If AI monetization gets pushed out and if it's not a 2027 or 2028 occurrence, but 2030 or whenever, then Oracle will have fundamental financial problems. You have a much riskier construction of the demand than you did in the telecom and dot-com era. The unprofitable companies as a percent of the spend is higher in this cycle than it was in the telecom cycle. As long as those unprofitable companies can keep raising money to keep paying those bills, that's fine. But if we get a credit crunch or we get a 2001-2002 pullback in sentiment, you're going to see a lot of that spending drop, and drop pretty quickly. It's just added risk. The risk levels for all these companies are much higher than they were. I think in a way that's worse than 1999-2000.
Today's episode is brought to you by Tukrium. Learn more about how agricultural commodities can play a role in your portfolio by clicking the link in the description. Today we've got a very special interview. I am joined by Jim Chanos of Chanos & Company, investor, short seller. Jim, welcome to Monetary Matters. The first question I want to ask you is: many of the investors in AI and in data centers and GPUs — what do you think are some of their widely held beliefs that they are convinced are true that you are doubting whether they are true or not?
Thanks for having me. So our view on data centers is somewhat bifurcated. We know the actual legacy data center business pretty well. We put a big short on it in mid-2022. And one of the things that struck us about the cloud data center business, the old legacy data centers, which are much smaller than the new ones we're talking about, is just what a crummy business it was. It was a very low return-on-capital business, highly capital intensive. The data center companies capitalize a lot of things that we believe should be expensed. They're negative free cash flow. I could go on and on and on, and we're still short those companies.
And so as the AI story has progressed, the latest hot area is data centers, including now a view that we're going to be putting data centers in space, which I sort of joked about a year ago. Of course, in this market, jokes become true. But what we really pointed out was that hosting GPUs is not a great business. It's basically a commodity business, particularly as everybody's building them. And our view to clients has been that the magic and the money is going to come from what the chips produce ultimately, not where they reside, and that if you're going to make a bet on AI, you should make it on either the pure AI companies like OpenAI, xAI, Anthropic, or you should bet on the hyperscalers — and more on that later. But investing in the Bitcoin miners that are converting to data center companies, or colo companies that are jumping in, or just the so-called neoclouds, which are basically just landlords, to me seems problematic for two basic reasons.
Hosting GPUs or CPUs is inherently a low-margin, low-return-on-capital business, number one. And number two, if you're actually buying the GPUs yourself, you have to make a bet on depreciable lives, which is one of the big controversies — how long until you have to replace the GPUs because your tenants want something bigger and better. And we can argue about that. I have my own views. We use a five-year life with 20% residual value. It could be six years, it could be three or four years. Time will tell.
So if we had to say there's three different kinds of stacks of this AI: there's the hyperscalers who are selling things that are produced in the data center; there are the companies that sell things to and make the data center, Nvidia being the most obvious of this; and then the third are the companies that own the data center. It sounds like a lion's share of your skepticism and perhaps bearishness is concentrated in that third sector. How much would that be concentrated in, let's say, the S&P 500? Because a lot of them are the hyperscalers, which is the first category, as well as Nvidia, which is the second.
So when we do our model short portfolio for clients, we advocate that it always be hedged, and that's how we've run our fund business since '96. So I always joke I'm sort of long those companies because we own the indices, and they just make up such a huge part of the S&P and the NASDAQ. And so in effect, I'm advising people to be long Microsoft and Nvidia and Meta and Google. So it's really, as you point out, the companies — sort of the second or third derivatives of the companies that are trying to either recast their business or build a new business around owning and operating the data centers. That's the area we know well and we think is not a particularly good business.
And where does Oracle come into this?
Oracle is kind of a hybrid. They're a hyperscaler, and of course they have a legacy software business that's profitable, but if you look at the five hyperscalers that are building out data centers — which are Microsoft, Meta, Google, Oracle, and Amazon — they are one of the two that don't earn their incremental cost of capital. And that includes the numbers that were released last night. So in their zeal to catch up to the others, they're expanding their balance sheet really, really quickly, but don't have the income and cash flow levels that someone like Meta or Microsoft do. Meta and Microsoft, by the way, according to our work, are the only two hyperscalers that are, right now anyway, earning above their cost of capital on their incremental capital spending. And that can change obviously as AI output hopefully gets monetized in a better way down the road, but Microsoft and Meta can easily handle their investment load. Whereas Oracle and, interestingly, Amazon are at the bottom of that stack in terms of their inability to at least monetize yet their massive investments in AI equipment and hardware and locations.
So tell me about Meta. I had it referred to as a hyperscaler. Because I thought that meant, you know, providing clouds to other people, I thought Meta was just spending all this money to improve Instagram and all these other things — or are they selling the data to other people?
Well, they're integrating AI and building out massive data centers, including I think the big one in Louisiana, if I recall, with the controversial accounting, the SPV accounting. So they too are building out, most likely for their own internal capacity, their data center capital plan.
And again, they're monetizing things through, in effect, selling ads to Instagram and Facebook users, and Microsoft is doing it through their corporate customers. So at least for the time being, they're the two that are able to sort of finance this from their own cash flows. The others are going to need external financing.
And that is a remarkable statement, Jim, because for so long these Magnificent Seven companies have just been so profitable and have grown so much with so little capital requirement. Now they're just chewing up capital and their free cash flow is steadily going down quarter over quarter. Are you envisioning a future where some of these companies' free cash flow might go negative, meaning basically that their operating cash flows are lower than the amount of money they need to invest?
Well, that's already happening with Oracle. So they've gone free cash flow negative. I think it will erode. And again, the idea is that AI monetization from output will hit an inflection point in 2027 or 2028, and then profits will flow. We'll see. I don't know, and I don't think anybody else really knows. We'll just have to see. That's the bet. The bet is a big one, and for some of them it's a bet-your-company bet, which they never had to do before.
The underpinning of your question is correct in that you basically had asset-light businesses based on intellectual property that just gushed cash flow, and now you have basically, you know, the second industrial revolution, if you will, where the next round of technology improvement is really capital intensive, and that's a change.
And how are you assessing that return on invested capital? You can sort of pick a company, but let's say 10 years ago when Microsoft was investing in building data centers for its earlier cloud clients, I presume the return on invested capital was quite high. What does the return on invested capital in this new AI data center world look like?
Most of them don't break out their actual pure AI revenue and costs, and you can only look at the aggregate and just make some assumptions about their spending and look at their stated returns. And it gives you an idea of magnitude and direction. It's not going to be precise. And Amazon would be the better example, I think, because AWS very early on was unprofitable for a number of years until it hit its stride about 10 years ago. And so the bet might be a correct one, we'll just have to see. And so what we're looking at right now in the analytics I just discussed is...
...what is the increase in adjusted operating income year over year, and then annualize that, and then what is the increase in the capital base year over year — with the increase in operating income annualized as your numerator. And how is that trending? Is it going up or is it going down? So for example, last night Oracle reported, and their incremental increase in operating income divided by their incremental increase in capital year over year was about 8.5%. That's below their weighted average cost of capital. It's positive, but basically it's telling you so far they're destroying value. Now, for Microsoft, that same number is almost 40%. So it gives you a pretty good idea of so far who's kind of winning in this and who's spending a lot of money for not a lot of return.
But again, like Amazon, that can change. So we'll just have to see and keep monitoring it. As I like to say, predicting the future is hard. People are doing it. The difference now is they're committing a lot of dollars to these projects about predicting the future, and that's a new risk.
Right. And 8.5% is positive, as you say, but doesn't sound terribly high. I think long-term Oracle bonds have a 6.5% yield. So there is a spread to harvest there, but I'm sure with cost of equity and stuff it's...
With cost of equity, it's above that.
Yeah. And that is what the return is for Oracle right now, where things are white hot and all of this money is being raised in venture capital to back companies like OpenAI and Anthropic that are the customers of these things, which is why the money is so good right now. So it could look worse when things cool down a little bit.
Yeah. So that's the sort of left side of the tail, right? Is that these returns are all based on, in many cases, selling to unprofitable companies, and as long as those unprofitable companies can keep raising money to keep paying those bills, that's fine. But if we get a credit crunch or we get a 2001-2002 pullback in sentiment, you're going to see a lot of that spending drop, and drop pretty quickly. So it's just added risk. The risk levels for all these companies are much higher than they were five or 10 years ago.
So of the giant companies in the S&P 500, it sounds like the ones you have the most concerns about are Oracle, Amazon, and Meta. If you had to rank those three from least concerned to most concerned, which would it be?
No, we would be most concerned right now about Oracle in that whole group, and that's pretty much our only negative view on the whole group is Oracle, because their spend plan is just so enormous relative to returns. Amazon has a lot of other businesses, and we think that our analysis of Amazon is skewed a little bit. And the rest, in effect, as I say, we're long through our hedges. So one of the only hyperscalers, of the mega caps, that we would be concerned about right now is Oracle. But we're short lots of others, I'd say secondary and tertiary plays.
And yeah, Oracle stock over the past few months has been suffering, and its credit default swap for five years has widened out significantly. How much of this do you think is a potential — you earlier said bet the company — a potential existential question? Because okay, if Google or Microsoft is going to waste hundreds of billions of dollars, ultimately they will be okay, but Oracle is borrowing all this money and it's investment grade, but for a long time it hovered on that investment grade line. I mean, do you think that there could be some substantial credit issues for Oracle?
I mean, if AI monetization gets pushed out and if it's not a 2027 or 2028 occurrence, but 2030 or never, then Oracle will have fundamental financial problems. It's not at that level yet, but certainly they will be under a lot of stress if they continue to borrow money to do their buildout as opposed to use equity. I have to think at some point they might have to reconsider and start issuing equity to keep the balance sheet in check. So far it's been mostly done with debt.
And what fueled the surge in Oracle shares during the second quarter, in the early summer, before the quite drastic correction we've had — what fueled that surge was, I think, it reported earnings and its remaining performance obligations surged from $137 billion to $455 billion. Which some people think of as, 'Oh, this is future demand and this is what's going to be revenue and earnings in the future, so it indicates incredible demand.' What do you think about that number? How concentrated is it among one customer, mainly OpenAI?
Yeah. It's not a GAAP metric. It's not a real backlog. It's, you know, if everything goes right and they continue to deliver, our revenue should be this over the life of the contracts. I mean, again, the stock reacted positively to it, but again, the data center and hyperscaler and AI stocks were all reacting in the second quarter to those kinds of announcements. It wasn't just Oracle — I should make that clear. Any company that came out and said, 'Oh, we're going to build data centers,' went up 50 to 100% during that period. So I look askance at the RPO number, and I think investors should give it a wide berth.
All right. Now let's talk about the lesser-known, smaller companies, the so-called neoclouds. These are CoreWeave, Nebius, former bitcoin miners like IREN. Tell us, what role do these sorts of companies play in the ecosystem, and why is your concern and skepticism concentrated in this area?
Because again, these are the guys that are hosting the GPUs, not benefiting from the GPU output. So it gets back to my first comment to you, that we think the money and the magic is going to be made from what the chips produce, not where they sit. And I just think that hosting GPUs or owning them and renting them out to third parties is just not a particularly good business. It's a commodity business, and these are landlords — they're not tech companies, and I think that's the basic difference here.
And with everybody rushing to build out data centers, it's a supply and demand story, right? You're going to have more and more people who will be having facilities. There seems to be this canard out there that, well, if you have land and power, you have a monopoly or you have an oligopolistic situation where you can charge ridiculous rents and nobody else can get them. And all you need to do is look at IREN's Microsoft deal, which was trumpeted at the time as this great deal, and if you sort of drill down and analyze the returns, it was a mid-single-digit return on capital deal. And so, if that's what a monopoly-type provider gets because they have access to power and land right now from a really good credit — I mean, yawn, who cares? And so I think that's the core of our negative view on these neoclouds and Bitcoin miners, is that they're all jumping into a business that's basically a commodity, low-return, ultimately business.
Tell us — so you're very familiar with the old-school data center stocks. I don't think you mentioned them, but they might be stocks such as Equinix, such as Digital Realty Trust. You said earlier that they capitalize expenses that should go through the operating line, and they overall earn low returns. So how would you estimate overall the returns of that business — not the stock price, but just the overall economic returns of those two publicly traded companies, as well as the privately traded, private equity landlords that are quite big in this business too?
Yeah. So again, the Digital Realties and the Equinixes are just colocation businesses, basically. Equinix has an interconnection business, but for the most part these are just simply companies that provide the real estate, provide the backup power, provide the power itself, the cooling, what have you. And you could just look at the stated returns, and that's based on a 15- to 20-year depreciable life for the assets. And using that figure, their returns are really, really low. They're low single digits for Digital Realty, mid-single digits pre-tax for Equinix.
But what really gets interesting, and the reason why we're still short them, is if you actually look at the cash flow statement, you'll see massive amounts of capex that continue, and it's not new bills. Most of it is for the existing data centers, but they have this little accounting game they play because they're REITs where they claim that 90% of that spend is growth capex and not recurring capex, and everybody sort of says, 'See, don't count that capex because it's for growth.' Well, the accountants let them call it growth capex. If it's in an existing system and you can either acquire a new client or raise prices on existing clients, then you can categorize it as growth capex, not recurring capex. But if the HVAC system goes down at a data center, it goes down — you have to replace it. How you categorize it is sort of semantics. And if you can persuade your auditors that, well, we'll be able to raise prices and maybe get one new customer by putting in a new HVAC system, you know, it's growth capex.
Meanwhile, these companies aren't really growing. They're growing revenues mid-single digits. They have 20% vacancy rates. It's ridiculous. And both Digital Realty and Equinix have capex greater than EBITDA. So they're financially stretched as well. So they're in effect borrowing money and issuing stock to pay interest and dividends. And that's just never a good position to be in.
And is the reason that these companies kind of exist because the large companies, the Microsofts of the world, basically don't want to have this on their balance sheet?
I think that's a really good point. Yes. The answer is that it's excess capacity, and in the cloud world you might have the right geography because of the need for clients to access data quickly — latency. So that does matter more so than in the AI data center business, for example.
However, you raise a really good point, and that is that increasingly we see with Microsoft's announcements, with the Meta announcement and Blue Owl and their financing, more and more of the hyperscalers want to get these assets off their balance sheet and lease the capacity from somebody else and let somebody else take the actual capital intensity risk. And I think that's a really interesting point that we've made to clients, is note how many of these companies are trying to not have this stuff on their balance sheet.
So let's now talk about CoreWeave, which I think is the biggest and the most indebted of these so-called neoclouds. Is their basis, their business, owning the data center, or what else is their business other than that?
Their business is not owning the data center. They're like a virtual — they're leasing the data centers from others and owning the GPUs.
And then selling that compute to the clouds?
To the hyperscaler, yeah. All right. And what do you think about the returns on those businesses?
Yeah. So that's a bet strictly — if you're basically the GPU middleman, if you will, that's just a bet on depreciable lives. And I think that's what you need to — that's the bet: that you're buying these things and leasing them out, and that five or six years is actually too conservative, that you'll be able to earn money on them for 10 years or 12 years. And that's the bet you have to make if you're a CoreWeave investor. The bet there is that those chips last for a long time, so that every year you're not depreciating that much. If the chips last for 10 years, you only depreciate 10% as a cost every year. But if they last for three years, you have to depreciate a third of what you spend, which is a lot.
Right. Right. And again, the question gets back to your comment, and that is: if these chips last so long, and Microsoft is depreciating its chips over six years, if these things last longer than six years, why is Microsoft leasing the capacity from you? That's a good question. So explain what the depreciable lives are that the CoreWeaves, the Microsofts of the world are putting on the chips that they're buying and then depreciating, and what do you think is a more appropriate depreciable life?
We're using something similar to them in our modeling. So they're using six years as I recall, and a number of the hyperscalers really increased that lately from five years to six years. We're using five years with a 20% residual value. So the numbers are somewhat similar ballpark, and again, I'm just looking at rental rates for GPUs and other things to get some third-party verification on this. There's an index on Bloomberg quoting spot rental prices. I think it's down 28% year over year currently. That's—
That's the Hopper GPU rental index.
Yeah. Yeah. And that's one check. You have some of the chip companies themselves, CEOs have said five years, six years seems right. People believe that if you use them 365 days, 24 hours a day, that the physical life is not much more than six or seven years. They only last 10 or 12 if you're not using them at 80 to 100% capacity. So we'll see. It's one of these things where we're going to see. I would not want to be betting my company on an 8, 10, 12-year life on these things. I think that's an asymmetric bet that you're going to lose on. And conversely, if it's two or three years, then all these companies are in trouble. So we'll have to see. It's something obviously we and other investors are monitoring pretty closely.
But again, that's the bet there: that simply they're smarter than everyone else and have figured out that these GPUs have incredibly long lives. And they said something on their last call about, well, we're still renting some of these chips out after the term is over at basically 90 to 95% of the last deal. And I'm highly skeptical of some of those comments, whether they're cherry-picked or whatever, but there'll be enough GPUs out in the marketplace within a year or two that we'll have a pretty good idea of what rental prices are.
Right. I don't really understand the technology at all, but Nvidia is just such a dynamo in terms of improving their architecture basically every year. So there was Hopper, and then this year we have Blackwell from Nvidia. Next year we're going to have Rubin, maybe the year after that we're going to have Feynman. What you were talking about on Bloomberg is the Bloomberg Hopper GPU rental index, so this is the old Nvidia GPUs, and it shows that the cost to rent these has gone down by so much. So what you're saying doesn't in any way contradict the very rosy outlook of Nvidia transforming the world with AI. I mean, it sounds like you kind of agree with that, but because they are so transformative, and year after year just the old GPUs are not that exciting and valuable anymore, and therefore the people who buy those should depreciate it at a faster rate and a higher depreciation cost. That's what you're saying, right?
And again, when CoreWeave or others say, well, we'll simply use the old chips for other tasks like inference and not training — that's great, but that should reflect itself in rental rates, and so we will have a check on this from the free market as to how valuable these chips really are three years out, five years out. So far, we're seeing depreciation, reductions in rents, consistent with a sort of four- to six-year life.
Talk about how indebted CoreWeave is. How is it financing this incredibly capital-intensive growth?
So they're doing what the Bitcoin miners are. They're using lots of converts. And I kind of laugh because when I see people throwing bricks at me on X for my comments, I often see the same people saying, 'Oh my god, they just did a 1% convert deal. It's like free capital.' And not understanding that converts for some companies can be really, really expensive when you factor in the option value that they're giving up.
A convert for your listeners is composed of really two components, right? A call option based at some premium price, the conversion price, for typically five years, seven years, whatever the term of the bond is. And then the remaining part is in effect a very low coupon or zero coupon bond that will mature at par at the maturity date. So once you get the option value of what you've given away in the convert, you can figure out the yield on these things, and they're quite high. The cost of capital is quite high when you factor that in. But retail investors seem to think the opposite, that this is just 1% cost of capital because that's the coupon on the bond.
And if you knew for a fact that the stock would never hit the strike price, maybe it would be a 1% cost of capital. But if it hits that strike price, there's going to be a very large dilution, which is basically just issuing shares.
