WEBVTT - Monologue: How OpenAI Kills Oracle

0:00:02.400 --> 0:00:08.240
<v Speaker 1>Cauzone Media. Hello on, Welcome to this week's Better Offline Monologue.

0:00:08.320 --> 0:00:18.360
<v Speaker 1>I'm your host ed z tron. Forgive me if I

0:00:18.400 --> 0:00:21.480
<v Speaker 1>sound a little congested. I'm sick. I promise I'm resting up.

0:00:21.680 --> 0:00:24.279
<v Speaker 1>But this is an important one today. I'm going to

0:00:24.320 --> 0:00:25.840
<v Speaker 1>talk to you about something I'm brewing over in my

0:00:25.880 --> 0:00:29.520
<v Speaker 1>premium newsletter. It'll be out later today, but this is

0:00:29.560 --> 0:00:31.400
<v Speaker 1>information you need to know and don't need to cough

0:00:31.480 --> 0:00:34.159
<v Speaker 1>up money to find out. There'll be more details, but

0:00:34.280 --> 0:00:36.920
<v Speaker 1>you're going to get a lot today. So I want

0:00:36.920 --> 0:00:40.080
<v Speaker 1>to be very clear about something. For open Ai to

0:00:40.080 --> 0:00:42.640
<v Speaker 1>be able to pay the deals it is signed with

0:00:42.800 --> 0:00:47.120
<v Speaker 1>all of these different companies Amazon, Microsoft, Core, with seah

0:00:47.120 --> 0:00:49.600
<v Speaker 1>Brus and Oracle, it will have to be ten to

0:00:49.680 --> 0:00:52.240
<v Speaker 1>fifteen times its current revenue and have raised at least

0:00:52.240 --> 0:00:55.320
<v Speaker 1>another one hundred and fifty billion dollars. If it fails,

0:00:55.680 --> 0:00:58.200
<v Speaker 1>Oracle will run out of money and not be able

0:00:58.200 --> 0:01:00.960
<v Speaker 1>to pay its debts and probably shut down. And I

0:01:01.000 --> 0:01:04.000
<v Speaker 1>know that sounds insane because Oracle is so large and

0:01:04.000 --> 0:01:06.560
<v Speaker 1>it has been around for so long, but let me explain.

0:01:07.240 --> 0:01:09.920
<v Speaker 1>Oracle is building around seven point one gigawatts of data

0:01:09.920 --> 0:01:12.240
<v Speaker 1>center capacity for open AI with the majority of it

0:01:12.280 --> 0:01:14.399
<v Speaker 1>looking like it will be finished somewhere towards the end

0:01:14.440 --> 0:01:17.440
<v Speaker 1>of twenty twenty eight at absolute best, if not the

0:01:17.520 --> 0:01:19.760
<v Speaker 1>end of twenty twenty nine or twenty thirty if it

0:01:19.800 --> 0:01:22.840
<v Speaker 1>ever gets built. Star Gay Abilene, the only one of

0:01:22.840 --> 0:01:25.520
<v Speaker 1>them anywhere near completion, started in the middle of twenty

0:01:25.520 --> 0:01:29.040
<v Speaker 1>twenty four. Sits two years into construction at two buildings

0:01:29.040 --> 0:01:32.080
<v Speaker 1>and around two hundred megawats of capacity out of the

0:01:32.240 --> 0:01:35.320
<v Speaker 1>eight that are being built eight buildings. I mean, based

0:01:35.360 --> 0:01:37.959
<v Speaker 1>on reports throughout the last two years, Star gay Abilene

0:01:37.959 --> 0:01:40.800
<v Speaker 1>was meant to be fully energized somewhere between I don't know,

0:01:40.959 --> 0:01:42.840
<v Speaker 1>the end of twenty twenty five, in the middle of

0:01:42.880 --> 0:01:45.640
<v Speaker 1>twenty twenty six. Yet sources on the ground tell me

0:01:45.920 --> 0:01:48.680
<v Speaker 1>they don't expect this thing to be finished anywhere before

0:01:48.680 --> 0:01:53.120
<v Speaker 1>April twenty twenty seven, phy a source familiar with Oracle's infrastructure.

0:01:53.320 --> 0:01:56.240
<v Speaker 1>Building three has finished construction and been handed over, but

0:01:56.320 --> 0:01:58.840
<v Speaker 1>barely has any gear in it. It'll be months before

0:01:58.880 --> 0:02:02.600
<v Speaker 1>it starts generating revenue now. Based on discussions with sources

0:02:02.600 --> 0:02:06.080
<v Speaker 1>familiar with Stargate Abeleine's infrastructure costs, the total cost will

0:02:06.160 --> 0:02:08.560
<v Speaker 1>run somewhere in the region of forty eight million dollars

0:02:08.560 --> 0:02:11.120
<v Speaker 1>per megawat, which is in line with an estimate of

0:02:11.160 --> 0:02:14.160
<v Speaker 1>around forty four million dollars per megawat for any data

0:02:14.160 --> 0:02:17.160
<v Speaker 1>center construction that I got from Jerome Darling and analyst

0:02:17.160 --> 0:02:20.640
<v Speaker 1>over at TD Cohen. In total, this puts the estimated

0:02:20.680 --> 0:02:23.760
<v Speaker 1>cost of Stargate Abilene at around fifty nine billion dollars

0:02:23.880 --> 0:02:27.079
<v Speaker 1>and the overall seven point one gigawa project so far

0:02:27.200 --> 0:02:30.520
<v Speaker 1>at least and estimated three hundred and forty billion dollars.

0:02:31.280 --> 0:02:33.680
<v Speaker 1>Sources have also told me the Oracle estimated in twenty

0:02:33.720 --> 0:02:36.160
<v Speaker 1>twenty four that it would spend over two point one

0:02:36.240 --> 0:02:39.240
<v Speaker 1>billion dollars a year on co location, lease and power

0:02:39.280 --> 0:02:42.240
<v Speaker 1>cost to land developer Cruso. And that's just because Crusoe

0:02:42.320 --> 0:02:46.040
<v Speaker 1>actually they leased the land from a company called Lancium.

0:02:46.400 --> 0:02:48.639
<v Speaker 1>They then leased that land and the stuff on it

0:02:49.040 --> 0:02:52.240
<v Speaker 1>to Oracle, and then Crusoe handles the power and then

0:02:52.280 --> 0:02:55.639
<v Speaker 1>passes through the cost to Oracle. It's complex, it's annoying,

0:02:55.680 --> 0:02:58.960
<v Speaker 1>but that's what it is. Stargate Abilene will, based on

0:02:59.080 --> 0:03:02.280
<v Speaker 1>estimates from Lando Lanceium and discussions with sources familiar with

0:03:02.360 --> 0:03:05.520
<v Speaker 1>hyperscalar billing, generate around ten billion dollars a year in

0:03:05.560 --> 0:03:08.000
<v Speaker 1>revenue when it's complete at a rate of around twelve

0:03:08.040 --> 0:03:11.160
<v Speaker 1>point five million dollars a megawatt of critical it so

0:03:11.280 --> 0:03:14.160
<v Speaker 1>eight hundred and twenty four megawatts. The way it works is,

0:03:14.240 --> 0:03:17.079
<v Speaker 1>it's a twelve hundred gig sorry, it's a one point

0:03:17.160 --> 0:03:20.640
<v Speaker 1>two gigawatt campus, but there's only so much of it

0:03:21.040 --> 0:03:23.760
<v Speaker 1>that's actually critical it that would be billed for, so

0:03:23.919 --> 0:03:27.120
<v Speaker 1>eight hundred and twenty four megawat's worth. Now, while this

0:03:27.280 --> 0:03:30.160
<v Speaker 1>might suggest that Oracle is making eight billion dollars in

0:03:30.200 --> 0:03:32.680
<v Speaker 1>profit a year, and that's not true, one has to

0:03:32.720 --> 0:03:35.320
<v Speaker 1>reckon with the astonishing cost of building this data center,

0:03:35.440 --> 0:03:38.440
<v Speaker 1>and of course other costs like people and keeping like

0:03:38.520 --> 0:03:42.960
<v Speaker 1>insurance and stuff. Anyway, each building contains around six hundred

0:03:42.960 --> 0:03:45.800
<v Speaker 1>and ninety four GB, two hundred NVL, seventy two ras,

0:03:46.000 --> 0:03:49.520
<v Speaker 1>seventy two GPUs each, each retailing around three and a

0:03:49.560 --> 0:03:52.280
<v Speaker 1>half million dollars, meaning that each building contains around two

0:03:52.320 --> 0:03:54.520
<v Speaker 1>point four to three billion dollars of GPUs, for a

0:03:54.520 --> 0:03:56.920
<v Speaker 1>total of nineteen point four to four billion dollars for

0:03:57.000 --> 0:04:01.960
<v Speaker 1>the entire campus. As Oracle appreciates those GPUs over six years,

0:04:02.160 --> 0:04:04.440
<v Speaker 1>that comes to around three point twenty four billion dollars

0:04:04.480 --> 0:04:07.360
<v Speaker 1>a year in depreciation costs, so spreading them out. I

0:04:07.360 --> 0:04:10.040
<v Speaker 1>can also confirm, based on discussions with sources familiar with

0:04:10.040 --> 0:04:13.160
<v Speaker 1>the Apelien project, that Oracle is footing at either large

0:04:13.200 --> 0:04:15.960
<v Speaker 1>parts of the entirety of the construction of these data

0:04:16.000 --> 0:04:19.400
<v Speaker 1>centers and is spent as of recent over five point

0:04:19.480 --> 0:04:22.839
<v Speaker 1>four billion dollars in construction and infrastructure costs, not including

0:04:22.880 --> 0:04:26.640
<v Speaker 1>GPUs on just the first two buildings, with an estimate

0:04:26.640 --> 0:04:29.240
<v Speaker 1>that they'll spend up to ten billion dollars despite it

0:04:29.279 --> 0:04:33.440
<v Speaker 1>being operational and there are tons more expenses, though it's

0:04:33.560 --> 0:04:37.000
<v Speaker 1>very weird now. On a strictly mathematical basis, this means

0:04:37.040 --> 0:04:39.960
<v Speaker 1>that Oracle will end up, assuming the project is ever completed,

0:04:40.240 --> 0:04:43.000
<v Speaker 1>spending around sixty billion dollars to build a data center

0:04:43.040 --> 0:04:46.960
<v Speaker 1>campus that will make it about ten billion dollars of revenue.

0:04:47.640 --> 0:04:50.159
<v Speaker 1>Inside said data center campus will be racts of seventy

0:04:50.200 --> 0:04:52.599
<v Speaker 1>two GB two hundred GPUs that are already a year

0:04:52.640 --> 0:04:55.520
<v Speaker 1>old and will be in theory two years old by

0:04:55.560 --> 0:04:59.440
<v Speaker 1>the time that the campus is complete. Now the later

0:04:59.560 --> 0:05:03.000
<v Speaker 1>data set, the ones that aren't in Abilene will be

0:05:03.160 --> 0:05:07.200
<v Speaker 1>likely using Nvidia's Vera Rubin GPUs, which means they'll pay

0:05:07.200 --> 0:05:09.000
<v Speaker 1>a little more. Well, I mean they'll be paid a

0:05:09.000 --> 0:05:11.880
<v Speaker 1>little more by open Ai again in theory, likely around

0:05:11.880 --> 0:05:15.000
<v Speaker 1>fourteen million dollars per megawatt, which means, again in theory,

0:05:15.240 --> 0:05:18.920
<v Speaker 1>Oracle will get paid assuming everything gets built, around seventy

0:05:18.920 --> 0:05:21.159
<v Speaker 1>five billion dollars a year in revenue from open ai,

0:05:21.920 --> 0:05:26.120
<v Speaker 1>if open ai could afford that, which it cannot. For context,

0:05:26.200 --> 0:05:29.279
<v Speaker 1>open ai estimates it will have around thirty billion dollars

0:05:29.320 --> 0:05:31.680
<v Speaker 1>in revenue this year and lose twenty five billion dollars

0:05:31.720 --> 0:05:34.279
<v Speaker 1>to make that. In twenty twenty seven, it projects to

0:05:34.279 --> 0:05:36.440
<v Speaker 1>pull in sixty two billion dollars in revenue and lose

0:05:36.480 --> 0:05:38.919
<v Speaker 1>fifty seven billion dollars to make that, and in twenty

0:05:39.000 --> 0:05:41.560
<v Speaker 1>twenty eight it will lose eighty five billion dollars and

0:05:41.680 --> 0:05:45.479
<v Speaker 1>allegedly make one hundred and fifteen billion dollars. Open Ai

0:05:45.560 --> 0:05:47.920
<v Speaker 1>has also agreed to spend twenty two billion dollars with Core,

0:05:48.000 --> 0:05:50.200
<v Speaker 1>with one hundred and thirty eight billion dollars with Amazon

0:05:50.240 --> 0:05:53.280
<v Speaker 1>Web Services, two one hundred and fifty billion dollars with Microsoft,

0:05:53.400 --> 0:05:57.320
<v Speaker 1>and twenty billion dollars with Cerebros on compute. Yeah, so

0:05:57.560 --> 0:05:59.760
<v Speaker 1>for open ai to pay all of its bills, it

0:05:59.800 --> 0:06:01.760
<v Speaker 1>will have to raise at least another one hundred and

0:06:01.800 --> 0:06:03.640
<v Speaker 1>fifty billion dollars, and I mean on top of the

0:06:03.640 --> 0:06:06.480
<v Speaker 1>money it just raised, if not two hundred billion dollars,

0:06:06.600 --> 0:06:09.560
<v Speaker 1>and that's only only if it makes over one hundred

0:06:09.560 --> 0:06:12.279
<v Speaker 1>billion dollars in annual revenue by the end of twenty

0:06:12.279 --> 0:06:15.280
<v Speaker 1>twenty eight. And if I'm honest, I'm not sure that

0:06:15.279 --> 0:06:17.560
<v Speaker 1>that maths even makes sense even if it does so,

0:06:17.880 --> 0:06:20.360
<v Speaker 1>because we don't know the precise year by year breakdown

0:06:20.360 --> 0:06:23.920
<v Speaker 1>of its other deals. I genuinely think they may be

0:06:24.040 --> 0:06:27.440
<v Speaker 1>spending if they actually kept the agreements one hundred and

0:06:27.440 --> 0:06:29.760
<v Speaker 1>twenty five billion dollars in compute by twenty twenty eight.

0:06:30.160 --> 0:06:32.919
<v Speaker 1>I mean, they don't have the they don't have the money.

0:06:32.960 --> 0:06:36.320
<v Speaker 1>They can't do that. But nevertheless, it's how is no

0:06:36.360 --> 0:06:39.520
<v Speaker 1>one else worried about this? This would, also, by the way,

0:06:39.560 --> 0:06:41.760
<v Speaker 1>mean that open AI grows its revenue by four times

0:06:41.760 --> 0:06:44.200
<v Speaker 1>in the next two years and by nearly ten times

0:06:44.200 --> 0:06:46.560
<v Speaker 1>by twenty thirty. And I think they need to be

0:06:46.760 --> 0:06:49.279
<v Speaker 1>ten to fifteen times to pay all these fucking bills.

0:06:50.680 --> 0:06:55.000
<v Speaker 1>Sounds implausible, right, Well, it has to happen otherwise Oracle

0:06:55.080 --> 0:06:58.320
<v Speaker 1>runs out of money. In its last quarterly earnings, Oracle

0:06:58.360 --> 0:07:01.560
<v Speaker 1>had free cash flow of negative twenty four point seven

0:07:01.640 --> 0:07:05.320
<v Speaker 1>billion dollars and has, through both completed and planned bond sales,

0:07:05.560 --> 0:07:08.520
<v Speaker 1>debt financing, and at the market share sales, raised over

0:07:08.560 --> 0:07:12.080
<v Speaker 1>one hundred and fifteen billion dollars, which is not sufficient

0:07:12.120 --> 0:07:15.080
<v Speaker 1>to complete the construction of the remaining Stargaate data centers.

0:07:15.520 --> 0:07:17.760
<v Speaker 1>Oracle has also raised a great deal of that money

0:07:17.840 --> 0:07:21.880
<v Speaker 1>using construction project financing, keeping the debt off balance sheet,

0:07:21.920 --> 0:07:24.840
<v Speaker 1>and tying its repayments entirely to cash flow from the

0:07:24.880 --> 0:07:28.280
<v Speaker 1>various projects revenues that are being paid by open Ai,

0:07:28.400 --> 0:07:30.560
<v Speaker 1>a company that's going to lose hundreds of billions of

0:07:30.640 --> 0:07:32.520
<v Speaker 1>dollars in the next few years if it doesn't fall

0:07:32.560 --> 0:07:37.200
<v Speaker 1>over and die Jesus fucking Christ. If those revenues don't

0:07:37.200 --> 0:07:40.200
<v Speaker 1>come through, say open a I just didn't pay them,

0:07:40.720 --> 0:07:43.000
<v Speaker 1>Oracle will be unable to pay its debts. This is

0:07:43.040 --> 0:07:47.920
<v Speaker 1>not an opinion, This is maths. It's other businesses Oracles,

0:07:48.000 --> 0:07:51.200
<v Speaker 1>by the way, In hardware and software licensing, they're plateawing.

0:07:51.240 --> 0:07:53.440
<v Speaker 1>They have been for quite some time, and its only

0:07:53.520 --> 0:07:56.320
<v Speaker 1>growth market is renting out AI compute using GPUs that

0:07:56.400 --> 0:07:58.880
<v Speaker 1>burn out fast and are upgraded on a yearly cycle,

0:07:59.120 --> 0:08:02.120
<v Speaker 1>meaning that by the time Stargate Abilene is completed, it's

0:08:02.120 --> 0:08:04.920
<v Speaker 1>black well GPUs will be two or even three years old.

0:08:05.320 --> 0:08:07.920
<v Speaker 1>In fact, all of Stargate will be full of years

0:08:07.920 --> 0:08:11.960
<v Speaker 1>old GPUs if it ever gets completed. Like think about it.

0:08:12.040 --> 0:08:14.600
<v Speaker 1>Pretty much every data center you know is gonna have

0:08:14.640 --> 0:08:17.440
<v Speaker 1>obsolete GPUs because if it takes two three years to

0:08:17.440 --> 0:08:19.920
<v Speaker 1>build one, and these ones are being filled with black

0:08:19.920 --> 0:08:22.720
<v Speaker 1>well gps, well, by the time the fucking things are done, well,

0:08:22.720 --> 0:08:24.680
<v Speaker 1>you can have a thing full of years old GPUs.

0:08:24.760 --> 0:08:26.960
<v Speaker 1>The same thing's gonna happen with Vera Rubin and I

0:08:27.000 --> 0:08:28.960
<v Speaker 1>don't know the next one. I think they're just gonna

0:08:28.960 --> 0:08:32.120
<v Speaker 1>call them red Foo. They're gonna name them after one

0:08:32.120 --> 0:08:35.439
<v Speaker 1>of the blokes from from LMFAO. He Jensen Wang is

0:08:35.480 --> 0:08:39.640
<v Speaker 1>a huge LMFAO fan. He's just blasting shots, shots, shot

0:08:39.720 --> 0:08:43.439
<v Speaker 1>shot shots constantly in video party rock Anthon plays at

0:08:43.440 --> 0:08:46.920
<v Speaker 1>every board meeting. It's very weird. If you've heard anything,

0:08:46.960 --> 0:08:48.880
<v Speaker 1>this is all made up. I'm just having some fun.

0:08:49.920 --> 0:08:54.640
<v Speaker 1>But look, look, look, look I'm deeply confused. Then nobody

0:08:54.640 --> 0:08:58.440
<v Speaker 1>else is on this. I am genuinely genuinely confused. Where

0:08:58.640 --> 0:09:03.040
<v Speaker 1>the fuck is my Michael Burry? Michael Burry came back

0:09:03.120 --> 0:09:06.800
<v Speaker 1>out of the shadows, Cassandra unchained. He came out of

0:09:06.840 --> 0:09:09.000
<v Speaker 1>the shadows. He was going to blog about this stuff.

0:09:09.080 --> 0:09:13.400
<v Speaker 1>He did one blog about depreciation with nvideo and Mata,

0:09:14.000 --> 0:09:17.320
<v Speaker 1>and then a real big, juicy wet kiss to Jack

0:09:17.360 --> 0:09:21.559
<v Speaker 1>Clark over anthropic in a supposed debate about the AI

0:09:21.640 --> 0:09:24.400
<v Speaker 1>bubble that mostly involved Michael Burry saying Claud's so good.

0:09:24.480 --> 0:09:26.600
<v Speaker 1>I love it so much. What the fuck is the

0:09:26.600 --> 0:09:29.720
<v Speaker 1>point of someone like Michael fucking Burry if he can't

0:09:29.880 --> 0:09:33.719
<v Speaker 1>look at something like this. The numbers are there. Oracle

0:09:34.040 --> 0:09:37.160
<v Speaker 1>is very likely going to die unless it starts backing

0:09:37.160 --> 0:09:40.160
<v Speaker 1>away from these data center projects, and all signs point

0:09:40.200 --> 0:09:43.800
<v Speaker 1>to it accelerating towards building them as fast as possible.

0:09:44.080 --> 0:09:47.640
<v Speaker 1>Even if it succeeds, open AI cannot afford to pay

0:09:47.640 --> 0:09:51.320
<v Speaker 1>for them. Seventy five billion dollars a year is so

0:09:51.600 --> 0:09:54.520
<v Speaker 1>much money. I think Microsoft's operating costs like one hundred

0:09:54.559 --> 0:09:58.040
<v Speaker 1>and fifty something billion, and Microsoft's very very, very profitable.

0:10:00.000 --> 0:10:01.680
<v Speaker 1>When I sat and thought about how I might be

0:10:01.720 --> 0:10:03.920
<v Speaker 1>wrong about this one or lot, because it's a huge

0:10:03.920 --> 0:10:05.760
<v Speaker 1>claim to make. But I cannot find a way that

0:10:05.800 --> 0:10:08.120
<v Speaker 1>Oracle actually makes this work. And every time I tell

0:10:08.160 --> 0:10:12.280
<v Speaker 1>somebody about it, they just respond with inshallah. Anyway, if

0:10:12.320 --> 0:10:13.960
<v Speaker 1>you like all the sound of this, check out the

0:10:13.960 --> 0:10:17.520
<v Speaker 1>premium newsletter that'll be out later today. There's a discount

0:10:17.559 --> 0:10:19.559
<v Speaker 1>code in the show notes a link. Click it. Please

0:10:19.640 --> 0:10:21.680
<v Speaker 1>a principal form of income now. But even if you

0:10:21.720 --> 0:10:24.320
<v Speaker 1>do not pay to subscribe, there is a generous free

0:10:24.360 --> 0:10:26.199
<v Speaker 1>section and a summary at the bottom of it that

0:10:26.200 --> 0:10:27.960
<v Speaker 1>will give you a great deal of the story. You'll

0:10:28.000 --> 0:10:31.079
<v Speaker 1>be able to get the highlights now. The meat will

0:10:31.120 --> 0:10:32.560
<v Speaker 1>be in there if you want to pay, but I

0:10:32.600 --> 0:10:35.320
<v Speaker 1>don't want to get too much of this information. I'll

0:10:35.320 --> 0:10:38.679
<v Speaker 1>be back next week. I think i'll have an episode

0:10:38.720 --> 0:10:41.640
<v Speaker 1>on Wednesday. But if I'm honest, if I'm too ill,

0:10:42.320 --> 0:10:44.640
<v Speaker 1>or if I'm just run down because I had family

0:10:44.640 --> 0:10:47.240
<v Speaker 1>in the hospital, I might skip. But I think I've

0:10:47.280 --> 0:10:50.400
<v Speaker 1>got one in session either way. Catch me on the Reddit,

0:10:50.520 --> 0:10:53.480
<v Speaker 1>shoot me a slub on pluk, whatever however you want

0:10:53.480 --> 0:10:56.240
<v Speaker 1>to contact me. I love hearing from you all. Cheers,

0:10:56.720 --> 0:11:00.680
<v Speaker 1>Love you all. Heyy