How OpenAI ends and takes Oracle with it | Ed Zitron

Eventually, everyone will have to do token-based billing. There's no avoiding it unless they want to just continue annihilating billions of dollars.

My source at Abene says there's barely any hardware in there. Open AI needs to take residency of this so Oracle can pay its project management fees. So, even if OpenAI makes all this money, which includes them projecting to make $284 billion in 2030, they will have to raise at least $218 billion worth of funding or debt or something. I don't know. These numbers are insane. They're insane. And they are published by The Information of the Wall Street Journal like they're normal. They are not normal. And by the way, if OpenAI can't make that $852 billion, Oracle dies.

On the Tech Report with me again is the writer of Where's Your Edit and the host of the Better Offline podcast, Ed Zitron. Thanks for coming on.

>> Thanks for having me. So, Anthropic has overtaken OpenAI and has hit $1 trillion valuation, at least on the secondary markets. How has Anthropic reached this price and more importantly, how detached is this valuation from the reality of the company right now?

>> So extremely detached from reality. These are private secondary sales. They are low liquidity, low volume. Remember, these aren't like a stock you'd see in the public market as billions, 10 billion, 19 billion in some cases of daily volume. This is however much volume is on the secondary market. So probably thousands, hundreds of thousands at best. So, yeah, one person probably bought it, maybe two. So, this has happened because I don't know, no one really knows why any of this is worth anything anymore. So, everyone's just jumping around to what might be worth more next. I mean, OpenAI released Spud, their 5.5 model, and uh, Sam Altman said, uh, I paraphrase, I think he said, "Well, I really like it." It's the kind of thing you say to a child when it does something mediocre. So the real answer is that the excitement is gone from OpenAI and there's still kind of some excitement for Anthropic. So people are moving stuff around their plates. But I remember a few weeks ago there was $600 million worth of volume of OpenAI sh um allotment of OpenAI shifts that nobody could shift. So, I just assume that this is just a little bit of fluff jumping around a bit rather than indication of anyone's true value. Cuz I don't know if Anthropic, I don't know if either of these companies goes public, but if they do, I don't think Anthropic is going at a trillion.

>> On that point, you have described Anthropic's valuation as a sham built on venture capitalist subsidies.

>> Yes. Just to explain that a bit. I know obviously we've talked about this kind of thing before, but what would happen if Anthropic charged for the actual running cost of their AI or or another way of putting it, implemented their underlying business model without subsidization? So, we don't know what the API cost them. What we do know is they spend, they let people spend anywhere between like $2 and $4,000 a month on their $200 a month plan in tokens. Now, people say, "Oh, there's 50% gross margins." No one has any proof about this. I think in reality, the only way they could make their plans work is by having two of them. One that's $500 a month, one that's a $1,000 a month. And even then, I'm not confident that those would be profitable. So, I mean, we'll get to GitHub Copilot in a second, but I know that when you look at the real cost, when you look at how businesses that need to make this actually work do things, such as Anthropic with their enterprise customers, by the way, they charge on a per million token basis. Just I'm hinting at, by the way, that Anthropic moved their enterprise customers onto token-based billing. I don't even think there is a way of doing profitable monthly subscriptions to this stuff without just straight up charging for the tokens. And even then, I don't even know if serving the tokens is profitable. We've never had any proof of that. We've had Dario Amodei reference twice in the space of two years, stylized facts, his words, that it could have gross margins over 50%. Reporting says gross margins non-GAAP, so not generally accepted accounting principles, 45%. I think their margins are in the toilet and I don't think there's any profitable product that Anthropic actually has.

>> Like I say, we we'll carry on with Anthropic in a second, but let's get on to the the Microsoft GitHub stuff. They're looking to reduce their service with their Copilot service on GitHub pretty much across the board and move it over to token-based billing. I mean, what what is the significance of that? That's the that's the second company that I can think of off the top of my head with Anthropic being the second that is moving to token-based billing.

>> So GitHub Copilot is Microsoft's premier AI coding tool. It was for a time a very popular product and still kind of is. That's because it allowed you to burn way more than your subscription. They obfuscated the actual cost of things with requests. Fun fact, GitHub Copilot is also one of the top two largest clients of Anthropic for compute costs. That's because GitHub Copilot customers were able to burn as many tokens as they'd like within their requests. On Monday, I believe Microsoft is announcing that they're moving all GitHub Copilot subscribers to token-based billing in June. The significance of this is is that Microsoft is the single best capitalized, single most profitable, and single best positioned infrastructure-wise AI player. They are very profitable. They're the only one of the Magnificent 7 that's doing any significant capex investment just out of cash flow. They haven't had to raise as much debt as Amazon or Google. So if Microsoft is doing this, this means that Microsoft is feeling the pinch and that they're realizing that they need to start making costs, well, lowering costs. Now the question is is whether anyone pays for GitHub Copilot after this because it's going from, "Oh, you pay $19 or $39 a month and you get a certain amount of requests, a decent amount of them, even though they pulled the rate limits, they tightened them." I mean, now it's going to be, "Oh, you pay $19 a month, you get $19 worth of AI tokens." Uh, not really sure. I I guess you trade $19 for $19 and you can use GitHub Copilot. It's a sign that things are getting tight. And remember, Microsoft can afford to actually do the subsidies longer than anyone. They could afford to do it in perpetuity. If they're choosing not to, it means that CFO Amy Hood has got Sachin Nadella in a room and she's locked the door and she's told him she needs to see numbers go down, or else.

>> Like you say, Microsoft is the one that can afford to be doing this. So if we assume that soon this would imply others would be following suit. Can can the AI bubble survive token-based billing?

>> No. I mean that's the thing. The problem with this whole bubble, one of many I should say, is that they never should have done monthly subscriptions. It was a mistake from the beginning, but they tried to follow the tried-and-true Silicon Valley thing of, "We'll make it cheap at the beginning, then we'll jack up the price." The problem is is I don't think that there is an economical price for this. So, every user right now for the most part has been trained to have a subsidized product. They're not trained to pay their token costs. They're not paid, they're not trained. They're not used to doing per million input and output tokens. That to them is considered very strange. And if you look on the subreddits for any of these products, whenever they do a rate limit shift, whenever they tighten them, whenever they lower requests, they scream like they're being electrocuted. So, I don't think these people are going to react very well. But we're going to see within like 3 months exactly how popular these products are on their actual value. When you stop subsidizing the underlying compute, suddenly we'll find out whether they actually matter. Now, to be clear, GitHub Copilot is the only one right now that has actually moved to full token-based billing. Anthropic only did it for their enterprise customers, but the fact that Microsoft is doing it with their only really popular AI product, I mean, we're going to see how their users react. But after I scooped Microsoft on that story, the GitHub Copilot subreddit won't they they keep deleting the story. It's funny because it it's kind of like watching the dinosaurs look up in the sky and be like, "Is that is the sun getting closer?" I assume dinosaurs are smart enough to know what the sun was in this case. But nevertheless, it is going to be a real test of the validity of the AI bubble. I look forward to it. I actually think it's a good thing because this whole time everyone's yap yap yap, "AI is so powerful. AI is so important. It's so much productivity." No proof of that. But all of these stories that have been written about the power of Claude Code and all that have been based on subsidized accounts. And Anthropic indeed this week has started, well, I've never seen this in my career, AB testing pricing for Claude. They briefly removed Claude Code from the $20 a month plan. They've briefly they've randomly been showing people prices that anywhere from 30 to 100% higher. They're not really explaining what's going on. It's very weird and it suggests that everyone's getting very nervous. Now, to be clear, as Anthropic did all of these price checks, OpenAI is just going out there saying, "We've reset rate limits again. Reset rate limits, baby. Come burn our pile of cash. Come waste Sam Altman's money. You can do it, folks." So, it's it's a real it's a very bizarre situation because eventually everyone will have to do token-based billing. There's no avoiding it unless they want to just continue annihilating billions of dollars.

>> One thing that I did see that was again, this is Nvidia's marketing. They said that their new Vera Rubin GPUs will be uh reducing the energy usage per token by up to 35, 35 times, not percent, 35 times cheaper. I mean, that sounds too good to be true and maybe one day in the far future we we'll get there, but that that is that is not that's marketing, right?

>> Yes. And the thing is, they said that it was 10x cheaper between Blackwell and H100s too. And also, these these benchmarks are always for very specific benchmark tests that Nvidia runs using open-source models. It would be, you know what, if they went out there, I would eat my words if they went out there with a defined study that actual technical people looked at and agreed with that said, "We have made GPT 5.5 10 times cheaper to run or 35 times cheaper to run." I I'd eat my words, but they're never going to do that because no one wants to talk about the actual cost of running these models because I think they're really, really high. We're still in a situation where we do not know how much it costs to run OpenAI or Anthropic's models. We know their overall cost, but we don't know the actual line costs. And we won't unless one of them files that S1 that I got to I got to see that S1. But I think that Nvidia is always going to say that that I what's funny is I swear to God when Vera Rubin was announced, they said 10X, but I read 35X too recently. So it's just like, ah, which one is it? Jensen. Jensen, what what are you talking about, mate? Like, is it 10, 35, or is it just a number you come up with every time you're asked?

>> So, if we imagine AI is now paid for by the token, user bases across the board have dropped, which were quickly followed by valuations. What happens next? I mean, are there any companies that are left alive still doing this kind of stuff? Is is there that techno-feudalism, the the rich people have AI and the rest of us are living in the dirt kind of thing?

>> I don't think that that's how that goes. Because here's here's the whole thing. Even the most belligerent AI booster will have trouble telling you how much it costs to do anything. Because the inherent unreliability of large language models means that you can't say with any guarantee how much or how long anything will take. One day a model might oneshot something. One day it might spin its wheels and come back and go, "I deleted everything." And I'm not even being facetious. This happened to someone on Twitter talking about Opus 4.7 yesterday. This is the large language model problem. And on top of that, most people who have used large language models have been trained to use it with a subscription with a rate limit. They don't know how to run a token-based economy. Every report I read of someone actually paying their rates, they're talking $300 to $500 a day. Anthropic themselves I think says $6 a day for the average user of Claude Code with 90% being below $12 a day. That's a lot of money. You're like, e even on a low level, you're talking hundreds of dollars a month and I think it's more like thousands, especially for an engineer. So what will happen is you will start seeing companies, I mean, you had the Uber CTO say they blew through their token budget, their AI token budget in a few months, their entire budget within a few months. At some point companies are going to go, "Wait, is this losing us more money than we're gaining? Is this inefficient?" Because they haven't had to before. When you eat the buffet every day, you don't have any idea how much a meal costs. And so you're going to see people make hard budgetary decisions and suddenly the gifted child of large language models is just not going to be that impressive anymore. It's just going to be like, "Why am I paying $100 a day? What am I doing with that?" There was a Goldman Sachs report that said that there were some businesses spending 10% of their headcount on AI tokens and that that could increase in the next few quarters to 100% of headcount. And isn't the whole point, we're not that I like this idea, isn't the whole point we're replacing people and removing the cost of a person? We've added more people except instead of having a person that can learn and grow and kind of interconnect into the organization and strategize with people, we have just money to Dario Amodei and Sam Altman. I'm not sure most people are going to like this so much, especially because no one can really explain the productivity benefits. Every study feels like it's written by the Riddler. So, it's I think we're just going to have a moment of truth. I think we're we're going to see an unvarnished evaluation of this. The real question is who is the last man standing? Because GitHub Copilot, GitHub Copilot is the first one that will be moving to token-based billing. Who's next? Because it's the question is who moves regular subscribers to token-based billing? Is it OpenAI? Is it Anthropic? It won't be Grok. I think Elon Musk's going to be the last man standing. Him and Michael Trull of Curser, hand in hand, as Elon does not pay them. I think it's probably OpenAI will be the last to drop. If Anthropic does token-based billing first for everyone, I honestly give it to them. But people are going to be so angry. This kind of brings us back to that valuation, which admittedly, like you say, is is not massively representative, but it does beg the question why people are buying it at that valuation when the company is struggling with capacity issues that we've mentioned, especially that it could never pay them off and is also there are growing complaints over deteriorating product, which is gathering more and more evidence from AMD big wig saying actually no, there are numbers to this, it's not just a bug or people not using it correctly. So the thing is with the getting dumber as well, they've they spent, Anthropic spent a month saying, "We've not changed the models at all." Boris and the the munch bunch over at Anthropic, they said yesterday, "Oh, actually, yeah, we tweaked something. It did make it dumber." Just lying. They just gaslit their customers for a month. They their compute capacity, and I mentioned this before, last week, I believe. It's like it's not a capacity issue. They would be fine with capacity if they just stopped accepting customers, but they'll take any money they can. And I think what it is is that there's nothing else to really invest in. There's not there's not any other big startups that are going to have any kind of multiples. There's nothing else that's really exciting unless you want to pretend that robotics is going to happen again. I thought we tried that with the iBO about 25 years ago, but we're back, baby. We're going to pretend that robots work. You can't really invest in the public markets. Uh, what other venture capital swindles are there? Crypto. We No, I can't do that again. Meta. I had someone mention the metaverse again the other day. That's not a good sign. Yeah. Uh, we didn't we didn't need a second Grinch. But it's it's interesting because I think people are just moving around just being like, "What are we all doing? Anthropic. I guess we're excited about that now." It all feels very like like very uh half-hearted at this point. You can't really find anyone who is telling you with any certainty why this is happening. You've got the new the new um plate spinning thing they're doing is they're claiming there's a CPU shortage now. They're claiming that agents use more CPUs. And it's just like the moment I saw that I'm like, "Oh, Christ. You're just going to in in 15 weeks you're going to be saying we have a storage crisis. We need more NAND storage and oh, we need bigger cases. We need cases. We need more water cooling." They're going to find every possible part of the hardware stack to get excited about before one of them finally starts the car in a garage. And it's like it's frustrating because the longer this goes on, the worse it gets. The more money that gets into this, the more that is wasted. It really comes down to going back to the token-based billing. Once one of the major labs goes full token-based, that's how you know it's over. That's how you know that they are saying, "Crap, we can't do this how we're doing this." And you know what? Maybe they'll pull it off. Maybe everyone instead of paying $200 a month, we'll pay $1,500 or more. I'm sure that people will do that for exactly the same, if not slightly worse, because that's what's coming. Because the crazy thing is is I hear that the API, Anthropic's API, is less stable than using it on the subscription. So you have the indignity of paying more and you get worse service. That this is the thing. Anthropic ostensibly worth $380 trillion dollars or bazillion dollars, however much they run their company terribly. Their service is inconsistent. They have less than 49s of uptime. They have a CEO that talks like a duck that was told turned into an old man that tells us that they're going to fire 50% of all white collar workers within a year or two years or 6 months or 18 months or 50 years. We don't know. They gaslight their customers. They have some of the most annoying people boosting them. It sucks. It's It's not fun. This isn't fun for anyone. I don't It It's the worst possible bubble we could ask for. It loses the most money. The people are boring and annoying and think they're philosophers, but they're just regular guys who have read like Atlas Shrugged and The Dictionary. They're just boring and the products are dull and the excitement is fake and the valuations are fake and the economy doesn't make sense. It's just so upsetting and annoying and wasteful. We we've mentioned this before a bit, but Mythos, the the too dangerous for public consumption thing that was now now that we're a little bit further on, we've had some time to sort of digest all of this stuff that was more about them not them not having the capacity to release a model like that to the public. Right.

>> Yeah. And also the reports have come out of organizations using it. Firefox came out and said, "Yeah, it was quick at find it's quick finding bugs, but didn't find any like crazy stuff. It's not better than a human." They were very like lukewarm on it. A security researcher did a big report on it as well that basically said that they kind of ginned the numbers a bit to seem a bit more impressive. The whole thing was a farce. The whole thing was just hype. And I'm actually really grateful to Anthropic for this because without Mythos, we would not have known who the most credulous people in the world were. We all Anthropic had to do was do a press release because let's be honest, half the people that went on TV talking about this didn't even look at the system card. They got they found the people. They went, "Our model is going to wake up and kill you. Our model is going our model is in your crawl space. It's in your walls. It's poisoned your dog." And these people were going, "I saw someone on the TV the other day in the UK going on one of the supposed godfathers of AI going, Mythos could hack every system. It's so dangerous. It's so..." You've not used it. Also, it can't hack. That's not what it does. It's a bug finder. It can't hack. Even though they claim it can do these exploits, it didn't actually action any of them. Ah, it that's the thing though. Anthropic is just it's a it is a credulousness magnet. It's just people that are ready to believe anything. If it's convenient to their narrative or convenient to their theoretical pocketbook or just an excuse to get on TV and seem scared, then yeah, Anthropic is great for that. If you want to find a dishonest company that runs a bad business that sells things through lying, Anthropic, this is the business for you. They lose billions of dollars and hope they to lose billions of dollars in the future. Would you like to invest now or in the FE? I'm just so tired of them.

>> Well, let's move on to data centers because you reveal in your premium newsletter that 15 months into Stargate now, much less of it is under construction than their public statements would have you believe.

>> and barely any of it is online. H just how far have they fallen off schedule and what does it mean for Oracle who's building this on razor-thin margins and is currently the only one footing the bill for it?

>> Okay, so Stargate Abilene, a 1.2 gigawatt capacity. So about 824 megawatt data center of eight buildings was meant to be done, was meant to be energized by the end of 2025. N sorry, I mean the first quarter of 2026, uh, middle of 2026, October 2026, not sure. There are now two buildings operational of eight. Two in December of last year, they had the chips delivered for them, not installed, delivered. They have finished a third building. Finished does not mean filled with anything. In fact, my source at Abilene says there's barely any hardware in there. Open AI needs to take residency of this so Oracle can pay its project management fees. Oracle will be spending when this thing is fully built about $2.14 billion a year, maybe more, and it will make them theoretically $10 billion a year in revenue. But nevertheless, they need to complete the bloody thing first. Based on discussions of people on the ground, Stargate Abilene will not be online anytime before Q1 2027. The overall Stargate data center project is about 7.1 gigawatt of data center capacity. Nothing to do with Donald Trump. That whole thing was a sham. No government money. Very unfair. Nevertheless, those data centers in totality will cost about $340 billion. They will theoretically make $75 billion a year in annual revenue. Few problems with that. These data centers are meant to be done end of 2028. If these things are done before 2029 to 2020, well 2031 at this point, I will eat my hat after slow smoking it on the pit. There is no way these things get done. And by the way, if OpenAI cannot pay Oracle, Oracle dies. This is not hyperbole. This is the financial situation Oracle has got itself into. There is no other tenant that can afford that. $75 billion a year is so much money. And based on projections from the information leaked from OpenAI themselves, OpenAI needs to make in revenue and raise in funding and debt $852 billion over the next four years. I'm not kidding. They also say they'll be making $673 billion of revenue. Now, Isaac, no you don't have a calculator, but 673 is is smaller than 852 billion. So even if OpenAI makes all this money, which includes them projecting to make $284 billion in 2030, they will have to raise at least $218 billion worth of funding or debt or something. I don't know. And um the only way this works, by the way, is their projection says they become cash flow positive to the tune of $39 billion in 2030. These numbers are insane. They're insane. And they are published by The Information in the Wall Street Journal like they're normal. They are not normal. And by the way, if OpenAI can't make that $852 billion, Oracle dies. Oracle cannot pay its bills. Oracle had negative cash flow of nearly $25 billion last quarter. It is having to raise debt in this weird Byzantine way where they Oracle doesn't actually take on the debt. It puts it off balance sheet on a project financing loan that is then put into a special purpose vehicle, which they are still on the hook for. So Oracle doesn't own anything, an SPV does, but Oracle has all of the risk and literally they cannot afford to pay for it unless OpenAI can pay them $75 billion a year. And got some more bad news for Larry Ellison fans. I know there are many of them in the audience. Larry Ellison is he has pledged 346 billion shares of his own shares in Oracle, about 60 something billion dollars worth, on personal loans based on his Oracle shares. If Oracle's deal falls apart with OpenAI, Oracle's share price will fall and when that happens, Mr. Ellison will face margin calls from hell itself and there is nothing to fix it. People are going to read hear this and they're going to go, "Oh, oh, Donald Trump's going to bail him out. Donald Trump is a bad man's going to..." No, because even if there was a bailout of Oracle, even if there was to somehow deal with these horrible project financing, the stock of Oracle is going into the toilet and being flushed aggressively because Oracle's other businesses, hardware licensing, software licensing, and the like are plateauing. Their cloud businesses are the only things that's growing. And by the way, they have negative to low margins. OpenAI is paying them theoretically $10 billion a year for Abilene on completion, but that's only if they got a good price. It could be lower. And just to be clear, Oracle needs $75 billion. And OpenAI needs to be able to pay $75 billion a year. And also towards the $138 billion deal it did with Amazon Web Services, towards the $250 billion it owes Microsoft Azure, the $20 billion it owes Cerebrus, the $22.4 billion it owes CoreWeave, there's some money they own owe Google as well. They're meant to be installing 10 GW of Broadcom GPUs by the end of 2030. None of this makes sense. None of this makes sense at all. Where is Michael Bur? Where are the people that Where are the I I love doing this job. It's great. But there's a certain level of like, "Hey guys, anyone else want to do the math?" Because this Oracle Oracle's going to die unless they walk away from Stargate or massively renegotiate these contracts. And I just don't see that happening. And I also don't see OpenAI actually being able to pay. And there's no one else, no one that can afford these deals unless you choose Anthropic, who has exactly the same problem as OpenAI. When you were going through the the ever-shifting deadline for Stargate, the one thing that obviously this is the question we're going to talk about it now. The GPUs that are supposed to be put being put in there with the construction rate and then you add in the installation time of those cards, are they even going to be able to be finishing these data centers before they have to start well, just taking the GPUs back out again?

>> Well, that's the funny thing, Isaac. They can't take them out. So Stargate Abilene, by the time it's complete, will be full of GB200 MVL 72 racks. 72 racks of 72 GB 2000 GPUs. Those things will be 2 years old by the time that Abilene is fully operational, by the time the other data centers are finished, and they're most likely going to be Vera Rubin GPUs. Those two will be about two, maybe 3 years old after I think there's two generations of Vera Rubin. I know, very confusing. First Vera Rubin generation fits in the same racks as Blackwell Oberon racks. Mr. Jensen Huang has already said that they are moving on to Kyber racks. Big massive completely different 600 kilowatt, about five times uh five times the power on them. Uh, yeah, completely different cooling required. Just you can't plug and play that. You'll have to demolish the inside of the thing, just like people had to do between the previous generation and Blackwell. So yeah, everything that you're hearing about data centers is already obsolete. And it just means that, and the worst thing is is that with Blackwell especially, all of the data centers you've heard announced recently are all Blackwell. So on top of the fact that you've got this nearing obsolescence within the next 3 years, I'd say GPU series, everyone's got them too. It's not like before where there were a lot of H100s and 200s. This is there are so many more Blackwells, millions of these fle these bloody things, and everyone has them. And so as data centers come online, which is taking forever anyway, it's going to be full of these decrepit chips. And you already have a situation with Super Micro, a uh embarrassed a a dishonored Neo, not NeoCloud, um ODM, original device manufacturer. They build the servers, put the GPUs in them. They have over a billion dollars worth of Blackwell B200 GPUs that they cannot sell. They're languishing in inventory. We're already seeing the signs. And that's kind of why none of this makes sense because you buy the GPU in 2025, you install it by what, 2027? 2 years later, you have the the two years earlier gear. This is everything is misaligned. Everything is misaligned and there's no aligning it unless we find a way to build data centers and power and also get uh transmit, sorry, um transformers are in low supply and also electrical grade steel and also the talent to build it, also there's not enough water, also we don't have the power, we're running out of cable, also there's a war in Iran. If we can solve all of that and invent a new way to do data centers, we might be slightly on on it's not happening. It's not happening at all.

>> You hinted at a little bit of news which I I think has generally gone kind of under the radar, which Oracle cancelled about $1.1 to $1.4 billion dollars worth of chips from Super Micro. I mean, you kind of hinted at it, but what what is that a sign of?

>> Well, Super Micro's co-founder, Wally VR, got arrested for selling GPUs to the Chinese. So, it could be related to the fact they're dodgy. It could also be the fact that Oracle can't afford them. They they ended up buying the servers from a company called Wistron, which is one of that's the name. Uh, it's out in Taiwan, one of these original uh design manufacturers like Hon Hai, Foxconn, and all them. That's where all the hyperscalers buy their GPUs. They buy what happens is a company buys the GPUs in Taiwan from Nvidia, puts them in a server, designs the server, sells them to a Microsoft and Oracle or what have you. Nevertheless, it could be they're dodgy. It could be the fact that Oracle didn't need the B200s anymore and they wanted the fancier new ones. It isn't really clear, but Super Micro is in a bad position. They had an accounting scandal two years ago. They've had people arrested before. This isn't even their second rodeo. They have a very bizarre situation going over there and they've claimed that it's all sorted now. Bob, if Oracle is canceling with them, it's either that they don't trust the company in general, they don't trust them to actually deliver them, or the gear wasn't what they needed. And I mean, Super Micro might be in real danger. A billion dollars of inventory for a company with a I think 10 or something billion dollars annual revenue. Forgive me for not remembering off the top of my head. That's not insignificant. And also it will be the first write-down of present Nvidia inventory in history as far as the AI GPUs go.

>> What happens once we we reach that point where these GPUs are being taken back out of the racks maybe before they've even been put in? Is there a mechanism for for companies to sell them back to Nvidia for a discount on the next round or are we going to be seeing a massive flood of them hitting the second grade market which

>> Oh, I don't think you're going to see them replaced. I think I think that these Blackwell data centers are going to turn on and they're going to do what they can to shift the compute. I heard a truly insane story just before coming on here from The Information that they're barely getting $3.70 an hour per Blackwell GPU renting them out, and that's on the higher end. I don't think that even covers the cost of building the infrastructure to house them and the debt associated. $3.70. I had a quote from a source over at Oracle that said Blackwell B300 GPUs cost about six bucks an hour a GPU. So yeah, I don't think $3.70 is going to cut it. And there's a capacity crunch cuz data centers take forever to build, but I think what's going to happen is they're just going to kind of sit in it because Nvidia's not going to buy them back. Why would they? Well, what possible reason would Nvidia buy them? So if we see these GPUs dumped on the market at scale though, that's when things start getting hairy. But I think what's more likely is we see one of my pale horses, which is an in a fully built data center goes under. I think one goes under and at that point that's when we start we start seeing everyone get a little bit worried, which they should have been before they built the things. If the GPUs do end up getting dumped, where does that leave the companies that have tens of billions of dollars in GPU back debt and and then the people that then gave them that loan on on the basis that these GPUs might be worth something when if it ever came to that?

>> It heavily depends on the nature of the relationship. If these are asset-backed securities that have some sort of margin call related to the value of the GPUs, that could lead to a margin call because GPUs won't be worth as much. If it leads to mass impairment of GPUs, you will see 5 to 10 billion chopped off a quarter off of multiple hyperscalers. It really depends how significant the sell-off is. It might take another 6 months for this to begin. It might be a case where I don't know, it just a random data center maybe Ellenale, North Dakota falls apart because they're losing a million dollars a day, just an example. It might be that they can't afford their actual loans. I think you're going to see a wave of refinancing within data centers. You're going to see these companies trying to get a better deal while the lights are still on. I think at some point someone's going to fall behind. Maybe it's CoreWeave. It might be Nebius. Nebius is building a data center out in Vineland, New Jersey. Other than the fact they're not and construction got halted, uh, they still claim that they're on schedule. A lot of that going around. A lot of companies not on schedule saying, "Oh, yeah, we're fine. Everything's good. We're not doing anything." But that just means we're more ready to do stuff when we can. And it's it's interesting because to this day, we still don't have a very good understanding of how much money these things make. We all of it still obfuscated. This thing. We have now built well invested at least a trillion dollars in. We do not know how much it truly costs. That's a sign that it the costs aren't good because everything else, any other business, we know where the margins are good. We know everything about it because they're boasting about it. So I think I don't know. CoreWeave's next earnings should be real interesting. I think that it's looking like barely any data centers are getting built before 2028. I don't think Stargate makes it, but I think that instead of these GPUs being ripped and replaced because the moment you put them in service, the clock starts on depreciation. And if you mark them down, you have to mark them down immediately. So, they're going to try and use them. They're going to do everything they can to sell that compute, even if it means taking a loss. And at some point, someone's going to take too big a loss and they're going to run out of money.

>> So, just finally, is a good time to ask this question, I think. Does Oracle have any destiny ahead of it after all of this other than bankruptcy, other than perhaps maybe recouping some cost from OpenAI, or is that just going to be bringing them down, bringing them crashing and burning down with them?

>> Oracle could theoretically get out of Stargate data centers. All of the data centers are funded under SPVs, separate entities. If they really wanted to play legal, I don't know, like the movie Duel, they just want to go straight at their debtors. They could just refuse to pay. They could attempt to swindle them. Oracle loves suing. They love they have a great series of lawyers. They're famous. I've heard journalists refer to them as a law firm with software attached. But the problem they have is that their only growth right now really is from AI GPUs. Without that, most of their businesses are barely above a straight line. So they don't really Oracle doesn't really have another choice. There are ways of recapitalizing that company. I think if it runs out of money, it will be passed out. It will rip bits of it off and send to Microsoft and whoever. I just don't know how it survives this. OpenAI needs $852 billion in the next four years. That is not an like that is an alarming figure to me, but it is not hyperbole. It's what The Information and the Wall Street Journal have reported. This is these are the numbers. If they cannot do that and Oracle builds these data centers, Oracle is screwed. The other way out is that Oracle just doesn't build them. That Oracle just theoretically the project could shut down. They could shut down Michigan. They could shut down all of these others. But they just got this debt and they Oracle has really strung themselves out to get this debt. So at this point, I don't know what saves Oracle from the void. I don't think Ellison has any other plans. And while I think the Paramount deal goes through, I think that in the end he they're going to end up selling Paramount. They're going to end up selling CBS because they're not going to have the money to pay for it. I think they're going to sell their share of TikTok probably back to ByteDance at this point. Could take a year or two. Could be really quick. It really depends on what breaks and how it breaks. And a lot of that comes down to OpenAI paying its bills and these contractors actually finishing these data centers.

>> Well, Ed Zitron, thanks for taking the time.

>> Thanks for having me. If you enjoyed today's episode of The Tech Report, please consider liking and subscribing. Also, you can get episodes of The Tech Report wherever you get your podcasts.