions.
> The decline marks a continued drop from June, when companies raised a record $113 billion for AI infrastructure including data centers and chips. In the US investment-grade market, where blue-chip tech groups borrowed around $306 billion between January and August, AI-related bond issuance stopped completely last month.
0: https://www.investing.com/news/stock-market-news/tech-compan...
The CDOs on some loans are getting to the point where it’s too expensive to even bet that the loans will fail as the CDO market seems to have concluded the loans will fail.
> The CDOs on some loans are getting to the point where it’s too expensive to even bet that the loans will fail as the CDO market seems to have concluded the loans will fail.
Did you mean CDS? As your sentence doens't make sense as its written.