Thursday's Diet TBPN

0:00There is a bunch of breaking news, the big news out of OpenAI. Of course, the OpenAI show continues. Sarah Fryer had some comments about ChatGPT's growth potentially slowing, and it's unclear. It's not in decline. It's maybe deceleration. We'll have to dig into that. It had me thinking about debt, and I was thinking about just the fact that the debt has come to tech for the first time, really. And this was sort of my take, and I'm a little bit, this is an area that I know the least about. And so I was doing some research, learning about a different industry since it's just so abstract. We keep going back and forth on the debt is coming to tech narrative as like, it's very scary.

0:46We live through the global financial crisis. And there's a lot of jitters when debt is around. It's like, oh, you could get wiped out. You could blow up. The backstop comes in. It just feels like all of a sudden we're talking about things with a much more serious consequence than like, oh yeah, a startup raised some money and it didn't pan out and it was a zero and it wound up being a write down. Even when Theranos blew up, it was only equity holders that were lost. It wasn't this entire industry and it didn't turn into this systemic issue. But now it feels like with the 1.4 trillion of backlog that OpenAI has kind of opened up across a whole bunch of different deals.

1:30There is this worry that maybe the level of indebtedness could be risky, the level of risk in the system, the level of investment in the system could be something that's bigger than just, oh, if you're in this one name, you're taking a big risk. Now it's maybe like, hey, we're all taking a risk. And if we're talking about backstops, at least. Data is the new oil. And back then in 2006, his point was, he was working as a data scientist at Tesco, which is this British grocery store. I don't know if you know this story. But he was working at this British grocery store chain. And his point was, we have all this data on a customer is in the rewards program.

2:10We see that they buy a Thanksgiving turkey before Thanksgiving. We see that they buy this type of paper towels or this type of milk or whatever. We have all this data, but we don't really do anything with it. The data is not valuable. We need to refine it, much like oil, into gasoline. And once we refine it into gasoline, then we can do things like targeted advertising. And so it was basically just a generic call to action for taking data science seriously. People have been saying data is the new oil for, I guess, two decades now. And it never really sat that well with me because unlike oil, data is not perfectly fungible.

2:44So one tranche of data is not equivalent to another. Like Reddit it is clearly very valuable since it provided the backbone for GPT-3. All the analytics data that flows out of some mobile game is basically useless. A lot of data is worthless. A lot of data is worthless. All oil has at least some value. Essentially. I mean, I guess there are different levels of crude, right? There are different grades. And I was actually trying to play out the metaphor more, and I was wondering, like, Like, can we get to a place where we can wring intelligence out of raw data, like the oil, and the result can be low-octane gasoline, kind of like mid-wit level, like slop, an AI slop.

3:26Or it can be jet fuel, like a deep research report that's actually pretty great, or some code that's really reliable and really useful. But it all depends on the processing methodology. But the more interesting data is the new oil take that I don't think was considered in 2006 is that maybe the tech industry is going to look like the oil and gas industry soon. Like I was looking up how much debt is in the oil and gas industry. It's over a trillion dollars of debt. It was two trillion like, you know, a decade ago. And then it went down and then went up. And it's like, it's all just a function of like how much oil and gas is going on.

4:00How many, where are the new projects? How big are the projects? How much debt goes in? The difference is that if you identify oil in the ground and you figure out how much it's going to cost you to extract it, and how long you think you'll be able to, like basically estimating how much oil is actually available in this site, then you can lend against that pretty predictably because you know that the price of oil is going to fluctuate. But in general, as long as it's in some range, it will be like a profitable operation to pull it out of the ground. And I think it's a little bit easier to lend against that than GPUs today.

4:37The big debate is around depreciation schedules. And we have a sense that a data center that has power and basically a box with a lot of power will be valuable in the future. But if a lot of the cost of a new data center is GPUs, it's harder to gauge on what the value of those GPUs will be in four years than it is. Okay, will this oil, like production site, still be producing oil in five years? I think that's a bit easier to answer and easier to lend against. The fact that you're walking through that math is very different than what the venture capitalists in 2000 were doing. Like Google was like the most pure play, just beautiful software business.

5:22So Google from 2001 to 2004 grew from$86 million in revenue to$3.2 billion in revenue. And net income over that period went from$10 million to$400 million. and that includes stock-based comp. So they were still making$400 million in profit with the stock-based comp. Googlers made a lot of money. They gave away a lot of stock. And so it didn't look like an oil business. There was not this big CapEx build out. There was not this big, or even this crazy R &D phase. There wasn't that much capital that went into Google before it became this monster cashflow machine. It was sort of an infinite money glitch.

6:01It was this beautiful algorithm that was just discovered and it was so elegant and it just produced this monopoly insane like growth rate for so long but for a long time like tech just meant take a bet on a company and it's either a zero or trillion dollars or something like that and so it's a lot different and i wanted to dig into like the actual structure of one of these deals you remember the hyperion release zuck went on threads and announced that he was going to be building a five gigawatt data center it's going to be as big as It's like somewhat of a Manhattan project. Somewhat of a Manhattan project, exactly.

6:35The crazy, crazy thing about that deal. So he spins up the, he puts out the announcement post on threads, says, hey, we're going to build this five gigawatt data center campus. It's gonna be online in a few years. It's gonna be as big as Manhattan. And he shares some of like where it's going to be, how many racks are there gonna be, square footage, stuff like that. But he's basically just announcing that like, hey, the project's financed, we're ready to go on this. Like you would expect that when that, it's a$27 billion deal. You would expect that, okay, Meta went down, they spent$27 billion, it's worth it, they're gonna.

7:13No, they got paid three billion. They got paid three billion. And the reason is because Blue Owl financed it with external debt and they are basically paying Meta upfront for the right to have them as a tenant, as a leaser, for a very long time. And so you have this massive data center project that's going to be paid for, even if it's not producing any valuable tokens, Zuck's still going to, he's not just going to default and be like, yeah, take the company. No way. He's going to pay. And so in exchange for that, they got 3 billion up front. This feels deeply important to the current AI build out boom, the tech story.

7:48It feels like an entirely new piece of the puzzle to understand where this technology is going. And I don't feel equipped to understand it at all. Private assets star, Blue Owl has been flying high. Is it too close to the sun? So the article says, suddenly Blue Owl Capital is everywhere. This past Tuesday, the upstart alternative investment firm with an aptitude for private credit announced a financing deal for Meta Platform's$27 billion AI data center in Louisiana. That is Hyperion that I was mentioning earlier. The week before at the Pact CAIS Alternative Assets Summit in Los Angeles, Blue Owl's co-CEO Co-CEO Mark Lipschitz called JP Morgan Chase's CEO Jamie Dimon cockroach warning about risk and private credit An odd kind of fear-mongering.

8:33So basically private credit has been growing a ton We've talked about this a few times Aries is massive now Blue Owl is really big and and there's basically been this little bit of a fight Emerging between where the debt is coming from do you do private credit or do you go with the traditional bank route? And so Jamie Dimon, at least I'm pretty sure he's going head to head against Blue Owl in a bunch of these deals. Jamie Dimon was cautioning investors about potential risks in the credit market by invoking a proverb. When you see one cockroach, there are probably more. And so he was referring to recent loan defaults, such as the bankruptcy of auto parts maker first brands and subprime lender Tricolor Holdings, as warning signs of broader credit issues.

9:15So, Diamond noted that J.P. Morgan took losses on some bad loans and implied that trouble in one corner of the credit market could mean undiscovered problems elsewhere, implicitly casting doubt on the booming private credit sector. And so, Mark Lipschitz fires back and he says, I guess he's saying that there might be a lot more cockroaches at J.P. Morgan. During a recent private call, OpenAI's investors asked about external signs that Chachapiti's growth is slowing. CFO Sarah Fryer. The external signs were, I think, like App Store data. There was some data out of Europe. Oh, yes, yes, yes. That's right.

9:50That's right. And it was hard to read into the European data because Europeans. They don't work ever. No. I mean, it was coming off of summer, right? And, you know, ChatGPT is a popular student. But European summer hasn't ended yet. There's been early warning signs. Yeah, walk me through some others. So I can walk through Alex's coverage. Sarah Fryer held a private quarterly earnings call with the company's biggest investors. As usual, the numbers she shared were mostly up and to the right. But behind the strong top-line figures, a quieter question hung over the call. Was ChatGPT's momentum starting to slow?

10:24During the Q &A portion of the call, sources say Fryer was asked to reconcile ChatGPT's meteoric growth in weekly users from 250 in September 2024 to over 800 million now, with external signs that the app's growth has slowed in recent months. Close followers of OpenAI's business have been whispering about these signals from research firms since late summer, but this was an opportunity for company backers to hear directly from leadership on the matter. After telling the investors to take third-party estimates with a grain of salt, Fryer acknowledged a chink in ChatGPT's armor. She said time spent had declined slightly in response to, quote, content restrictions the company rolled out in early August.

11:04She then referred to the loosening of those restrictions that CEO Sam Altman has said will be implemented for adults in December. And Sarah says, And so I don't think them announcing that they're getting into erotica is a sign of strength. Felt like something that they would do in order to stimulate growth while they get a bunch of other monetization online, right? So like commerce, ads, et cetera. The reason I reacted strongly to it was that there had been messaging, you know, around the same time of, I don't want to be in a world where we have to decide between curing cancer and free education for the world.

11:46Yes. And so then at that same time, deciding we're going to do erotica. It was very weird timing. It was very weird. It was very weird timing that those two statements like came out one after another. OpenAI has decelerated revenue before because they, I think they tripled and then they went to a doubling or they were quadrupling and then they went to a tripling. And so they actually decelerated in 2024 and then they re-accelerated in 2025. And so I was kind of saying like, well, there's a good chance that you could see deceleration in the future. It's happened before. To be accelerating forever is basically impossible.

12:23but it would be interesting to track exactly how how chat gpt's growth is slowing there certainly feels like there's just a level of saturation in a month over month change in total visits to leading gen ai tools chat gpt is at the bottom of a list that includes gemini deep seek perplexity grok claude copilot and meta ai the key difference here is that like ChatGPT is just so much bigger than these other platforms. They could still be adding more users on a per user basis than these other tools, even if their growth is slower. Like, Meta is not accelerating top-line users. They have, like, 3 billion users.

13:11Like, no one's expecting them to accelerate top-line users. Maybe, like, randomly, one quarter they accelerate, but not continually. And so, I don't know. It just feels like an odd thing. It's going to be very funny when LLMs plateau around 120 IQ, and what we've created is just a digital guy, not a god. This doesn't make any sense. If we have infinite digital guys, that's literally a guy is just a worker. If we have infinite workers, that's insanely bullish. He didn't say it's gonna collapse the economy when we just get a digital guy. He said it's gonna be funny. And I agree. If you get a digital guy, that's pretty powerful because guys can do a lot of stuff.

13:52You need a guy for everything. The middle class has apps. The wealthy have guys. Tyler, did you get a chance to read Fiji Simo's latest blog post, moving beyond one-size-fits-all? Yeah, nothing, I would say, super substantive in it. I think with 5.1, they were going for a lot. Like, we made our digital guy faster, better. I mean, it is crazy following this company so closely because in here there's a line that says, with more than 800 million people using ChatGPT, we're well past the point of one size fits all. And 800 million sounds amazing, except I feel like I heard the 800 million number like two months ago, and I feel like they have been accelerating so fast.

14:31You would expect them to be at 900. Exactly. And so the fact that they're repeating the 800 number is like, what's going on? They're like, sorry, we can't add a third of the United States every month. ChatGPT is officially in its Fiji-SIMO phase. If you're wondering why the upgrade doesn't come with benchmarks, have fun. Rune says, you are confidently wrong about the internal dynamics of this. It could be better summarized as an infra cleanup. And Nir says, the source for my top tweet is Fiji's blog post from today, which discusses the release and its goals. I don't really know what else to say.

15:07Is there hunger for benchmarks anymore? I might actually take the other side of this here. I like that they're getting away from benchmarks. I wish they didn't do a 5.1. Just make it better and don't do a release. And certainly don't tell people because what if people imagine - Easy for you to say because you're not in love with a specific version, John. I'm in love with 5. I'm in love with 5, Rune. Bring back 5. I don't like 5.1. When GPT-4, like GPT-4, like not 4 or anything, when that was the best model, they would do updates. They wouldn't say, oh, this is a new model. And people could definitely tell.

15:41Oh, sure, sure. Okay, they released the model, it's worse. So you think putting a version number actually helps fight back against that? I think it's more, it's just easier for people to tell it that it was actually a change when they're noticing something that they've been depending on. Yeah, yeah, yeah. Sources say that sources say. Okay. Sources says that sources say CEO Mark Zuckerberg joined an internal employee Q &A and shared a warning about the AI bubble. First, he shared a breakdown of how different players from startups to big tech names like Meta should think about timing their bets.

16:13He described three camps in the industry, optimists who see superintelligence emerging within two to three years, moderates who expect breakthroughs by the end of the decade, and pessimists who think it'll take well into the 2030s. Each outlook, he said, dictates how aggressively a company invests. Then he expounded on a version of the answer he gave me recently in our last interview. He noted that while unprofitable startups like OpenAI and Anthropic risk bankruptcy see if they misjudge the timing of their investment. Meta has the advantage of strong cash flow. He also made the point that while big tech has historically been relatively debt-free compared to large companies in other sectors, the AI infrastructure race is leading Meta and its peers to start using leverage in a more normal way relative to their size.

16:56Like he told me in September, Zuckerberg acknowledged to employees that Meta's market cap could suffer if his timing is wrong in the bubble bursts, but the message was clear, we'll have the balance sheet to survive and emerge stronger than most on the other side. SuperDario was quoting that and said, the obvious end game in the next two to three years is that Microsoft acquires OpenAI, Google acquires Anthropic, and Tesla acquires XAI. Only the large caps survive. That's a nuclear hot take. In contrast, OpenAI employees stayed for two plus years, sold$6.6 billion of equity last month. Many hit the$20 million cap.

17:31Morale and vibes are high, but so is the turnover rate. New OpenAI hires are often shocked by how many Slack accounts get deactivated each day. There are dozens or perhaps a couple hundred OpenAI, XAI, Google, DeepMind researchers founding companies in the current climate. And this was talking about. The simple answer, the liquidity of anthropic options is the worst among those frontier labs. This is talking about how a lot of people have been leaving various labs. Less people have been leaving anthropic. And so Liang Chen is saying the simple answer. What would be interesting is if companies started offering liquidity in the form of annuities.

18:13But I don't know. There's some way to deal with this. Like, you know, if you don't get employee liquidity, like they'll leave for something else. They'll just go somewhere else that pays them a higher salary. If you give them too much liquidity, they'll leave and start new companies. very very tricky to manage the manage the team but that is the nature of these these companies one of my strongest beliefs is that it's going to take 20 plus years to get AI penetrated into the real economy I filled out a piece of paper at the doctor's office last week I filled out a piece of paper the doctor's office last week too I finally realized why DocuSign has so many employees because you need to go to every doctor's office in person apparently for decades to get them to use online form-filling technology.

18:59Like general SaaS really does not, has not permeated as much of the economy as people think. There's a company that makes paper receipts that's worth$20 billion,$20 billion. There are fax machine companies. The fax machine industry is still over a billion dollars. It's still a billion dollar industry. In my opinion, the entire AI field switched from explore to exploit two years early. Everyone convinced themselves, no, this isn't the case. Look at our exploration. and it's like watching someone go on a 50-foot walk and find a cool tree when the entire continent is still covered in fog of war. Now that the terrain seems known, it should be harder to convince yourself.

19:36I suppose this makes sense given a lot of people hint at being good as gone as soon as they have enough money. But no, not me. I've been gone for ages already. Weren't we talking about this yesterday, this idea of like, where will the next innovation come from? Where will the next breakthrough come from? Will it come from any of the, like, will it come from XAI? Will it come from DeepMind? Yeah, I mean, it's really tough right now. You can stay in a university system and be a student and be taking on debt, or you can go work at a lab and have a good shot, at least if you did this a few years ago, have a good shot of making$20 million in a few years.

20:16And it's hard to give up that kind of opportunity. This Wall Street Journal article has given more context on the AI boom. says the AI boom is looking more and more fragile. AI stocks have swung downward as doubt rises about sustainability and payoff. Perfect isn't good enough, and any sign of weakness is a disaster. This is what's happening. It's like you double revenue and your stock trades down. It's very, very odd, but everything's been priced to perfection. CoreWeave, who again is the only neocloud in the platinum tier semi-analysis, is down 45%. Recent history suggests that the gloom won't last, but the shakeup serves as a strong reminder that the early years of AI pose a challenge for investors.

20:56Companies at the heart of AI are now talking about years, plural, of all major investments still ahead. There is, of course, real reasons to worry about the sustainability of the boom. Chief among them is that there is far more AI computing infrastructure spending than there is AI revenue. A gulf widening by the day. OpenAI says it's planning to spend$1.4 trillion dollars in the next eight years but is only pulling in around 20 billion of annual revenue today has the mood become that ceo sam altman felt the need last week to defend the company on x saying the spending was understandably causing concern wow he says he understands your concerns if you're down today you're a certified beta bubble boy you literally bit up sandisk wtf Alternatively, you can call yourself a bad beta, B-I-T-C-H.

21:48The White House last night tweeted, we are so back in all caps. What did that mean? What did they mean by that? In what way were we back? In other news, Anthropic disrupted a highly sophisticated AI-led espionage campaign. The attack targeted large tech companies, financial institutions, chemical manufacturing companies, and government agencies. We assessed with high confidence that the threat actor was a Chinese state-sponsored group. I guess they were using Claude, I think. Yes, I think they were using Claude code, actually. Weird. So they were vibe coding espionage. Yeah, it was pretty funny.

22:26I read through some of the blog posts, and it was like some of the interactions of the hackers, and it was like, this is what they were saying to the model. They were like, okay, good job, Claude, but I think this part is wrong. You can see the actual transcript. Very bullish for anthropic. Michael Burry appears to be shutting down Sion asset management. He said, dear investors, with a heavy heart, I will liquidate the funds and return capital, but for a small audit slash tax holdback by year's end. My estimation of value in securities is not now and has not been for some time in sync with the markets.

23:00With heartfelt thanks, but also with apologies, I wish you well in your future investments. I do suggest investors contact my associate PM. Did he really quit right before the market started correcting? Is this one of those 90 % of gamblers quit right before they finally call the top correctly? Yeah, it does seem odd. His memory will live on through meme images from the big short. Vine is being rebooted under the name Divine with funding from Twitter's former CEO, Jack Dorsey. The app plans to feature more than 10 ,000 previously archived Vines and does not allow AI-generated content. That's remarkable.

23:49There have been so many Vine revival attempts. Elon was talking about bringing it back at one point. I believe the founder of Vine was talking about bringing it back and did a number of different projects. There was a project called V2. We will see you tomorrow. Can't wait. 1 a.m. Pacific. Sure. Goodbye. you