transcript

Speaker 1:
[00:05] Hello and welcome to this week's Better Offline monologue. I'm your host Ed Zitron. Forgive me if I sound a little congested, I'm sick, I promise I'm resting up, but this is an important one. Today I'm gonna talk to you about something I'm brewing over in my premium newsletter. It will be out later today, but this is information you need to know and don't need to cough up money to find out. There will be more details, but you're gonna get a lot today. So I wanna be very clear about something. For OpenAI to be able to pay the deals it has signed, with all of these different companies, Amazon, Microsoft, CoreWeave, Cerberus and Oracle, it will have to be 10 to 15 times its current revenue and have raised at least another $150 billion. If it fails, Oracle will run out of money and not be able to pay its debts and probably shut down. And I know that sounds insane because Oracle is so large and has been around for so long. But let me explain. Oracle is building around 7.1 gigawatts of data center capacity for OpenAI, with the majority of it looking like it will be finished somewhere towards the end of 2028. An absolute best, if not the end of 2029 or 2030, if it ever gets built. Stargate Abilene, the only one of them anywhere near completion, started in the middle of 2024, sits two years into construction at two buildings and around 200 megawatts of capacity out of the eight that are being built. Eight buildings, that mean. Based on reports throughout the last two years, Stargate Abilene was meant to be fully energized somewhere between, I don't know, the end of 2025 and the middle of 2026, yet sources on the ground tell me they don't expect this thing to be finished anywhere before April 2027. Per a source familiar with Oracle's infrastructure, building three has finished construction and been handed over, but barely has any gear in it. It'll be months before it starts generating revenue. Now based on discussions with sources familiar with Stargate Abilene's infrastructure costs, the total cost will run somewhere in the region of $48 million per megawatt, which is in line with an estimate of around $44 million per megawatt for any data center construction that I got from Jerome Darling, an analyst over at TD Cohen. In total, this puts the estimated cost of Stargate Abilene at around $59 billion and the overall 7.1 gigawatt project, so far at least, at an estimated $340 billion. Sources have also told me that Oracle estimated in 2024 that it would spend over $2.1 billion a year on co-location lease and power cost to land developer Crusoe. And that's just because Crusoe actually, they lease the land from a company called Lansium, they then lease that land and the stuff on it to Oracle, and then Crusoe handles the power and then passes through the cost to Oracle. It's complex, it's annoying, but that's what it is. Stargate Abilene will, based on estimates from landowner Lansium and discussions with sources familiar with hyperscaler billing, generate around $10 billion a year in revenue when it's complete, at a rate of around $12.5 million a megawatt of critical IT, so 824 megawatts. The way it works is it's a 1.2 gigawatt campus, but there's only so much of it that's actually critical IT that would be billed for, so 824 megawatts worth. Now, while this might suggest that Oracle is making $8 billion in profit a year, and that's not true, one has to reckon with the astonishing cost of building this data center, and of course other costs like people, and keeping like insurance and stuff. Anyway, each building contains around 694 GB 200 NVL 72 racks, 72 GPUs each, each retailing around $3.5 million, meaning that each building contains around $2.43 billion of GPUs, for a total of $19.44 billion for the entire campus. As Oracle depreciates those GPUs over 6 years, that comes to around $3.24 billion a year in depreciation costs, so spreading them out. I can also confirm based on discussions with sources familiar with the Abilene project, that Oracle is footing either large parts of the entirety of the construction of these data centers, and has spent, as of recent, over $5.4 billion in construction and infrastructure costs, not including GPUs, on just the first two buildings, with an estimate that they'll spend up to $10 billion, despite it being operational. And there are tons more expenses, though. It's very weird. Now, on a strictly mathematical basis, this means that Oracle will end up, assuming the project is ever completed, spending around $60 billion to build a data center campus that will make it about $10 billion of revenue. Inside said data center campus will be racks of 72 GB 200 GPUs that are already a year old and will be, in theory, two years old by the time that the campus is complete. Now, the later data centers, the ones that aren't in Abilene, will be likely using NVIDIA's Vera Rubin GPUs, which means they'll pay a little more. Well, I mean, they'll be paid a little more by OpenAI, again, in theory, likely around $14 million per megawatt, which means, again, in theory, Oracle will get paid, assuming everything gets built, around $75 billion a year in revenue from OpenAI, if OpenAI could afford that, which it cannot. For context, OpenAI estimates it will have around $30 billion in revenue this year, and lose $25 billion to make that. In 2027, it projects to pull in $62 billion in revenue and lose $57 billion to make that. And in 2028, it will lose $85 billion and allegedly make $115 billion. OpenAI has also agreed to spend $22 billion with CoreWave, $138 billion with Amazon Web Services, $250 billion with Microsoft, and $20 billion with Cerebrus on Compute. Yeah, so for OpenAI to pay all of its bills, it will have to raise at least another $150 billion, and I mean on top of the money it just raised, if not $200 billion. And that's only, only, if it makes over $100 billion in annual revenue by the end of 2028. And if I'm honest, I'm not sure that that math even makes sense even if it does so, because we don't know the precise year-by-year breakdown of its other deals. I genuinely think they may be spending, if they actually kept the agreements, $125 billion in compute by 2028. I mean, they don't have the money, they can't do that, but nevertheless, it's... how is no one else worried about this? This would also, by the way, mean that OpenAI grows its revenue by four times in the next two years and by nearly 10 times by 2030. And I think they need to be 10 to 15 times to pay all these fucking bills. Sounds implausible, right? Well, it has to happen, otherwise Oracle runs out of money. In its last quarterly earnings, Oracle had free cash flow of negative $24.7 billion and has, through both completed and planned bond sales, debt financing and at-the-market share sales, raised over $115 billion, which is not sufficient to complete the construction of the remaining Stargate datacenters. Oracle has also raised a great deal of that money using construction project financing, keeping the debt off balance sheet and tying its repayments entirely to cash flow from the various projects revenues that are being paid by OpenAI, a company that's going to lose hundreds of billions of dollars in the next few years if it doesn't fall over and die. Jesus fucking Christ. If those revenues don't come through, say OpenAI just didn't pay them, Oracle will be unable to pay its debts. This is not an opinion. This is maths. It's other businesses, Oracle's by the way, in hardware and software licensing, they're plateauing. They have been for quite some time, and its only growth market is renting out AI compute using GPUs that burn out fast and are upgraded on a yearly cycle. Meaning that by the time Stargate Abelane is completed, its Blackwell GPUs will be two or even three years old. In fact, all of Stargate will be full of years old GPUs if it ever gets completed. But think about it, pretty much every datacenter you know is going to have obsolete GPUs. Because if it takes two, three years to build one, and these ones are being filled with Blackwell GPUs, well by the time the fucking things are done, well you can have a thing full of years old GPUs. Same thing is going to happen with Vera Rubin. And I don't know, the next one, I think they're just going to call them Red Fu. They're going to name them after one of the blokes from LMFAO. I hear Jensen Huang is a huge LMFAO fan. He's just blasting shots, shots, shots, shots, shots, constantly at Nvidia, party rock anthem plays at every board meeting. It's very weird. If you've heard anything, this is all made up. I'm just having some fun. But look, look, look, look. I'm deeply confused that nobody else is on this. I'm genuinely, genuinely confused. Where the fuck is Michael Burry? Michael Burry came back out of the shadows, Cassandra Unchained. He came out of the shadows. He was going to blog about this stuff. He did one blog about depreciation with NVIDIA and Meta, and then a real big juicy wet kiss to Jack Clarke over Anthropic in a supposed debate about the AI bubble that mostly involved Michael Burry saying, Claude's so good. I love it so much. What the fuck is the point of someone like Michael fucking Burry? If he can't look at something like this, the numbers are there. Oracle is very likely going to die unless it starts backing away from these data center projects and all signs point to it accelerating towards building them as fast as possible. Even if it succeeds, OpenAI cannot afford to pay for them. Seventy five billion dollars a year is so much money. I think Microsoft's operating costs like 150 something billion, and Microsoft's very, very, very profitable. And I sat and thought about how I might be wrong about this one a lot, because it's a huge claim to make, but I cannot find a way that Oracle actually makes this work. And every time I tell somebody about it, they just respond with inshallah. Anyway, if you like all the sound of this, check out the premium newsletter that will be out later today. There's a discount code in the show notes, link, click it, please, for principal form of income now. But even if you do not pay to subscribe, there is a generous free section and a summary at the bottom of it that will give you a great deal of the story. You'll be able to get the highlights. Now, the beat will be in there if you want to pay, but I don't want to gate too much of this information. I'll be back next week. I think I'll have an episode on Wednesday, but if I'm honest, if I'm too ill, or if I'm just rundown because I've had family in the hospital, I might skip, but I think I've got one in session. Either way, catch me on the Reddit, shoot me a Slub on Plurk, whatever, however you want to contact me. I love hearing from you all. Cheers. Love you all.