EP 020: Is Open AI in trouble?

Gil, how is OpenAI? Well, it depends on if Sam Altman drives him off a cliff or not. He is definitely driving fast towards the cliff and he's gambling on the fact that somebody else in the car is going to grab the wheel and swerve. He's making commitments he can't possibly live up to. And not just verbal, informal commitments. He's signing contracts that there's very little chance OpenAI will ever be able to live up to. And he's spreading OpenAI, which is a brand new organization, way too thin. So if he continues down the path, yeah, they will not succeed. In fact, they could fail spectacularly. I don't think that's the likelihood though. I think there's too many other responsible people in the room in that car that will grab the steering wheel and swerve. They will make him focus on the two things that OpenAI really has. A top notch frontier model and a big lead in the AI consumer chat race. And then the byproduct of that of having good APIs as well. All the extracurriculars, building data center himself, making his own chips, making hardware, enterprise software, salesforce, all, all that stuff has to go away. Otherwise they will drive off a cliff. One of the core pillars of the deep research we do is field research and using expert networks. I always start with AlphaSense. They have a great expert library and they also have AI technology that allows me to get to the information a lot quicker. They allow me to access local periodicals and uncover information that I otherwise wouldn't have access to. They are used by 75% of the world's top hedge funds and 85% of the S&P100. Not only do you have access to the library, but they'll also go find experts for you based on what you're looking for. It is one of my go to ways to do primary research. Make sure to check it out with the link in our show notes. Hello and welcome to Pitch the pm. My name is Doug Garber and I'm here with Gil Lauria from DA Davidson, the head of technology research Google. Gemini is reportedly better and I heard Sam Altman put out a code red. What's the reality on the West Coast? The reality is that these models are a lot closer to each other in performance than people have realized. Gemini didn't go from being terrible to great. Gemini went from being very good to very, very good with this particular iteration. That doesn't mean that the next model from OpenAI won't be better, that the next Grok Won't be better. Anthropic in some ways their opus 4.5 is better. So this race is really just in its early stages and we're at the fluid part of the race. Things haven't congealed yet. In terms of leadership, it has very little meaning that Gemini right now as we sit here today is mostly the best model because it's not by much and it's going to change. Let's not forget there's a long shot from Meta. The next Llama could be very good. The Chinese models are getting excellent and they're open source, which is a huge advantage. So it's way, way too early to crowd Google as the winner. Although it is good that they're at least in the race. A couple of years ago they were caught flat footed. Let's not forget Google had GPT before chatgpt. They just put it in the pantry. When they have great technology and they don't want to disrupt their own business, they put it in a pantry. Their word, not mine. And that's where it was when ChatGPT came out. And because Satya and Microsoft were so ready to commercialize it, they took a huge lead on Google. Google fumbled multiple times. It's great that they caught up, but their lead is not insurmountable. Market sentiment swings wildly from Google is the biggest loser to now Google is the biggest winner. I think the truth is still somewhere in the middle. Where is the value in the AI chain? Is it going to be in the hardware? Is it going to be in the model? Is it going to be in whoever has the Most data, the APIs and the proprietary content? Is it going to be in the distribution? Who already is in the workflow? Who's losing money now? Who's going to ultimately win when the models go from spend, spend, spend to I want a return on capital that's the most efficient. So let's be realistic. Right now, the vast majority, vast majority of the profits are accruing to one Jensen Huang to Nvidia. Then you do have a decent amount of profit going to Amazon, Microsoft and Google, who are taking his ships and parking them up to rent them out. They are making money. Anybody beyond that is not in fact. And then at the bottom of the chart you have OpenAI, which is losing colossal amounts of money every time you ask it a question. Forward looking, it's unclear where the value is going to accrue, but what we do know is there is economic value being created by the chat. So when people, sophisticated users of AI are now Replacing labor with AI. Now that's not broad. It's not enough to move the needle yet for AI, and that's part of the bubble conversation. But sophisticated users of AI already generating significant economic value, which they are not paying for. Because let's not forget many free users, most people are at the $20 a month mark. Some people are at the $200 a month mark. At some point we're going to be using this is going to be so ingrained in our lives that we'll be willing to pay the economic value. At which point a lot more value is going to accrue to the chat, to the consumer interaction, and then there's going to be an element here of accruing the employee interaction because we'll be using this more at work. There's going to be some level of a rebuilding of the IT stack around AI type solution. That's longer term. The value right now is all Nvidia, but it will start moving down the chain to where it creates productivity and where it creates value for individuals that will be willing to pay for that value. It'll take a few years. And importantly, to put your two questions together about OpenAI and Google, we don't know who the winner is in the chat race, right? There was a point where Yahoo had the search lead. There was a point where MySpace had the social lead. They were the places where people went to and that's when things were fluid. But as we moved to Google and as we moved to Facebook, they congealed and competition was over and you had winner take most. There will be a point in chat where that happens, where things will congeal and we're going to use one chat over all the others. It might take a while for us to get there because the stakes are so high that neither Google nor Microsoft through OpenAI, nor Amazon through Antropic, nor Meta, nor GROK are giving up anytime soon. So you have five participants that know how high the stakes are and are willing not to make money. Because remember, Google initially didn't make money on those searches, and then it did. Facebook initially didn't make money on the interaction, and then it did. Uber again charged $5 to $10. It lost a ton of money until it didn't. Because when you are a winner take all or winner take most, you can raise prices to the economic value. There's so many big participants, it may be years before somebody's willing to charge the economic value. Streaming is another recent example. Initially we were getting Netflix at $8. Right now, I'm paying 22 for whatever reason. And now I'm paying for all the streaming and I'm paying a lot more than I used to. They're charging us the economic value. Now they're profitable again. It took them several years to get there. And that's a uni market compared to the market we're talking about here. For a while, there won't be economic value there, but the winners, one or two winners will be able to extract very significant tolls once it's congealed into a one or two winner in the chat market. How does OpenAI finance their $1.4 trillion contracts? What is their roadmap? How much equity? How much debt? How much cash flow? Do they have a plan? I mean, I heard they were asking for government backing for this. They don't. The likelihood of them spending $1.4 trillion over the next five or six years is less than 1%. Right. If they achieve AGI first or superintelligence, an AI tool that does anything a human can do at least as well as a human. Yeah. Maybe they're going to be able to generate the revenue and economic value to charge for it so they can pay for that. That's less than 1% chance. I'm not saying there's less than 1% chance we hit super intelligent. I'm saying there's less than 1% chance that OpenAI gets there first and therefore capitalizes on all the economic value. Less than 1% chance they actually spend that. Now, that doesn't mean they won't be able to raise capital and again succeed in what they're very good at. Frontier model that leads to APIs and chat. That's what they're very good at. That's where they have a lead and they can maintain that. And they don't need 1.4 trillion to do that. They need tens of billions of dollars to cover their losses in order to do that. And some combination of equity and debt should get them there. And we should hear from them sometime this year. Now, what you heard the whole blow up of AI around their over commitment when he went on the podcast and he attacked Brad Gerstner, asking him the softball of all softball questions. And then subsequently, Sarah Fryer goes to the conference and really lets us know that they were seeking federal guarantees. And then they gaslit us about us. And then the letter came out that literally said that they are asking for federal guarantees. That's happening because they're having a hard time raising capital. They thought they could just go to the market and Raise hundreds of billions of dollars of equity and debt and everybody would continue to bow to them. And they went into the market and they had the hard reality that there's not that much capital available to take that big of a risk, which is when they started asking for federal loan guarantees. But again, that doesn't mean they can't raise tens of billions, which will finance the continuation of more frontier models and chat and, and API access to their product, which is what they should be focusing on. Oracle is almost back to the price. When they announced the deal, it was at one point a couple of weeks ago, well below the price originally it was up billions of dollars of value creation. What are they actually on the hook for? Yeah, and how much financing do they need? So, you know, it was some combination of. OpenAI played Oracle and then Oracle tried to play its investors. Only you don't play investors and don't pay a price for that. Right. OpenAI. So the sequence of events. Important to understand the sequence of events because the news flow is so crazy and comes at us so fast that it's sometimes easy to forget what actually happened. What actually happened is OpenAI didn't have very many commitments on September 10. All it had was the commitments to undisclosed commitments to Microsoft and about $20 billion of commitments to Core we, which was proportional to the growth rate and to what they need. They show up on September 10th and Oracle announces on their earnings call that we've gone from 150 billion of remaining performance obligations to 450 of remaining performance obligations. Performance obligations is when you have a contract, you can include that as future revenue. So Oracle goes and makes this one of the most dramatic announcements in the history of corporate America. Right. We went from 150 billion of remaining. This is a company with less than 60 billion of revenue. So they went from 150 billion of performance obligation to 450 in an instant. Right. And what they said is, oh, it's a series of customers, it's meta, It's X, it's OpenAI. It's all the. It's the largest hyperscalers. We, you know, we're winning so big. I wanted to use the other word, but because, because there is a reference here to the relationship with the President. It's so bigly that, you know, we, we are by far the winner in AI. And for about five minutes it looked like they were the winner in AI. And then we wake up the next day. Stock goes to 350, right? From 250 to 350. We wake up the next Day, the Wall Street Journal discloses that Oracle signed a $300 billion deal with OpenAI. And then you do the math. Wait a second, 150 to 450. How much is that? That's 300 billion. So this is all OpenAI. But even then it seemed like on September 11th, it seemed like, oh wow, Oracle, they beat Everybody to this $300 billion contract. Right? The 300 billion contract for the next five or six years, which is what OpenAI is going to need for its compute. Oracle won that. So the Stock lingered at 325. It went off from 350 to 325. Yeah, it's only one winner. But hey, they won the OpenAI sweepstakes, which is great. What we've learned subsequently is that OpenAI made another $1.1 trillion of other commitments, at which point we realized it's ridiculous. And then again, he bit Brad Gerstner's head off. And Sarah Fryer said they want federal loan guarantees. And it became increasingly clear there's no way they do this. And if they can't do 1.4, they can't do the 300 billion to Oracle. And so then we get to the answer. So where's Oracle right now? Because Oracle has a contract with OpenAI that says they need to build this capacity so OpenAI can have it. So they need to borrow money so they can build this capacity, so OpenAI can have it. Only they don't know if it's going to materialize. In fact, it now looks increasingly clear that it won't materialize. So what's Oracle to do? Are they going to stop construction? Are they going to give back the money they borrowed? They're in a pickle. Which is why we are now lower than where we were before that September 10th announcement. Because not only is OPODI probably not going to materialize, Oracle's on the hook for the build out until they're told that it won't materialize. So we have to again, back to the original questions. Is OpenAI maybe? Because now they have to do renegotiation across the board. They have to go to Oracle and to CoreWeave and to Amazon and then crawl back to Microsoft and say we actually don't have an AMD and Broadcom and tell them we don't actually have $1.4 trillion. You kind of knew it, but we signed the contract knowing full well we can't live up to the contract. So hey guys, you can let me drive off the cliff. OpenAI will be gone. But you're kind of screwed if that happens. So you should probably renegotiate with me and then we get into what the consequences of that renegotiation are. So this is why Oracle is lower than they were even before announcing this deal. Does Oracle get milestone payments? Is that a firm commitment or is it in layers where they can maybe build some at a time and as layer it in so they don't have all of that counterparty risk at once? Right, so it's back end loaded. It's back end loaded. So again, the likely scenario is that OpenAI just scaled back its ambitions and spreads around what it has next year. So they'll raise enough money to support the compute commitments they made maybe next year, 2026 again the likelihood they ramp it to 2030 is very, very low. But the likelihood they could get enough capital to support their growth in 2026 is quite possible. They're just growing from 20 billion of run rate revenue to somewhere between 30 and 50 billion of run rate revenue. That's doable. That's doable with tens of billions of dollars and that they could probably raise. So Oracle may be fine for next year, but they have to then start doing the math of, well, do we keep buying land, do we keep expanding, do we keep borrowing more? Get into that whole conversation of their debt now is trading at a big discount, right? The insurance on their debt is so high that their debt's trading at a discount. So when they next time they go raise capital it's going to be at a higher rate. Not to mention the fact their leverage now is high. It's not just they have leverage, it's high. It was high before because they wasted money on Cerner and now they borrowed all this money that they're sitting there and maybe they shouldn't deploy or maybe they just deploy part of it. So they're in a pickle. So likely there'll be milestones where they could get paid and slow and extend. They're on the hook for the debt, they borrow and essentially the finance in half of the data center finance and half a trillion has really come from called the big four with amazing balance sheets and free cash. And they're getting that financing because the debt is triple secured, right? You have a contract, you've got a corporate guarantee and you've got GPU collateral value. So the residual value on a GPU in year two, three, four or five, not as relevant because you got the corporate guarantee from these, the biggest companies in the world. OpenAI has meaningful counterparty risk are we going to see? Because one of the things I've researched since we've started they say your real research starts once you put your position on. We talked about Nvidia earlier in the year and now I've realized there's a very strong correlation to the year forward capex revisions to the Nvidia stock price or the customers makes logical sense. The fact that it's 90%. So those four or five customers are super important just to understand where the stock's going. Is there capex risk now in the system that if OpenAI are they going to have to renegotiate down blend and extend lower? Is there risk now? There is risk, but let's walk through it. First of all, I just want to clarify, I'm not catastrophizing Oracle. I actually think they will build the data centers and they'll have customers for them. Because we don't have enough data centers, they'll end up just reselling them to Meta and Microsoft at a discount if they overbuild. But I don't actually think Oracle's going to go bankrupt or anything like that. They have a very good core business that generates a ton of cash and they will build some data centers and they will be able to use that capacity, whether if it's for OpenAI or somebody else. So I'm not catastrophizing Oracle, I'm just saying they're not out of the woods yet. But to your conversation, let's talk about what's real and what's not real and where the risks are. Microsoft, Amazon, Google have all the customers. Everybody is their customer, all businesses are their customers and in fact all consumers are the customers. And all the chat goes through them. Right? So when they're building a data center, they can sell it. They've already pre sold it. In fact, they don't have a problem. They're using cash on hand that they had nothing to do with. All they could do is buy back stock or pay a dividend, which are really low return endeavors. Or they could do this thing which they know is going to pay off one way or another. They're not sure how well it'll pay off, but they know. So I refer to that as healthy behavior. Those three companies know what they're doing. They're in charge. They will get the value they are right now and increasingly they'll get more and more value when they borrow directly. It's treasury management. They're not borrowing because they need leverage. They're borrowing because sometimes they have the cash in one geography and they don't want to move it to another geography. And by the way, their credit worthiness is such that they're literally borrowing at 5% and their returns are better than that. So we're all good there. We start getting in trouble when other entities that don't have customers and are borrowing money, that's when we start getting in trouble. And yes, there's some level of guarantee around that. So for instance, let's use Core Weave, right. When Corey was borrowing money to support build out for a Microsoft contract, we were good, right? Microsoft is good for those commitments by the way. We'll definitely need that compute. So they're good for it either way. So yeah, so that's good to secure a contract. The GPUs, we get into that. Not worthy of the securitization. But the contract with Microsoft is. But then when Core Weave is dealing with OpenAI. So let's go back to that scenario, right? OpenAI just made $1.4 trillion of commitments. 20 billion of that is to Core Weave. If they have to go back and renegotiate with everybody. Where do you think Core Weave stands in the rankings of Microsoft? Amazon, Oracle, coreweave is the last that's going to get paid. So is their contract worthy of securing and raising debt for? Definitely not. So then all you have is securing it to the chips so we can get into that. It's a good conversation about what it means to secure a loan to the chips. Lease obligations, new accounting rules, lease obligations are on there and essentially Metta has backstopped this. It's essentially a corporate guarantee for the people building it. Except it's not through a contract, it's through the lease. Right. Only we don't. My understanding is that it's four year terms renewable, meaning somebody's going to go borrow a ton of money. I think they probably already have like 27 billion on the 2 billion of equity from Meta that they didn't even have to use cash, they just contributed the data center and they're building out a data center where it seems like Meta's only contracted for four years and the guarantee is maybe only four years. So could work out. It very well could not work out. In which case Blue Owl gets wiped out. Meta doesn't care. $2 billion is a rounding error. But Blue Owl's equity, which is more and a bigger proportion of who they are, gets wiped out certainly in a second. And then the debt holders are left owning again. 4 year old GPUs I don't know what GPUs are going to be worth in 4 years. The debt holders have a guarantee from Meta. I would imagine they have a four year guarantee. We don't know that it extends beyond that. Why would Blue Owl do something that wouldn't with that much residual risk? Because underwriting right now is not very strict. That's the answer that if people are borrowing money, if people are lending money to CoreWeave to build out data centers for OpenAI, that's all you need to know about where we are on underwriting right now. A big thank you to street account by FactSet for helping in the research process for these episodes. I've been a big fan of this product since it came out many years ago. It's a staple in my process and serves as the top of funnel for all names on my watch list. I just enter the ticker and I'm instantly in the flow for all key news flow, allowing me to ramp up on the name while I manage the rest of my book. And I also use it to monitor customers and competitors for names that I have in the book. They also have value added analysis in their earnings summaries with historical valuation and they report consensus earnings revision, which saves me a lot of time on gauging what level to add to a position or not. Their gen AI tool Mercury is cutting edge and provides access to one of the largest databases for top tier analysis. You should definitely check it out. Meta has just been penalized for overspending and you know, great, great social media apps. Not quite a monopoly, but essentially one depends on where he sits at the president's table and the market told him hey, this is too much. And I think you called it and said hey, you're spending less on his last moonshot moved it into this moonshot and I think it's two different things. So how much of their capex is going towards just creating a better, more productive user experience for their customers, for the ads, for the users and then how much of their capex is going for this AGI thought of call it ego. Yeah, most of it is for his sandbox. Right. He's going to spend $100 billion next year. He probably needs 10% of that to make Instagram ads better. He's going to use the whole rest of it to try to achieve superintelligence. So again investing in Meta is saying I trust you Mark, you've done unbelievable job so far. You are clearly still in founder mode and I trust you and you will be a great shepherd of shareholder value long term for me, just like we used to do for Jeff Bezos. Jeff Bezos would show up sometimes and report these massive losses in a quarter and he'd tell people, deal with it. I don't really care. I'm growing a business, deal with it. And that's what Zuckerberg is doing right now. And he's still in that mode. Once in a while he gets, he gets enough pressure that he does the year of efficiency. Right now he's cutting some of the Metaverse, but at the end of the day, we're in the same spot, which is he has voting share majority. He could do whatever he wants and you have to trust him to do it. And he's done a remarkable job, so it's hard to question him. So let's talk about the two big investments, the two moonshots, right? There's the Metaverse investment, which people are making fun of, but he's been explaining to us very clearly for a while now that he's sick and tired of paying Apple and Android from Google 30% off the top. And he will own the next platform, come hell or high water, even if it means losing $18 billion a year. He will own the next platform. Which, by the way, he's not just telling us, he's telling everybody else, right? He's telling every, all, all other companies. You will not beat me in the race to glasses and goggles because I'm willing to lose $18 billion. Are you? So that's what he's been doing. And by the way, he has a remarkable set of products there, right? The Ray Ban AI glasses are hugely successful and are well on their way to being what they need to be, right? But now what he said is, look, I thought if you asked me two, three years ago, I would have told you we're going to have a new platform, Goggles and Glasses, before we have AI. And what I've learned over the last couple of years is, no, it's actually going to be the other way around. We're going to have AI first and then we're going to have platform. And by the way, the two are very related. But now that means that I need to take some investment off the table for the new platform. I'm still pursuing it, but it's now a longer timeframe and put all that investment in AI, because just like any other big technology cycle, there's only going to be one winner, one or two winners. It's going to be a winner take all or winner take most market. And Again, I'm doing the same thing I've been doing, which is telling everybody I am not going to lose if it means I need to pay Alexander writing $14 billion. Are you willing to do that? Are you willing to put again out of the 100 billion? I'm telling you the 90 billion is for developing a better model. Nobody else will. And that's his signaling to the market that he intends to win. And he's trying to compel everybody else to stop this race by signaling that he's not willing to lose. This is founder mode now. I don't know that he's going to win. My colleague Alex Platt's in the weeds here and he tells me that it's unlikely that Llama 5 will be the winner. But Mark is going to do everything he can to make sure that it is. And if it is, then again, the pot is so big that it's worth the gamble. A lot of money to lose there. When you have a pretty steady cash cow, it seems more like Howard Hughes taking the cash cow from the oil tools and investing it into airlines or a business that kind of was cool and a hobby and fun, but not necessarily a great return. But I guess time will tell. But that's exactly right. He could lose. If he loses, he loses big. But again, it's not going to devastate Mehtum because they'll have this data center capacity. They'll be able to resell it again. It will have value. They'll resell it to other customers. They'll resell it to Microsoft, Amazon, Google. The capacity isn't going to go away. It's not a waste of resources. He needs some of that capacity for the growth of Instagram, WhatsApp and Facebook and whatever else he buys along the way. It's a huge gamble. But it's not like there's infinite downside, there's finite downside. How would you frame the amount of downside if he decides to pivot away from? AGI says, all right, I need to re rent some of this stuff. I've overbuilt. He's still got the cash cow of the other businesses. That business is not, you know, it's an infrastructure asset now. It's obviously not a huge high return. Near monopoly. How much downside is there if he has to make that pivot? And would the market at this point reward him for taking his foot off the capex? It has in the past and it will in the future. I mean, we really just have to Add up that capex number, right? Because if they put probably, let's call it 40 billion that they didn't need to last year or 50 billion last year and they're going to put 90 billion next year that they don't need and they do one more year of that, let's say all in, he will have spent $200 billion and let's say he'll salvage a portion of that. So only 100 and 150 billion would have been wasted. That's 10% downside, that's 5% downside. It's not devastating. And again this is a stock to your point. Let's talk about the business. Meta just grew its business 25% twice as fast as Google. Guess what Google's trading at 28 times, he's trading it 22 times. Right. He has a better business than Google. Let's not forget that the core business, the digital advertising business, he's eating their lunch day in, day out. Let's not forget that. Maybe that should be, maybe the AGI data center business, the asset intensive one should be, you know, NewCo number two spinoff. But maybe. Well, we'll let. I'm sure he's been pitched that I. Think he's even he's hinted at it that at some point if he's overbuilt he could just get into the hyperscaler game to compete with Amazon, Microsoft and Google. Yeah, it's a different type of asset, right? Very different. The hard asset type of it. Google, you made a great call in March. Google was the have not. They're losing the top of the funnel. ChatGPT is replacing them. They were the cheapest in the group. The core business is in trouble. Two things happened. One, now they have the best, the better model LLM, but also the hidden value from the TPUs. You were onto it early and can you walk us through where we are in that cycle? They've sold some to Meta, cheapening up their capex. Where are they in monetizing that? I've heard engineers at Meta complain but how much is it worth to Google and what is the path to further unlock on that? Or is that already baked in at this point? I don't know that it's baked in. So I mean look, the pendulum swung wildly, right? 6 months ago Google was dead. Everybody say, oh Google's dead. Right? Because search is dead and Google's dead. And the point we were starting to make is well search is challenged and there's much debate over whether or not we'll succeed. We don't have actually that strong of an opinion on that. We have an opinion, but it's not very strong. There's plenty of other people debating that question, but we were pointing out. Forget that, look at all the other stuff they have. They have the TPUs, which is the second best chip to Nvidia, and I don't think people realized it at the time. Then they have the cloud. On top of that, they have YouTube, which is valuable, a media property as any media property. They have an ad network that's the most valuable ad network. And they have Weibo, which is by far, which is as valuable or more valuable than what Tesla has. And what we were arguing was they need to split it off. If they're trading at 18 times earnings, they're trading at 18 times earnings. So people are worried about search and are ignoring all the other good things that we said. We thought that if they stayed at 18 times, eventually that's what would have happened. But as it happens, the pendulum swung all the way back, really, based on vibes. The catalyst, interestingly, was the government not forcing them to spin off Chrome or Android, which is a completely tactical issue. The government didn't force them to do that. And for whatever reason, something switched in the market's head to say, oh, that means they're winners. And then again, Gemini went from being here to here in the rankings, a very small difference. And that helped the pendulum continue to swing. And that's when people realized, oh, wait a second. Part of this is because they have the second best chip that they could actually train the model in, that's theirs. And they know how to use it because it's been used for search for decades, for a decade, and they've gotten really good at it. Where generation six and seven and now the pendulum swung all the way to Google is the winner, right? Not a winner, the winner, which is probably a little too far. Now, the TPU business, yeah, when we started talking about the beginning of the year, they were mostly just using TPUs internally, meaning to run search and YouTube and all that stuff. And we were saying, wait a second, people in the industry are telling us this is the second best AI accelerator. What if they just started selling it externally and this was fantasy. When we started talking about it at the beginning of the year, it was total fantasy. Nobody even occurred to them that they would allow other users to use TPU in their cloud. That was a new thing. And certainly now that they would sell it. And something happened at Google. I think they probably had the same epiphany of wait a second. If we have the second best chip and Nvidia can't sell them fast enough, we should start selling them. And so they started going to the market and it started with a small with Fluidstack in London and then they negotiate a bigger deal with Meta. When was that Fluidstack deal? It was probably a couple of months ago. And how much cheaper are they selling it versus Nvidia? What's the savings to the customer? Not much. Not much. This really isn't the motivation here isn't even to save money. I'm not even sure. There's a lot of debate now about total cost of ownership and I don't think anybody knows the answer to that quite yet. What the actual difference is if you were to buy TPU from Google, Google has a lower total cost of ownership of operating TPUs than GPUs. But that doesn't mean that you as a third party, as a Meta will have a lower cost of ownership because there's a lot of other pieces there. What's more important to Meta? Well, that's the key debate. Has Meta commented on that Anthropic mentioned this in a conference in September, that their investment in all the other Asics was finally paying off. That was kind of the first clue that of the commerciality of the tpu. I mean semiannalysis has been talking for years about the outperformance from a technical perspective. That didn't matter until it became a commercial reality here with those two contracts. That was the switch. Yeah. So I think that what's hidden here is the true motivation, which isn't actually total cost of ownership, it's diversification. It's companies like Meta and Microsoft are very concerned about being this dependent on Nvidia. Google realized this well in advance, which is why it has independence and negotiating leverage. Amazon realized it a little later, so they have a little more negotiating leverage, although their training is not as competitive. And Microsoft and Meta just arrived there very late and it's going to take them another five or 10 years before their own Asics are competitive. So they're saying if we're going to get any diversification in the short term, we have to make some bet on amd. Maybe or maybe not. It may or may not turn out, but we have to make some bet there and we have to make another bet now on tpu. So Meta already sucked it up at some point. Microsoft may suck it up as well because we have to diversify. None of our other. There's nothing else. This is the most important thing we buy, we can't buy it from one vendor. So the diversification here is a lot more important than the cost savings. Is Meta going to be able to operate GPUs at a lower cost than GPUs at some point? Maybe. Maybe for some stuff. I think with all the analysis from Semianlysis and the analysis that they're doing internally, I don't think they know because they've never done it. And again, chips are hard. Chips are hard. Building data centers is hard. You don't just know ahead of time what your total cost of ownership is going to be. You have to do it for a while before you actually have the real numbers. But their motivation is to diversify and then maybe also they can have lower total cost of ownership. You mentioned earlier that OpenAI loses money on the search at this point. Does Gemini and Google, do they lose money every time I have AI do something within my Google workstation? Oh yeah, well in fact Google. Anything else apart from searching on Google and watching a YouTube? Google loses. Right. Google Cloud is profitable. I'm probably exaggerating a little bit, but Waymo and all these other things are money losers. They're still very valuable. They're money losers. When you do an AI search, the incremental AI search costs the money and you're not paying for it because you're not paying per usage. Either you're a free user and they have to pay for that compute or you're a $20 user. They get the $20 subscription and then every search that you do costs them money. So they're losing money on the comp because their price points are the same as ChatGPT. And we know ChatGPT loses money, right? It's a negative gross margin. They argue that it's not. It's negative Gross margin costs are a lot lower. Right? They don't have to. Their capex is a lot lower on that 50, 60% GPU. They're buying a TPU without that 60, 70% margin and that depreciation. It's got to be 20, 30% lower. Yeah, maybe they're not losing as much. Let's not forget OpenAI is going to probably lose 20 billion on the 20 billion of revenue they're doing this year. So maybe Google is just marginally gross margin negative in this business. But by the way, again, they can absorb it. They have no problem absorbing it. They have to your point about Meta earlier, it's true about Google. They have such a cash rich core business that they can absorb Losses in chat for years. This is part of the reason this race is going to be so long is that the players here could lose money on chat searches for years before they charge the true economic value of it. And Google is certainly there in terms of making sense for them to lose money on the chat until they can charge the economic value. The way I think about it is what's given forward the expectations for 15, 20% growth from the 500 billion or so that's baked in. So we really need to grow, we need to revise that up and up for the Nvidia of the world to work right and then for the capex to trickle down through the economy really. But this is the seems like the first time where there's a little bit of counterparty risk on OpenAI. Maybe they had to dial it back. They over promised they couldn't get the debt. It seems like the next two key parts here are the OpenAI and the Anthropic IPOs. Any thoughts on them going around with testing the waters, how much they're looking to raise, what the appetite is? I mean these deals more and more now are getting kind of pre marketed, pre funded. By the time they come they kind of know what the book looks like. It's a shorter roadshow. What's the update there when you start. Getting to these kind of numbers, it's a little tricky to ipo, right? Because you're going to need a tremendous amount of influx of capital on day one, even if you do a small float. We're even dealing with this on databricks. Databricks is now worth more than $100 billion and they're kind of waffling on IPO because that transition from raising private capital at convenient terms without short term accountability is really tempting. Stripe just stayed private, right? Databricks is really waffling. Anthropic is going down the same path as databricks, only it's already bigger. And so an IPO is not trivial. But anthropic is being managed by responsible adults. So they have actually played this really well. They've now worked with all three of the most important players, right? Google, Amazon and Microsoft. They've gotten good capital from them. They're not over levering, they're not over promising, they are doing a better job. It's what's really important in AI, which is math, right. If we get to superintelligence through this technology, the transformer technology, large language models, it's going to be because we get really good at math and that's what Anthropic is focused on. That's why you're seeing them do so well in programming. That's why they're outgrowing the corporate business of OpenAI. So they have a well thought out strategy, they're executing it well and they could IPO in a reasonable way. It'll still be challenging because again, even a 10% float is going to require a ton of equity to flow out of other things and into it. But maybe they'll be able to pull it off again. OpenAI is going down fantasy land of we're in the business of building data centers and we build chips and we build hardware. With Jony I've and we have an enterprise salesforce and we're doing all the things for everybody and that's why we're worth 500 billion and we're going to be worth a trillion and we're going to IPO at a trillion. I don't think that's a reasonable path. Again, I think there's a point of reckoning that's going to happen. I hope that they just replace Sam Altman. I've actually put this in writing that it doesn't need to be like this. If you just put a responsible adult there, they will do the right thing, which again is focusing on the model, the APIs and the chat and then they can IPO successfully. But that would require the step in the middle of going back to Oracle and coreweave and AMD and Broadcom and breaking the news to them that the contracts that they have signed are meaningless and they need to scale those back. So we need to first focus, renegotiate all the deals. Then OpenAI will be in a position that Anthropic is now which is on a path to an IPO that could do well. So what you're saying is for OpenAI to fill this $200 billion gap over the next few years in these hockey stick projections, either Altman has to do it or someone else. The board has to get someone else to say hey, you messed up, you over promised, you need to go back, reduce your commitments, refocus. But that's a capex draw. That's a downward move in the capex. What happens to the stocks in the OpenAI? I mean they've already been. You look at the OpenAI versus the Google system, there's a huge dispersion in a very tight competitive landscape based on the counterparty risk. I don't think it drives down the capex. I'll tell you why. Because those Things that I do think that they'll stick to are what's driving the capex. It's chatgpt that's driving the capex. It'S training those big models that's driving the capex. So that stuff will still need to happen. Whether it happens on Core, Weave, Oracle or on Microsoft, it still needs to happen and we still need that compute. In terms of, I think you're spot on about what's Nvidia driven by its CapEx estimate changes. Right? We just went through an increase in capex and it's because remember I said Amazon, Microsoft, Google, they have all the customers. It's not just OpenAI. They have all the customers and they're customers who are telling them I'm willing to sign three year contracts on this AI compute capacity so I can build AI tools internally and I'm going to make that commitment to you Microsoft. And so Microsoft knows they can build the data center. And Microsoft, by the way, the reason we love Microsoft the most is because they have the best ability to say okay, three years of AI rental but that's not going to do it for you. You also need Data Fabric and Cosmodb and Business Intelligence tools on top of that. And by the way, you need a higher subscription level to Office. So I'm going to move you to tier 5 from tier 3. Microsoft is able to capitalize on all this, which again is why Microsoft, Amazon will continue to do this capex. And again they just told us that the demand inflected positively during the quarter in terms of their customers. That's not just chatgpt, that's all the companies in the world because their customers are all the companies in the world. Their demand for compute inflected positively so their capex is safe. The marginal players will go away. The marginal players are disposable. Microsoft is offloading on Oracle and coreweave because they don't want to just have a Blackwell fleet and then have an outdated fleet in a year or two. They want to manage their fleet in a way that they buy enough of each generation of Nvidia chips so they're not stuck with one generation so they can roll the demand across every generation of chips because the leaps in every generation are significant. They're saying instead of me just buying Blackwell for all this demand right now I'm going to buy some Blackwell, then I'm going to offload demand to coreweave and Oracle and others and in three years I won't need it. But right now it allows me to build more Gradually and more thoughtfully. But the capex is still needed to serve the AI demand by all these companies that are booking three years in. Advance last quarter, right? The Nvidia stock moves with their big customer CapEx revisions. It's a pretty logical equation. The big four or five are 50ish percent. We can kind of piece that together. The number went to 526 billion in November from September. 450, right. 17% increase. They revised up based on real contracts from real customers. Except meta in that group as well. Nvidia stock didn't react with the logical 17% positive revision. This is a logical correlation. What am I missing? If these guys continue to ratchet up their capex 15% a quarter, I'd expect Nvidia to continue to do well. Why did it not react well during the quarter or during their customers reports? The short answer is Vibes and market digests information in interesting ways. You and I, last time we talked, we talked through all these issues for Nvidia, right? We talked through Capex scrutiny, we talked through China challenges, we talked through depreciation, we talked about models getting smaller, we talked about Chinese competition, we talked about all these things. Again, there's been a conversation about it for more than a year. But as it happens, the market sometimes just wakes up to things all of a sudden. You woke up to some things in January when we had the deep sea concerns. And then it woke up to the fact that Nvidia has competition from TPUs, which again we've been talking about for two years. It woke up to that all of a sudden. And so Nvidia is of digesting all the concerns that you and I have talked about in the past. And that's just how the market is. You never know when information gets reflected in a stock and it's happening right now as we speak. But the other part of it is Vibes AI trade. Nvidia was the AI trade for a couple of years because it ascended in value more than anything else. And then as growth rates went from 300% to 70% to 50% then the AI iTrade started looking for other things to do and it started doing nonsense like Oracle and Coreweave and Oclo and IonQ, which isn't even AI, and it started doing all this nonsense. And then again Google went from here to here and everybody all of a sudden decided that Google was the only winner. And then in a huge swing, $2 trillion of value accrued to Google. So the AI trade just migrated from Nvidia to nonsense to Google in a matter of weeks. And that's where we are today with Google capturing really most of the value this year on the AI trade. But that's all Vibes. Google's numbers didn't go up. Google's numbers haven't gone up in six months more than any other company. And the stock's doubled. It's Vibes. Well, their capex has gone up, which is actually a fundamentally positive thing for Nvidia. I mean Nvidia is strongest sequential growth, even take out China. That hiccup kind of played out in April and actually for two months it was actually great. Right. And then it hit almost your downside target. And then after the numbers are just the fundamentals are getting better, right? The customers are spending, their customers are spending more because their customers are contracting it. How Short Are we GPUs or TPUs today versus six, nine months ago? Just supply demand on the ground. How much is term versus spot? I actually think Nvidia can deliver all the chips that are being ordered right now. Really what you hear from the big customers is they just don't have the warm shell to put them into. Right. It's actually the bottlenecks are actually in the construction, the equipment, the electricians. That's where the bottlenecks are right now. And so the chips, they're making enough chips. There's really the supply demand there. I won't say that it's balanced, but it's getting balanced. By the way, the bottleneck is going to increasingly become TSMC. Right. Because when you talk about TPUs, Google again, we think that they should be selling them and they could probably get 10, 20% of the market if they do. But they still have to get that capacity from tsmc. That's not trivial. TSMC makes everybody's chips. What about the software? The software, I mean people like to use what they use. And how much is CUDA really a protective moat. I put it in the category of it's not just the chip. So there's a whole level of sophistication you need to do AI compute that involves networking equipment, base layer software, and then other parts of the software stack on top of that. So when we say TPU is the only other relevant competitor to Nvidia, we mean the whole stack. We don't just mean the chip, we mean the networking, we mean the software. And the reason is that Google has gotten really good at using TPUs because they've used it internally. Right. And so they have the competitive product. It's a segue to amd. Right, amd. Supposedly their chip next year will be performant compared to Nvidia GPUs. It will be in the ballpark, which it's never been. But that's not enough. Again, you have to solve the networking, you have to solve the software. And their customers that again are super eager to diversify, will go through a journey of doing that just so they can diversify. Not because they actually need AMD chips, not because there's a shortage of Nvidia GPUs, because they're so eager to diversify that they're trying to prop up amd. It makes sense. So Gil, this is. I've learned so much. So thank you for everything. I'm trying to summarize and ingest what I've learned here and what are the actions I would hypothetically be bringing or discussing with a in a real portfolio situation. So first I should also tell people, you know my positions, which I've Talked about on LinkedIn. I still have, I've Google and met along. I have Nvidia short, it's still hedged. And it seems like the next things that could happen is if Meta cuts capex because they kind of reacts to what the market is telling them. If Zuckerberg actually cares, that could be positive. I don't know that that's a likely scenario in the near term. Maybe it's oversold there. The other thing that could happen, it Sounds like is OpenAI has over promised and they're going to have to renegotiate. But it also seems like some of that has already started to penalize called the neoclouds, you know, the oracles, the, the core weaves. So when that happens, is that something people are short into that expecting that to happen? I mean the logic I'm putting out here is not. There's a lot of other people who can put this together. Are people short the. I guess they are short the OpenAI ecosystem. So when they start to renegotiate, is it actually bad or is it actually time to. Yes, it depends on the magnitude of it. Right. But if the capex is blended and extend, it's not a big revision comes out. It's a speed bump. It's not actually a problem, it's just a perceived problem is a potential thing. But it seems like right now that seems to be the fast money trade. I mean these guys are switching it every quarter, every two quarters sort of thing. What are you seeing as the consensus trade right now? Yeah, I think we're very much the market is very much overweight. Google, it's a little underweight Microsoft and Meta, it's a little underweight Amazon and it's underweight Nvidia. So right now the market is because I mean, just look, let's talk about the multiple historical multiples, right? Google's been trading 20 to 30, but closer to 20 for a while. There are 28 right now. Microsoft, Nvidia and Amazon have been trading at more like 25 to 35. They're all in the high 20s, so they're below their median trading range. Google is above its median train rate. Nvidia has traded more like 35 times and right now it's at 26. So Nvidia is very much below its historical trading range. So the market right now is very, very much long. Google, I think there's two potential negative. So a they could win. The market could be right. They could win. There's a ton of upside. There's two potential hitches along the way for that. One is somebody Grok, Meta, OpenAI comes up with a model that's clearly better than Gemini. Three and then how do you say that Gemini is the only that Google's the only winner if there's a model that's clearly better than it. That could happen in weeks, could happen in days. The second thing that's going to happen is Again, if OpenAI does the right thing, which is to focus on ChatGPT, it will turn on ads. At which point there's going to be a quarter or two where a lot of digital advertisers are going to say right now I'm only spending on Google, Meta and ByteDance, TikTok. Maybe I should rebalance at that point. ChatGPT is going to have a billion weekly active users. Maybe I should rebalance across the four. There's going to be a quarter where digital advertising dollars get spread around a little different quarter or two. That could happen as early as next year to Google. So those are the two things that I'd be worried about long Google is those two things could happen in days or months and neither of these things is going to be very good to Google because if you show up and there's a Google quarter where search advertising growth goes from the current low to mid teens to 10% or 9% because there was a rebalancing of advertising budgets across those four platforms, that's Armageddon. It may never happen, but if it does, that's Armageddon. I hear what you're saying. Has OpenAI built out the sales and Implementation infrastructure. You don't just turn that on in a night. Have they invested in that infrastructure to be able to turn it on? Should the people say, hey, I want 10, 15% of my ads on here? Now they're investing in it. They hired a chief revenue officer from Pinterest. That's how Pinterest makes money. Right. In fact, in the reports about them starting to think about focusing, which I think are very positive, it said that they're going to hold off on advertising, which tells you that they've been working on it. So again, I think it's an eventuality. You have to capture the economic value. You can't lose $20 billion on $20 billion of revenue every year. You have to at some point capture some of the value in order to pay for all this. What is the best way to get into that channel of digital advertisers to understand the magnitude of the shift that's going on? I mean, I used to work at a company that their go to market was all digital advertising. It was like 60% of revenue or something and funnels everywhere. But what would you suggest people do? We're looking at AlphaSense transcripts. Are there other sort of places where you can get insights from what's going on? Yeah, we plan to do work on that next year. We plan to lay a baseline of Google advertisers and their behavior and then revisit once OpenAI ads get turned on. ChatGPT ads get turned on. We'll revisit to contrast those. That's our plan for next year. Again, we still think we have time because we would have heard a lot more chatter about ChatGPT taking that step. Until we hear that chatter, we're still pre that check. So I want to get as close to that to say, the baseline. So then when they announce it, you could contrast those two sets of behaviors from digital advertisers to say, how much is that hit going to be to Google? Maybe they just go from 13% growth to 12% growth. Everybody, nobody cares. That could be very much the case, at least in the short term. Or there could be a bigger adjustment than people are really going to care. Is there a way to scrape data to not just do channel checks and try to figure out how material they are, but to actually quantify. Is there another way? I'm just trying to think if it could be more systematic of an approach. There are data sets for this. Nielsen collects a good data set for this. I'm sure everybody's going to be on top of that we're not subscribers necessarily. But I think everybody will be on top of it after the fact. Right. Again, the goal is to predict this with surveys and kind of quick follow ons because by the time it hits that data, it'll be too late. But that's definitely going to be part of what people do around this conversation. Again, it's inevitable. It's happening. It's just a matter of when the channel checks. Now you need to do the channel checks to front run the alt data because everyone has the alt data. That's table stakes. That's right. Which front runs the company's actual announcements. And then six months later the market wakes up to it. Just like with TPUs and Nvidia's challenges. You never know when things hit the market consciousness and get processed. Our job is to still try to get ahead of it. Even if it means being too early on some stuff. It's better because they get our investors. You know, you used to be at some of the biggest, most powerful hedge funds in the world. Those are our customers. They appreciate it if you're early. Even if you're too early. Yeah, you're typically they want to know the other side of the bet. And sometimes if you're right, you're going to be early. You have to have conviction in your thesis and be willing to take, I guess they call it take a little bit of pain if you're better to be a little early. Some people will chase on the downside, but I don't. Yeah, you better be early if you have conviction. So Gil, this is. I'm very grateful for your time and coming on the show. You're our first guest to come twice and just want to tell the audience. Again, this is not investment advice. This is for educational purposes. Do your own homework. This is for professional investors. And thanks for tuning in. Please click the like button and spread the word. And everyone have a great holiday season and a new year and we'll talk soon. Thank you, Doug. Enjoyed the conversation as usual. Thank you. Please make sure to follow Pitch to PM on your favorite podcast app and never miss out on episodes. This show is for entertainment purposes only. Please consult a professional before making any decisions. This show is copyrighted by Pitch to pm. Written permission must be granted before rebroadcasting.