China Is About To Pop The AI Bubble — Transcript

Full transcript

1. 0:00So the whole US stock market including
2. 0:02your 401k, your index funds and the
3. 0:04value of your retirement is based on a
4. 0:07story that might be coming to an end.
5. 0:09>> You mentioned earlier the dot bubble.
6. 0:11Are we doing bubble 2.0 right now?
7. 0:14>> Oh, this is much bigger. The AI the AI
8. 0:18buildout relative to the TMT buildout of
9. 0:20992000 is multiples even as a percent of
10. 0:24the economy.
11. 0:25>> Okay. So, one of the big reasons why the
12. 0:28stock market is being held up right now
13. 0:30is because there's a story that American
14. 0:32companies are going to make trillions of
15. 0:34dollars in profits forever because the
16. 0:37world will be forced to use America's
17. 0:39technology.
18. 0:40>> You have a lot of data that you look at.
19. 0:42Do you think China is getting better at
20. 0:44AI? How is the
21. 0:46>> two there are two relevant tech centers
22. 0:47on two and a half America, China,
23. 0:49Israel. Those are the tech centers of
24. 0:51the world. They will win or we will win.
25. 0:54Now, some people say that that story is
26. 0:56coming to an end because of all the
27. 0:58lies, the spending, and the competition.
28. 1:01Now, on June 12th, something interesting
29. 1:03happened. A letter was sent to a company
30. 1:05in San Francisco. That letter was from
31. 1:08Howard Lutnik, the commerce secretary of
32. 1:10the United States. And by the end of the
33. 1:11night, the most advanced AI in the
34. 1:14world, was shut down. Andropic, which is
35. 1:17the company behind Claude, was ordered
36. 1:19to cut off two of its most powerful AI
37. 1:23models from every foreign national in
38. 1:26the world, not just China, by the way.
39. 1:28That order included countries like
40. 1:30France, Germany, Japan, and even
41. 1:33anthropics employees. If they weren't
42. 1:36American citizens, they were also locked
43. 1:38out. Now, 4 days after that letter,
44. 1:41France fires Palunteer. The French prime
45. 1:44minister said that we can't depend on
46. 1:46partners who are capable of turning off
47. 1:48the tap. But Germany already walked
48. 1:50away. Spain told its companies to stop
49. 1:53signing deals. The same goes for
50. 1:55Britain. And it's because the world
51. 1:58found out that it has a choice. And the
52. 2:01choice is literally 7 and 12 times
53. 2:04cheaper. Cuz while America is spending
54. 2:06$1 trillion a year building AI, which is
55. 2:093% of the whole US economy, China is
56. 2:13spending a fraction of that and giving
57. 2:16it away virtually for free. So in this
58. 2:19video, I want to explain how China is
59. 2:21competing with the US and how this AI
60. 2:24story might be coming to an end and some
61. 2:26of the things that you can use to
62. 2:28potentially see this bubble popping
63. 2:30before anybody else. So with that said,
64. 2:32let's get into it. Hi, my name is Henri
65. 2:34Jick. Hope you're doing well. Come for
66. 2:36the finance and stay for the AI bubble
67. 2:38everyone saw coming. So, let me just
68. 2:40start with a basic question. Do people
69. 2:42really want this AI technology? Because
70. 2:45there's a theory that says the reason
71. 2:47that this is such a prevalent story in
72. 2:49the market is so that these tech
73. 2:51companies could justify their insanely
74. 2:54high stock prices because in reality
75. 2:56they've run out of really good ideas. In
76. 2:59fact, there's a really good interview on
77. 3:00CNBC with Ed Zitron who brought up a lot
78. 3:03of really great points.
79. 3:04>> But fundamentally, large language models
80. 3:06are not the future. The only reason big
81. 3:08tech is investing in this is that
82. 3:09they've run out of hyperrowth ideas.
83. 3:11They don't have a next iPhone. They
84. 3:13don't have a new Google search. So,
85. 3:14they've put over a trillion dollars with
86. 3:16trillions more to come into a kind of a
87. 3:18deadend industry because when they when
88. 3:20that ends, they'll have to admit that
89. 3:22they don't have anything else. Now,
90. 3:23throughout the video, I'm going to show
91. 3:25you more clips from that interview, but
92. 3:27there was also an interview with Alex
93. 3:29Karp, who is the CEO of Palunteer, which
94. 3:32if you don't know is the company that
95. 3:34works closely with the government and
96. 3:36pretty much every three-letter agency in
97. 3:38the world. And Alex also brings up the
98. 3:41fact that nobody really trusts AI right
99. 3:44now.
100. 3:44>> Who owns the data? Where is it cached?
101. 3:47Are the prompts secure? Is this being
102. 3:49transferred to you? Are you being comp?
103. 3:52Okay, if it was so valuable, let's say I
104. 3:54can make you a billion dollars right
105. 3:55tomorrow. Wouldn't I say I'll make you a
106. 3:58billion dollars and I want 30%. Why are
107. 4:00they charging for tokens if it's so
108. 4:02valuable? He is saying if the promise of
109. 4:05AI is as good as they are marketing it
110. 4:08to be in its current form, they would
111. 4:11not be charging us for tokens. Instead,
112. 4:13they'd be charging us for building a
113. 4:15billion dollar business idea where they
114. 4:18would take 30% of the revenue. Cuz think
115. 4:20about how you pay for anything in
116. 4:22business. You pay a lawyer to win a
117. 4:24court case. You pay a contractor to
118. 4:26remodel your house. The price you pay is
119. 4:29attached to a specific result. Now, AI
120. 4:33companies do not work that way. They
121. 4:35charge us per what's called token usage.
122. 4:39Now, a token is basically a word. Every
123. 4:42word the AI reads and every word it
124. 4:44writes for you, we pay for that. whether
125. 4:47the answer was good or was really bad.
126. 4:49And Alex Karp is basically saying why
127. 4:52would they price their business that
128. 4:54way? Why wouldn't Open AI chat GPT just
129. 4:57say only pay me when it works? If we
130. 5:00create a good idea for you, give us a
131. 5:02cut of your income. But I'm telling you
132. 5:05in this country at every single
133. 5:07enterprise I deal with they these people
134. 5:09are livid. They're like I am paying for
135. 5:12tokens that create no value. These
136. 5:14people are stealing the weights and
137. 5:16alpha of my business and they're
138. 5:17creating a wealth tax that does not help
139. 5:19the poor. It just punishes starts with
140. 5:21the billionaires. Every single person at
141. 5:23this table is going to be paying a
142. 5:24wealth tax only to punish us.
143. 5:27>> If they were confident that this thing
144. 5:28created this value, that would be the
145. 5:31easiest sales pitch in history. Pay us
146. 5:33nothing unless we make you money and
147. 5:36unless we build you a billion dollar
148. 5:38idea. But they can't offer that because
149. 5:41these models do what's called
150. 5:43hallucinate. Right? This is where they
151. 5:45confidently make things up. And nobody,
152. 5:48including the people who built them,
153. 5:50could tell you when or really why it
154. 5:52happens. Right? No one's been able to
155. 5:54figure out how to fix it completely.
156. 5:56>> You'll notice that both Anthropic CEO
157. 5:58Darama Day and Sam Wman have both said,
158. 6:00"We can't wait to see what you build
159. 6:01with this." Well, that's because they
160. 6:04don't know what you can build with this.
161. 6:05They want everyone else to do their
162. 6:06innovation for them. spend as much as
163. 6:08they can on tokens and then take
164. 6:09whatever's left except they lose too
165. 6:11much money for that strategy to actually
166. 6:13work.
167. 6:14>> And that puts every corporation in
168. 6:16America in a very awkward situation
169. 6:19because let's say you're the CEO of a
170. 6:21corporation, right? You just spent $50
171. 6:23million on AI this year. So your board
172. 6:26of directors asks you a question.
173. 6:28They're like, "What did we get for this
174. 6:30$50 million we just spent? What's the
175. 6:33ROI?" And you're like, "I don't know,
176. 6:35right? We don't have a number. Nobody
177. 6:37has a number. Corporate America's paying
178. 6:40subscription fees on a technology whose
179. 6:42outcome we cannot measure. And it gets
180. 6:46worse though because not only can we not
181. 6:48measure it, we are also risking our
182. 6:52company secrets and potentially creating
183. 6:54a competitor. Here's what Alex has to
184. 6:57say about that.
185. 6:57>> But something has gone completely wrong.
186. 7:00And the basic view among enterprises in
187. 7:03this country is I'm going to chill lax
188. 7:06uh and waste my time with tokens. I'm
189. 7:08going to get no value and they're going
190. 7:10to get my IP.
191. 7:11>> The fear for all these CEOs is that when
192. 7:14your company uses these models, your
193. 7:17data flows through them, right? Your
194. 7:19process, your trade secrets, the special
195. 7:21sauce that makes you profitable, which
196. 7:23Alex Karp calls the alpha. So what
197. 7:27happens when the AI company that you use
198. 7:31learns from your business? It just
199. 7:33becomes your competitor. And this is not
200. 7:34a hypothetical thing by the way.
201. 7:36Anthropic launched a design product
202. 7:38called Claude Design while having a
203. 7:41relationship with a company called
204. 7:43Figma, which is a design company. Figma
205. 7:46CEO publicly said he was shocked. So,
206. 7:50picture being a business and then
207. 7:53watching that. You're paying your vendor
208. 7:55millions of dollars a year to use their
209. 7:57AI, but what you're actually doing is
210. 8:00paying them and training your own
211. 8:03replacement. So, what's the solution
212. 8:06then? This is where it gets really
213. 8:07interesting.
214. 8:08>> But what is happening among the most
215. 8:10technical players is they're saying, "I
216. 8:13want something I own. This is my
217. 8:15business. I want to own the GPUs. I want
218. 8:18to own my data. I want to own the model.
219. 8:20I want to control the alpha. Why would
220. 8:22they get access to my data? If they're
221. 8:24going to build my alpha, why wouldn't I
222. 8:26control the weight?
223. 8:27>> Right? What that means is instead of
224. 8:29renting an AI from someone, instead you
225. 8:32just download one, you run it on your
226. 8:35own computers using your own data where
227. 8:37no one can see it and no one can learn
228. 8:39from it and also no one can shut it off.
229. 8:42He then goes on to say though that most
230. 8:44businesses don't even need the latest
231. 8:46and greatest cuttingedge AI because you
232. 8:49don't need the smartest one to process
233. 8:52insurance claims, right? You need one
234. 8:53that's specifically really good at
235. 8:55insurance claims. And what's interesting
236. 8:57is that Alex Karp profits from this
237. 9:00business model as well. So why would he
238. 9:02be saying all this? What's his motive?
239. 9:04Cuz remember France, Germany, Spain,
240. 9:07they're canceling their contracts.
241. 9:09Palanteer has been losing those
242. 9:10contracts across Europe all year. Now, 2
243. 9:13days before that interview, Palanteer
244. 9:16announced a partnership with Nvidia to
245. 9:18sell open models in sovereign
246. 9:21environments.
247. 9:22Basically, that interview was a product
248. 9:25launch for his new service, which is why
249. 9:28he's out there telling other companies
250. 9:29to download and own their own AI. So,
251. 9:33that is the first problem with AI.
252. 9:36Nobody trusts it. But there's a second
253. 9:38problem that's even bigger. Now, before
254. 9:40I explain the second problem, everything
255. 9:42in this video, like the AI spending and
256. 9:44whether this is a bubble at all, depends
257. 9:46completely on where you're reading it.
258. 9:47And that's where today's sponsor, Ground
259. 9:49News, comes in. Ground News is an app
260. 9:51that shows you the same story from
261. 9:52hundreds of different outlets at the
262. 9:54same time. And it tells you which ones
263. 9:56are left-leaning, right leaning, or
264. 9:57center, so you can see how the story
265. 9:59changes depending on the source. Perfect
266. 10:01example, South Korea just announced a
267. 10:03huge national AI and chip investment.
268. 10:05Over 159 news sources covered it. And
269. 10:08here's what ground news shows.
270. 10:10Left-leaning outlets frame this as a
271. 10:11historic industrial strategy, stressing
272. 10:14the huge scale, and they lead with the
273. 10:16market's outcome. Right leaning outlets
274. 10:18lean into the urgency and survival using
275. 10:20phrases like race against time, framing
276. 10:23it as existential for South Korea's chip
277. 10:25industry. Now, center outlets just skip
278. 10:27the drama and focus on policy execution.
279. 10:30Even the headline number changes
280. 10:31depending on the source. Some report a
281. 10:34thousand trillion one, others up to
282. 10:352,000 trillion, some just say 1.2
283. 10:38trillion, but it's the same announcement
284. 10:39with three different narratives. That's
285. 10:41what ground news makes visible. I use it
286. 10:43when I'm researching for these videos so
287. 10:45I can separate what's actually happening
288. 10:46from how it's being told. If you want to
289. 10:48see the full picture of global events
290. 10:50and not just one side, go to
291. 10:51ground.news/jick
292. 10:53or click the link down below. You'll get
293. 10:5540% off the Vantage plan, which is the
294. 10:57one I use, and you'll be supporting
295. 10:58independent journalism and this channel.
296. 11:00Thank you to Ground News for sponsoring
297. 11:02this segment. And now, let's get back to
298. 11:03it. Now, the second problem with AI is
299. 11:05that even if every company in America
300. 11:07trusted these AI companies, the money
301. 11:10still does not make sense because the
302. 11:12business model is broken in a way we've
303. 11:14never seen from tech companies. Because
304. 11:17here's how software is supposed to make
305. 11:19money. Software is the greatest business
306. 11:21model ever invented because you spend a
307. 11:23lot of money building the thing at once,
308. 11:26right? And then every new customer is
309. 11:28basically free money. That's why tech
310. 11:30stocks have done so well over the past
311. 11:32few decades. When you buy Microsoft
312. 11:35Excel, right, Microsoft doesn't spend
313. 11:37anything extra to sell you that copy.
314. 11:40Their costs stay the same, but their
315. 11:42revenue goes up. And it's the gap
316. 11:44between those two lines that is their
317. 11:46profit. And that gap is why tech
318. 11:48companies, the most valuable companies
319. 11:50on Earth. Now, AI broke that model. And
320. 11:54how they broke it was every single time
321. 11:57you ask Chad GPT a question right now,
322. 11:59it costs Soap AI money cuz it uses
323. 12:02electricity. Chips are being worn down.
324. 12:05So more customers does not translate to
325. 12:08free money anymore. More customers means
326. 12:11more cost dollar for dollar. So AI is
327. 12:15not a software business. It's more like
328. 12:17a restaurant, right? Every time a meal
329. 12:19gets served, somebody has to buy the
330. 12:21ingredients every time. Except this is a
331. 12:24restaurant that loses money every time
332. 12:26it serves food. And its plan to fix it
333. 12:29is to serve more food. Now, let me give
334. 12:31you some context. In 2025, Open AAI
335. 12:35burned over $20 billion in just one
336. 12:38year.
337. 12:38>> Well, they'd be the first to be this bad
338. 12:40other than we work. And even then, this
339. 12:42is so much worse than that. Open AAI
340. 12:43burned $20.9 billion in 2025. That's the
341. 12:46auditive financials that the FT and II
342. 12:48reported. And the problem with these
343. 12:50companies is their margins are getting
344. 12:51worse and they actually their costs
345. 12:53increase linearly with their revenues.
346. 12:55>> So he's basically saying that costs
347. 12:57increase linearly with revenues, right?
348. 13:00The two lines are going up together and
349. 13:02the gap never really opens up. Now for
350. 13:0525 years, every investor has been
351. 13:08trained to be patient with this cuz they
352. 13:11say, "Well, they're losing money now,
353. 13:12right? But at scale their margins will
354. 13:14get better, right? Amazon lost money for
355. 13:17years. Except with AI, we just keep on
356. 13:20waiting and the margins are getting
357. 13:22worse cuz every new model costs more to
358. 13:26run than the last one. And the market is
359. 13:29starting to notice it.
360. 13:30>> There is no proof that they can improve
361. 13:31their margins. No amount of specialist
362. 13:33silicon or supposed Vera reubans will
363. 13:35bring these costs down. And we're at a
364. 13:37point now where OpenAI is now
365. 13:39potentially pushing their IPO to 2027
366. 13:41because they couldn't get a trillion
367. 13:42dollar valuation. It's clear that people
368. 13:44are wising up to the problem of
369. 13:46generative AI, which is there's not
370. 13:48really a business there.
371. 13:49>> Now, all of this, by the way, is not
372. 13:50just open AI cuz look at who's paying to
373. 13:53build all of this. This is from Oracle's
374. 13:56annual report.
375. 13:57>> Oracle is a particularly scary one
376. 13:58because they are building 7.1 gawatt of
377. 14:01capacity just for one customer. And they
378. 14:03even said in their annual report that
379. 14:04the risk was they might not get paid.
380. 14:06Open AI only loses money and I think I
381. 14:09estimate it's like $75 billion of
382. 14:11revenue annually that they will have to
383. 14:13pay for the full Stargate data center
384. 14:14project in annual compute revenue. Open
385. 14:16AAI can't afford that and if they can't
386. 14:18Larry Ellison can't afford to pay back
387. 14:20those bills and Oracle stock will be in
388. 14:22jeopardy along with the margin loans
389. 14:23that Mr. Ellison holds. It's genuinely
390. 14:25dangerous.
391. 14:26>> So Oracle is building the equivalent of
392. 14:29several nuclear power plants worth of
393. 14:31electricity for basically one customer.
394. 14:34a customer that just lost $20 billion.
395. 14:37Here's my favorite one, though. Nvidia
396. 14:40sells its chips to a group of smaller
397. 14:43cloud companies. They're called
398. 14:44NeoClouds. Now, those companies borrow
399. 14:47billions of dollars to buy Nvidia's
400. 14:48chips and Nvidia rents them back.
401. 14:52>> I think companies like Core and
402. 14:54especially Nebius and Iron and Cipher
403. 14:56Mining and all of them, Terowolf as
404. 14:58well, they are all very they're
405. 15:00basically outgrowths and they're
406. 15:02subsidiaries of Nvidia. Nvidia is now
407. 15:05according to the information going to be
408. 15:07paying them to rent back their GPUs when
409. 15:09they install them in the data center.
410. 15:11This is the this is something that only
411. 15:12happens in an industry without diverse
412. 15:15and real demand.
413. 15:16>> So what he's saying there is that
414. 15:17Nvidia's sales are partially funded by
415. 15:20Nvidia. That's like a car dealership
416. 15:22lending you money to buy a car and then
417. 15:25paying you to borrow the car back for
418. 15:27the weekend and then reporting all of it
419. 15:29as demand. Now look how much profit
420. 15:31we're making, right? Yeah, because you
421. 15:33are buying back your own equipment.
422. 15:35There's a name for when an industry
423. 15:37starts doing this. It's called not
424. 15:39enough real customers. So for companies
425. 15:41investing trillions in AI like
426. 15:43Microsoft, Google, Amazon, Meta, what is
427. 15:47the ROI from all their spending? They
428. 15:50won't tell you. These companies report
429. 15:52everything. Cloud revenue, ad revenue,
430. 15:54YouTube revenue. But AI revenue, they're
431. 15:58not telling us that. Microsoft, Google,
432. 16:00and Meta, and Amazon are all doing a
433. 16:02funny little I don't want to call it a
434. 16:03scam, but it's a a trick where because
435. 16:05their other businesses are still
436. 16:07growing, but they never disclose their
437. 16:08AI revenues, everyone conflates that
438. 16:10with AI driving their growth. In
439. 16:11reality, their other businesses are
440. 16:13growing and AI is losing them money
441. 16:15across the board. You'll notice that
442. 16:17neither Microsoft or Amazon, who both
443. 16:19share their run rate of AI, will share
444. 16:21the actual AI revenues. That tells you
445. 16:24that these companies are afraid. Public
446. 16:26companies love good news. If they had
447. 16:27good news, why wouldn't they share it?
448. 16:29That's because they've only got bad news
449. 16:30here.
450. 16:30>> Now, as of right now, the stock market
451. 16:32is still patient and investors are
452. 16:34saying, "Okay, give it time still." But
453. 16:37all of it really depends on one big
454. 16:40assumption, which is that if and when
455. 16:43the profits do come, it's going to be
456. 16:46the American companies that will make
457. 16:48the profits because the world has no
458. 16:51other option. But the third problem with
459. 16:53AI is that the world has another option.
460. 16:56That option is called China. So, let me
461. 16:58show you what China is really doing.
462. 16:59Remember this chart from the beginning
463. 17:00of the video. The trillion that America
464. 17:03is spending. Well, here's something
465. 17:04interesting. This is China. On that same
466. 17:07scale, America, $764 billion this year,
467. 17:12and then 1 trillion next year, 3% of the
468. 17:16whole US economy. China, 102 billion
469. 17:20this year, 123 billion next year. 0.6%
470. 17:246% of their economy, which means
471. 17:27America's outspending China almost 10
472. 17:30to1. Why? It's cuz China figured
473. 17:33something out. A developer took the
474. 17:36exact same coding task and gave it to
475. 17:39two AI models, Claude Opus, which is one
476. 17:41of the top American models made by
477. 17:43Anthropic, and GLM, which is a Chinese
478. 17:46open model. Both models finished the
479. 17:49same task, and both took about 5 1/2
480. 17:52minutes. The American model charged
481. 17:55$2.33
482. 17:57and the Chinese model charged 31. That's
483. 18:017 12 times cheaper. Now, before you
484. 18:04think that that's a cherrypicked test,
485. 18:06here is the industrywide data. This is
486. 18:10called the artificial analysis
487. 18:12intelligence index. And this is
488. 18:14basically the official rankings of every
489. 18:17AI model in the world. Now look at the
490. 18:19top. The best American model scores 60.
491. 18:23Now look right here. This is GLM. The
492. 18:26best Chinese open model 51. And look at
493. 18:29how much of this chart is from China.
494. 18:32Deepseek, Quen, Kimmy, Miniax. They're
495. 18:35not at the top. They are everywhere.
496. 18:37They are filling the whole middle of the
497. 18:40global rankings. Now to be fair, the US
498. 18:43still has the smartest AI in the world.
499. 18:45And that's true. But if you ask the
500. 18:47question that every business is asking,
501. 18:50do I need the fastest AI model to manage
502. 18:53my business? The answer is no. For most
503. 18:56businesses, that's things like answering
504. 18:58their customer service emails to
505. 19:01basically do the boring work that is 90%
506. 19:04of what companies actually use AI for.
507. 19:08Based on that logic, China's winning.
508. 19:10The US is winning the race for the most
509. 19:12dollars spent, and China is sort of
510. 19:13winning the race for the customer. Now
511. 19:15the question is how is China doing this
512. 19:18while spending 10 times less money? And
513. 19:20the answer is distillation. Okay, here's
514. 19:23how it works. When you train a frontier
515. 19:26AI model from scratch, that means
516. 19:28spending billions of dollars teaching it
517. 19:30everything the hard way. But there's a
518. 19:32shortcut. You can train your model by
519. 19:35studying the answers of a model that
520. 19:37already exists. This is basically like
521. 19:39copying someone else's homework and then
522. 19:40you know the answer for almost no money
523. 19:42spent, by the way. So the US spends
524. 19:45trillions of dollars doing the hardest
525. 19:47research in human history and then China
526. 19:49just sort of copies it by compressing
527. 19:51the results into smaller cheaper models
528. 19:53and then it gives them away for free. It
529. 19:56open sources them which means anyone can
530. 19:58download them. And if you think about it
531. 20:00every dollar of US AI spending it's kind
532. 20:03of like a donation to the Chinese AI
533. 20:06industry. And this has become common
534. 20:09practice for China. So much so that they
535. 20:12are doing it as a side hustle. Check
536. 20:14this out. This is a model, for example,
537. 20:16called LongCat. Look at the benchmarks.
538. 20:18It's going toe-to-toe with Google's
539. 20:20Gemini, and it's beating older versions
540. 20:23of Anthropic's flagship models on
541. 20:25realworld agentic tasks. But the thing
542. 20:28is, do you know who built Longat? It's a
543. 20:31company called Mtoan. Do you know what
544. 20:33Mtoan does? It's a food delivery
545. 20:37company, right? It's the Chinese Door
546. 20:39Dash equivalent. And they built an AI
547. 20:41that competes with the smartest labs in
548. 20:44the United States. Which means when a
549. 20:47company like that can do what a trillion
550. 20:50dollar US company is doing, is that
551. 20:53company still actually worth trillions
552. 20:55of dollars? Maybe not. Cuz remember the
553. 20:58assumption that's holding up the whole
554. 21:00AI stock market is that let's be patient
555. 21:03guys. the profits will come and when
556. 21:05they do all the US companies will
557. 21:08collect those profits because the world
558. 21:09has no other choice. But here's what the
559. 21:12actual cost is when there is a choice.
560. 21:15Right? This is the same type of work.
561. 21:17The American model shows 18 1.5 cents
562. 21:20per task and the Chinese model 4 cents.
563. 21:24Right? Within a few quality points of
564. 21:25each other at a 76%
565. 21:28discount. This is sort of the chart that
566. 21:31destroys the whole story because you
567. 21:34cannot make back a trillion dollars
568. 21:37selling something that your competitor
569. 21:40is giving away at 90% of the quality for
570. 21:44just 10% of the price. So, let me tie
571. 21:46all of this together. If all of this is
572. 21:48true, then when does this bubble pop, if
573. 21:51ever? And logic says it's when
574. 21:54corporations stop their capex, their
575. 21:57capital expenditures. It's when they
576. 21:59stop spending money building all these
577. 22:01data centers. But believe it or not,
578. 22:04that is not when the bubble pops.
579. 22:07According to the data, data shows that
580. 22:09the bubble could pop a lot sooner. And
581. 22:12here's why. During the dot bubble, the
582. 22:16NASDAQ index peaked in March of 2000.
583. 22:19Now, all the companies that were laying
584. 22:21the fiber optic cables at the time,
585. 22:23which are the data centers of that era,
586. 22:25they kept spending billions of dollars
587. 22:28well into 2001, even though the stock
588. 22:32market collapsed a full year before
589. 22:35their spending stopped. So, the market
590. 22:38did not wait for companies to stop
591. 22:40spending and to admit to anything. The
592. 22:43logic of the market changed when enough
593. 22:45investors stopped believing in that
594. 22:48story. So the trigger this time around I
595. 22:51think will be something a lot more
596. 22:53subtle. Something like a big tech
597. 22:56earnings call where a CEO says something
598. 22:58like we are moderating uh the pace of
599. 23:01our infrastructure investment or some
600. 23:03boring small thing like that. And that's
601. 23:05because the first company that gets
602. 23:08rewarded by Wall Street for cutting
603. 23:10their AI spending that will give every
604. 23:14other CEO permission to do the same
605. 23:16thing. In fact, according to Ed Zitron,
606. 23:18Goldman Sachs recently said that the
607. 23:20first hyperscaler to pull back on
608. 23:23spending will get rewarded by the
609. 23:26markets.
610. 23:26>> So, I heard Goldman analysts say
611. 23:28recently that the first hyperscaler to
612. 23:30pull capex will get rewarded by the
613. 23:32markets. I think the capex pullbacks are
614. 23:34they're the sign. I also think any
615. 23:36financing falling through Baro and AI or
616. 23:38anthropic would be a sign, but I think
617. 23:40we're going to start seeing AI companies
618. 23:42kind of start falling out of favor and
619. 23:44not being able to raise money. But the
620. 23:46big thing is debt. When data center debt
621. 23:48stops being issued, that will be when
622. 23:50it's bedtime for this industry because
623. 23:52even if they think AI is going to win,
624. 23:54we've got 100 gawatt or so of data
625. 23:56center capacity allegedly under
626. 23:57construction or in planning. That's
627. 23:59trillions of dollars of money that needs
628. 24:01to come from somewhere and we are
629. 24:03tapping out the debt markets. We saw
630. 24:04that with Google raising that $85
631. 24:06billion equity raise.
632. 24:08>> That's going to be one of the early
633. 24:10signs. Now, another sign that we could
634. 24:13be at the peak of the bubble is the bond
635. 24:15market. That's because unlike the stock
636. 24:18market, which runs on stories of hopes
637. 24:20and dreams, the bond market doesn't work
638. 24:23like that. All bond investors care about
639. 24:27is will I get paid my interest payment.
640. 24:30Right? The moment they get scared, they
641. 24:33start to demand a much higher interest
642. 24:35rate. Now, how we measure their fear is
643. 24:38something called a credit spread. Here's
644. 24:41how that works. In the world of
645. 24:43investing, there's a concept called the
646. 24:45riskfree interest rate. It's called that
647. 24:49because it is set by the US government
648. 24:52which is considered to be the safest
649. 24:54borrower on earth. That's government
650. 24:56bonds, right? Whatever they're at,
651. 24:58that's the risk-free rate. Okay? But
652. 25:01remember, companies can also issue
653. 25:04bonds. Except because companies are
654. 25:06risky, cuz they can go out of business,
655. 25:09their bonds pay that risk-free rate plus
656. 25:15something extra to compensate you for
657. 25:18the risk that their company could go
658. 25:19broke and never pay you back. Makes
659. 25:21sense, right? Well, that extra between
660. 25:24the risk-free rate and their rate, that
661. 25:28is called the spread. Think of it as an
662. 25:30insurance premium. When lenders feel
663. 25:33safe, the premium is small, meaning the
664. 25:36spreads are what's called tight, meaning
665. 25:38corporate bond rates are close to the
666. 25:40risk-free rate. But when investors feel
667. 25:44like there's some market risk, the
668. 25:47premium explodes, right? The spread
669. 25:51increases. That is one of the early
670. 25:53signs that we could start to see that
671. 25:55this is going to fall apart. Now, here's
672. 25:57an example. By the way, see this
673. 25:59increase in 2008. Spreads hit almost
674. 26:0222%.
675. 26:04Lenders started to charge very high
676. 26:06prices. Credit shut off completely and
677. 26:08companies that ran on borrowed money
678. 26:10just collapsed. Also see the jump in
679. 26:132020. Now look at today. The spreads are
680. 26:16very tight. 2.6%.
681. 26:19That is close to the lowest and the
682. 26:21calmst readings in recorded history.
683. 26:24What this means for now is that either
684. 26:27bond investors see no problem and
685. 26:30everything in this video is completely
686. 26:31wrong or bond investors are wrong and
687. 26:35they can be wrong. Look at early 2007.
688. 26:38The housing crisis was already underway.
689. 26:41Bear Sterns was months away from blowing
690. 26:43up and spreads were only 2 1/2%. They
691. 26:47were super calm right where they are
692. 26:50today. Right? The fear gauge didn't
693. 26:52predict 2008, though. That's because
694. 26:54spreads don't really measure what is
695. 26:57true. They measure what lenders believe.
696. 27:00In 2007, lenders believe the housing
697. 27:04market was safe. Which now we know
698. 27:05obviously that it wasn't. But the point
699. 27:08is is that when you see someone on the
700. 27:09news say, "Hey, don't worry. AI is
701. 27:12totally safe. It's doing great. Credit
702. 27:14markets, right? The spreads are not so
703. 27:16worried." Right? If you hear that,
704. 27:18remember that the credit market wasn't
705. 27:20worried in ' 07 either. Credit markets
706. 27:22can be wrong and they have been wrong
707. 27:25before.
708. 27:25>> Although you you'd agree that spreads do
709. 27:27not do not imply that that moment is
710. 27:30anytime soon.
711. 27:30>> Spreads have been wrong before. That's
712. 27:32the thing. And I think that perhaps the
713. 27:34timing isn't going to be immediate, but
714. 27:35at some point a hyperscaler is going to
715. 27:37pull back capex. And when that happens,
716. 27:40well, this is an industry of followers.
717. 27:41The tech industry doesn't have ideas.
718. 27:43They just copy each other. Everyone
719. 27:44copied Sachin Nadella when he put chat
720. 27:47GPT in Bing. And I think that whoever
721. 27:49breaks capex first, they'll follow them,
722. 27:51too.
723. 27:51>> And finally, I just want to show you
724. 27:53what Michael Bur posted a few days ago.
725. 27:55And remember, he's the guy who predicted
726. 27:57the 2008 financial crisis. So, in chart
727. 27:59one, he shows chip stocks are trading at
728. 28:02the top of their 15-year valuation
729. 28:06range. Basically, the same peak that
730. 28:08they hit right before the 2024
731. 28:09correction, which is marked with those
732. 28:11red circles. The market is basically
733. 28:13pricing chips like the trillion dollars
734. 28:15has already been made. Now chart two is
735. 28:18even more interesting. This tracks the
736. 28:20three groups of AI stocks since last
737. 28:23year. Now the gray and white lines going
738. 28:25up to 200% are the AI winners, right?
739. 28:28The companies selling the chips and the
740. 28:30equipment. But the orange line way at
741. 28:32the bottom that's barely above zero are
742. 28:36the hyperscalers, right? Companies like
743. 28:38Microsoft, Google, Amazon, Meta. What
744. 28:41does that mean? It means the market is
745. 28:43telling us that the companies that are
746. 28:46doing the spending, the trillions of
747. 28:48dollars, right, they're getting almost
748. 28:50no credit for it. Their stock values
749. 28:52aren't really going up. Wall Street
750. 28:54instead is rewarding the companies that
751. 28:57are getting that money and it's ignoring
752. 28:59the companies spending to build it.
753. 29:01Right? That's basically the market
754. 29:02admitting it doesn't believe the
755. 29:05spenders will make it back. And in the
756. 29:07third chart he posted, it shows the
757. 29:10Silicon Data LLM token expenditure
758. 29:12index. It's a fancy name, but what it
759. 29:15shows is it shows us the price that
760. 29:17people pay for AI tokens. This index is
761. 29:21the price of AI itself. And look at it.
762. 29:24It's down almost 20% from its high in
763. 29:25May. Now, the question is, why would the
764. 29:28price of AI be going down during the
765. 29:30biggest AI buildout in history?
766. 29:32Bloomberg says either it's because
767. 29:34demand is going to cheaper models or
768. 29:36buyers are just not willing to pay more.
769. 29:39Look at the middle one. Demand is
770. 29:41shifting towards cheaper models. And
771. 29:43that is the China theory that's showing
772. 29:46up in this data. Now, to be fair though,
773. 29:48Michael Bur's been early before. And
774. 29:50when people say early in the market,
775. 29:52that's a polite way of saying he's been
776. 29:54wrong, right? He's made market crash
777. 29:56predictions over the years quite a lot.
778. 29:58That didn't really come true. And even
779. 30:01this index has dips that have recovered.
780. 30:04Basically, Bloomberg says that the
781. 30:05signal for all of this is ambiguous,
782. 30:07right? We can't really learn anything
783. 30:09from this data. It can mean anything.
784. 30:11So, basically, the real answer to how
785. 30:14long it will take for the AI bubble to
786. 30:15pop, if ever, is that no one knows. But
787. 30:19those are some of the early signs to
788. 30:21look for based on the data from the
789. 30:24past. Now, if you're interested in
790. 30:26seeing how I'm personally preparing and
791. 30:27more of my thoughts, those videos live
792. 30:29in the premium member section where
793. 30:30you'll also get access to my videos
794. 30:32earlier. And if that's valuable, the
795. 30:33link is down below. It allows me to make
796. 30:35more videos like this one and take on
797. 30:36fewer sponsors. Thank you for watching
798. 30:38and being a premium member. I hope you
799. 30:40have a wonderful rest of your day. Smash
800. 30:41the like button, subscribe if you
801. 30:43haven't already. Would love to see you
802. 30:44back here next time. See you soon.
803. 30:46Bye-bye.