---
title: 'The Next 2008? Why Citadel Just Liquidated 80% Before the $10T AI Crash'
source: 'https://youtube.com/watch?v=efWYw0ruXl8'
video_id: 'efWYw0ruXl8'
date: 2026-08-21
duration_sec: 1500
channel: 'Meet Kevin'
---

# The Next 2008? Why Citadel Just Liquidated 80% Before the $10T AI Crash

> Source: [The Next 2008? Why Citadel Just Liquidated 80% Before the $10T AI Crash](https://youtube.com/watch?v=efWYw0ruXl8)

## Summary

This video analyzes several major developments in the AI and tech markets, including Citadel's liquidation of 80% of its situational awareness stock portfolio, the upcoming Anthropic IPO, rising off-balance-sheet debt in the AI buildout, and the potential for robotics to drive the next growth S-curve. The host, Kevin Paffrath, provides a detailed breakdown of these events and their implications for investors.

### Key Points

- **Citadel Liquidates 80% of Situational Awareness Portfolio** [00:42] — Citadel offloaded 80% of its aggregate risk associated with the situational awareness stock portfolio, completing nearly 100 block trades totaling over $4 billion in market value. This was likely due to a quick 20% rally in two days, making it profitable to dump the stocks immediately.
- **Anthropic IPO and Adjusted Operating Profit Concerns** [02:07] — Anthropic is preparing for an IPO, following SpaceX's record $75 billion raise (up to $86.2 billion with green shoe). Anthropic is disclosing an 'adjusted operating profit,' but the host questions what is being adjusted, suspecting training costs are being excluded, which are a critical component of their cost of goods sold.
- **Compute Cost vs. Revenue Mismatch** [03:17] — Anthropic is estimated to have 1-2 gigawatts of compute and is at $65 billion ARR, implying they earn $32-65 billion per gigawatt. Elon Musk wants to sell SpaceX compute for $40-50 billion per gigawatt, which is upside down compared to Anthropic's current earnings, indicating a potential profitability problem.
- **$10 Trillion AI Buildout Debt** [04:44] — There is $3 trillion in off-balance-sheet commitments, including $1.2 trillion in not-yet-commenced leases and $1.9 trillion in purchase commitments. Combined with on-balance-sheet liabilities, the total AI buildout debt could approach $10 trillion, eerily similar to the 2008 financial crisis when the $10 trillion mortgage market was collateralized.
- **Broadcom's $100 Billion SPV Raise** [06:10] — Broadcom is seeking to raise up to $100 billion through a special purpose vehicle (SPV), in negotiations with Blackstone and Apollo for over $60 billion in senior secured and $30-40 billion in subordinated tranches, with Broadcom offering guarantees that don't appear on their balance sheet.
- **SEC Removing Regulations on AI Debt** [07:24] — The SEC removed asset-backed securitization rules for data centers, meaning they don't have to disclose utilization or require skin in the game. This is going in the opposite direction of regulation, contributing to rising credit default swaps (CDS).
- **Credit Default Swaps Skyrocket** [08:38] — Oracle, Broadcom, and Nvidia have all surpassed their previous CDS peaks, while SpaceX has actually derisked. This indicates rising market concern about the debt levels in the AI sector.
- **SpaceX's $1.3 Trillion Spending Projection** [09:35] — SpaceX ended Q2 with 1.4 gigawatts of AI compute capacity, with a stretch goal of 10 gigawatts by end of 2027. At $50 billion per gigawatt, that's $500 billion needed, but their balance sheet is only $87 billion, requiring $414-430 billion in extra financing. Wall Street projects SpaceX will spend $1.3 trillion over the next 5 years.
- **Micron's EPS Growth Rate Problem** [12:16] — Micron's EPS growth rates are projected to be flat or declining, even if you 5x their forecast earnings. The host argues that a new S-curve, likely from robotics, is needed to reignite growth in memory plays like Micron and SK Hynix.
- **Robotics as the Next S-Curve** [14:31] — Robotics could be the next S-curve for AI, driven by large action models (LAMs). Agentic LAMs are text-based, but robot LAMs require synthesizing video, lidar, and sensor data, which is GPU-heavy and requires high bandwidth memory for training.
- **Timing Gap Between LLM and Robotics** [16:55] — The host believes robotic compute won't scale until 2030 or later, creating a potential lull between the decline of LLM training growth and the rise of robotics. This gap could lead to a bear market in AI stocks if it plays out before robotics take off.
- **Anthropic IPO Bullish Short Term** [23:24] — The host thinks the Anthropic IPO is bullish short term because it will generate capital that gets recycled into hardware, potentially fueling a 'hardware 2.0' rally. However, the turning point of the Anthropic LLM cycle will be a key red flag to watch.

### Conclusion

The video concludes that while the Anthropic IPO may provide short-term bullish momentum, the massive debt accumulation and potential timing gap between LLM and robotics growth pose significant risks. Investors should watch for the turning point in the Anthropic LLM cycle as a key indicator of the broader AI market's health.

## Transcript

about, including how Citadel just paperhanded their investment into well paperhanded their investment into well 80% of the stocks that they bought from Leopold who ended up getting Leo folded. We'll talk about that. We'll talk about
the anthropic IPO and why this matters for the AI bull market. Uh we'll also for the AI bull market. Uh we'll also talk about Micron and the problem but potential upside with Micron. So, this will be really exciting. Uh, and then
we'll talk about robotics, which actually really potentially helps Micron about this in large action models. There's a lot. So, let's try to break this down as simply as possible. So, first things first, Citadel. Citadel
first things first, Citadel. Citadel already offloaded 80% of their quote unquote aggregate risk associated with a situ uh situational awareness stock situ uh situational awareness stock portfolio. They completed nearly 100
portfolio. They completed nearly 100 block trades totaling over $4 billion in block trades totaling over $4 billion in market value. So my guess is even though their institutional analysts are coming out
like I was literally last week reading a piece and it was Citadel 10 reasons why we're bullish on the market and and then you hear oh but JK we've actually liquidated a lot of our positions uh that we bought from situational
awareness. My guess is what happened here is when they picked up the portfolio on July 29th, we believe it was the morning of July 29th, we had such a large rally back in some of these stocks coming up, you know, 20% in the
stocks coming up, you know, 20% in the span of 2 days. The ROI to basically dump the stuff right back on the market was just too great to pass up in such a short period of time. They would have made more money if they held more, but I
paperhanded. And so that's led to a lot of people wondering, did they paperhand because you have the next big suckening coming up? Maybe. So Anthropic is now trying to beat the SpaceX IPO. So SpaceX IPOed and raised $75 billion, biggest
IPO ever. Uh, plus the green shoe they ended up raising up to $86.2 billion. That was really impressive. But now everybody's kind of on pins and needles waiting for the anthropic S1 because we want to see what their profit actually
want to see what their profit actually looks like. See, they're disclosing that they're actually seeing an adjusted operating profit. But a lot of people hear adjusted operating profit and they're like, "What are you adjusting?"
The assumption is that what they're adjusting is all of their training costs. And so a lot of people are like, "All right, whatever. we can't, you saying, you know, here's here's the quote. Uh, Anthropic saw a positive
quarter. Yeah, that's great. But if you're adding back in your training costs, which in my opinion are like a pretty critical component of your costs of goods sold, you know, you wouldn't have one of the best frontier models if
keep training to stay the best, otherwise you'll fall behind. The S1 is going to reveal the true story. So, even though they're telling us, you know, they're they're at $65 billion of ARR, that's great, but think about what that
that's great, but think about what that means for a moment. Elon wants to sell means for a moment. Elon wants to sell SpaceX compute uh for, you know, SpaceX compute uh for, you know, somewhere around 40 to50 billion a
gigawatt. Well, in order for you to sell compute at 40 to50 billion dollars per gigawatt, you have to have way more revenue to justify that. Some people come out and say, "Oh, well, you know, they're making a hundred billion of ARR
per gawatt." But wait a minute. Anthropic is estimated to have between 1 Anthropic is estimated to have between 1 to 2 gawatt of compute and right now they're at 65 billion which means right now they're spending 30 or they're
now they're spending 30 or they're earning 32 to 65, right? 65 if they're only at 1 gawatt. I think they're closer to two. So they're probably only earning closer to $32 billion a gigawatt. And Elon is trying to sell them compute for
40 to50 billion a gigawatt. that's upside down. Over here, it's really it's decently low margin. And you know, if they're over here, they're upside down. So, there's a little bit of a problem with how much Anthropic is expected to
be able to like earn compared to what their compute costs are. That's why that's probably why you're getting a lot of people going, "Let's let's see what this filing actually looks like, especially since debt levels are going
absolutely crazy." Look at this piece right here. We sit at $3 trillion of offbalance sheet commitments, which represents 1.2 trillion in not yet
commenced leases and 1.9 trillion in purchase commitments. In other words, we got $3 trillion of offbalance sheet debt that we're not even aware of right now. And when you factor in all of the other debt that already exists out there and
debt that already exists out there and is getting uh getting added together, we could probably start knocking on the door of $10 trillion of AI buildout debt. So that would be a combination of the off-balance sheet
liabilities plus the onbalance sheet liabilities, the money uh companies like liabilities, the money uh companies like Google and Meta, SpaceX are raising. You knocking on the door of$10 trillion dollars of debt just for AI and the AI
buildown, which is eerily similar to the top of the 2008 do uh uh 2008 financial top of the 2008 do uh uh 2008 financial crisis when the $10 trillion mortgage market was collateralized into multiple different Wall Street assets. And so
so did all those derivatives. And well, we all know how that point played out. we all know how that point played out. But right now, the level of debt growth on this hope that Anthropic is going to be able to keep paying their bills keeps
expanding. Look at this. We just heard that Broadcom is seeking uh to raise up that Broadcom is seeking uh to raise up to $100 billion through a special purpose SPV vehicle. So, special purpose vehicle. Broadcom is in negotiations
with Blackstone and Apollo group management to raise over $60 billion management to raise over $60 billion in senior secured and 30 to 40 in subordinated tranches with Broadcom offering guarantees.
So here's the problem with companies offering guarantees. Those guarantees don't show up anywhere on their balance sheet. It's not like when Nvidia guarantees uh you know a hundred billion dollar
project for OpenAI. I could go open up the balance sheet for Nvidia and go ah guarantee. We don't see that. It doesn't show up. And so one of the things that this article is arguing on Broadcom is that
raise a bunch of offbalance sheet debt quote before the SEC potentially changes the rules. But so far and lately, the rule changes are actually going in the opposite direction. You know, just what a week to 10 days ago or so, Jensen
a week to 10 days ago or so, Jensen announced this $500 billion incentive financing structure with KKR and and uh you know, these other suits on Wall you know, these other suits on Wall Street. And that came immediately after
the SEC removed the assetback securization rules for data centers. The bottom line of that meant data centers don't have to disclose their utilization assetbacked securities. You don't require skin in the game anymore. Like
going in the opposite direction. We are removing regulations and limits on this debt. And that is in part why we're seeing credit default swaps rise. So you've got a lot here. You know, Citadel dumping, not a surprise. They're
traders. Then you've got this big Anthropic S1 coming up which could potentially reveal massive holes in this whole profitability argument that oh you know, they're going to make a hundred billion ARR for every 50 billion
they pay in compute. Well, the numbers right now don't seem to justify that. Uh maybe they'll grow into that. We'll see. But that's also why we're seeing credit default swaps now skyrocket. So you could obviously see this credit default
swaps right here. These are the CDS's. Uh Oracle is up here at the top which has surpassed its peak of uh right before or right during the Leo fold event. So Oracle surpassed its peak.
You've got Broadcom which just rapidly surpassed its peak. Uh and then you've got Nvidia which is roughly at peak slightly above uh peak credit default swap. Interestingly, the only company that actually rerated lower or derisked
was SpaceX. SpaceX seems to have derisked in the credit default swap market. Interesting. Maybe because of their money raise or whatever, but I personally think SpaceX is going to have to raise a lot of
money. And that's not trying to be bearish on SpaceX. It's just Elon wants to generate uh this or or get to 10 gawatts of compute. But how are we going to get to 10 gawatts of compute? See, here it is. Right now, SpaceX ended the
second quarter with 1.4 gawatts of AI uh compute capac capacity. Uh they the stretch goal is getting to 10 gawatts of compute by the end of 27. Well, at 50 you're spending on the leading edge chips, which you likely will, the Vera
Rubins and otherwise, that's 500 billion they got to come up with. So, what's their balance sheet? Well, their balance sheet is sitting at $87 billion to work with. So we need, you know, 414 to roughly $430 billion of extra dollars.
So where are we going to get extra $30 billion of financing? SpaceX has actually made that clear that their expectation is to finance the majority uh of their their capital needs, which is in part good for
see a lot of stock dilution by them issuing stock, especially $400 billion of debt and rates are high and somebody's got to pay this and rate, you know, debt keeps getting more expensive. You know, Tesla or SpaceX actually had
to increase how much they're willing to pay for their debt um to to in one of their latest bond offerings, but they're still only paying like 7% on their debt, whereas Coreweave is paying 9 to 10% on their debt. So, you know, SpaceX is
paying somewhere around 1.5 above Treasury, somewhere around there. So, anyway, the Wall Street Journal is projecting, and this blew my mind. Like my estimate was, oh, okay, SpaceX is going to raise, you know, $400 billion
compute. You ready for this? This blew my mind. The Wall Street Journal projects, well, actually, it's not Wall Street Journal. It literally just says Wall Street. Wall Street apparently is projecting SpaceX is going to spend $1.3
trillion over the coming 5 years. That's crazy. So in other words, like the spending is just continuing to go and that's obviously what's leading to so much enthusiasm at companies like Micron. And
so we actually put together uh some charts. You could get all these charts download that for free. Uh you could also get your alpha wire over in the and running right next to me. Uh and this is sort of what it looks like. Oh,
look at that. SpaceX AAI expands Grotbot availability to super grot plus cursor pro and cursor teams, you know, whatever. So, if you want to see this media, I think it's kind of cool. And download that in the uh Google or
Android app store. Sorry, Apple or Android app store. Just uh type in me let's talk about Micron for a moment. So what's fascinating here is that uh this is a chart of Micron's earnings per share historical on the left side and
share historical on the left side and their forecast for 2026 78 and onward. their forecast for 2026 78 and onward. Uh except this is their growth rate on earnings per share. So as you could see historically we have these booms and
busts where you get 100% up on growth rates and then it comes down negative 30% and then it's up 100%. And then it's flat right just to show you the magnitude of growth rates that we have seen for these high bandwidth memory
plays like skhinix or micron which represents about 18% of their revenues. magnificent that you've gotten this insane spike over here to where just to
put it into perspective, if I take their forecast earnings 5 years out and I just jack up their earnings. So, I'm going to go like jack them up as much as I can
and I'm going to literally 5x their earnings. The blue line I'm about to show you is going to show you their EPS growth rate if you 5x their current forecast for five years from now. Right here. That's it. This little jump is all
here. That's it. This little jump is all you would get. So instead of getting a relatively flat blue line right here, the green represents what you know we could see potentially with uh or sorry the the green would represent without
buybacks. The blue represents with buybacks. Getting a little granular mekevin.com. I don't want to over complicate this video. But the point is the growth rates are projected to be
this on the blue line. If I go in and manipulate the growth rates and 5x it in 5 years from now, which is not projected there, they are projecting growth but not a 5x, right? So I go crazy here. This is all I get for the spike. So, the
reason I say this, like simply put, the reason I put these charts together is we need a really big next wave of enthusiasm to really get growth rates to
enthusiasm to really get growth rates to skyrocket for especially memory plays. Now, how could you get that to happen? Robotics. That's where robotics come in. See, robotics are really interesting because in robotics you can maybe have
because in robotics you can maybe have the next Scurve coming from large action models. Now you have to divide this. So first of all, this is sort of the uh the outline that we have here, right? Let's go get my
Citadel, anthropic, micron, and robotics. Now we're going to talk about interesting about robotics. With robotics, you have two types of large action models. So they apparently what they call them these days. That's
what the kids are calling them these days. Two types. You've got large action models that are agentic. But agentic is really just a textbased LLM going through an order of operations. Okay, if this, then this, then this, then this,
then this. So it's like a prompt getting fed in over and over and over again. That's very different. uh if there are two types of lamb large action models two types of lamb large action models that's very different from a uh robot a
robot is going to require uh synthesizing video data lar data sensor synthesizing video data lar data sensor data from like fingertips or the feet if you will whatever that's a lot of data and in my opinion that is going to be
significantly GPUheavy and require a crapload of high bandwidth and require a crapload of high bandwidth memory for training. And so this is the memory for training. And so this is the next Scurve for uh the AI play. It's
There's just a little problem with that. And there are obviously some stock names benefit from this, but I want to hit the problem first. Not because I want to be a bear, because I just want to be realistic. So I drew this out. Ignore
realistic. So I drew this out. Ignore all the red text for a moment. So if this is where I think we're probably here on the LLM frontier. So LLMs are really good, but I think the growth rates are collapsing in how much LLMs
rates are collapsing in how much LLMs can improve. That's cool, but you know, it's not great. It means eventually we're going to see negative growth rates and that's already actually getting priced into some stocks uh in in the AI
trade that in four years from now you see negative growth. a company, you know, some of the um storage plays like a SanDisk or uh an Nvidia, you start negative unless there's another S-curve. The next scurve could be robotics, this
green line. Best case scenario, this green line Best case scenario, this green line shifts over and it picks up at the peak. The problem is I don't actually think we're going to scale robotic compute
until the end of the decade. Like realistically, I mean, companies are saying 2028, 27, maybe even. I think it's more realistically going to be 2030 or even into the 2030s. I mean, Elon's talking about, you know, Optimus robots
doing surgery. I can't wait. I think that's great. I think it's awesome, but it's probably 20 to 30 years out. That would be really bad for this cycle,
that long cuz if we have to wait that long to get our GPU training over here, you're going to have a crappy lull in between. So there is a very real between. So there is a very real potential weak out on frontier LLM AI
and the spending and the money raise that's possible. You get a bare market. If we get that to play out before robotics take off, that would be the robotics take off, that would be the best time to buy GPUs and other related
stocks in the stack. Think about it. Nvidia GPUs, AMD GPUs, Broadcom and integrated circuits. You don't have to remember that. Just little chips that do a specific purpose. Uh Marll for edge compute, base station AI for 5G and 6G.
This is what uh Jensen's talked about. Qualcomm uh communication chips, Qualcomm chips right here in your uh in your phones, you know, uh internet of things, robotics, Qualcomm. Pterodine, the stock's been on a tear. Uh, Unitatry
is really your Chinese competitor for for robotics. Uh, this is it's ticker six836 apparently. I don't know if it's all
just Chinese. I think I hear Chinese are really like 8s or 88. Like you go to the 88. I guess it's like a good luck thing. I don't know. I I've never been to China. But uh high bandwidth memory place for
Micron SKH highix that could be what actually really helps you boost that growth rate at the end of the decade which is now in decline. Uh so right now which is now in decline. Uh so right now where we sit with EPS projections for
the company uh buybacks are not okay. Uh blue is assuming they buy back 40% of the company. The EPS growth rates don't move that much compared to the baseline move that much compared to the baseline because the growth is so massive right
now. Like if you remove this big bump, these buybacks would be huge. Buying back 40% of the company in Micro would be so cool. UBS is projecting that, right? That'd be amazing. But compared to the growth where we are now here in
2026, it's just you can't get there again unless you get another S-curve. And maybe that's robotics. Uh, some people say uh CGNX. I don't a apparently robotics vision AI play. Cognix,
don't know anything about that one. I'm just being transparent. Samsung and LG potentially LG has a partnership with Nvidia. Uh, and then what would also win or get bailed out depending on when it comes uh would be compute, right? So
like SpaceX, Cororeweave, uh you know, NBIS, uh you know, NBIS, Iron, whatever. The the question though is if we end up, you know, we S-curved up with our LLMs, if we have this big
trough before the Scurve for robotic training comes, this is going to be very painful right here in a gap. Best case scenario
to kind of keep the economy booming and to keep the whole bubble going if you will, you know, the party going to keep it going. Best case scenario, maybe you don't see an LLM cop out and instead you see robotics pick up right here. So
robotics start picking up on GPU training before you get that decline in um in in the LLM training side. And we'll see anthropic IPO will be a big we'll see anthropic IPO will be a big sucking that will test markets for this.
Nvidia is a big player too with what they call a unifier model. Their Groot model unifies textbased tokens and action tokens. You know there are a lot Tensor to basically incentivize training. Uh this is also why you know
there is some lack of training data for this. This is why there's talk about Amazon and Meta putting glasses on their employees and even Tesla so they could track the, you know, how their employees operate and and create more data for
robots in the future. It's like you're training your own replacement, which is kind of eerie. There's even a rumor, or maybe call it an inference, that maybe call it an inference, that Microsoft bought Activision because the
gaming labs literally have some of the best potential training data maybe for robotics. Think about like Call of Duty characters, uh, you know, um, what's
characters, uh, you know, um, what's that company? Uh, Uni Unity. Unity. There it is. you know those Unreal engines, those are really movements and you could actually use those in robotic simulative trainings.
there's a lot of hope for that next frontier coming. There are a lot of requirements to get to that. Uh we also put on the MK app
under the data tab, you could see the where's the Nvidia page. We did a uh took a little video or a screenshot from the Nvidia presentation on this
and uh yeah we actually covered most of this already but this is sort of Jensen ends up narrating how this real and simulator data comes in to train robots.
This training phase is going to be huge. I think right now it's a tiny fraction of the compute that's being done. But as robotics go more mainstream, that's that's going to be the future. The next 50 years, there's going to be a crapload
of training robots. So, if you could have that really long-term time horizon, man, it's juicy. The question is, how much of a debt bubble 2008 do you go through between then uh and now? Or now
and then, I guess, would be the way to put it. So, um, you know, my take on all of this is I think the anthropic IPO is bullish short term because even though it's going to create a suckening, it's going to generate so much capital that's
just going to get recycled into hardware. So, I still think you'll see this sort of hardware 2.0 rally post Iran ceasefire. I still think software bottoms in Q3, Q4. Kind of already been seeing quite a bit of that bottoming. uh
plays that we're really a big fan of in the um uh course member trade alerts have done very well uh you know Palunteer uh Salesforce and we got others as well uh from uh certainly up from their lows like substantially up
Salesforce somewhere around 160 you know it's broken through I think 210 now or something like that 205 but anyway um you know the anthropic IPO could still benefit markets between now in let's say
6 months from now. The question is when when does that curve start turning? That's going to be your big red flag on the cycle because when that anthropic LLM turn starts, you're going to want to pay attention to that and hopefully
robotics have taken over before then. All right, cool. So, told us here? I feel like nobody else knows about this. >> Congratulations, man. You have done so much. People love you. People look up to
>> Kevin Praath there, financial analyst and YouTuber. Meet Kevin. Always great and YouTuber. Meet Kevin. Always great to get your take.