Sam Altman JUST Accidentally Exposed the AI Collapse.

Sam Alman just warned that compute might be in a bubble, which is weird because he's running Open AI, you know, chat GPT. How how could you argue that while simultaneously saying you need more compute and you're buying Mac Studios? So, in this segment, we're going to try to reconcile the shift that's happening. And I think it's really important because it helps us explain potentially why we are seeing some shifts in stocks including Nvidia up 3.7% today. Cerebras up 5% today, Dell up 9% today while at the same time Marll is down 2%. But more importantly, Credo is down 20% after earnings. How do we reconcile all of this? How could this potentially all make sense? Well, I think it's useful to start with what's going on with OpenAI. And no, I don't mean the relatively basic news that was shouted out by Elon Musk where Elon Musk says a leopard can't change his spots. because after all, Elon hates Sam Holman. Here you have uh the article that Elon Musk was responding to where Apple is accusing OpenAI of destroying evidence in the lawsuit they have between each other. That's not the issue at hand here. The issue at hand, which just so we can verify it, there's the Elon tweet, "A leopard can't change its spots." That's not the issue at hand. The issue at hand actually has to do with this particular interview and this about 55 minutes in clip right here that suggests somehow simultaneously as there being a compute shortage there could also be a compute uh or lack of compute- which is weird like how do you have both at or sorry a surplus of compute how could you have both at the same time let's listen to this and see how we can reconcile to make sense of it here we go >> astronomical numbers associated with this that you feel is at risk at all when you look at all of this? >> I'm not worried about our compute buildout plans. I am worried about the world's compute buildout plans. Like I think we are going to be able to use all of the compute very profitably that we are planning to build. Um but I am seeing the first signs of what feels to me like unsustainable silliness of you know random new NeoCloud popping up. people claiming that they're going to build gigantic amounts of compute next year that I think they don't have the revenue to support or a buyer. Uh yeah, I definitely feel like some fear about what the world is doing as a whole. Although I >> or a buyer, too much compute, random Neo clouds popping up. Let's try to reconcile a little bit of this because Nvidia is probably popping today on this reportedly nearly complete $14 billion acquisition of Hugging Face. It's about a $ 122.9 billion acquisition plus about a billion dollars in employee retention. You know, got to give got to give the employees a billion dollars just to stick around. A lot of people believe all of this could be connected. What Sam just said, Nvidia's acquisition of Hugging Face and the Neocloud issue. So, how do we pull this together? Well, first we have to remember that Nvidia has got 15-year lease commitments for some uh data centers uh or or for some compute that they're trying to actively pawn off on third parties. They say this here in this CF letter, "We signed data center lease agreements with terms of approximately 15 years that are expected to commence between 2028 and 2029 uh fiscal year. uh and we expect to reassign those data center leases to third parties. So given that we're in fiscal 27 right now, this basically means these leases are starting next year and the year after that. They're trying to pawn them off on third parties, you know, maybe private credit, but they're signing these contracts here saying, "Hey, we'll lease a bunch of compute." A jaded view of the hugging face acquisition by Nvidia is that they have all this compute that they've just secured and that maybe there is a little too much compute being built. And so Nvidia is coming in swooping some of this surplus compute. One of the companies they bought surplus compute from or promised to buy surplus compute from was Coref. And they said, "Hey, let us know if you ever need us to buy your surplus compute." And then they did. Core we've said, "Yeah, we've got surplus compute. We can't get rid of Nvidia. You said you would guarantee it and you would buy it. Here you go. Buy it." So now Nvidia is like, "Great. We just wanted to sell the chips. Now we have compute that we don't really need. What if we bought a business that needs compute? Huh? Welcome to HuggingFace. Hugging Face widely considered the GitHub of AI. They host like 3,000 different models and they're, you know, ARR businesses associated with hugging face subscription models and a lot of information about artificial intelligence and a lot of use for artificial intelligence. Hence why they have a nearly 14 billion uh acquisition. Some jaded views are that Nvidia oversimplifying here but that Nvidia can now utilize this power of controlling hugging phase to guide model adoption towards CUDA or Nvidia compute and potentially enable or support that varian compute with the data center leases they have now secured. So, in other words, you have this complete circle. Uh, and maybe it's a jaded point of view, but maybe that's exactly why Nvidia is going after hugging face. And does Sam Alman have a point about unsustainability in the Neoclouds? The best way to analyze that is to look at the financials for Coree, which we'll do in just a moment. Quick reminder, we did extend that coupon code to Friday over at meetke.com. So, make sure to join us over in the alpha membership. Remember, it could be a tax write off. You just go to meet Kevin.com. Join us there. If you want to bundle up the meet Kevin membership and the reinvestomes artificial intelligence software that we've created, you can do that at meetinvest.com. You buy both of those using the same coupon code expiring Friday. Uh if you have any issues, just email us at staffme.com. But that will be the last time the lifetime subscription will be available to lock in. So, how does that circular nature work? And then what are we actually seeing at Coreweave? And what's going on with companies like Credo? Okay, so the idea basically is if a company like Coreweave says, "Hey, we're going to build it and they will come. They build a bunch of compute." If they're only able to rent out a portion and they can't rent out all of their compute, uh then they don't need to buy any more Nvidia chips. Okay? So, if this is the unused uh portion right here, the dark portion, we'll call it kind of like dark fiber in the internet days, then they end up going to Nvidia and saying, "Look, we need you to take this over. We need you to do something with it." Well, now Nvidia is like, "Great. We were not expecting to have to take that compute so early. So, we need to put this to work. So, what can we do?" Well, let's buy something that needs compute or could utilize compute. Why don't we buy a software company like hugging face and we can put that compute to use. Now all of a sudden you can kind of compute complete the cycle where the additional jade comes in that says if you could then guide models to be a little bit more CUDA friendly which is that uh you know programming layer on the uh graphic cards that Nvidia uses massive moat uh in the artificial intelligence space then you could potentially drive more compute demand to the neoclouds and kind of complete that circle anytime they have excess capacity Nvidia will find a use for it. Fine. It's obviously weird that they talk about trying to offload this risk to third parties and it suggests that you don't really want this exposure. And I don't think it's a surprise that in their 10Q filing, they kind of purposefully downplay these leases. See, when you look at this chart, they have this chart written in here as they purposefully put data center uh supply and capacity. In other words, like, hey, uh, this is what we're promising to still supply. These are our supply commitments. Increasing our commitments for critical components for data centers. We're saying this is how much we're still going to spend on core components, high bandwidth, memory or whatever going into these chips. Then they purposefully try to bury this right here. Data center leases not yet commenced. These are those leases that they're committing to over a 15-year period. And because they can write this in billions, because they're comparing to their commitments, it doesn't actually look like that big of a deal. But starting next year, they literally have a one over $1 billion. We don't know because there's rounding because again, they're showing this in billions. They have over a billion dollars in data center leases expenses coming next year. That's a lot. So, they need a purpose for that. Again, this is the only chart they really show in billions uh that that where the numbers get so tight and small and it's fine. I get it. They've got a lot of big revenues. It makes sense. But a lot of other charts like we go over here and we want to look at, you know, gains that they have. Oh, what a surprise. These are being shown in millions. So, why are we trying to minimize those leases and make them hide? You know, we go to the cash flow statement. Oh, wow. It's shown in millions, but those leases are shown in billions because it makes the numbers look a little smaller. Okay, so now we kind of get the circularity of this. But is there a point? Does Sam Alman have a point? Because after all, the information ran a piece on Coreweave and they were actually pretty bullish on Cororeweave. They argue the following. They say that Cororeweave is the best Neocloud buy right now. And they make some fair fair points here. They say they're trading at just 2.2 times revenue. There's an opportunity for risk tolerant investors looking to bet on the AI boom. We've got uh 12.9 billion in revenues coming this year. That's four times that of Nebus, which is four times that of IRM. We've got uh you know 1.5 gawatts of active power, another 3.7 gawatts under contract. Okay, fine. So we've got a lot of enthusiasm that they argue is uh coming to coreweave but they argue is being discounted because of customer concentration risk notably Microsoft which is now going and buying their own compute. So, is there a risk that if Meta and OpenAI producing their Jalapeno chip uh and Meta acquiring their own hardware and Microsoft acquiring their hardware, is it possible that maybe they don't need Coree and is that the risk? They think that Coree is diversifying away. They've got Jane Street to sell to. Hey, there are plenty people who are going to rent this capacity. And maybe that's true. But I actually think there's a bigger story going on here because we have to somehow consider those data centers might need Nvidia to buy back some of the data centers. But if there's over supply in that compute, why is Core or why is OpenAI buying Mac Ultra and at the same time Coreweave trading for a discount? Well, I think there's a divergence that's starting. I think what's happening uh is the AI divergence is breaking AI into two various different components of chips. Uh you have specifically enterprise inference chips and then you have frontier inference chips. I'm not saying training and inference, okay? That ship is already sailed. That's old news. We already know there's training and then there's inference. What I'm arguing is there's frontier inference which is your biggest model, your fable 51s from anthropic, right? The latest and greatest thinking models from GPT. And then there's enterprise inference, which is exactly what the Max Studio Ultra would be perfect at and could potentially argue that why would OpenAI pay for enterprise inference if the vast majority of agentic needs that people have can run on a Mac Ultra? Well, then we don't need to pay for the GB300, the GB200 or H100s or other Blackwell chips. We don't need these. We don't need the Vera Rubin for that enterprise inference. And that could potentially explain the stock movement that we're seeing today. Although we want to be careful just to use a day uh in the stock market to argue, oh, that's it. This is definitely the trend that's happening. But it is quite interesting because look at the difference of these companies. Credo, which is down 20 almost 21% today, is a supplier of data center copper and photonics. Cerebrus is a supplier of inference chips that are high-speed tokenbased for enterprises. So enterprise inference up 5.5% today. Enterprise data centers Dell up 8% today. They don't supply the hyperscalers. They're not trying to sell hyperscalers Vera Rubin chips. They're trying to go to businesses and say, "Look, we'll get you localized data center compute in a box, in your own box. We'll we'll deliver the rack, the power supply, the switching equipment, the CPUs, the GPUs. We'll put it all together in a box for you and we only end up taking like a 6% net margin. So, why do it yourself? We'll set it up for you, and then we'll sell you our software to go with it, so it's nice and easy for the customer." So enterprise AI exploding, enterprise AI inference at least moving up today, though Cerebras has been a little bit of a lagger. Data center technologies like Credo slowing on margin issues. I'll show you those in just a moment. And then of course Coreweave the information argues is a buy, but then you get people like Sam Alman bagging on them going, I don't know, man. They might be getting a little speculative in their buildout. Well, this is where what's interesting is what kind of capital uh these companies have actually deployed. So, you look at a core, the majority of their chips, at least based on public filings and reporting on this, the majority of their chips are expected to be uh frontier level training style chips. Yeah, they're moving into inference. They're moving into the RTX 6000. They moved into that last year. Expectations are that's a very low exposure and enterprise inference is the next frontier. But if you're holding the bag on a lot of frontier data center chips, is there a debt risk that's keeping a company like Coreweave down? And personally, when I go look at the core financials, I think the answer is yes. We don't need to know that they're losing money. Technically, that's already old news. If you add back in depreciation, they're making money. But this is a return on the investment on their cash flow, which they're spending massive amounts of money on capex. They're spending 14 billion on 3.6 of operating cash flow. So, they're obviously needing to borrow money or raise money, which is exactly what they're doing. Here you go. They were they spent $14 billion on capex. They raised $3 billion in stock and net 11 billion in financing. Fine. Companies do this. They're trying to expand. They're trying to invest. But the issue with Coreweave is their balance sheet isn't that great. Their balance sheet compared to a company like Nbis is really poor. If you type into YouTube meet Kevin uh Nebius, you'll see my Nebius breakdown and their balance sheets actually in a pretty decent place. Coreweave, on the other hand, they have $5.5 billion in cash with $18 billion in bills. They have $3 in short-term debt for every $1 of cash they have. That's not good. that slows the build out and makes financing more challenging, especially if Sam Alman suggests, hey, maybe on the frontier level, we're starting to get a little excessive on compute. Jalapeno is actually potentially a discounted way to outperform some of that frontier level compute, which then suggests you have Jalapeno and potentially an overupp of frontier compute that could be weighing down the NeoClouds who are heavily exposed to frontier level compute where instead enterprise AI is sort of where it's at, the next big rally. Who knows, maybe that rally's already come. I mean, Dell's up like 3x, right? Maybe it's too late. But that's also where Nvidia rejecting 227 again today is interesting because a lot of people say, "Okay, so Nvidia wins either way. Nvidia wins with Frontier Compute. Nvidia wins with uh Enterprise Compute, does it?" Well, that's also where things are a little blurry. See, if you actually look and break down uh the graphics income uh which is exactly where Nvidia says that the graphics segment includes GPUs for gaming and enterprise workstations. So the Blackwell 6000, the 5090, whatever those are included in graphics. Graphics revenue is broken down right here compared to data center revenues, comput and networking right here. We can compare that to their operating income for those segments and we can come up with margins for those segments. We can see that data centers for Nvidia are pushing a 71% operating income margin, which obviously explains why Nvidia really wants to sell more data center chips, which explains why they might want to buy Hugging Face to kind of keep the Neocloud supported and keep those leases operating because they want to keep selling data center chips because they've got a 71% operating margin on that. Really, really good margin. the operating margin from the graphics side, so gaming and workstation AI and more enterprise AI only 49.2%. Still really good, but a fraction of the 71%. I mean, in fairness, nearly 50% operating income is still very good. It's just a whole lot lower than 71%. And when Nvidia forecasts even like a 1% miss on gross margins, people freak out and the stock ends up falling. So the margins matter a lot to Nvidia. So let's try to put all of this together because there's a lot of information here. Uh oh, after one more thing, I got to show you the Credo uh financials because again, remember how I just said margins matter a lot. Let's just do a quick look on Credo. Take a look at this. Credo, their their revenue goes up 9.6%. Oops. Revenue up 9.6%. But their cost of revenue went up 22.3%. So, their costs went up way higher than their revenue, leading their gross profit to only rise by 3% compared to May. Now, in fairness, these are not year-over-year numbers, but it shows you a weakening from August to May when you divide these numbers here. If I look at sales and administrative, like what they're spending on selling these their hardware, you're up 43%. The R&D part is up about uh 32%. Which then somewhat suggests or or sorry, they only pitch the R&D portion and total operating expenses are up 32%. Which is obviously way higher than their revenue growth. So their margin really got hit this quarter. I don't think they're necessarily overpriced. I actually think they're, you know, with net margins around 50%, they're selling relatively close to a one peg on, uh, you know, 26% 4-year forward growth could actually be a whole. It could be a buying opportunity, but I think the stock is down. I don't own it. I think the stock is down 20% today because the market hates when margin starts faltering. And so if you shift to enterprise compute, you're probably going to see more margin weakness at NVIDIA in the future. If you shift away from the hyperscalers and into enterprise compute, it's not margin accreative to Nvidia. It hurts margin for Nvidia. So let's put all of this together. Let's try to simplify this because this is a whole lot of information. Sam Alolman is basically saying there's excess supply in a portion of the market. So that's the thesis, excess supply. Our argument is where that excess supply is being seen based on his comments around the Neocloud and how they're going to use all their compute. Our argument is that excess supply is probably in the frontier inference. I'm not even talking about training frontier inference level. So he argues excess supply, we argue it's in frontier inference. If that is true, if that comes to bear, then yes, Neo clouds that have a lot of exposure to frontier inference are going to be sad. But also in the longer term, Nvidia's longerterm future margins are going to be sad if you shift to enterprise inference. the ones who win uh well also uh a company like Credo would be sad because you're seeing less of those optics or cabling going into big data centers and more going into enterprise artificial intelligence. The company that wins on enterprise AI right now is Dell. The problem is how long is that growth going to keep up? But then again, that's what people have been saying for the last two years. It's like how does a company with 4% margins go to five, go to six, go to seven, go to 8% net margins? Well, it's because of the AI boom, but how long does that go? So, that's a little bit of the warning broken down that Sam Alman's talking about. He doesn't specifically mention that it's frontier inference. That's my conclusion that there's this difference between frontier inference and enterprise inference and that's why we're seeing these moves in the stock market. And as an investor, it I've regularly been thinking, you know, where do you move money to get exposed to enterprise artificial intelligence? And the real answer is it's unless you're willing to take the low margin plays like a Corsair for example, which isn't exactly the cheapest stock, but a lot of workstation components go into a company uh or come from a company like Corsair. power supplies, even RAM sticks themselves. And honestly, if you look at a company like Corsair, IT'S IT'S BOTTOMED OUT AROUND, you know, uh the ceasefire days of uh the beginning of April. You know, the puppy was trading for like $5.60. It's now trading for 12 bucks. Corsair could be another low margin winner, much like Dell is this low margin winner of enterprise compute. If you want to get out of the infrastructure stack, then you have to get to the value providers and that's going to be software, which is really a topic for a whole another video. So that's my take on what Sam just said. Lot of info, but I think it's quite fascinating. >> Why not advertise these things that you told us here? I feel like nobody else knows about this. >> We'll we'll try a little advertising and see how it goes. >> Congratulations, man. You have done so much. People love you. People look up to you. >> Kevin Praath there, financial analyst and YouTuber. Meet Kevin. Always great to get your take.