Is the AI Boom About to COLLAPSE? — Transcript

Full transcript

1. 0:09Hello and welcome to Why is this happening with  me your host Chris Hayes and welcome back to our
2. 0:13ongoing series why is this happening the mind body  problem about all the implications of the AI boom
3. 0:19um you have probably seen well I don't know  if you have but maybe you've read the book
4. 0:23uh the big short by Michael Lewis or seen the  phenomenal movie by Adam Mccay which is like I
5. 0:27think genuinely a classic the big short and  the tale there is a chronicle of a disperate
6. 0:34group of what you might call our kind of financial  dissident who in the sort of era of 2006 2007 as
7. 0:42the housing boom is going and everyone's making a  ton of money start to sniff out there's something
8. 0:46very deeply a miss deeply wrong and the more  information they get the more convinced they
9. 0:51become that everyone in the market more or less  is wrong that this entire money-making machine is
10. 0:56about to collapse in on itself because they're  people in finance they take bets that it will
11. 1:00collapse and those bets prove to be accurate.  It's a very satisfying tale and a satisfying
12. 1:05um book and movie because it's a kind of David  and Goliath and you sort of get to follow the
13. 1:09profit as the prophecies come true. And the reason  I bring all this up is we're in a boom right now
14. 1:15with AI. I mean, the amount of money that's being  put into it is staggering. And I think the broad
15. 1:22amount of the financial system and the press  thinks that that it's going to pay off. um at
16. 1:28least you know that's basically been what markets  have priced things at. If you look at the stock of
17. 1:33the big AI companies and particularly Nvidia that  makes the chips this all depends on it keeps going
18. 1:38up and up and up. It's been a little down recently  but there are just like there were at the housing
19. 1:42bubble there are dissenters and dissident talked  to one today. Now the thing about being a denter
20. 1:48dissident against this sort of conventional wisdom  is you could be wrong. You know, there were people
21. 1:53that thought, for instance, um, the internet was  never going to amount to anything or that personal
22. 1:57computers were a ridiculous technology. And in  the long view of history, it's always a fine line
23. 2:02between crank and profit. Sort of depends on what  happens in the end. But given the fact that there
24. 2:07is so much money at stake, and because I think  the the maximalist view that this is going to
25. 2:12end in tears is one that people get the least  amount of exposure to, I wanted to spend some
26. 2:18time reckoning with that view today with my guest  Ed Zitron. Ed is the CEO of EasyPR. He's the host
27. 2:24of the Better Offline podcast and he writes the  Where's Your Edat newsletter on Ghost and I would
28. 2:30say that he is one of the foremost um AI bears  or even AI haters on the whole internet. Is that
29. 2:35fair? I think that's fair. But I really hate bad  software like that really. I love technology. I
30. 2:41re I owe my life to technology, but the current  state of the tech industry sucks. You're a tech
31. 2:46person. You were a gaming journalist for a little  bit. Yes, I was. I wrote about video games in in
32. 2:50England for goodness about five six years. Moved  to America in 2008. Great year. Fantastic year to
33. 2:56move to America. And yes, um it's strange watching  what's happening. And it feels like as I watch the
34. 3:04pieces fall together. It's all been kind of  working up to this because the AI bubble is
35. 3:10a symptom of a larger problem with the software  industry. The the hyperrowth era is ending. We
36. 3:16have seen software as this thing will always grow  exponentially. So every single bubble is looked at
37. 3:22through the same lens. You say metaverse will grow  exponentially. NFTts will become will take over
38. 3:27all culture. We won't buy physical things anymore  as far as collectibles go. We'll only have these
39. 3:33digital versions. Everything seen through what  I call the rot economy, the growth at all cost
40. 3:37mindset. And I think we're coming to the close  of it. Le let's stay there because I think one
41. 3:42of the reasons that I hear skepticism about AI  from people and I think is irrational is that
42. 3:48we did just go through this huge hype cycle.  Yeah. that was I mean it's a little bit wiped
43. 3:54from memory but before AI during COVID there was a  huge amount of money slloshing around the economy
44. 4:01um because of all of the money was being directly  injected both through fiscal stimulus and the Fed
45. 4:06and people were spending all their times in front  of their screensh um and there was this huge boom
46. 4:13around the metaverse the metaverse blockchain  and like nonfgeible tokens NFTs Yes. What do you
47. 4:22think the takeaways from that sort of boom bust  cycle are? So a little bit before the metaverse,
48. 4:28there was another bubble that other people forgot  about and that was Clubhouse. So Clubhouse was
49. 4:32this audio only social network that if you talk  to venture capitalists at the time, oh my god,
50. 4:37this was the biggest thing ever. It was going  to be worth a bazillion dollars. It was radio
51. 4:41too. Now nobody went and looked up how much radio  makes or what the revenues were really anything.
52. 4:46But they got celebrities on. They did everything  they could and they it was very clear the venture
53. 4:50capitalists were pushing this so that they  could get a big acquisition now never happened
54. 4:55and clubhouse has kind of fallen into irrelevance  with NFTTS with the metaverse. People forget that
55. 5:01Meta used to be called Facebook. We still call it  Facebook. They changed their entire company name
56. 5:07to Meta. One of the craziest things in history  and we just don't talk about it. It's insane that
57. 5:13happened. You had people on CBS News being like,  "Yep, the metaverse is here. We're all going to
58. 5:17live in the metaverse. It's gonna it's very real.  It's going to happen. And the actual experience
59. 5:22was a very bad virtual reality experience. But  I feel like the tech industry has kind of been
60. 5:30laring for the last 10, 15 years. They had laring  meaning live action role playing pretending going
61. 5:37through the motions because you had the era of  smartphones and mobile apps, huge deal. You had
62. 5:43the era of software as a service, SAS. These were  the big revenue drivers of the tech industry. It
63. 5:49was a way to get more money out of people because  you had a subscription service or you had an app
64. 5:53you could buy on your phone. Most of those all  have monthly subscriptions now. Yeah. Convert
65. 5:57to subscriptions. Exactly. So it was a way of  getting people away from that troublesome thing
66. 6:01where they only paid you once. Nevertheless, this  worked for a while. And so that worked and then
67. 6:08uh apps worked and then um nothing else really  worked. we kind of started running out of kind
68. 6:14of hyperrowth ideas. We haven't really had one  since like I don't know LinkedIn maybe there are
69. 6:19probably some examples but we just stopped having  big things that worked. So the tech industry did
70. 6:25what it did before hire a lot of people put a lot  of money into things buy a lot of things by little
71. 6:30companies. Uh the Activision Blizzard acquisition  from Microsoft was claimed as a metaverse thing,
72. 6:36which was very silly because if the metaverse is  all video games, that's just so much. But it's the
73. 6:42tech industry doing what it thinks works. And I  think we're getting to the end point of that. So
74. 6:48here's what I I want to talk a little bit. I want  to wait to talk about the tech and then and just
75. 6:54and hive that off from the economics of it. But ju  just stay on this for a second because to me the
76. 6:59big difference is no one could ever really explain  to me the use case of the metaverse. You know
77. 7:06people would sometimes be like yeah I'd be like  well what's what what do I do with it? And people
78. 7:10would say you know with crypto it's like well can  I pay for a cup of coffee with it? No you can't
79. 7:14really do that. No it's it's essentially um there  was a little bit of talk of like the blockchain is
80. 7:18going to replace contracts and I was like but is  anyone actually doing that? No. Yeah. and and even
81. 7:24the metaverse, you know, you kept sort of saying  like what's what's the use case? What's it do?
82. 7:29No one can ever answer. I don't feel like that  way with AI. Like there are use cases. In fact,
83. 7:34I've used it for use cases. And it seems to  me like there's a much clearer like, okay,
84. 7:40um do this doc review. Here's a thousand  documents. Um it would take a human to go
85. 7:45through them. It can read and synthesize. Now,  the question of whether it could like do it well
86. 7:49or not is a different question. But to me, the  difference is that you can at least articulate
87. 7:53what it's for in certain circumstances or reasons  or places it might be useful in a way that for me
88. 8:00the metaverse never did. So there are uses for  large language models. If you remove all of the
89. 8:06hype, there are things they can do. The problem  is is here's a little challenge for you. Go and
90. 8:12talk to a bunch of AI boosters and ban them from  speaking in the future tense. Don't allow them to
91. 8:18say anything about this will it might or it could  say what it does today because when you do that
92. 8:25it's not really clear what's changed. So yes, you  can use it to review documents. The results are
93. 8:31not great or maybe they are. You actually don't  know because you didn't read the documents a thing
94. 8:36that is based on statistics did. The idea that  you can rely on it is inherently broken because
95. 8:42OpenAI's own research says hallucinations are  a part of these things. They're never going
96. 8:46away. You have people in the AI industry claiming  hallucinations are going away. They're just wrong.
97. 8:52Open AI said it. You going to argue with them?  There there's a recent paper I was just looking
98. 8:56at yesterday that says that even when you cuz the  the when I use AI, I do, you know, use it where
99. 9:00you sort of like gate the sources. And there's a  paper I saw yesterday that like even when you gate
100. 9:05the sources, right? When you're saying just use  these sources, you cannot purge a hallucination.
101. 9:09No. And even with coding LLMs, because this  is one of the most annoying debates ever,
102. 9:16is the usefulness of LLMs within coding. And I  think the big thing with the AI LLM coding debate
103. 9:23is you're beginning to find out that there are I  don't want to say a lot, but there is a contingent
104. 9:29of software engineers that might not know a lot  about code or might be getting by with not a ton
105. 9:34of information. To them, this might seem magical.  There's an amazing writer called Nick Suresh who
106. 9:39did a piece he did this amazing blog called I will  effing pile drive you if you mention AI again and
107. 9:45he made the point that this is something that has  its uses for the little things but the moment you
108. 9:50start expanding it to building entire things for  you and writing all this code for you you are just
109. 9:55kind of kicking the can you're still going to have  to read all this code to make sure it makes sense
110. 10:01or alternatively you could not read it and just  hope it works software function when the code
111. 10:07is bad doesn't mean it's secure or stable or  efficient or indeed that someone else coming
112. 10:13along in the future to read it can understand the  intention because there was none because the large
113. 10:17language model wrote it right the sort of quality  control question which you get in research seems
114. 10:22to be a big thing right but let's put that aside  for a second right is this is that solvable right
115. 10:27is the quality control question solvable is sort  of we'll put to the side it seems to me that it's
116. 10:31worthwhile to just for for for the next part of  this to distinguish between is a tech useful or
117. 10:36even transformational and is the current financial  investment in it justified as distinct questions.
118. 10:44Yes. And two examples come to mind. There  was an enormous railroad boom that happened
119. 10:49in the 19th century, right? In which an enormous  percentage of the country's entire GDP went into
120. 10:54railroads. It completely overbuilt and it led  to a huge crash and that crash led to a great
121. 10:59depression. It doesn't mean that railroads weren't  useful. In fact, railroads are quite useful,
122. 11:04right? But it also is the case that you can have  a useful technology that leads to a boom and bust.
123. 11:11The internet being another example, right? It  isn't the case that the internet proved not to
124. 11:16be useful. It's also the case that like there was  a huge boom in 1999 and 2000. Right. Right. And I
125. 11:22have this thing I've been saying the beginning  of history. It's not fully connected to Fuki,
126. 11:26but nevertheless, it's the I don't think it's  instructive. I understand why people do it. It's
127. 11:31how human beings work. I don't think it's useful  to look at these previous booms because when you
128. 11:36put trains on the railroad, sometimes the train  didn't just randomly go up or into the ground.
129. 11:41I with the original internet systems uh back in  the middle of 2025, a guy called Jim Cavell from
130. 11:46Goldman Sachs did a great piece along with some  other analysts called Genai, too much spend for
131. 11:53not enough gain. And I paraphrase the title there.  and he made the point that when the the internet
132. 11:58did cost a lot of money at the $64,000 some  micros systemystem service nevertheless that
133. 12:03capital outlay was completely different but there  was also a very clear path to utility it would be
134. 12:10the dispersion of fiber optic cable it would  be the access points the actual things being
135. 12:16built so that people could get to the internet  and high-speed internet on top of that the same
136. 12:21thing happened with smartphones notes that in the  early 2000s There were clear road maps, smaller
137. 12:27Bluetooth radios, smaller GPS's, smaller chips,  smaller batteries. That would lead to smartphones.
138. 12:33No such path exists for large language models.  For that example to make sense, you would have
139. 12:38to have a way in which the cost came down and the  hallucinations went away. Neither of those appear
140. 12:44to be happening. And indeed, the efficacy of these  models, their actual outcomes, it's actually very
141. 12:50difficult to measure them. The benchmarks are  deliberately created for them and all of the
142. 12:56benchmarks for software engineering are focused  on one programming language, Python, and very
143. 13:00common GitHub issues. So to train for more things,  they're having to create specialized data. They're
144. 13:06going to have to do that forever. And even then,  it isn't obvious if it's actually fixing things,
145. 13:10right? I mean, this is this problem of are they  just basically are they training on the test data,
146. 13:14right? Are they are you basically saying here take  a look at all this this data and then we're going
147. 13:18to test you on it and oh lo and behold your  your performance is good fun fact about that
148. 13:23they actually found that one of the anthropic  models had just started going and looking for
149. 13:27the solution wasn't trying to solve it just went  on GitHub and did it now people mistake this for
150. 13:31intelligence no you asked the thing to do a thing  and it did a thing right it's just it's doing
151. 13:36the functions it was told to do okay but that's a  great example because like a year ago it couldn't
152. 13:40do that I mean it is doing something new even if  it's going to GitHub Right. GitHub for people that
153. 13:45don't know is a sort of this sort of open source  library where where people share and where you can
154. 13:49host projects. Yes. Where you host projects and  people share code. But like a year ago it didn't
155. 13:53do that. Right. It used the web search tool. It  used the tool it's had for a while. Perhaps it did
156. 13:58something new I guess. But it it's a relative it's  a lateral improvement. It's an improvement on the
157. 14:05thing it's already doing. It's not making unique  software. Even the clawed code things you're
158. 14:10seeing where people are spitting out websites.  There are tens of thousands of website templates
159. 14:16and open- source software projects that they're  replicating. It did, this is a little bit, I won't
160. 14:21get too in the weeds of it. They did something  called a C++ compiler and Anthropic said, "We
161. 14:26made this, we did this, was a clone of an open-  source project and it was less efficient. It was
162. 14:31something like 10,000 times less efficient." Which  is crazy. And it's these things don't make novel
163. 14:38ideas because if you just look at what has already  happened and say well based on this this will
164. 14:43happen right you'll never you can't get out past  the arithmetic statistical average essentially.
165. 14:48Exactly. Yes. So let's talk about the the scope of  the money here. Basically paint a picture of how
166. 14:55big this bubble is that you say is a bubble where  the money's coming from and how it's flowing.
167. 15:01So it's around a trillion dollars now I think by  the end of the year if you think about all of the
168. 15:07venture capital funding all of the money that's  been put into data centers all of the capital
169. 15:12expenditures from Microsoft Amazon Google Meta  and the money flowing through Taiwanese server
170. 15:18companies like Honhe so Foxcon and Quanta and  all that the money is coming from a few places
171. 15:23it's coming from venture capitalists and I can  get into the crisis there soon private equity
172. 15:29and specifically private credit and actually a lot  of the money is coming from Japan uh Suttomo. So,
173. 15:35SNBC and Mitsubishi, MUFG, I swear I'm going  somewhere with this, but the money is coming from
174. 15:42private equity, private credit, venture capital,  and in some cases the hyperscalers themselves, but
175. 15:48and most of it's flowing to like three companies.  It's also I mean it seems to me that it's also
176. 15:53right. So when you're when you're talking about AI  and anthropic, right, they need to raise capital,
177. 15:57right? Absolutely. But places like Google or  Microsoft are spending I mean Google just throws
178. 16:02off a ton of cash, right? So they're that's  a place where they've got this arguably the
179. 16:07most profitable business in the history of human  capitalism and they they can just sink that cash
180. 16:13into more and more investment. The problem  is that's slowly not becoming true. Amazon,
181. 16:18I think, is raising tens of billions of dollars  of bonds. Google already did the same thing.
182. 16:21Microsoft. I probably will at some point. They're  no Microsoft, I think, is the only one out of them
183. 16:27that is no longer that is uh not not using debt.  That's no longer just using cash flow to pay for
184. 16:34this. I see. Because none of these businesses  are profitable. Not a single one of them. What's
185. 16:38really interesting is none of them talk about  the AI revenue. None of them. Microsoft mentioned
186. 16:44it in two quarters. Last quarter of 2024, first  quarter of 2025, and then stopped mentioning it
187. 16:50entirely. IBM just stopped mentioning their AI  revenue. It's are they shy? And so when nobody
188. 16:57wants to talk about the money and nobody can  really precisely describe the outcomes, that's
189. 17:02when people should get a little concerned. There's  a B. So there's there's if we talk say a trillion
190. 17:09dollars, right? And the idea is you're investing  all this money and what does the investment go to?
191. 17:14Like what is it what needs to be built that all  this money is sunk into? So there's two things
192. 17:20to look at. There are the AI companies. So the  open AIs and anthropics of the world and then
193. 17:25the hyperscalers. And so let's talk software and  hardware. Okay. So AI companies like Anthropic. It
194. 17:32just came out. Krishna Ralph the chief financial  officer of anthropic in their case against the
195. 17:36department of defense just said that anthropic  through March 2026 for its entire lifetime made
196. 17:43exceeding $5 billion. They've spent $10 billion in  that period on training and inference. Inference
197. 17:50is the creating of the output. Fancy word for  that. Training is this word that's meant to
198. 17:55conjure up in your head this idea of research  and development. Training in large language
199. 17:59models can mean everything from pre-training, so  feeding a bunch of information, to post-training,
200. 18:05which is everything from we're going to give you  some stuff and test the outputs to minor tweaks to
201. 18:10stop something called model drift, which is just  when a model that is trained on static information
202. 18:17will eventually become irrelevant. So, you need  to keep updating it to make sure when you feed
203. 18:21it something, it understands it. And this there's  actual, you know, huge human intervention here,
204. 18:25which is like, no, that's wrong, that's wrong,  that's wrong. Because you have to kind of train
205. 18:28the model to to to learn. Exactly. You've got  human human trainers who are training the models
206. 18:34themselves as in model gives an output and they go  that's a good one, that's a bad one. Then you've
207. 18:39got people literally creating training data. Now  where do they spend that money? So this is the top
208. 18:45layer, the AI labs. Those ones are spending it  renting GPUs from Nvidia which are usually in
209. 18:51the case of Anthropic held by Amazon or Google or  in the case of Amazon and Google their own custom
210. 18:57silicon TPUs for Google and uh trainium and  inferentia for Amazon. If I'm anthropic I got
211. 19:03I got labor costs right I got employees and then I  have to to do all all the stuff that I want to do
212. 19:11run these models is very very computation  intensive. Yes. And in order to do that
213. 19:16computation, I need physical hardware. Yep. Um the  the so-called GPUs, which are the chip that Nvidia
214. 19:24and others make, which is this sort of um sort of  frontier next generation processing chip, right?
215. 19:30Yes. And the way that it works is that Claude and  AI rent rent that that hardware. So it's crazy
216. 19:38how much it costs as well because you may think  they want to say that inference is profitable.
217. 19:44No one's actually proven this. It's actually quite  expensive to provide a user a service. The other
218. 19:48problem is coding models especially are incredibly  computationally expensive. You've got one user who
219. 19:55might be tying up 6 to 12 GPUs, each one costing  50 grand a piece or more. You've got and the
220. 20:00more what's crazy and what really makes this  different to most software eras that your most
221. 20:06excitable customers are the ones that cost you  the most. And in all of the cases of the AI labs,
222. 20:13they're subsidizing them. Claude code, crazy fact,  researcher called Shell found this. For every
223. 20:19dollar that someone is spending on an anthropic  subscription, when they use Claude code, they can
224. 20:24spend anywhere from 8 to$13.5 worth of compute  costs because Anthropic is subsidi subsidizing
225. 20:30them, right? So cl let's let's stay on cloud code  because this is important on the business model,
226. 20:34right? So cloud code is people are have been you  know crowing about it and and every almost every
227. 20:40engineer I talks about is using it. Uh you can  do things where you're basically giving it plain
228. 20:46language instructions and it's coding for you.  The back end of what it's doing is extremely
229. 20:50compute intensive. Yes. And the expense of that  is renting the GPUs the electricity right the
230. 20:56storage the server space right those are the basic  usually you pay the company like Google or Amazon
231. 21:01directly but that's the cost. So that's the the  business relationship is I'm I'm anthropic and I'm
232. 21:07paying some other company that's doing all that  backend stuff, right? And the cost of that thing
233. 21:13I'm paying them for, right, is like can be it  $13 for every $1 I'm getting in revenue. I mean,
234. 21:20I think about this a lot. I use this example in  another conversation just with with Google and
235. 21:25Gemini where if you say, "What's a good Korean  restaurant in Brooklyn?" Google will show me a
236. 21:32Gemini response at the top and then like there's  a reddit thread that's like great Korean in
237. 21:38Brooklyn, right? the computational the like actual  resource cost of just going to the Reddit thread
238. 21:44is essentially zero basically tiny amount but the  the Gemini cost was like pretty significant to
239. 21:51go generate all that computation in the back and  you've scrolled right past that and gone straight
240. 21:56to Reddit because you trust a person way more than  you're going to trust Gemini right but but but so
241. 22:00the point is even if this thing is producing use  like in the cloud code case one of the things one
242. 22:06of my understandings of your main argument here  is that the current model is they are wildly
243. 22:12subsidizing because the compute is so expensive  and intensive in order to make it work. They have
244. 22:19to wildly subsidize it on the on the consumer end.  Yes. So really simple explanation. Anthropic has
245. 22:27two and OpenAI has this as well. Two different  kinds of customers. You've got a customer that
246. 22:31pays you a monthly subscription and you pay  through an API. It just means connect the model
247. 22:37to your thingy. Right now, when you use Claude  Code, you're just paying a monthly subscription,
248. 22:42$20, $100, or $200 a month. And then you have  arbitrary limits that Anthropic doesn't really
249. 22:49specifically say. But if you were paying on the  API, so if you're paying for the tokens directly
250. 22:53from Anthropic, you would be paying not $200 a  month, but $2.5 $2,700 a month. Gotcha. Right. So,
251. 23:01so these subscriptions are essentially massively  marked down to get customers who subscribe. Right.
252. 23:07But there doesn't appear to be a way that you can  con I don't think anybody that's paying 200 bucks
253. 23:12a month is going to go, "Yeah, I'll pay three  grand. That sounds great." I don't think that'll
254. 23:17happen. And it what it is is an attempt to graft  the previous business models and use the previous
255. 23:23growth trick which is the initification you the  cheap monthly fee that they can then rise and then
256. 23:29they'll find ways of undercutting you. Yeah, that  was going to be my next my next question. Right.
257. 23:33So this idea of you subsidize users on the front  end, you sort of m lose money on every customer,
258. 23:39you get enough market share that you could then  get price power and increase. This is famously
259. 23:44what Amazon used which you know lost money on  every customer and every book it sold for a
260. 23:48shockingly long period of time um and achieved  pricing power. And it's also in Uber is another
261. 23:55example right where you know people remember this  time when Uber came about where you could take an
262. 23:59Uber like five bucks. Yeah. I remember landing  in cities when I was doing like business travel.
263. 24:04I mean, cuz New York it was always like Yeah,  there was muddy. It was relatively expensive,
264. 24:09but but but still pretty cheap. But then sometimes  you'd land somewhere and be like a $6 ride from
265. 24:14the airport to the hotel. You think to yourself,  wait, how is this make any sense to me? This can't
266. 24:20possibly be the case that anyone's making money  out of this. But that in comparison, it would be
267. 24:25if like every Uber driver cost Uber $50,000 a day.  It's the economies of and the economies are just
268. 24:32completely different. When Uber was subsidized,  I think between 2019 and 2022 when they became a
269. 24:39kind of messy profitable like not a great one, it  was maybe 32 33 billion which is a lot of money.
270. 24:45Amazon Web Services arguably one of the most the  single most important technological innovations
271. 24:52ever mostly done through just money and time.  Though this isn't adjusted for inflation,
272. 24:56in the 11 years from I think 2003 onwards, they  spent 38 or 39 billion in capex. For some context,
273. 25:04OpenAI raised $42 billion in 2025. So you're  saying the scale of the subsidy here is just
274. 25:11way bigger than those previous ones. That's the  point. Yes. And the underlying infrastructure,
275. 25:15everything is more expensive and it's not getting  cheaper is so if Okay. So if we talk about the
276. 25:21the the front-end model makers right that they're  they're subsidizing you know even in this filing
277. 25:26right anthropic 5 billion of revenue 10 billion  in expenses obviously that's not profitable and
278. 25:31that's just the compute that's just the compute  are so then there's the hyperscalers right which
279. 25:37are the the the physical owners of that are  built and these ones to be the ones that are
280. 25:42building the data centers right in some cases  there's a lot of independent ones now that are
281. 25:46building data centers in the hopes that AI  demand arises that doesn't exist and people
282. 25:51like coreweave and nebus and such who are just  things called neoclouds they just are warehouses
283. 25:57full of GPUs that are technically data centers  and do they so let's say I'm one of those and I
284. 26:01build a data center do they then have a business  relationship with one of the intermediaries like
285. 26:06Amazon or do they directly contract with anthropic  the answer is yes so some of them do some of them
286. 26:11it gets even more complex we don't need to go  into it there are people that rent the data
287. 26:15centers who then sell the stuff but nevertheless  that's actually kind of the problem When you look
288. 26:20at who's paying for AI compute and you actually  really go and look at like who's paying the money,
289. 26:25there are really only two kinds of customers,  Anthropic and Open AI or Hyperscalers. Meta,
290. 26:32oh sorry, Nvidia. Nvidia has agreed to spend in  the next 5 years $26 billion in AI compute deals.
291. 26:40And I don't think it's a good sign that the shovel  seller is also paying for the digs. Wait, no,
292. 26:47wait. Take a second because it's going too fast.  So you got so this this is why so Nvidia buying
293. 26:53compute is weird for this reason. I just want to  walk people through this real soon. Yeah. Go on.
294. 26:57They make the chip, right? The chip is the thing  you sell to the person that's going to say set up
295. 27:03a data center, right? So in the ideal world, I'm  in Nvidia. I sell a chip to the data center. The
296. 27:08data center buys it from me because I make the  useful thing. And then the data center sells its
297. 27:12compute or rents it to one of the models. Right?  If you're selling the chip, why would you want to
298. 27:18be also buying the power of the data center? And  Nvidia has made a deal where they're basically
299. 27:24going to support the construction of a lot of  data centers. Yep. Meaning they're going to be
300. 27:29buying their own product essentially. Yes. They're  feeding money to themselves, right? So here's some
301. 27:34money for data centers so you can buy a bunch of  our chips, which is a little bit like it seems a
302. 27:39little bit like you're just paying yourself for  something. So, I'll give you the the really the
303. 27:43one that I think is going to blow up nasty. A  company called Corewave, AI compute company.
304. 27:48They're a public company. I think they they lose  money handover fist and they have tens of billions
305. 27:55of dollars of debt and they're making what are  they doing? They just build they have buildings,
306. 27:59they fill them full of GPUs. Nvidia invested in  them. Okay. Nvidia propped up their IPO. Nvidia
307. 28:05bought $2 billion worth of stock recently and  Nvidia is also one of their largest customers.
308. 28:10Yeah, that's their other customers are Open AAI,  Microsoft for Open AI, and Google unsurprisingly
309. 28:17for OpenAI. I'm not kidding you. Google is renting  compute from Coreweave to rent to Open AI. Oh,
310. 28:24wow. So, there are people that are doing compute  middlemen where they they rent they rent and then
311. 28:30they rent it out to someone else. And I don't  want to get too deep into it because we'll be
312. 28:35here forever. But there are also colloccation  companies who build data centers to rent to core
313. 28:40to rent to someone else. It's it's really bad when  you actually look at the non-hyperscaler or open
314. 28:46AAI compute. There's less than a billion dollars  of revenue on $178.5 billion of data center credit
315. 28:53deals done in 2025. Say that again. There's le who  who has less than a billion dollars of revenue?
316. 28:59Everyone. As far as people paying to rent GPUs,  when you remove all of the hyperscalers and open
317. 29:06AAI and Anthropic, right, it's less than a billion  of revenue last year. But doesn't that just mean
318. 29:10that the big ones are driving all the business?  Yeah, but the big ones are also not talking about
319. 29:17how much money they and in fact the big ones are  losing all the money and the one spending the most
320. 29:21money. Open AI is also burning so much money they  need to constantly raise billions of dollars. some
321. 29:28of it coming from Amazon and Microsoft and Nvidia.  At some point, you got to wonder if it's just the
322. 29:35same billions being cycled again and again, right?  And nobody making a profit other than Nvidia.
323. 29:41Nvidia is just printing money. Okay, that's the  one place. So, there is one place in this that
324. 29:46that people are genuinely making a profit, which  is, you know, I always use this example. I make
325. 29:52a sandwich for $2 and I sell it to you for $4.  Right? They Nvidia makes a chip for X dollars
326. 29:58and they sell it to someone for 2x or X plus Y.  They are definitely making a lot of money. Yeah,
327. 30:03the panini press guy, the panini press maker, they  are making the money, but the sandwich costs a
328. 30:08dollar and they're it cost them $10 to make. It's  really bad. And their only customers appear to be
329. 30:15themselves or very small amount of AI companies,  all of whom are terribly unprofitable. Right. So
330. 30:22they're they're definitely making a profit  and what you're identifying as the weakness
331. 30:26is the people they're selling to are not making a  profit. So Nvidia can make a is definitely making
332. 30:31a profit. They're booking profits. That's great  gross margins as well. Inarguably true. Their
333. 30:37stock is gone up hugely because they're doing  that. What you're saying is the people they're
334. 30:41selling to are not making a profit and a certain  point they can't keep buying if they're not making
335. 30:45a profit. Yes. And then the people that are  buying their chips that they're selling their
336. 30:51compute power to, which are the models, are also  not making a profit. And so at a certain point,
337. 30:58the music ends and people go diving for the  chairs. Because if the models aren't profitable,
338. 31:04then they don't need the data centers. And if  the data centers aren't profitable, then no one
339. 31:07needs the chips and the whole thing collapses.  There's also one abstraction that makes things
340. 31:11a little worse, which is when I say there isn't  there's less than a billion dollars last year of
341. 31:16AI compute revenue outside of the hyperscalers.  What I mean by that is it doesn't suggest there's
342. 31:22actually much revenue potential in renting an AI  data center. 178 Bloomberg reported at the end of
343. 31:28last year, $178.5 billion of data center credit  deals. So debt were done in America alone last
344. 31:34year. May even be more. That's a lot higher than  a less than a billion. The other thing is all of
345. 31:40these data center debt deals are basically done by  new companies. So all of the debt's kind of crap,
346. 31:46right? So these new companies, basically what's  happening is a bunch of new entrance are saying,
347. 31:50"Hey, I can find a warehouse, get a bunch of GPUs,  find electricity source, make a data center."
348. 31:58They're entering the market and they're floating  debt to make these new data centers with the idea
349. 32:03that when this all takes off, they're going to  have a steady diet of customers they can sell the
350. 32:08compute to because compute demand is going to go  up and up and up. But if the demand doesn't go up,
351. 32:16then that collapses and they won't be able  to pay the debt that they raised from private
352. 32:22credit that already has issues with uh people not  paying their debts cuz their due diligence wasn't
353. 32:28so good. Right. Right. So the so so the way that  this your understanding of of the sort of vector
354. 32:33that this gets into something that's a larger  financial problem is of how much of this paper
355. 32:38you know that there's there's a lot basically your  contention your thesis is that there's a ton of
356. 32:43bad debt floating around. Yes. And also this is  all happening in a historic downturn in venture
357. 32:50capital and in private equity. Since 2018 p uh  venture capital has failed to on average. There
358. 32:57are still some success stories of course to have a  TVPI total value put in of higher of higher than8
359. 33:04to 1.2 sounds complex. It just means for every  dollar you invest you get somewhere between 80
360. 33:09cents and $120 back. That's not very good. The  S&P 500 are beating the crap out of that. That's
361. 33:15happening with venture capital. Private equity  is also having the other problem which is private
362. 33:20equity is having trouble selling their companies.  There was the massive rush in the kind of software
363. 33:27era, the runup there where private equity bought  an absolute crap ton of software companies. 30 to
364. 33:3340%, the co-president of Apollo said this recently  of private equity deals between 2018 and 2022 were
365. 33:40for software companies, which means the private  equity firm and the software company took on debt.
366. 33:45And after that, of course, we had the well,  we had the 2021 era, the massive amounts of
367. 33:50insane crazy deals. The metaverse era, ton  of really bad companies got bought for 30,
368. 33:5640% higher than they should have been. So, you've  got private equity and venture capital sitting at
369. 34:02this time with a bunch of stuff they can't sell,  which means they don't have liquidity, which means
370. 34:08that they can't invest quite as much, and indeed  they themselves might have debt they have to pay.
371. 34:14This is happening at a time when technology and  the infrastructure behind it referring to AI needs
372. 34:20more money than it's ever needed ever. That's  that's the thing. So there's two parts of this
373. 34:25I want to push on. So one is if if you think about  this idea that look, we're going to take on a lot
374. 34:32of debt to build something out in the future that  isn't profitable now but will be. Okay, fine. That
375. 34:38people do that all the time. That's like that's  kind of the risk of investment. That's the risk
376. 34:41of investment. People do that all the time. That's  the basic model here. So then the question is okay
377. 34:46um one is can the co right now it's very  expensive and computer compute intensive
378. 34:53to do this but maybe it won't be in the future and  what I think is interesting about that question is
379. 35:00that might be a really good thing if that's true  for claw or open aai but if that were true it's
380. 35:10going to be a bad thing for all the data centers  and the GPUs right like the the principle right
381. 35:15now is you need a lot of computing power. The  computing power is being populated with these
382. 35:20huge physical infrastructures and enormous amounts  of investment. But maybe we'll figure out a way
383. 35:25there's some evidence that you know one Chinese  model has done this that you don't need all that
384. 35:29computer power and you can still get really you  can still get the same results. Even though that
385. 35:34would seem like a great innovation at some level  if that were true, it means that all of that
386. 35:40physical infrastructure is no longer needed or  valuable. Right? So we can get back to the fact
387. 35:46that it isn't getting cheaper. What DeepS did was  they trained cheaper but the cost of inference is
388. 35:51still going up because even if the model is  what do you mean by the cost of inference? So
389. 35:55the cost of inference is when you the amount  of money that it costs to create an output.
390. 36:00So you will see that models, some models have  got cheaper. People conflate that with it with
391. 36:06the companies themselves finding a cheaper way of  doing this. They've never said that. They've just
392. 36:11brought the price down. They can afford it when  they can raise 5, 10, $30 billion at a time like
393. 36:15Anthropic just did. What Deep Seek did was they  were able to train a model for cheaper. Right.
394. 36:20That's the Chinese company that sort of shocked  people and there was this big hit that happened
395. 36:24to the market because of it. Yes. Well, like they  they they were able to sort of shortcut this this
396. 36:29process. Yes. Because they couldn't access  the latest chips. But putting all that aside,
397. 36:33the other problem is that pre-training, which when  you shove all the data in, stopped having the same
398. 36:39results. We kind of hit the diminishing returns  point. So their only way to make these models do
399. 36:44more was to burn more tokens. So even if a model  cost comes down, you're using more tokens to do
400. 36:50the same thing. You're spending more money as a  user. We don't know what it costs them. They're
401. 36:54all unprofitable. But to your point, you're  completely right about these data centers. They
402. 36:59also have another problem which is takes about 2  years, 3 years to build an AI data center. Nvidia
403. 37:05is selling new chips every year. This seems like a  big problem, a depreciation problem, right? Well,
404. 37:10the depreciation problem is one in that they they  burn out in 3 to 6 years. We don't really know
405. 37:16yet, but I've heard crazy failure rates like 10 to  20% within a year. But we truly don't know that.
406. 37:22It's both the depreciation problem and the fact  that let's take Blackwell released kind of in 2024
407. 37:28but really in 2025. We still have data centers  being built like Stargate Abalene out in Texas
408. 37:33for Open AAI and Oracle that are using Blackwell  GPUs that by the time that bloody thing's built
409. 37:38which will be 2027 they will be 2 to three years  old right you will be have an entire data center
410. 37:44full of obsolete GPUs and all of the GPU data  centers being built last year are going to be
411. 37:50blackwell so you have just blackwell is what the  the it's the current gen the new gen is ver Ruben
412. 37:57this is just the GPUs you've Nvidia's. Yes. Yes.  Sorry, I should have said that. So, you've got all
413. 38:02these these data centers and now you've got this  flood of supply of an obsolete chip. I don't know.
414. 38:08I I I ain't no economics knower or anything, but  generally when the supply increases, they have to
415. 38:14lower the price cuz everyone's got it. And you're  already seeing the price of renting those GPUs
416. 38:19come down, right? Because they're because they're  they're older chips. And so they're going to the
417. 38:24same way that like you know a newer car sells for  more than a than a used car. But also there are
418. 38:30more and more of them coming online any every day.  Right. Right. Right. Right. And also we don't know
419. 38:34it. We don't I don't even think they're profitable  for the providers to run. Like it's really we
420. 38:39don't there is compelling evidence that no one's  making a profit renting them which is crazy. It's
421. 38:46crazy we're all doing this and we don't know  that for sure. Wait. meaning the the folks that
422. 38:50have the the centers, the actual data centers. I  hear I it's ruminance, so I I can't confirm it.
423. 38:55I heard of a data center out in North Dakota that  was losing a million dollars a day. That's that's
424. 39:00not a good business and it's crazy because so  all right. So so so let's say so one problem is
425. 39:07the time scale for building the data centers is  being outpaced by the new chips. You're building
426. 39:12things that are obsolete. There's also the threat  that happens of um that you actually come up with
427. 39:18more maybe you find more efficient ways in which  would you you have sort of stranded assets right
428. 39:22that you have all these data centers it turns out  you don't need all this compute because we've come
429. 39:26up with a a more efficient way to do it but again  the the the story that the AI people are telling
430. 39:33invest now it's not profitable keep building  and if we get to something that can for instance
431. 39:43do what a firstear associate at a law firm does.  Then you have a situation again I'm just this is
432. 39:50the this is the case right the case is you got  a situation we hire first- year associates they
433. 39:56largely do things like doc review and they draft  memos and we're going to have a model it's going
434. 40:02to be trained on legal stuff it's going to be a  you know enterprise system that Claude charges
435. 40:10$60,000 a year for a huge amount of revenue  would be like the most expensive software
436. 40:14basically the business case here is that  It's you. You hire a first year associate for
437. 40:22$120,000. We charge you $60,000, right? We're  making a ton of money. You're saving $60,000. And
438. 40:29it's a bummer that the first year law sort of gets  out of a job. But if we could do that at scale,
439. 40:35if there's millions and millions of these kinds  of jobs that people are making high five figures
440. 40:40to six figures that we can sell you software to  replace, I mean, again, this is the contention
441. 40:44of why it would be valuable. This is the core  contention like if you if you look into what
442. 40:50these companies are saying. I guess the question  then becomes is that a plausible outcome because
443. 40:57I think if it is plausible you could probably  make the math work and if it's not plausible then
444. 41:02you can't. So I actually the law firm example is  great. The problem is I don't think enough people
445. 41:08know what people do at jobs. Law firm associates  make law firms work. law firm associates are doing
446. 41:16the work that partners don't want to do. And if  any partners are listening, you know I'm bloody
447. 41:21right. So yeah, if you if you what you're  describing there would be AGI that just this
448. 41:26conscious computer, which by the way, everyone's  real excited to control a conscious creature
449. 41:32that's just describing slavery. It's important to  say what AGI is. It is describing slavery and it's
450. 41:38right. You're saying if you achieve what they call  artificial general intelligence, we actually just
451. 41:43had a conversation about this about with David  Chomers about consciousness. Um that then you're
452. 41:48actually there's all sorts of moral implications  of what that device is once you slave, right? Yes.
453. 41:54It it's literally But anyway, back to back to the  law slave. So this theoretical thing, yeah, if it
454. 42:00could do literally what a associate did, sure. But  an associate does much more than just dock review.
455. 42:08They're doing a bunch of research. And it's not  just I found a thing, right? Look, it's drafting
456. 42:14motions. If you get a motion wrong, a judge will  sanction you and you will embarrass the partner.
457. 42:20What you're paying for with employees in many  cases is actually risk management and judgment
458. 42:27and judgment and taste and culture and also risk  management. you're handing the risk off to a human
459. 42:33being that you can rely on and train. Also,  how are we going to make partners if we can't
460. 42:38make associates? We're just going to hire a law  student to become a partner. I mean, I don't know.
461. 42:43I could sit around handing other people's work and  talking. No, that's that's not true. Partners do
462. 42:46all sorts of work, I'm sure. But ne nevertheless,  yeah, in theory, if you could replace 10 $60,000,
463. 42:53$150,000 in the case of a law student, you could  replace 10 of them with $60,000. Sure. It isn't
464. 42:59doing that. And large language models are sold  as the reason I mentioned the thing earlier with
465. 43:04AI boosters. They need to be legally banned  from saying in the future it will could right
466. 43:08we need to talk about what's happening today. It  isn't doing it. It isn't doing it. And in fact,
467. 43:13every single example I hear of in specifically law  it large language models being used ends up with
468. 43:20someone getting in trouble with a judge. I think  they just had a DOJ person that this happened
469. 43:25to as well. Well, I don't think that's true that  every example because the people are using AI all
470. 43:29over the legal world, I can tell you. But there  definitely have been um hallucinated citations
471. 43:34that have been filed and I think in some cases  even by government lawyers, the DOJ uh that have
472. 43:39been caught that where they're they're citing to  a case that literally was invented by the AI. So,
473. 43:44but to your point, yeah, if you could do the  thing it doesn't do and has no proof of doing,
474. 43:48yeah, sure, grandmother had wheels should be a  bicycle and so on and so forth. A lot of this,
475. 43:54in fact, all of this is really sold on the coulds  and shoulds and wills. Yeah. It's not sold, it's
476. 43:58a bet about what its future capabilities are based  on what I would describe as semiotic knowledge. I
477. 44:07logic even. It's this idea that because things  have worked this way in the past, it'll happen
478. 44:12before. There was a time when the internet was  slower, then everyone's internet connectivity
479. 44:15went up. Not really the same thing because the  technology was always there to get it faster.
480. 44:20Fiber optic cable was there. The massive over bill  was there. This is not a problem that you solve by
481. 44:25having more compute. It is not. I mean they think  it is right. I mean that is the I mean just to
482. 44:30be clear about what the distinct disagreement  is. Their contention is that they have found a
483. 44:36reliable and straightforward law of scaling which  is that the more compute you have uh the better
484. 44:43it gets and the more that it starts to act in ways  that are intelligent. Except the scaling laws are
485. 44:48broken. That diminishing returns I mentioned  earlier. It's no longer getting the same kind
486. 44:52of improvements just by pre-training them. But  isn't it get I mean I just got to say like this is
487. 44:57where my experience of use of the models is that  they're getting much better at what specifically
488. 45:04multi-step tasks in research. So I if if you  use it for research so for here's a here's a
489. 45:10great example. You go to Claude and you say I said  this the other day. I'm trying to figure out the
490. 45:16relative homicide rates in major American cities  in the 1890s. I want to look at New Orleans,
491. 45:23which is what I'm writing about, and compare it  to New York and Philadelphia. A year ago, with
492. 45:30previous models, you would have gotten essentially  nonsense, or you would have gotten like, well,
493. 45:33here's the Wikipedia. Here's a few things. In this  case, it like went through it found like there's
494. 45:38two like real sources on this. Like there's a  book about southern homicides. There's another
495. 45:43book about northeastern policing. I know this  because I've actually done the research. Right.
496. 45:49Right. Um it goes through it basically does find  in one of the books because it's in public domain
497. 45:55what the New Orleans homicide rate is. It talks  about what the data difficulties are in New York
498. 45:59Philadelphia. And it basically spits out an answer  that I can check because it's citing it. That's
499. 46:04basically correct. Okay. In New Orleans, it's  25 out of uh, you know, I forget, 25 out of 100
500. 46:10thousand. And in New York and Philadelphia, it's  five, something like that. A computer could not do
501. 46:16that a year ago. Like, no, it just couldn't like  it. Now, there's all sorts of ways in which I can
502. 46:23check it because I have the expertise. But this  was like a sophisticated multi-step thing that it
503. 46:28had to go through and and and sort of use a bunch  of powers that it just didn't have a year ago.
504. 46:34I mean, you had to check every step though, didn't  you? You had to go and check all the data. I did
505. 46:39have to check the citations. Yeah. So, at some  point, I But so, what you're describing there is
506. 46:44an improvement. They have found ways to connect  them to web search tools, right? These things
507. 46:49are able to drag stuff, but what you're ultimately  describing is more sophisticated but less reliable
508. 46:56search. It is an improvement because they're able  to post train it in that case. and say this result
509. 47:04is bad, this result's good. I've used even the  most sophisticated ones used by hedge funds, the
510. 47:08searches. The problem is is that yeah, it will get  some things right and it will find the occasional
511. 47:14thing. Oh, you didn't see this in a 10K from 2  years ago. Problem is you have to check every
512. 47:19single bloody thing. You can't rely on anything.  You can't rely on a single thing. You perhaps it
513. 47:24helped you get in the right direction. Is that  worth this much money? is the and is I mean what
514. 47:30you're describing you were seeing in models middle  of 2025. I guess it's I guess we did something wow
515. 47:36we have better search more sophisticated but less  reliable search well or multi-step things. I mean
516. 47:42that the thing the thing to me was that this is  a fairly like compound task, right? So it has to
517. 47:46do it's it's got to do a bunch of stuff and and  the thing that I thought was striking was that it
518. 47:51actually it did a good job of finding the right  source which that was sort of interesting to me
519. 47:55like oh that is the book you know that is the  book where this is contained. You didn't just
520. 47:58like go to the Wikipedia page. I think the thing  that I I kind of come back to and this is the
521. 48:04sort of horns of the dilemma and many people have  sort of talked about this is that it seems to me
522. 48:10that there's no way out of some kind of cataclysm  for this reason. How do you mean either your case
523. 48:18is correct in which case it's just not going  to be profitable and the whole thing is going
524. 48:25to collapse in on itself. Yeah. or you're wrong  and they get a lot better and they are profitable
525. 48:33and what being profitable means is that they can  replace the labor of tens of millions of people,
526. 48:40right? That's another cataclysm. Like the point  is that like if they're right, if the thing that
527. 48:46they're promising, which is like, oh yeah, we  could start getting rid of all these people that
528. 48:50do all these jobs and replacing it with AI, that's  good for the profitability of these companies,
529. 48:56but I think it's probably insanely destructive  to America, the macroeconomy in American society.
530. 49:02Sure. And I'm not afraid of that because just you  just don't think that's going to happen. You see
531. 49:07no signs of it, right? I think that but it is the  only way like the only way it would make sense.
532. 49:12That's the point is the point is that like for  the math to work out it has to be something pretty
533. 49:18darn revolutionary and it has to be trillions  of dollar like I worked it out mathematically
534. 49:22by 2030 for any of this to make sense for  Microsoft Meta Google and Amazon they need
535. 49:27$2 trillion of new revenue not enhanced revenue  I mean brand new brand spec and new dollars in a
536. 49:33software industry that has never been higher than  $700 billion of yearly revenue in in an annual US
537. 49:39GDP that's like $30 trillion right so you're  talking about just 10% of the world, right? An
538. 49:44enormous part of the entire economy and it needs  to happen in the in the next 6 months. There is
539. 49:50one other thing though. Yeah. I think that there  is a social contam contagion that will happen with
540. 49:56this. Look around the world of bosses right now  and the amount of them who are like, "Oh yeah,
541. 50:02I can't wait to replace everyone. I'm going to  replace all the actors in my movies. I'm going to
542. 50:06replace all my workers. He's just going to give  me money and then I'm gonna have all the money
543. 50:10and the pieces of crap I sell things to as hogs  praying for my slop. The excitement in it. But
544. 50:17also, how many of them are just wrong? How many  of them just say things that aren't true? We have
545. 50:22people in newspapers saying things about AI that  aren't true. We have bosses claiming things about
546. 50:27AI that aren't true. We have people lying about  it. It's truly obscene. And regular people know.
547. 50:33regular people like if you go and talk to like  electricians, HVAC people, hairdressers, teachers,
548. 50:39their reaction to this is horror. There are some  who are using it to cut corners. Everyone wants to
549. 50:45do that. Human beings do that. Well, there's also  a lot of people that are like I mean there's also
550. 50:49a ton of people that have like crazily intense  parasocial relationships talk to it all the time.
551. 50:53I think that there should be criminal tribunals  for the companies that it's disgusting. Anyway,
552. 50:58I think that we are going to see something happen  before the economic stuff as an outcome of it
553. 51:04actually where regular people have seen who how  their bosses think of them and it's happened for
554. 51:11years. You saw it with remote work where bosses  were like, "Hey, you got to get back to the
555. 51:14office, man. I you got to get back there. I got to  be able to look at you every day. I got to be able
556. 51:18to stomp around so you can feed off my mood.  You got to go to the metaverse now cuz that's
557. 51:22where I'm going to be. Have fun staying poor.  You weren't in crypto. Also, I'm replacing you
558. 51:26with AI." So you've got that and then you've got  the other thing which is it needs to make all this
559. 51:32money now now I'm next 6 months open AI even then  raised $ 110 billion actually they only raised 15
560. 51:42billion 35 billion of the money from Amazon is due  when they achieve AGI or go public and uh both the
561. 51:4930 billion from Nvidia and the 30 billion from  SoftBank are being paid in $10 billion tranches
562. 51:54like like Cler um Not literally though. And what's  funny is SoftBank has to raise $40 billion in
563. 52:01loans to pay for their part. Everything that's  happening is a stress test of debt and equity.
564. 52:08How much can venture capital spend? How much can  hyperscalers afford? How much money is left in the
565. 52:14coffin? So then what's what is that out of your  theory? What emerges as a prediction of the first
566. 52:19place that you'll see a crack a fault? Like what  what would be the first sign? We're already seeing
567. 52:24it with private credit. So, I kind of hinted at it  earlier. Private credit, private equity, massively
568. 52:30bought so many different software companies and  when they bought them with these leverage buyouts,
569. 52:35they bought them, pumped them full of debt,  and then took on debt to buy them. I've heard
570. 52:39something ridiculous like hedge uh sorry, private  equity firms are leveraged to four to six times
571. 52:44the value of their assets. So, you're already  seeing it. There was a stat that came out the
572. 52:49other day. There's $42 billion of software loans  just for software companies that are in distress
573. 52:54status. So not likely to be paid, right? You  have across the board, you can go and look
574. 52:58inside there. They have to publish this private  equity firms, private credit firms and BDC's,
575. 53:03business development companies. Basically the same  thing. They're suddenly starting to take payment
576. 53:08in kind for loans as in you get stock, you get  given stock, and then you just kind of put all the
577. 53:15cost onto the end of the loan. They're not getting  paid on these loans. These loans are going,
578. 53:19they're starting to default. So, what are the  So, you you think you're going to start to see
579. 53:22loan defaults rip like the private credit market  is going to ripple first. We're already seeing
580. 53:26it. And then I think and my real my three horsemen  are you're going to see a data center project fall
581. 53:31apart before it's complete. You're going to see an  inconstruction one collapse and then you're going
582. 53:35to see a a fully constructed one that has to shut  down cuz it runs out of money. Because remember,
583. 53:41these things are debt. They they are heavily debt.  They are full of debt. There's not a single one
584. 53:47of them that is even close to profitable before  the debt and then you add the debt on top. I So
585. 53:52that's that's interesting. So those three things  to look for that in data centers. Exactly. And I
586. 53:56think that because the AI bubble in my opinion is  a symptom of the largest death of software as a
587. 54:03growth model because the assumption was software  eating the world Mark Andre uh was that every
588. 54:09industry could be software ties which is true and  that as a result all of them could grow forever.
589. 54:14private equity, venture capital bought into this.  They invested in all these companies except now
590. 54:19nobody will buy the companies. M&A has died and  it's harder to take them public. And if you can't
591. 54:24take it public and you can't send sell someone  else, what do you do? Well, the answer is you sell
592. 54:28them to another private equity firm. So, there's  currently a game of hot potato called continuation
593. 54:33funds or secondaries depending on who you ask.  Eventually, no one's going to have the money to
594. 54:38buy the thingy. And well, the other thing is that  at the core of all this, whatever financialization
595. 54:43you're doing, M you have to make things that are  profitable at the bottom of it. Yes, you do. I
596. 54:48mean that's that like whatever financialization  happens you can you know you can have periods
597. 54:53of where you buy an asset and you bloat it  with debt and you sell it to someone else,
598. 54:57right? And there are ways in some intermediary  sense that people can make money off passing
599. 55:02things hither and yan. But in the end you got  to make money. Things have to be useful and make
600. 55:08money for everything to keep going. And the thing  is all of these investments and Apollo's John ZTO
601. 55:14I think his name is he said that the whole thing  is is that these software companies were acquired
602. 55:20or invested in based on the idea that they would  grow like they did between 2005 and 2018. No,
603. 55:26they grow like after 2018 there's less money and  there's only so many software companies you can
604. 55:30build, only so many people you can sell to. The  last thing to me that also seems possible just
605. 55:38to end it here is that you know there's this  sense that like sometimes they'll talk about
606. 55:44um their own vision of artificial intelligent  intelligence being like a utility like electricity
607. 55:50or water. And what's interesting about that is  that like utility companies aren't that profitable
608. 55:57and they're not that sexy and they're not like  it's a strange thing because it seems to me like
609. 56:02maybe it is possible this becomes like a utility  that like everyone sort of has it or has access to
610. 56:07it. But if if that's the case then it's basically  just a commodity. Like it's not it's not actually
611. 56:11like utility companies are not super profitable.  In fact, they're regulated. And that's the thing,
612. 56:16the utility thing doesn't really make sense  because you water is water. An AI model is
613. 56:20many different things run by many different  companies that needs constant maintenance to
614. 56:25avoid model drift because otherwise it's just a  static object that cannot respond to new things.
615. 56:30You don't need to continually make sure power is  power before it turns into something else. Yes,
616. 56:35there are power of regulation things. I know. But  one other very scary thing to add, I don't like
617. 56:41scaring people, but some of the money that's going  into data centers now more and more in fact from
618. 56:47like Blackstone and so on and Harry's I hear is  from insurance funds and retirement funds because
619. 56:53they've needed more liquidity than they had. So  they've started investing in private credit loans,
620. 56:59the data centers and the sales pitch is simple.  This is the future. This is a good yield.
621. 57:05You'll get paid more on your money. without the  worrisome thing of will they always be able to pay
622. 57:11back their debt and we mentioned Oracle earlier.  Oracle is a mess but Oracle has taken over hundred
623. 57:18billion of debt. They had negative cash flow of  $24 billion and they are building 4.5 gawatt of
624. 57:25data centers. hundreds of billions of dollars and  they're building them for one company, Open AAI,
625. 57:31who lost at least 8 billion, more like 10 or  15 billion last year and expects to burn $230
626. 57:38billion by 2030. Oracle will die if or if OpenAI  dies. And this is not this is not a this is not
627. 57:48me catastrophizing. Mathematically speaking,  Oracle cannot pay the debt if OpenAI does not
628. 57:54make more money than Nvidia does right now by  2030. Yes, that's the the bet and they'll say
629. 58:00it themselves. Dario will say this and Sam Dario  Maday is that their their bet is on a revenue
630. 58:06trajectory that is essentially unprecedented in  human capitalism. They think they can achieve
631. 58:10that and if they don't, it doesn't work. Ed  Zitron, host of the better offline podcast,
632. 58:15writes the Where's Your Ed at newsletter on Ghost.  Great to have you here. Thank you for having me.