As noted in other comments, similar to crypto, this whole routine is a conscious and deliberate setup for bailouts.
Gary Marcus had pretty much nailed that most of the current iteration of AI via LLMs is not going to scale in any sort of way that justifies the current infrastructure buildout. The AI gurus all know this. So why, beyond keeping the NVIDIA farce going, is everyone still pretending the naked emperor is magnificently adorned?
I’m not the only one who theorizes that the massive data center capacity is being built out to accommodate the upcoming security apparatus (domination) that Thiel et al. have explicitly identified as their long term goal.
We are essentially funding the construction of our own panopticon, and via repeated bailouts we will continue to shovel enough cash into it until a sufficiently robust system exists to make political or systemic change impossible.
Everyone knew AGI was not coming for at least a few decades, if ever, and certainly not via LLMs at all.
Thanks for pulling all these points together in a palatable form. Well done.
I wonder if broad divestment from authoritarian-tech is possible. In my view, that could thwart their effort.
Pension funds successfully lobbied to divest from tobacco and fossil fuels. Most of VC money comes from pension funds and university endowments. And most of these guys personal net worth comes from publicly traded stock. Would be good if there was a way to opt-out of their companies in 401k holdings, but is tough given US laws that suggest investment managers must only consider financial risk & returns. But given this article I think there is an argument these companies present real financial risk.
At one point I worked for a university in Texas. We were required to put 6% of our salary each month into the Teacher Retirement System of Texas. When I left, I asked for all my funds to be rolled over (before the cratering in November 2002). Once it looked like Enron was cratering, I remember getting a notice from them stating they still had faith in Enron and would continue to invest in them. I had to go look this up to see if I was remembering this right: https://www.trs.texas.gov/sites/default/files/migrated/trs-historical-timeline.pdf. According to their timeline:
"2002 – The Trustees of TRS disclosed major losses in the TRS investment portfolio due to large investments in Enron Corporation. At the end of 2001, it was revealed that its (Enron’s) reported financial condition was sustained substantially by an institutionalized, systematic, and creatively planned accounting fraud, known since as the Enron scandal. As a result of these losses in the TRS investment portfolio, TRS Trustees suspended all annual Cost of Living (COLA) adjustments to retired teachers receiving pension benefits from TRS. A small adjustment of one extra pension check to retires was paid in 2007. A small COLA was also paid in 2013. Suspension of COLAs continued in 2014 and no change to suspension announced as of March 27, 2016."
We are seeing a combination of the dotcom bubble and the leadup to the Global Financial Crisis.
What was missing in the dotcom bubble, but present in the GFC was the pure Ponzi element, in which obvious rubbish was sold as AAA-rated gold. Unlike Enron, the fraud was out in the open, but people wanted to believe.
In the current situation, the most obvious Ponzi elements are crypto and Tesla rather than AI. In both cases, there is no story more plausible than "number go up"
And related to this is the huge growth in unregulated private lending, based on things like mortgaginf receivables.
One thought here, although the FAANGs have stock exposure here, the real losers are going to be private equity and there's going to strong pressure against any bailout.
I have more of a Nortel feel than an Enron one regarding the AI circlular money flows. This reminds me a lot of the vendor financing that ended up being totally fraudulent. But either way it isn't good
This is such a great analysis. It’s been so frustrating hearing all the talk of “well even if it’s a bubble, those data centers will still exist” like they are going to be useful for another purpose.
What other purpose? Something that develops over the next 10-20 years?
I think the thing that is missing from this bubble talk though are the negative externalities (e.g environmental costs) which are going to be paid one way or another, by all of us; and the opportunity costs. It really bums me out to think of all the other things we could have done with this money in this moment.
Our crumbling infrastructure, education, healthcare and utilities are going to be a drag on the economy for decades, and the recession that follows the popping of this bubble is going to make it that much worse.
I was gonna write something shallow like "of course there will be bailouts" but I though "just who gets bailed out, and how much?". I have no idea. Firms like a16z, Nvidia, Oracle? I imagine it would make that $20b to Argentina look like chicken feed.
I don't know if there has ever been such a widely acknowleged pre-burst bubble as the AI bubble. There was widespread concern over the valuation of dot-com companies, but nowhere near the same "this can't go on" fatalism of today. Maybe because its not retail investors pumping money into tulip bulbs, its a handful of billionaires making bets that look insane from the outside. It feels like standing on a beach watching a tsunami slowly rolling toward shore.
I'd just like to point out that even if we are in "merely" a telecom-style bubble, this time **IS** different: neither rails nor optical fibre lost value as fast as GPUs
Good point about Enron. It's byzantine complexity was not a nightmare only for investors. It turned out that, on the day the company declared bankruptcy, it was not, in fact, bankrupt. But the accounting was so abstruse it took months to figure that out. Caught in the trap of their own devising.
And guys 'who pay attention to history' are important people. When I was writing Wilful Blindness, I wanted to demonstrate that it was not just expert analysts who saw the looming crisis in the mortgage market; one of the first to spot and write about it was just a guy, so to speak, who throught through the numbers. The fog of complexity can be used as an effective defense against questioning, which is why curiosity and scepticism are our friends.
Thanks for the perspective. I had been thinking along the lines of the Global Crossing model because data centers can be repurposed to do other things, and so if it's not AI it'll be something else. So I figure the people who buy the data centers out of foreclosure sales will be fine. But the difference between data centers and fiberoptic cables is that data centers take an astronomical sum to operate. Wiping away construction debt in a foreclosure is a good first step but whoever comes away holding the property will still need to find a use that can pay operating costs.
I don't know the specific costs involved but data center capex has a much shorter depreciation cycle than fiber optic cable capex, so unlike cables and railroad tracks the hardware being installed on data centers won't really be that useful after the bubble bursts
Yes and even the chips involved are so specific to AI that they can’t just be repurposed for other uses. They also account for more than half the cost of the data center.
Well they do have other uses like image processing and gaming graphics, but we do not need so many of them if we’re not using them for AI.
I'm told that many of the servers will be useful for inference for several years even though they may be over the hill for training. Inference (i.e. generating output) is less GPU intensive and is, anyway, the bit people are willing to pay something for. I can see a vulture capitalist fund buying up the data centers and some of the intellectual property for cheap after a crash and running them profitably at that point. They won't improve the quality of the output, but AI slop may be adequate for enough people to pay a subscription to produce
Enron is a great comparison point. I like to articulate where we are as being ‘05 not ‘07. Although in the context of Enron I guess you can say it is like ‘98 not ‘00
"Sure, the specific companies went bankrupt, but look at how valuable the infrastructure they left behind was!": However, the infrastructure left behind when companies like CoreWeave go bankrupt will age poorly. Roughly speaking, Nvidia introduces a new generation* of GPUs every couple of years, so today's "wired" is two-years-from-now's "tired"**. Optical fiber is more durably valuable. Moreover, GPUs aren't great for most non-"AI" and non-"crypto" kinds of computing. Optical fiber is more versatile too.
*A fuzzy and overused term, hence "roughly speaking".
As noted in other comments, similar to crypto, this whole routine is a conscious and deliberate setup for bailouts.
Gary Marcus had pretty much nailed that most of the current iteration of AI via LLMs is not going to scale in any sort of way that justifies the current infrastructure buildout. The AI gurus all know this. So why, beyond keeping the NVIDIA farce going, is everyone still pretending the naked emperor is magnificently adorned?
I’m not the only one who theorizes that the massive data center capacity is being built out to accommodate the upcoming security apparatus (domination) that Thiel et al. have explicitly identified as their long term goal.
We are essentially funding the construction of our own panopticon, and via repeated bailouts we will continue to shovel enough cash into it until a sufficiently robust system exists to make political or systemic change impossible.
Everyone knew AGI was not coming for at least a few decades, if ever, and certainly not via LLMs at all.
Thanks for pulling all these points together in a palatable form. Well done.
I wonder if broad divestment from authoritarian-tech is possible. In my view, that could thwart their effort.
Pension funds successfully lobbied to divest from tobacco and fossil fuels. Most of VC money comes from pension funds and university endowments. And most of these guys personal net worth comes from publicly traded stock. Would be good if there was a way to opt-out of their companies in 401k holdings, but is tough given US laws that suggest investment managers must only consider financial risk & returns. But given this article I think there is an argument these companies present real financial risk.
At one point I worked for a university in Texas. We were required to put 6% of our salary each month into the Teacher Retirement System of Texas. When I left, I asked for all my funds to be rolled over (before the cratering in November 2002). Once it looked like Enron was cratering, I remember getting a notice from them stating they still had faith in Enron and would continue to invest in them. I had to go look this up to see if I was remembering this right: https://www.trs.texas.gov/sites/default/files/migrated/trs-historical-timeline.pdf. According to their timeline:
"2002 – The Trustees of TRS disclosed major losses in the TRS investment portfolio due to large investments in Enron Corporation. At the end of 2001, it was revealed that its (Enron’s) reported financial condition was sustained substantially by an institutionalized, systematic, and creatively planned accounting fraud, known since as the Enron scandal. As a result of these losses in the TRS investment portfolio, TRS Trustees suspended all annual Cost of Living (COLA) adjustments to retired teachers receiving pension benefits from TRS. A small adjustment of one extra pension check to retires was paid in 2007. A small COLA was also paid in 2013. Suspension of COLAs continued in 2014 and no change to suspension announced as of March 27, 2016."
Apparently TX wasn't the only state to do so: https://www.mrt.com/news/article/State-investment-funds-lost-62-million-from-7850189.php
You were in TRS, not ERS?
The calls of an AI bubble are reaching a crescendo, and they're coming from everywhere now.
To abuse a quaint expression ("what's a photocopier?"): By the time the hot stock tips reach The Great Unwashed at the photocopier, it's too late.
When you hear big bankers and financial companies talking about it, you know they've already priced their positions in.
We are seeing a combination of the dotcom bubble and the leadup to the Global Financial Crisis.
What was missing in the dotcom bubble, but present in the GFC was the pure Ponzi element, in which obvious rubbish was sold as AAA-rated gold. Unlike Enron, the fraud was out in the open, but people wanted to believe.
In the current situation, the most obvious Ponzi elements are crypto and Tesla rather than AI. In both cases, there is no story more plausible than "number go up"
And related to this is the huge growth in unregulated private lending, based on things like mortgaginf receivables.
One thought here, although the FAANGs have stock exposure here, the real losers are going to be private equity and there's going to strong pressure against any bailout.
I have more of a Nortel feel than an Enron one regarding the AI circlular money flows. This reminds me a lot of the vendor financing that ended up being totally fraudulent. But either way it isn't good
This is such a great analysis. It’s been so frustrating hearing all the talk of “well even if it’s a bubble, those data centers will still exist” like they are going to be useful for another purpose.
What other purpose? Something that develops over the next 10-20 years?
I think the thing that is missing from this bubble talk though are the negative externalities (e.g environmental costs) which are going to be paid one way or another, by all of us; and the opportunity costs. It really bums me out to think of all the other things we could have done with this money in this moment.
Our crumbling infrastructure, education, healthcare and utilities are going to be a drag on the economy for decades, and the recession that follows the popping of this bubble is going to make it that much worse.
I was gonna write something shallow like "of course there will be bailouts" but I though "just who gets bailed out, and how much?". I have no idea. Firms like a16z, Nvidia, Oracle? I imagine it would make that $20b to Argentina look like chicken feed.
I don't know if there has ever been such a widely acknowleged pre-burst bubble as the AI bubble. There was widespread concern over the valuation of dot-com companies, but nowhere near the same "this can't go on" fatalism of today. Maybe because its not retail investors pumping money into tulip bulbs, its a handful of billionaires making bets that look insane from the outside. It feels like standing on a beach watching a tsunami slowly rolling toward shore.
I'd just like to point out that even if we are in "merely" a telecom-style bubble, this time **IS** different: neither rails nor optical fibre lost value as fast as GPUs
Good point about Enron. It's byzantine complexity was not a nightmare only for investors. It turned out that, on the day the company declared bankruptcy, it was not, in fact, bankrupt. But the accounting was so abstruse it took months to figure that out. Caught in the trap of their own devising.
And guys 'who pay attention to history' are important people. When I was writing Wilful Blindness, I wanted to demonstrate that it was not just expert analysts who saw the looming crisis in the mortgage market; one of the first to spot and write about it was just a guy, so to speak, who throught through the numbers. The fog of complexity can be used as an effective defense against questioning, which is why curiosity and scepticism are our friends.
AIUI the same may have applied to FTX. It was illiquid and suffering essentially from a bank run, but it may not technically have been bankrupt.
I disagree on the prosecution thing for a couple of reasons
(a) it will be an easy and popular hit for the Trump admonistration (the claimed reason)
(b) it will be an easy way to transfer assets and wealth to the Trump clan during quote restructuring unquote (the real reason)
expect no one went to prison that time.... Typo
Thanks for the perspective. I had been thinking along the lines of the Global Crossing model because data centers can be repurposed to do other things, and so if it's not AI it'll be something else. So I figure the people who buy the data centers out of foreclosure sales will be fine. But the difference between data centers and fiberoptic cables is that data centers take an astronomical sum to operate. Wiping away construction debt in a foreclosure is a good first step but whoever comes away holding the property will still need to find a use that can pay operating costs.
I don't know the specific costs involved but data center capex has a much shorter depreciation cycle than fiber optic cable capex, so unlike cables and railroad tracks the hardware being installed on data centers won't really be that useful after the bubble bursts
Yes and even the chips involved are so specific to AI that they can’t just be repurposed for other uses. They also account for more than half the cost of the data center.
Well they do have other uses like image processing and gaming graphics, but we do not need so many of them if we’re not using them for AI.
I'm told that many of the servers will be useful for inference for several years even though they may be over the hill for training. Inference (i.e. generating output) is less GPU intensive and is, anyway, the bit people are willing to pay something for. I can see a vulture capitalist fund buying up the data centers and some of the intellectual property for cheap after a crash and running them profitably at that point. They won't improve the quality of the output, but AI slop may be adequate for enough people to pay a subscription to produce
That’s a good point. Existing models are good enough at a lot of things, they’re just too expensive to run profitably.
Enron is a great comparison point. I like to articulate where we are as being ‘05 not ‘07. Although in the context of Enron I guess you can say it is like ‘98 not ‘00
"Sure, the specific companies went bankrupt, but look at how valuable the infrastructure they left behind was!": However, the infrastructure left behind when companies like CoreWeave go bankrupt will age poorly. Roughly speaking, Nvidia introduces a new generation* of GPUs every couple of years, so today's "wired" is two-years-from-now's "tired"**. Optical fiber is more durably valuable. Moreover, GPUs aren't great for most non-"AI" and non-"crypto" kinds of computing. Optical fiber is more versatile too.
*A fuzzy and overused term, hence "roughly speaking".
**I've read some of those old tech magazines too.