Accounting Mismatch in AI Profits | Jim Chanos and Val Zlatev on Long/Short Alpha in AI & Semis

There is a disconnect in the profitability accounting. The companies that are selling the picks and shovels are recognizing revenues and profits immediately. The hyperscalers and others who are spending those very same dollars are capitalizing those costs.

They're cautious.

Well, as Rick Rieder said in the past panel, previous panel, I mean, it is dominating everything in the financial markets right now. It's really a unifying concept when it comes to particularly the equity markets, but increasingly the credit markets as well. The one macro comment I'll make since we're at the Macro Minds is I think people should be a little bit careful about extrapolating much broader impacts to global economic growth and earnings growth. We took a look at the decade that preceded the introduction of Netscape in late '95, early '96, to the decade, the 10 years that had happened after Netscape, the post-internet era, and was not penalized by the GFC. It ended in '06, '07, and US economic growth was virtually the same in the decade before the introduction of the internet versus after the internet was introduced. And interestingly, corporate profitability, which I would have thought would have had a meaningful increase with productivity, did not increase its growth rate at all. It was 6% per year, which is the long-term average in the decade before the internet versus the decade after the internet. Now, of course, there's a lot of dynamism in that, and that's the long-winded way of saying there's going to be a lot of dynamism in the winners and losers in the AI economy, but whether it'll contribute to overall economic growth and/or long-term increase in corporate profitability remains to be seen.

But if you really want to be scared, I'll tell you that I'm net long AI versus my short.

Because that ought to terrify you all.

So we're, I mean, we, you know, I'm going to make one preface comment before I go into that. One of the other things you have to also keep in mind, and we did see this parallel in the late '90s, is when you get these type of capex, technology capex booms, and there's no doubt in anyone's mind, bulls and bears agreed, we're in a capex boom in high-tech. There is a disconnect in the profitability accounting between the companies that are selling the picks and shovels, in this case the chips and data center equipment and construction companies, what have you, and the companies that are spending those dollars. The companies like Nvidia and GE Vernova and Vertiv and what have you that are building out this giant capital-intensive business called AI are recognizing revenues and profits immediately.

The hyperscalers and others who are spending those very same dollars are capitalizing those costs. And that's a really important thing to remember when you're looking at the profit boom that we're seeing in the high-tech area right now. We saw this from '98 to '01. From the middle of '98 to the middle of 2000. The earnings peak in that cycle. S&P operating profits went up 30% over those two years. It's basically going at that clip, a little higher maybe even right now. And then does anybody know how fast it dropped from mid-2000 to 2001?

Dropped 40%.

Yeah.

When order books dropped and costs continued, and particularly depreciation, but order books collapsed and profitability of the S&P 500 dropped as much in that year, which was a mild recession, as it did during the global financial crisis. S&P earnings were down about 40% both periods. So we really have to kind of watch that. But what we're focused on, Jack, is what we think are inherently unprofitable business models that are attached to this AI ecosystem where any way you kind of look at it on a best-case basis, the returns on capital are going to be diminutive. So we would look at things like the Bitcoin miners turned data center developers, even the Neoclouds, where if you make just heroic assumptions on profitability and you give them 10-year life on the chips, you still get 4, 5, 6% returns on capital in the out years. And I just think that those are going to be winnowed away over time. I've joked with my clients that you want to own what the chips produce, not where the chips reside. I think that that's probably still a valid investment thesis going forward.

So there's two problems. One is that an awful lot of this capital spending for the people spending the big money like Alphabet and Microsoft and Amazon and Oracle, a lot of that is going into what's called construction in progress right now. And all of those costs, the cost of the chips, the cost of the labor, the interest cost, all of that is capitalized. It's not expensed until the data center comes online and is producing revenue. So that's an important thing. So for setting aside the life of the chips themselves, there's now increasingly because of lags, there might be 12 to 18 months where you've spent money on the data center, but it isn't producing revenue yet and you aren't depreciating those assets. Now as a cynic on this stuff, in order to be conservative, what we are using in our modeling is 10-year life on the GPUs. There's a misconception that the AI bears are saying, 'Oh, it's two years or three years.' We're using 10-year life, which is basically you running these things 365, 24 hours a day, you're not going to get physical life much more than 10 or 12 years out of them. So to be safe, I'm looking at business models and assuming that you write the GPUs off over 10 years. I think that's a safe bet.

It has changed the valuations of the Neoclouds. I don't know that their contracted prices have changed that much.

Because the hyperscalers aren't dumb themselves. And remember, in this business model, these are equipment leasing companies. In effect, if you are buying chips from Nvidia and then renting data center space from somebody else and then renting the chips out to Microsoft or Google or Meta, you're an equipment leasing company. You're not a high-tech company. You're a finance company in effect, and you're making a bet on the life of the chips and what you can get over the contract. But when some of these companies that are truly, and many of them are run by former finance people, the CoreWeave guys are the old Magnetar guys.

You remember them from the global financial crisis.

And so, you know, these are, if Blackstone is in your business, Blackstone just got into this business with a new REIT. You know, you're in the finance business. And that's a really important point to remember. And always remember that the hyperscalers can buy the chips themselves. They're choosing to rent them from the Neoclouds. Why?

Well, I think they're spending money on both, clearly.

They didn't spend money on both.

The amount of money that the hyperscalers are spending directly is massive.

But again, it's a gold rush mentality. So whoever has capacity will probably sign a deal. The problem, Jack, is that given that dynamic right now, you should be getting really good ROIs. If you have capacity of a powered data center right now and you have the chips or someone will bring the chips, you should get incredibly high returns on invested capital right now. This is if not now, when, right? And these deals where you get granularity on the deals where they actually give you quite a bit are working out, penciling out at 7%, 6%, 5%, 8%. They're all single-digit ROICs pre-tax. And so again, it gets back to my point. If that's the best you're going to do now, I'd much rather own other parts of the chain than just the middleman, if you will, the financial middleman.

Equipment leasing companies.

That's the value add.

They will stay, but eventually the value does move to the technology drivers. I don't think we're going to have power bottlenecks, for example, three years from now. I don't think we're going to have labor bottlenecks three years from now. We may have them for the next 18 months, but ultimately equities are long-duration assets, right?

And you should be looking at the core business over the whole cycle or over long periods of time and pricing an equity off current spot prices. And a shortage can be exciting and has been, but it can also be volatile.

Those are the legacy guys. Yeah.

Understood. But also keep in mind this is the technology space and technology can change and we could see inference going to our phones or to our desktops. I know that people say, no, no, it's not economic and never will be economic, but some of those same people are also then telling me that we're going to put them in space. So, you know.

Uh-huh.

What do you want to know about data centers in space?

Well, you're going to get a big chance next week. Because the only way that that thing works is if there's lots of data centers in space and pile drivers on the moon and colonies on Mars. But anyway, look, so the costs of putting a data center in space are obviously considerable, have a lot to do with launch costs. But a couple of observations knowing the data center space pretty well. Power costs are actually, despite the bottlenecks, power costs are a very small percent of data center costs. They're about 5 to 7% of revenues. So if you're doing this because the sun is a free source of power, you're starting on the wrong foot. And that power is not the problem. And in fact, I think power will be, as we discussed, less of a bottleneck going forward. So the big costs in space are radiation because it's hard to radiate in a vacuum. So the space station, for example, has these enormous radiators. So that's number one. Number two is radiation itself and complex systems exposed to space radiation over long periods of time tend to break down. But then you get into simpler things like the idea of redundancy and insurance, right? Like if stuff breaks in data centers all the time. If you look even at the simple old legacy data centers, their maintenance capex is through the roof. You know, stuff breaks, the HVAC goes down, this goes down, that goes down. Things are always needing replacing. So you send a tech out with the right part, the right equipment, they replace it and you're back up and running. In space, you got to send a launch, you know, hopefully with a humanoid robotic to do it. But you have another launch. And so you begin to get into issues of redundancy, insurance, and whatever. And then the cost, whatever cost savings you might be getting begin to immediately erode dramatically. And then of course there's the problem that the vehicle that the proponent of this that's coming public next week, their Starship hasn't made Earth orbit yet in 12 flights. I keep reminding people of that, that all these great promises are built on a rocket that has not yet achieved Earth orbit and has blown up, I think, six or seven out of the 12 flights. So we'll have to see. Obviously it's an amazing story. As I said, the TAM of space is infinite.

There's a lot of space there.

You know, so look, I mean, the numbers don't work on the existing business. Even with Starlink, Starlink is profitable. Starlink is a decent business. Its growth has slowed dramatically. They've had to cut price, and the prospectus points that out, to drive unit growth. But it's a profitable business. It's earning about $4 billion right now annually operating on, we think, about 25 to 30 billion of invested capital. So it's a good, it's not an insanely good business. It's a good business. The problem is the launch business still loses money.

Yeah, the launch business is still losing money after spending billions and billions and billions, and part of it is, you know, trying to get Starship to work. And also the launch business subsidizes Starlink. So you have to be a little careful. Starlink may not be as profitable as it says it is because it's getting cheap rates to launch from its parent. And then xAI is the wild card, right? It's, I mean, it's losing lots of money. It's spending lots of money. It cut a very short-term deal with Anthropic for space, for rental space. But it's just a sinkhole right now in terms of cash. So you have to believe in Mars and the moon and data centers in space to justify almost $2 trillion. I mean, it's like Tesla itself, right? Tesla, you can't justify on selling automobiles. It's all the stuff that's going to come. Like I said, bull markets, you know, put a premium on forecasts and bear markets put a discount on reality.

Yep. You're absolutely right. That should be taken with a grain of salt, especially knowing from whose mouth that's coming from. A couple of thoughts on — I think we should finish the '99-2000 comparison because that's an interesting comparison. You're absolutely right. That was like much slower growth than what MCI or whoever it is was talking about. I think —

Yeah. But exactly. Right now the graph actually you can track directly yourself by just looking at, for example, OpenRouter token counts, and you can see the growth of tokens being tracked, which is a small percentage of the industry token usage, but you can at least see it as opposed to waiting for some corporate CFO to show up once every three months to tell you something that he made up in the background. So you can track that much more directly.

And the reason the GPU rental price is going up is because the token usage is exploding and they just don't have enough GPUs to run the tokens through. This doesn't mean that will continue in perpetuity. We can discuss what can break that, but at least for the time being the real facts are suggesting that you don't need to listen to CFOs. You don't need to listen to some accountants to tell you what the growth is. You can just see it for yourself, and you can, by the way, you can see it in your own usage in your own offices. So that's one thing.

The other one is when you talk about '99-2000, there's actually two technology differences very different from right now. Number one, of course, all the spend was on fiber, right? Fiber into the ground by the Verizons that you mentioned.

Then Cisco routers and switches that you had to hook up to the stuff.

But PCs were fine. Let's put the PCs aside. Everybody talks about the fiber glut, right? The dark fiber. At that point in time,