원문 보기
Hello everyone, this is Kevin. So, there's a certain vicious cycle of failure in the markets right now, and it's almost entirely due to some uncertainty around CoreWeave and CoreWeave's debt. I mean, we've probably all heard about these charts and how scary it is: "Oh my God, CoreWeave credit default swaps are skyrocketing." And the reason people are panicking so much is because it reminds us of 2008, when people were trading CDS and these big crashes that eventually occurred when real estate credit default swaps paid out a hundred to one.
And so people say, "Yes, buy CDS." But there's an interesting thing happening here, and it's essentially a vicious cycle, and we're going to explain that in this video. But first, I want to add a little more context to this situation. So, if you're not familiar, CoreWeave is an inference cloud. About 26 percent of their active capacity is in inference, but 74 percent of contracted capacity is in rural third- or fourth-tier markets, which is very good for inference.
Okay, there 's a lot to sort through here. Let me try to explain this in more detail. Basically, if you are doing inference, you have the luxury of time. You don't need an instant Cerebras-style response, do you? So you go and build a data center in a more rural area , which is considered tier three or tier four .
Where the first is a metropolis, the second is a secondary market, more suburban, and the third, fourth is more land, cheaper energy, etc. But because they are further away, you have a longer delay with them. More inference means good for CoreWeave, because 74 percent of their contracted capacity is in markets suitable for inference.
Good. But here's the problem. Lately, with this agent-based AI phase that everyone's excited about, there's been a bigger focus on AI, if you will, like, "oh, let's invest in things that are useful for agent computing." This morning I was chatting a bit with my team about how much Mac Studio is needed to host the Meta Muse model, the latest model codenamed Watermelon, or also known publicly as Wonderful, on a Mac Studio.
And this is actually very interesting. In fact, it 's worth getting out the board because we can take a look at the scale of what Apple is doing here. And this will help us understand what is happening with CoreWeave and the debt death loop. So, this is a bit of a long introduction, but that's okay .
The first thing you should know: for a metamodel , even for agent purposes, if you want to run the latest and greatest model on it, you will need about 700 GB of RAM. That's quite a lot. Compare that to something like a quantized model that can be run on a 32GB machine like the 5090 and you'll be fine.
It just won't work here. But notice the difference for some of these new models. If I want to run a 700GB model, load it up, and have a spare to work with, I can do that with nine H100s. This is roughly what CoreWeave will be like. If you want to teach such a model—great. If you want to use it for inference—fine.
You can also use CoreWeave. This is a little less profitable because again you are targeting learning markets rather than inference markets. So, with inference, which is a big focus right now, how do I get this inside the company? Dell sells a lot of hardware that you can install in an office, but you can also use 22 5090 GPUs.
That's crazy. These are your gaming video cards. Right now, it will cost you $155,000 just for the chips, which is pretty crazy, right? And that will cost you around $220,000, or even closer to $ 300,000, depending on the price you can get. But this is where Apple comes into the picture. This is simply incredible.
I can run all of this on four Mac Studios, Mac Studio 256. And it's going to cost me about $40,000. The reason I mention this is because there's so much focus in the AI investment community right now on where can I, first of all, run models locally? Oh, that's good for Dell, that's good for Apple.
Okay, but where can I run agents? Well, surprisingly, you can run agents on very small computers. Many Muse models use only about 2GB of video memory, so they can be run on laptops or smaller devices, effectively on edge computing. And so CoreWeave finds itself facing a double- edged sword: agent computing and even large models that people want to have locally can now run on Mac Studio.
So then it turns out like this: okay, so we need less Coreweave for this. But on top of that, there's the concern that maybe we don't need to train that much , we just need more agents. So the border will matter less, and certainly H100 or even Blackwell may matter less. Let's get out of the Coreweave trade.
This is what some hedge funds and institutions are doing, and I think it is driving down the price of Coreweave. But Coreweave has more problems than just the fact that it relies heavily on training. They have balance problems. Their balance sheet problem is that if you add up the cash they have and the accounts receivable , assuming everyone pays their bills, we end up with $8 billion in cash and accounts receivable.
If I now add another $3-3.7 billion in debt that they've raised, they'll have about $11.7 billion in debt, or, excuse me, available cash if everyone pays them. But they need to pay $18 billion in bills over the next 12 months. So, they do n't really have liquidity in the near term. So now let's combine both of these things.
The shift towards Gen AI, a slight departure from training— all of this hurts CoreWeave, plus probably the worst balance among Neo Cloud. So what happens then? Then the stock price falls . Well, what happens when the stock price falls ? People get the impression that if the stock price falls, the company is at greater risk, and perhaps its debt is at greater risk.
So what happens then? People are hedging positions in the AI space using credit default swaps. So think of buying a CDS as an instrument that goes up when the stock or company falls or defaults. So, the worse CoreWeave's stock does, the harder it is for them to raise cash through a share issue. So you say, "Okay, no problem." "They'll just issue more debt ." But the problem is that the worse stocks do, the worse bonds usually do, because people think, "Oh my God, someone is pricing in a balance sheet problem." So the bonds fell in price by 20%.
And when bonds fall 20%, they're trading at about 80 cents on the dollar, then people say, "Oh my God, CoreWeave is n't going to be able to raise the money to pay its bills." They are already in a difficult situation. Now they simply won't have the money to pay their bills. So, in other words, it will be harder for them to sell the stock because the stock price is falling.
It will be more difficult for them to raise debt funds because bonds are becoming cheaper on the secondary market. We see this departure from the previous model. I still think we'll be doing training for quite some time, but we have a short-term shift away from training that is being superimposed on the move to using AI locally or to Gen AI computing that can run on smaller models.
All four of these things together show CoreWeave a big advantage. And this is very beneficial for people who bought credit default swaps on CoreWeave. Because remember: the worse things get for CoreWeave— and all four factors put it in a worse position—the better for this tool, the higher it soars into the sky.
But here's the craziest part. Are you ready for this? The higher this graph goes , the more companies like LSEG, formerly known as Refinitiv—I'm not sure if LSEG bought them, I don't know all the details, but anyway , LSEG says that this company now has a 64% chance of bankruptcy or default within 5 years.
I don't know if it's all that bad. Yes, I know their balance sheet is crap, but they are still above their IPO price. That means they will still be able to issue additional shares. So , I think what's happening is a so- called "death loop ": stock prices fall, CDSs rise, CDSs rise—bonds fall. When bond prices drop, people start to panic .
CDS are rising even more. Plus all the other factors I just mentioned, it just keeps growing and growing and growing. So when people see a line going up, what do they do? What does everyone do when the stock is skyrocketing? They buy it even more. And then the graph becomes vertical. And then people say, "Oh my God, we knew it.
This balance is really terrible." Or worse than we previously thought. We've already seen from afar that they have a bad balance, but this just drives the situation into a loop where people say, "Oh my God." CDS are rising because meme traders are driving them up. This probably means that things are getting even worse at the company.
And so you get into this vicious circle, where CDSs grow even more, and the more they grow, the more actively bonds and stocks are sold off. The more they sell, the higher the CDS become. So yes, this is where a company that probably doesn't have a 64% chance of going bankrupt looks to the market as having a 64% chance of going bankrupt.
I would actually take the opposite position. I would say that Nvidia, which owns 10% of the company, could lend them the $7 billion needed to keep them afloat , and they probably would. In fact, CoreWeave literally stated: “ CoreWeave’s position is strengthening every day, and I am optimistic. Their cost of capital is decreasing.” Good.
Um, their position may be strengthening every day as they increase their supply and computing capacity and connect their power supply. This may be true. But this quote: "Their cost of capital is declining." I don't know where you got that from, Jensen, because if the secondary market goes down, the value of your new equity goes up.
So that's not very good. But they hold 47.2 million shares. Next in line is Magnetar Financial, and this all sounds like we're writing the dot-com recession book again, does n't it? Um, I am, and don't get me wrong. I think this is all a "bubble " and it will all burst sometime in the future . These are things we constantly talk about during live broadcasts for course participants, how to prepare for this.
You can join us at reinvest1.com. But the bottom line is this: you get lifetime access to all the material, the real estate terminal or the reinvest terminal, which will have real estate, stocks, and research coming out at the end of this year. This will be cool. Plus everything else. Just go to reinvest1.com and you will see everything.
I don't need to "sell" it to you. You already know. But anyway, if you put it all together , yes, eventually it will all turn into a big bubble that will burst, collapse and explode, but I don't think CoreWeave is a "black swan." I think it's the labor market. People are convinced that everything is fine with the labor market .
Everything is fine now. But this won't last long, given how good agent-based AI is becoming. This wo n't last long. This is your "canary in the mine." CoreWeave? This is a secondary show. It's a vicious circle. It's like the opposite of GameStop. And I want to make it clear here. I have no position on CoreWeave.
I always like to talk openly about my prejudices. I don't care if it depreciates to zero, because if it does and rates go down, it will still help real estate, for that matter —that's where my bias lies. So, let it depreciate to zero. But I actually hold the opposite opinion and believe that it is not.
You know, Bernstein calls them the most vulnerable neo-cloud. Maybe that's true. At least on the public market. There are other neo-clouds that wanted to go public but postponed it because things are looking a little weird right now and people are saying maybe AppHarvest will postpone going public again.
But for me, it's not all over yet , and they won't necessarily go bankrupt. They have a bunch of other options for how to ask for a bailout and who to call before declaring bankruptcy. They will simply give away shares and dilute the capital of their shareholders. I expect exactly that. So I don't consider this a systemic problem.
I see this as an opportunity to buy on the dip. This is not the beginning of the end. And if you want to see us buy on dips or make deals, join us at reinvestone.com. We provide transparent portfolio reports and transaction notifications. With an alpha membership, you . If you're not yet familiar with agent-based AI, we'll be releasing entire modules on the topic early this year, and it'll be very helpful.
This will help you quickly figure everything out. So, yes, join us. You get lifetime access. This is great. And you'll never have to pay a penny for it again . This is just great. No regular payments at all. So . With you was .
Comments 0