YouTube2Text

A BOLHA DA COREIA DO SUL ACABOU DE ESTOURAR (explicado) — Transcript

by O Primo Rico · 6,542 words · 1,024 segments · language en · Watch on YouTube

Full transcript

  1. 0:00What is happening on the other side of
  2. 0:02the world, in South Korea, could say a
  3. 0:04lot about the future of artificial
  4. 0:05intelligence and the global economy.
  5. 0:08The South Korean stock market is
  6. 0:09experiencing one of the biggest crashes
  7. 0:10in its history. In June, it hit an
  8. 0:12all-time high, and since then, it has
  9. 0:14fallen over 40%. There have been so many
  10. 0:17days of panic that the Korean exchange
  11. 0:19has had to trigger circuit breakers more
  12. 0:22than 10 times this year alone.
  13. 0:24Can you imagine experiencing what we
  14. 0:26went through in October, but 10 times
  15. 0:27over?
  16. 0:28And the headline going around the world
  17. 0:30is that the Korean bubble has burst. But
  18. 0:32wait a second, what bubble?
  19. 0:33Did you know that the world's entire
  20. 0:35artificial intelligence infrastructure
  21. 0:37depends on a type of memory that comes
  22. 0:38primarily from South Korea, and that
  23. 0:40every time you ask GPT a question,
  24. 0:41there's a good chance Korean memory is
  25. 0:43making that answer happen?
  26. 0:45That out of every 10 dollars the world
  27. 0:47spends on AI chip memory, more than
  28. 0:49eight go to two Korean companies, and
  29. 0:52that one of them, Samsung, just had 80
  30. 0:54billion dollars in operating profit in a
  31. 0:56single quarter: more than Nvidia, double
  32. 0:59Google, and more than double Apple?
  33. 1:01If the companies that manufacture AI
  34. 1:03memory have never made so much money,
  35. 1:05why is their stock market crashing like
  36. 1:07this? Is this really a bubble? Does this
  37. 1:10reach American giants, and does it reach
  38. 1:12your money? That is exactly what I am
  39. 1:14going to explain to you in this video.
  40. 1:16What is really happening in Korea, why
  41. 1:18this memory has become the most
  42. 1:20sought-after piece of artificial
  43. 1:21intelligence, and what it all means not
  44. 1:23just for your money, but for the future
  45. 1:25of the world economy.
  46. 1:27And for this, I have divided this video
  47. 1:29into four pillars. First, I will explain
  48. 1:30the piece at the center of this story,
  49. 1:32which is the memory. Second, how it
  50. 1:34became the most sought-after piece of
  51. 1:36artificial intelligence and made these
  52. 1:38companies profit like never before.
  53. 1:39Third, why the Korean market plummeted
  54. 1:41even with that. And fourth, what this
  55. 1:43all means for your money. But before I
  56. 1:45continue, man, I'm going to ask you
  57. 1:47something, because this video took a lot
  58. 1:48of work. My team and I dedicated a lot
  59. 1:50of time to this video, studying company
  60. 1:52reports, translating Korean articles,
  61. 1:54and comparing numbers from different
  62. 1:56sources. So if you made it this far,
  63. 1:57leave a like on the video and subscribe
  64. 1:59to the channel. This really helps
  65. 2:01YouTube understand that this type of
  66. 2:02video is worth it.
  67. 2:04And it helps our team a lot to keep
  68. 2:06bringing this kind of investigation to
  69. 2:08you. So let's go. Starting with memory.
  70. 2:10To help you understand, let's build an
  71. 2:12analogy together here. Think of a
  72. 2:14professional kitchen. There's the pantry
  73. 2:16, where everything is stored. There's
  74. 2:17the countertop, where only the
  75. 2:19ingredients for the current recipe are
  76. 2:21kept, and there's the cook. Every device
  77. 2:23, from your phone to the largest data
  78. 2:25center in the world, works this way.
  79. 2:28The cook is the processor, the chip that
  80. 2:30does the math; the pantry is what we
  81. 2:32know as storage, where your photos,
  82. 2:34files, and apps stay even when the
  83. 2:36device is off.
  84. 2:38And in a computer, this is the HDD or
  85. 2:41SSD. And the countertop is the working
  86. 2:43memory, the one that appears as RAM on
  87. 2:45the device's spec sheet. It is much
  88. 2:47faster, but it clears out when you turn
  89. 2:49it off. Its technical name is DRAM, and
  90. 2:51it is the most important part of this
  91. 2:53video. A model like ChatGPT is made of
  92. 2:56billions of numbers.
  93. 2:58To write each word of the response you
  94. 3:00receive, the chip needs to read a huge
  95. 3:02portion of those numbers, word by word.
  96. 3:05In the kitchen analogy, it's as if the
  97. 3:07recipe had billions of ingredients and
  98. 3:08the cook needed almost all of them at
  99. 3:10every step.
  100. 3:12The larger the countertop, meaning the
  101. 3:14memory, and the faster the ingredient
  102. 3:15gets to it, the more the cook can
  103. 3:17produce.
  104. 3:18And the AI cook is so fast that often
  105. 3:20the one setting the kitchen's pace is
  106. 3:22the countertop.
  107. 3:24That is why AI doesn't just use any
  108. 3:26memory. It uses a special type called
  109. 3:28HBM.
  110. 3:29The acronym stands for High Bandwidth
  111. 3:32Memory.
  112. 3:33In practice, it is the memory that
  113. 3:35delivers the most data per second. HBM
  114. 3:37is made of the same DRAM as conventional
  115. 3:39memory, only it is assembled in a much
  116. 3:42more difficult way. You take several
  117. 3:44memory chips, eight, twelve, and then
  118. 3:45you stack them one on top of the other,
  119. 3:47like in a building.
  120. 3:49Then you connect the floors with
  121. 3:51thousands of tiny holes that go through
  122. 3:53the chip itself, and you glue this
  123. 3:55building right next to the AI chip.
  124. 3:58In the kitchen, that would be equivalent
  125. 3:59to gluing the countertop to the cook and
  126. 4:01having it deliver the ingredients
  127. 4:02through thousands of paths at the same
  128. 4:04time.
  129. 4:06This memory was created over 10 years
  130. 4:08ago for computer graphics cards and was
  131. 4:10a niche product for a long time, until
  132. 4:13November 30, 2022, when OpenAI launched
  133. 4:15ChatGPT, and in two months it reached
  134. 4:18100 million users.
  135. 4:21No application in history had grown so
  136. 4:23fast, and each person asking a question
  137. 4:25meant a chip in a data center reading
  138. 4:26billions of numbers in memory to be able
  139. 4:28to answer.
  140. 4:30That year, Nvidia's main AI chip, the
  141. 4:33H100, came with 80 GB of HBM.
  142. 4:36Two years later, the next version
  143. 4:39already had 141, then 192, and the chip
  144. 4:41Nvidia just started shipping now carries
  145. 4:44288 GB.
  146. 4:46In 4 years, the memory inside each chip
  147. 4:48has more than tripled, and it wasn't
  148. 4:49just the chip that grew.
  149. 4:51The way of using AI has also changed. In
  150. 4:53the beginning, you asked a question and
  151. 4:55received a short answer.
  152. 4:57Today, AI reads entire documents, keeps
  153. 4:58long conversations, and instead of just
  154. 5:00answering you, it often thinks and
  155. 5:01writes out an entire reasoning process
  156. 5:03you don't even see.
  157. 5:04And all of this needs to stay in memory
  158. 5:06while it works. With more people using
  159. 5:08larger models and each answer demanding
  160. 5:10more. AI has become the biggest consumer
  161. 5:12of memory the world has ever seen. In
  162. 5:142025, the world spent about 35 billion
  163. 5:17dollars just on HBM, and one HBM
  164. 5:19manufacturer projects that this figure
  165. 5:22will reach nearly 100 billion dollars
  166. 5:25per year by 2028. And this is the memory
  167. 5:27that I mentioned at the beginning of the
  168. 5:30video comes mainly from Korea. In the
  169. 5:32entire world, only three companies know
  170. 5:33how to manufacture HBM on a large scale
  171. 5:35. SK Hynix, Samsung, and the American
  172. 5:38company Micron. And combining the two
  173. 5:41Korean companies, SK Hynix and Samsung,
  174. 5:43we are talking about two firms
  175. 5:45responsible for 80%of the HBM memory in
  176. 5:47the market. This shows up in the numbers
  177. 5:49for the entire country. In September,
  178. 5:51South Korea exported 60 billion dollars
  179. 5:53in chips alone in a single month, and
  180. 5:55this is more than triple what it had
  181. 5:57exported in the same month last year and
  182. 5:59half of everything the country sold to
  183. 6:00the world. To give you an idea, that is
  184. 6:03more than double everything Brazil
  185. 6:05exports in a month, adding up soy, oil,
  186. 6:07iron ore, meat, everything. But this
  187. 6:10jump in export numbers didn't just come
  188. 6:12from selling more chips; it came mainly
  189. 6:14because the chips became much more
  190. 6:16expensive. And to understand why the
  191. 6:18price went up so much, you need to know
  192. 6:19one thing about this market. For decades
  193. 6:22, memory was treated as if it were a
  194. 6:24commodity, like soy or iron ore. The
  195. 6:26brand mattered very little. A memory
  196. 6:28chip is very similar to gold, and what
  197. 6:30defines the price is supply and demand.
  198. 6:32That is why the sector has always lived
  199. 6:34in cycles. When prices rise, everyone
  200. 6:36builds new factories, the market gets
  201. 6:37flooded with memory, prices crash, and
  202. 6:39companies start losing money. Then they
  203. 6:41cut production, and after a while,
  204. 6:43memory becomes scarce again.
  205. 6:45The last bottom was in 2023, and in just
  206. 6:48the first nine months of that year,
  207. 6:50Samsung and SK Hynix combined lost over
  208. 6:5221 trillion won on chips, about 16
  209. 6:54billion dollars in losses.
  210. 6:57So, the manufacturers cut production to
  211. 6:59try to keep prices up. Notice the timing
  212. 7:01, then. GPT had just been released. The
  213. 7:04biggest wave of demand for memory in
  214. 7:06history arrived just when the sector was
  215. 7:08losing money and producing less.
  216. 7:10And with memory, you don't fix that
  217. 7:12overnight. A new factory costs tens of
  218. 7:14billions of dollars and takes years to
  219. 7:16finish. To make matters worse, HBM comes
  220. 7:19from the same factories as standard
  221. 7:21memory and uses about three times the
  222. 7:23capacity to deliver the same amount.
  223. 7:26So, the more manufacturers shifted
  224. 7:28production to AI, the less there was for
  225. 7:30everything else: phones, computers, and
  226. 7:32regular servers.
  227. 7:33The result was the price of memory
  228. 7:35skyrocketing. According to this second
  229. 7:37site, in just the second quarter of 2026
  230. 7:40, the price of DRAM rose 44%in three
  231. 7:42months. And storage, the chip in your
  232. 7:45phone and SSD, went up 53%.
  233. 7:48In other words, the world is paying much
  234. 7:50more for memory because there isn't
  235. 7:51enough for everyone who wants to buy it
  236. 7:53.
  237. 7:54In September alone, the global
  238. 7:56semiconductor industry—including memory
  239. 7:58, processors, and sensors—sold over 170
  240. 8:00billion dollars, nearly two and a half
  241. 8:02times what it sold a year earlier.
  242. 8:04And by August, total sales for the year
  243. 8:06had already surpassed 1 trillion dollars
  244. 8:08. And who is on the other side of the
  245. 8:09counter? Well, the same Korean companies
  246. 8:10. Because Samsung and SK Hynix don't
  247. 8:13just make HBM. They are also the largest
  248. 8:15manufacturers of standard memory, the
  249. 8:17kind in your phone and your computer.
  250. 8:19Together, they make nearly two-thirds of
  251. 8:22all the DRAM on the planet. Meaning they
  252. 8:24win both ways: with HBM, where the
  253. 8:26competition is, and with standard memory
  254. 8:28, which got more expensive precisely
  255. 8:30because of it. And so, the numbers for
  256. 8:32these companies changed completely. Do
  257. 8:34you remember the losses from 2023?
  258. 8:36In the third quarter of 2026, SK Hynix
  259. 8:39reported 104.6 trillion won in revenue.
  260. 8:42The first time it surpassed 100 trillion
  261. 8:44in a single quarter. And this is more
  262. 8:45than four times what they had earned a
  263. 8:47year before. And for every 100 won that
  264. 8:50came in, about 80 turned into operating
  265. 8:52profit. The highest profit in the
  266. 8:53company's history.
  267. 8:55To give you a benchmark, at Apple, one
  268. 8:56of the world's most profitable companies
  269. 8:58, that number is around 35.
  270. 9:00At Nvidia, it's 65. And Samsung went
  271. 9:02down the same path.
  272. 9:03In the third quarter, it hit its record
  273. 9:06profit, that 80 billion dollars I
  274. 9:07mentioned at the start, more than any
  275. 9:09major American company last quarter.
  276. 9:12In other words, two companies that were
  277. 9:13losing money on chips three years ago
  278. 9:15have become profit machines. At Samsung
  279. 9:17, the chip division had a margin of
  280. 9:19about 72%, and at SK Hynix, around 80%.
  281. 9:22Both of them even higher than Nvidia.
  282. 9:24And market projections suggest that in
  283. 9:262026, Samsung will end the year as the
  284. 9:27most profitable company on the planet.
  285. 9:30In fact, in early July, at an internal
  286. 9:32meeting with Samsung's chip division
  287. 9:34employees, one of the division
  288. 9:36presidents said that this year's profit
  289. 9:38alone would exceed everything the
  290. 9:39division had earned in the previous 40
  291. 9:41years combined.
  292. 9:43And here is something very bizarre.
  293. 9:44Everyone knows Samsung for its phones,
  294. 9:46the great rival to Apple that it is. But
  295. 9:49in the third quarter, the mobile and
  296. 9:50electronics division reported a loss.
  297. 9:52Meanwhile, the chip division brought in
  298. 9:5575 out of every 100 won Samsung earned,
  299. 9:58and virtually all of the company's
  300. 10:01profit—106.5 out of 107.4 trillion won.
  301. 10:04In practice, Samsung has become a memory
  302. 10:06company that happens to also sell phones
  303. 10:08. And that prosperity has even reached
  304. 10:10those who work there. SK Hynix shares 10
  305. 10:12%of its profit with employees, and in
  306. 10:14February, that resulted in an average
  307. 10:16bonus of 140 million won per person,
  308. 10:18which is about 100,000 dollars. And the
  309. 10:20forecast for the next one, in early 2027
  310. 10:22, is about half a million dollars per
  311. 10:24employee.
  312. 10:26If the profit was reaching the employees
  313. 10:27like that, imagine what was happening to
  314. 10:29those who held the stock, right?
  315. 10:31In 2025, Samsung stock rose 127%, and in
  316. 10:34May 2026, for the first time in history
  317. 10:36, Samsung became worth more than 1
  318. 10:39trillion dollars.
  319. 10:40And a few weeks later, it became the
  320. 10:42first Korean company worth 2 quadrillion
  321. 10:45won.
  322. 10:46SK Hynix went even further. At the end
  323. 10:48of 2025, it was worth about 300 billion
  324. 10:50dollars. By June 2026, it surpassed 1.3
  325. 10:52trillion, more than quadrupling in less
  326. 10:55than six months. A company that almost
  327. 10:57went bankrupt in the early 2000s
  328. 10:59overtook Samsung to become the most
  329. 11:01valuable in Korea. A position that
  330. 11:03belonged to Samsung for over 25 years.
  331. 11:05Two Korean companies each worth over 1
  332. 11:08trillion dollars. And even after all
  333. 11:11that growth, they still looked cheap.
  334. 11:14There is a number every investor looks
  335. 11:15at to know if a stock is expensive or
  336. 11:16cheap, which is the P/E ratio, the stock
  337. 11:18price divided by the company's earnings
  338. 11:19.
  339. 11:20In plain English, how many years of
  340. 11:22earnings would you need to recover what
  341. 11:24you paid for the stock?
  342. 11:26Tech companies usually have a very high
  343. 11:28number, as the market pays a premium
  344. 11:30because they grow fast.
  345. 11:32Nvidia, for example, in July was trading
  346. 11:34at about 22 times the earnings expected
  347. 11:36for the next 12 months. Now look at
  348. 11:38memory chips.
  349. 11:40In May, SK Hynix was trading at around
  350. 11:425.6 times expected earnings.
  351. 11:45And as you will see in a moment, after
  352. 11:46all the decline, that number got even
  353. 11:47lower.
  354. 11:49To bring it to our reality, here in
  355. 11:50Brazil, Weg, one of the companies the
  356. 11:52market most loves to overpay for, trades
  357. 11:54at almost 30 times earnings.
  358. 11:56SK Hynix, with higher margins than
  359. 11:58Nvidia and profiting like few others in
  360. 12:00the world, cost less than six.
  361. 12:02For investors, buying shares of these
  362. 12:04companies was an obvious choice. But
  363. 12:06this logic completely changed their
  364. 12:07stock market. The Korean stock index,
  365. 12:10the Kospi, is calculated by the market
  366. 12:12value of the companies. The more a
  367. 12:13company is worth, the more weight it has
  368. 12:15in the index.
  369. 12:17So, the more Samsung and SK Hynix rose,
  370. 12:19the more they pulled the whole market up
  371. 12:21with them.
  372. 12:22At their peak, the two alone generated
  373. 12:24more than half of the index. This
  374. 12:26coincided with another movement in the
  375. 12:27Korean market. For decades, Korean
  376. 12:29stocks were cheaper than those of
  377. 12:31similar companies abroad.
  378. 12:33Investors were wary of the families that
  379. 12:34control the country's large
  380. 12:35conglomerates, right? And the market
  381. 12:37ends up calling this the Korean discount
  382. 12:39. And the government that took office in
  383. 12:412025 enacted reforms to give more voice
  384. 12:44to small shareholders, precisely to end
  385. 12:46this discount.
  386. 12:48Even while still a candidate, the
  387. 12:49president of Korea had promised to take
  388. 12:51the stock market to 5,000 points, when
  389. 12:53it was near 2,500 points.
  390. 12:55It was a promise to double it.
  391. 12:56Artificial intelligence on one side,
  392. 12:58reform on the other, Korea became a
  393. 13:00trendy market. And in 2025 the Korean
  394. 13:02stock market rose 75%, one of the
  395. 13:04biggest gains in the world. And in June
  396. 13:062026 it surpassed 9,100 points. In just
  397. 13:09over a year, it nearly quadrupled. And
  398. 13:11the market thought there was much more
  399. 13:13to come. In May, JP Morgan predicted the
  400. 13:16Korean stock market would hit 10,000
  401. 13:17points by the end of the year. Nomura
  402. 13:19predicted 11,000.
  403. 13:21But curiously, on the day the stock
  404. 13:22market closed above 9,000 points for the
  405. 13:24first time, out of every 10 traded
  406. 13:26stocks, more than eight fell.
  407. 13:28The index was rising, but what was
  408. 13:29really going up were Samsung and SK
  409. 13:31Hynix. It was a warning sign. The entire
  410. 13:33market depended on these two companies,
  411. 13:35but for those looking from the outside,
  412. 13:36all that appeared was the index hitting
  413. 13:38record after record. And when the market
  414. 13:40hits record after record, everyone wants
  415. 13:42in. And for many Koreans, it was
  416. 13:44actually the only way out. In Seoul, a
  417. 13:46house costs almost 14 times what a
  418. 13:48family earns in an entire year. The
  419. 13:50president himself even said that real
  420. 13:52estate used to be the only possible
  421. 13:54investment and now there was an
  422. 13:55alternative. Imagine the mindset of a
  423. 13:57Korean in their 20s: home ownership is
  424. 13:59out of reach, impossible. And right next
  425. 14:02door, two of their own country's
  426. 14:03companies hitting record after record.
  427. 14:05Man, the choice that made the most sense
  428. 14:06wasn't a house. And the Koreans went all
  429. 14:08in. Many people started borrowing money
  430. 14:10from brokerages to buy stock. It’s a
  431. 14:13huge thing in Korea.
  432. 14:15In June, Koreans already owed 38
  433. 14:17trillion won to brokerages, more than 25
  434. 14:19billion dollars, and over a quarter of
  435. 14:21that had come in just six months.
  436. 14:24And that is what we call leverage.
  437. 14:26In May, Korea released a new product:
  438. 14:28funds that double the daily movement of
  439. 14:31Samsung or SK Hynix.
  440. 14:33If the stock goes up 5%, the fund goes
  441. 14:35up 10; if it falls 5%, the fund falls 10
  442. 14:37. And on the first day, about 3 billion
  443. 14:39dollars flowed into these funds. In less
  444. 14:42than a month, it was already over 9
  445. 14:43billion. But these funds are
  446. 14:45double-edged swords.
  447. 14:47To deliver double the performance every
  448. 14:48day, this fund needs to buy more stock
  449. 14:49at the end of the day when it rises and
  450. 14:51sell when it falls.
  451. 14:52On the way up, this helps push the price
  452. 14:54higher; on the way down, it pushes it
  453. 14:56lower.
  454. 14:57It was a market sort of hanging on two
  455. 14:58companies, with the average investor
  456. 15:00increasingly leveraged and in debt,
  457. 15:01betting everything in the same direction
  458. 15:02.
  459. 15:03And it was exactly there, at the peak,
  460. 15:04that everything began to turn.
  461. 15:07While the small Korean investor was
  462. 15:08entering the market, large foreign funds
  463. 15:10were going the other way.
  464. 15:12In May alone, at the height of the rally
  465. 15:14, they sold 44 trillion won in Korean
  466. 15:16stocks.
  467. 15:17It is the largest sale in a single month
  468. 15:18in the country's history. And a large
  469. 15:20part of those buying these stocks were
  470. 15:22local investors, often with money
  471. 15:24borrowed from the brokerage. That
  472. 15:26borrowed money worked very well as long
  473. 15:28as the stock market was rising. The
  474. 15:30problem is when the tide began to turn.
  475. 15:32And turn it did. On June 22nd, the
  476. 15:34Korean stock market closed at a record
  477. 15:379,114 points. The very next day, it
  478. 15:39plummeted 9.99%. Samsung and SK Hynix
  479. 15:42both fell more than 12%each. Their worst
  480. 15:45day since the 2008 crisis. And as shares
  481. 15:47dropped like that, brokerages began
  482. 15:49forced liquidations of positions bought
  483. 15:52on margin.
  484. 15:53And every forced sale pushed the price
  485. 15:55down a little further, which forced
  486. 15:56brokers to sell even more people's
  487. 15:58positions.
  488. 15:59And that creates a domino effect. And
  489. 16:00that is how Korea arrived at the July
  490. 16:02crash.
  491. 16:03By the 30th, the market had accumulated
  492. 16:06a 34%drop for the month, more than in
  493. 16:08the worst months of the 1997 and 2008
  494. 16:10crises.
  495. 16:12In that month alone, brokerages forcibly
  496. 16:14sold over half a billion dollars in
  497. 16:15shares from regular investors.
  498. 16:17People who didn't want to sell but were
  499. 16:19forced to. And at the end of the month,
  500. 16:21both Korean exchanges had to trigger
  501. 16:22circuit breakers at the same time for
  502. 16:24two days in a row, something that had
  503. 16:25never happened there before. Now, you
  504. 16:27know what’s curious? At the beginning of
  505. 16:29July, while the market was entering one
  506. 16:30of the worst months in its history,
  507. 16:32Samsung announced its biggest profit
  508. 16:33ever. So why were the world’s major
  509. 16:35funds selling off exactly the companies
  510. 16:36making the most profit on the planet?
  511. 16:38And the answer begins with that story
  512. 16:39about the cycle. Memory has always
  513. 16:41worked like this. Memory is scarce,
  514. 16:43prices go up.
  515. 16:44Manufacturers make a lot of money, build
  516. 16:45new factories, and a few years later the
  517. 16:47factories are ready, there’s a surplus,
  518. 16:49prices collapse, and profits collapse
  519. 16:50with them.
  520. 16:51So, for those who have followed the
  521. 16:53sector for a long time, record memory
  522. 16:55profits usually appear right near the
  523. 16:56peak. In other words, when profits are
  524. 16:58high, it’s time to get out. And that was
  525. 17:00when the market split in two. On one
  526. 17:02side, those who believe that this time
  527. 17:03it’s different. This group has strong
  528. 17:05arguments. The first is that the memory
  529. 17:07shortage continues, right? In May, the
  530. 17:08CEO of Micron said that the company can
  531. 17:10only deliver between half and two-thirds
  532. 17:12of what its main clients ask for.
  533. 17:15Second, TrendForce, which is one of the
  534. 17:17leading consultancies in the sector,
  535. 17:18right?
  536. 17:20Where they said that the shortage should
  537. 17:22increase in 2027 and the new factories
  538. 17:23being built should only make a real
  539. 17:25difference in 2028, and the situation
  540. 17:27has reached a point where Nvidia itself
  541. 17:29is considering putting less memory in
  542. 17:31one of its next chips, right?
  543. 17:33Because there might not be enough memory
  544. 17:35to buy.
  545. 17:36In fact, the president of ADATA, one of
  546. 17:38the world's largest memory module
  547. 17:39manufacturers, even said that the entire
  548. 17:41production from the three major
  549. 17:43manufacturers is already sold out
  550. 17:44through 2027.
  551. 17:46And according to industry sources, some
  552. 17:48clients are only receiving 60%to 70%of
  553. 17:49what they ordered. And at the end of
  554. 17:51October, the Korean companies themselves
  555. 17:53reinforced this. Samsung said it has
  556. 17:54already closed its HBM contracts for
  557. 17:562027 and that production is sold out.
  558. 17:59And SK Hynix said the memory shortage is
  559. 18:01expected to last until 2030. The second
  560. 18:03argument is that AI memory has stopped
  561. 18:06functioning as a commodity.
  562. 18:08HBM is custom-made for each client, with
  563. 18:10a long-term contract.
  564. 18:12Micron, for example, already has
  565. 18:14agreements covering about 35%of its
  566. 18:16sales through 2030, and this is the side
  567. 18:18where most banks and brokerages stand,
  568. 18:20right? At the end of September, the
  569. 18:23brokerage Yuanta set a price target for
  570. 18:25Samsung at 630,000 won.
  571. 18:28Mirae Asset, one of the largest in Korea
  572. 18:30, set SK Hynix at 3.1 million, and in
  573. 18:32early November, Samsung was trading near
  574. 18:34200,000 won and SK Hynix near 1.2
  575. 18:36million.
  576. 18:38In other words, according to these
  577. 18:39analysts, these stocks are worth between
  578. 18:41two and three times today's price. On
  579. 18:42the other side, there are those who
  580. 18:44think it's the same old cycle, just
  581. 18:46bigger.
  582. 18:47This group also has a strong argument
  583. 18:48that Peter Lynch, one of the most famous
  584. 18:50fund managers in history, explained back
  585. 18:51in the late 80s.
  586. 18:53In cyclical companies, the P/E ratio is
  587. 18:55lowest precisely at the peak, when
  588. 18:56profits are at their maximum. And that
  589. 18:58five-times P/E ratio that looks like a
  590. 19:00bargain is only a bargain if the profit
  591. 19:01stays where it is. So do the math with
  592. 19:03me. If a stock costs five years of
  593. 19:05profit and the profit drops by half, it
  594. 19:06starts to cost 10. If the profit drops
  595. 19:08by 75%, it starts to cost 20. This is
  596. 19:11without the stock price changing by a
  597. 19:12single cent. And the memory industry has
  598. 19:13already been through this. In 2018, at
  599. 19:15the peak of the previous cycle, Samsung
  600. 19:17made record profits. The following year
  601. 19:19, memory prices plummeted and the
  602. 19:20company's profit fell by more than half
  603. 19:22. And this is the side where the big
  604. 19:24foreign funds are.
  605. 19:26In 2026, they have already sold more
  606. 19:28than 200 trillion won in Korean stocks,
  607. 19:30and almost 90%of that was just Samsung
  608. 19:32and SK Hynix.
  609. 19:33And it makes sense. When a big fund sees
  610. 19:35a stock quadruple in six months in a
  611. 19:37sector that has always been cyclical, it
  612. 19:39does what it always does. He pockets
  613. 19:41some of the profit and reduces the bet
  614. 19:43size.
  615. 19:44And it was this profit-taking that
  616. 19:45triggered the nearly 10%drop on June
  617. 19:4723rd.
  618. 19:48But in a market full of leveraged
  619. 19:50players, a normal sale becomes an
  620. 19:52avalanche, because since they were
  621. 19:54launched in May, the funds that double
  622. 19:56the movement of the two stocks have
  623. 19:58received 12.5 trillion won from small
  624. 20:00Korean investors. And between early June
  625. 20:03and early July, even before the July
  626. 20:05crash, they had already lost an average
  627. 20:07of 27%. And it wasn't just the small
  628. 20:09investor who was leveraged. The most
  629. 20:11famous case is Situational Awareness,
  630. 20:13the fund run by Leopold Aschenbrenner,
  631. 20:16an ex-OpenAI researcher in his twenties
  632. 20:18who became a market star betting on
  633. 20:21artificial intelligence. The fund
  634. 20:23managed over 20 billion dollars, and one
  635. 20:25of its bets was precisely SK Hynix.
  636. 20:28When the stocks plummeted in July, the
  637. 20:30banks that had lent money to the fund
  638. 20:31demanded more collateral.
  639. 20:33With no cash to cover it, on July 30th,
  640. 20:36the fund sold most of its shares to
  641. 20:38Citadel, another giant in the US market
  642. 20:40. And the very next day, with that
  643. 20:42seller out of the way, SK Hynix rose 24
  644. 20:44%and the Korean stock market 14%in a
  645. 20:46single session.
  646. 20:48You can't prove a single fund crashed
  647. 20:49the market, but you can see the weight
  648. 20:51that forced selling from leveraged
  649. 20:52positions can have.
  650. 20:54And notice the irony.
  651. 20:55The day after the sale, the stock he was
  652. 20:58forced to sell rose 24%.
  653. 21:00His thesis on artificial intelligence
  654. 21:02might still prove right down the road,
  655. 21:03but leveraged the way he was, he didn't
  656. 21:05have time to wait for the market to
  657. 21:07catch up. And well, after July, the
  658. 21:09market showed signs it would recover. In
  659. 21:11early October, the Korean stock market
  660. 21:13was around 7,000 points and companies
  661. 21:15kept breaking records. Korea set a
  662. 21:17record for chip exports in September,
  663. 21:20Samsung announced record profits again,
  664. 21:22and SK Hynix soon after. But two more
  665. 21:25waves of decline followed, one in
  666. 21:26October and another in early November.
  667. 21:28And this time, the arguments from the
  668. 21:30pessimists gained strength. In '26, the
  669. 21:32most profitable memory for these
  670. 21:35companies was standard memory, for
  671. 21:37phones and computers, not AI HBM.
  672. 21:40The chairman of the SK Group, which owns
  673. 21:42SK Hynix, actually said that HBM
  674. 21:44provides a margin of about 60%, while
  675. 21:46regular memory reached around 80%. And
  676. 21:48it makes sense. Remember that HBM comes
  677. 21:50from the same factories as regular
  678. 21:52memory and uses three times the capacity
  679. 21:54to deliver the same quantity? To meet AI
  680. 21:56demand, manufacturers have been
  681. 21:58dedicating an increasingly larger
  682. 22:00portion of their factories to HBM. And
  683. 22:02today, they already occupy about 20%of
  684. 22:05all DRAM production capacity in the
  685. 22:07world. Next year, it should reach close
  686. 22:09to 30%, but regular memory goes into
  687. 22:12many more places, right?
  688. 22:14Phones, computers, cars, and even the AI
  689. 22:16servers themselves, which use a lot of
  690. 22:18regular memory in addition to HBM.
  691. 22:21And with fewer factories left for it,
  692. 22:22the shortage became even greater exactly
  693. 22:24there. And there is one more difference
  694. 22:26. HBM is sold in long-term contracts
  695. 22:28with prices agreed upon with each client
  696. 22:30for the entire year. Regular memory has
  697. 22:32its price renegotiated every quarter. So
  698. 22:34, when the shortage tightened, its price
  699. 22:36went up immediately. For 2026, the price
  700. 22:38of HBM went up, and that increase
  701. 22:40applies to the whole year. However, the
  702. 22:42price of regular memory rose nearly 100
  703. 22:43%in just the first three months of the
  704. 22:45year.
  705. 22:46In other words, a huge part of the
  706. 22:48memory companies 'profits came precisely
  707. 22:49from the most commoditized part of the
  708. 22:51business, the part that has always lived
  709. 22:52in cycles.
  710. 22:53And that is exactly where China appeared
  711. 22:55.
  712. 22:56In October, at a conference in Shanghai
  713. 22:58, the president of CXMT, China's largest
  714. 23:00DRAM manufacturer, said they would
  715. 23:02launch a regular server memory by the
  716. 23:04end of the year with technology that
  717. 23:06shouldn't reach Samsung before 2028, and
  718. 23:08CXMT, which had 4%of the DRAM market a
  719. 23:10year ago, already has 10%.
  720. 23:13The company hasn't proven it can produce
  721. 23:15this at scale yet, but the message to
  722. 23:17the market was clear. There is a large
  723. 23:18Chinese competitor growing fast exactly
  724. 23:20in the memory that is currently the most
  725. 23:21profitable. Three days later, Samsung
  726. 23:24and SK Hynix fell more than 7%each. And
  727. 23:27on another front, HBM also took a hit.
  728. 23:29It depends on tech giants continuing to
  729. 23:32build AI data centers. And this
  730. 23:34investment has become gigantic.
  731. 23:36Microsoft, Google, Amazon, and Meta
  732. 23:38alone are expected to spend more than
  733. 23:40700 billion dollars in '26. And Google
  734. 23:42and Amazon have already warned that they
  735. 23:44will spend much more in 2027. But this
  736. 23:46money no longer fits within their cash
  737. 23:48reserves. In 2024, for every 100 dollars
  738. 23:50they invested, about nine came from
  739. 23:53loans. Today that number is 32, and in
  740. 23:55the last quarter, Google spent more cash
  741. 23:58than it generated.
  742. 23:59Meta invested more than its operations
  743. 24:01brought in, and Amazon borrowed 48
  744. 24:03billion dollars in three months.
  745. 24:06And you might think, well, borrowing to
  746. 24:07grow is normal. The problem is that this
  747. 24:09borrowing has become much more expensive
  748. 24:10.
  749. 24:11The rate the U.S. government pays to
  750. 24:13borrow money for 10 years, which serves
  751. 24:15as a benchmark for much of the world's
  752. 24:17debt, exceeded 5.3%in October, the
  753. 24:19highest level since 2002.
  754. 24:21And even after pulling back slightly, it
  755. 24:24remains near 5%. And behind this is
  756. 24:26inflation. With the war in Iran driving
  757. 24:28oil prices up, gasoline in the United
  758. 24:30States became 34%more expensive in a
  759. 24:32year. And U.S. inflation reached 3.7%.
  760. 24:35The Federal Reserve raised interest
  761. 24:37rates again in September, for the first
  762. 24:39time since 2023, and the market is
  763. 24:41already betting on another hike in
  764. 24:42December. And those who lend to tech
  765. 24:44giants were already charging more.
  766. 24:47According to a Reuters survey from July
  767. 24:48, of the debt securities these companies
  768. 24:50issued in 2026, nearly nine out of 10
  769. 24:52were already paying higher interest than
  770. 24:54when they were sold.
  771. 24:56In other words, building data centers
  772. 24:58depends increasingly on borrowed money,
  773. 25:00just when that money has become more
  774. 25:02expensive. And now, besides being more
  775. 25:04expensive, building data centers may
  776. 25:05become more difficult.
  777. 25:07In early November, during the U.S.
  778. 25:09midterm elections, the Democrats won
  779. 25:11control of Congress.
  780. 25:13And on the same day, several cities
  781. 25:15voted on whether they wanted data
  782. 25:16centers near their homes.
  783. 25:18There were 41 local votes, and in 40 of
  784. 25:20them, the side wanting to restrict
  785. 25:22construction won with 72%of the votes.
  786. 25:25That same week, the U.S. chip company
  787. 25:27index fell more than 12%in a single day
  788. 25:30. And the Korean stock market hit
  789. 25:31another circuit breaker.
  790. 25:33And the revenue that should pay this
  791. 25:35bill also wavered, because on the same
  792. 25:37day Samsung announced the record, OpenAI
  793. 25:39, the owner of GPT, warned investors
  794. 25:41that annual revenue was around 50
  795. 25:43billion dollars, not the nearly 80
  796. 25:45billion that had been disclosed the
  797. 25:47previous month.
  798. 25:48And in early November, it managed to
  799. 25:51raise 11 billion dollars from investors
  800. 25:53, much less than the 30 billion it was
  801. 25:55seeking. And it's not just OpenAI.
  802. 25:57Anthropic, owner of Claude, is going
  803. 25:59public on the U.S. stock exchange. And
  804. 26:01in the documents it filed with the SEC,
  805. 26:02it showed that it generated 17 billion
  806. 26:04dollars between July and September alone
  807. 26:06, more than 12 times what it generated a
  808. 26:08year earlier. But the company has
  809. 26:10already committed to spending around 900
  810. 26:12billion dollars, almost all of it on
  811. 26:13computing and infrastructure.
  812. 26:16And the market is no longer going to pay
  813. 26:17for this IPO the way the market was
  814. 26:19expecting back in August, right?
  815. 26:21There was talk that Anthropic's IPO
  816. 26:23wanted to value the company at 2
  817. 26:25trillion dollars, and now the IPO price
  818. 26:27range points to 1 trillion dollars. And
  819. 26:29if the bill for artificial intelligence
  820. 26:31gets more expensive, the revenue takes
  821. 26:32longer to arrive. Building data centers
  822. 26:35becomes more difficult, and that is a
  823. 26:37disaster for HBM memory. So, returning
  824. 26:39to the title question: has the
  825. 26:41artificial intelligence bubble burst in
  826. 26:43Korea? The honest answer is that it's
  827. 26:45difficult to pin down an answer. On the
  828. 26:47optimistic side, analysts are moving in
  829. 26:48the exact opposite direction of the
  830. 26:50market.
  831. 26:51In early October, D.A. Davidson, an
  832. 26:53American brokerage firm, raised Micron's
  833. 26:55price target from 2,100 to 3,000 dollars
  834. 26:58. The stock was around 1,000 dollars,
  835. 27:00and according to this brokerage, it
  836. 27:01could practically triple.
  837. 27:03And the Korean brokerages, as I showed
  838. 27:05you, still see Samsung and SK Hynix
  839. 27:06being worth two or three times today's
  840. 27:08price.
  841. 27:09And as I said, the argument is that this
  842. 27:11cycle is different. On the pessimistic
  843. 27:13side, the rise in the price of standard
  844. 27:15memory is losing steam.
  845. 27:16It rose almost 100%in the first quarter
  846. 27:18of the year, about 60%in the second, and
  847. 27:20about 15%in the third.
  848. 27:22For the fourth, TrendForce projects
  849. 27:24between 10 and 15%. It continues to rise
  850. 27:26, but by less and less. This weighs more
  851. 27:28than it seems. At Micron, for example,
  852. 27:31according to analyst estimates, HBM is
  853. 27:33less than 10%of revenue, and everything
  854. 27:35else depends on standard memory prices
  855. 27:37staying high. Now, there is one thing
  856. 27:39that is very clear. Regardless of who is
  857. 27:41right, the expectation is that the
  858. 27:43Korean market will continue to
  859. 27:45experience volatilities that feel like
  860. 27:47crises. And after all the tumble in July
  861. 27:49, Koreans started borrowing money again
  862. 27:51to buy stocks. In October, debt with
  863. 27:54brokerages returned to 34 trillion won,
  864. 27:56and with the new drops, it shrank to
  865. 27:58about 28 trillion by early November.
  866. 28:00And many people continue to see the drop
  867. 28:02as a chance to buy cheap. In early
  868. 28:04November, while foreigners spent seven
  869. 28:06straight trading sessions selling, the
  870. 28:08small Korean investor was buying. And
  871. 28:10forced liquidations have started to rise
  872. 28:11again. There was even an emergency
  873. 28:13government meeting, right? In early
  874. 28:15November, they relaxed the loan
  875. 28:16guarantee rule for stock purchases until
  876. 28:19the end of January, and brokerages
  877. 28:20started offering to postpone forced
  878. 28:22sales.
  879. 28:24And leveraged funds remain in the middle
  880. 28:25of this mess.
  881. 28:26Besides Leopold's fund, another AI fund
  882. 28:29, Vara, which managed about 20 billion
  883. 28:32dollars, lost 44%in July and another 33
  884. 28:34%in October.
  885. 28:36The fund said it remains stable, but
  886. 28:37these numbers show just how much this
  887. 28:39market is shaking. In other words, the
  888. 28:41Korean stock market remains hung up on
  889. 28:43two companies with many indebted
  890. 28:44investors and leveraged funds in the
  891. 28:46middle of it. And with these pieces on
  892. 28:48the board, any good or bad news turns
  893. 28:50into a giant move. And the craziest part
  894. 28:52is knowing that this moment of
  895. 28:53volatility and uncertainty isn't just in
  896. 28:55Korea. Abroad, US interest rates started
  897. 28:57near the highest level in over 20 years
  898. 28:59, and that weighs on the entire world's
  899. 29:01economy, including Brazil's. And here in
  900. 29:03Brazil, we just saw this up close. One
  901. 29:05of the weeks in October was, well, the
  902. 29:06worst week for the Brazilian stock
  903. 29:08market since the start of the pandemic.
  904. 29:10The Central Bank had been cutting rates
  905. 29:12, stopped at a Selic of 13.75%, and even
  906. 29:14increased the inflation projection. And
  907. 29:16it's in moments like these of great
  908. 29:17uncertainty that many people end up
  909. 29:19making the worst decision of the year
  910. 29:20with their own money. And a good part of
  911. 29:22these bad decisions stems from the
  912. 29:24attempt to guess what will happen, where
  913. 29:25the stock market is headed, how far US
  914. 29:27rates will go, and how far the dollar
  915. 29:29will go. But to get through a moment
  916. 29:30like this, what helps is knowing what to
  917. 29:32do with your money in any of these
  918. 29:33scenarios. And when everything becomes
  919. 29:35uncertain, having a team by your side
  920. 29:37studying each company and telling you
  921. 29:39what price is worth paying makes a big
  922. 29:41difference. And that is why I cannot
  923. 29:42fail to talk to you about Finclass.
  924. 29:44Inside, you have access to recommended
  925. 29:46portfolios—there are two, appreciation
  926. 29:48and income—put together and monitored by
  927. 29:50our team of analysts.
  928. 29:52You enter how much you have to invest,
  929. 29:54you see which assets to put it in, how
  930. 29:55much to put in each, and what the
  931. 29:57maximum purchase price is.
  932. 29:58If the asset has passed that price, you
  933. 30:00know it's not time to buy. And this
  934. 30:02makes a big difference on days like the
  935. 30:03ones the Brazilian stock market just
  936. 30:05lived through.
  937. 30:06The day after the second round, when the
  938. 30:08market fell nearly 14%, the Finclass
  939. 30:09appreciation portfolio fell about 3%.
  940. 30:12Just as important as knowing where to
  941. 30:13put your money is knowing where not to
  942. 30:15put it. You saw in this video the damage
  943. 30:17that leverage caused in Korea. And
  944. 30:19inside Finclass, we steer our
  945. 30:20subscribers away from that path before
  946. 30:21they even start. In May, Alexandro asked
  947. 30:23in the community about leveraged ETFs.
  948. 30:25And the response from Vinicius, one of
  949. 30:27our moderators, was short: it's best you
  950. 30:29stay away. And when the market drops,
  951. 30:30guidance also makes a difference. In
  952. 30:32July, the same month the Korean market
  953. 30:34crashed, Diego, a Finclass subscriber,
  954. 30:36wrote in the community that he was
  955. 30:38watching his stocks fall sharply and had
  956. 30:40already thought about selling, but he
  957. 30:41held on. In his words, it's impossible
  958. 30:43for so many studious people to see a
  959. 30:45future in this and be wrong. And on the
  960. 30:47day he wrote that post, the stocks had
  961. 30:49already recovered all their losses. And
  962. 30:50that is the kind of guidance you get
  963. 30:52when you join Finclass. And man,
  964. 30:53Finclass goes way beyond just portfolios
  965. 30:56. You know all that stuff that appeared
  966. 30:57in this video? US interest rates,
  967. 30:59inflation, the dollar, elections
  968. 31:00affecting the stock market. Inside, you
  969. 31:02have economic courses to help you
  970. 31:03understand all of this at your own pace
  971. 31:04.
  972. 31:06There are over 100 pieces of content,
  973. 31:07including classes, series, documentaries
  974. 31:09, and "finbooks," covering everything
  975. 31:10from your first investment to stocks,
  976. 31:12real estate funds, fixed income,
  977. 31:13international investing, crypto, and
  978. 31:15retirement.
  979. 31:17And right upon your first access, you
  980. 31:18receive a study plan that tells you
  981. 31:20where to start.
  982. 31:21Even better, when you have a doubt, you
  983. 31:23have live sessions with our analyst team
  984. 31:24every business day, plus the community.
  985. 31:26It's where the moderators answer you,
  986. 31:28just like you saw in the cases I showed
  987. 31:30you. Today, there are over 130,000
  988. 31:31subscribers doing this right along with
  989. 31:33you. And I have some news for you. On
  990. 31:35November 27th, Finclass is holding its
  991. 31:37Black Friday. One day only, with the
  992. 31:39checkout open for 24 hours. And man, we
  993. 31:41are preparing a lot for Finclass next
  994. 31:44year. It’s going to change significantly
  995. 31:46; it will be much more complete and the
  996. 31:47price will be adjusted. That’s why this
  997. 31:49Black Friday will be the last time
  998. 31:51you’ll be able to subscribe to Finclass
  999. 31:53for 12 installments of 39.95. Even if we
  1000. 31:56offer a 50%discount in the future, like
  1001. 31:58we usually do on Black Friday, it will
  1002. 31:59be more expensive. Black Friday is on
  1003. 32:02November 27th, one day only, 50%off, 12
  1004. 32:04installments of 39.90, plus a new course
  1005. 32:06on AI for investors and bonuses. The
  1006. 32:08link to join the list is here in the
  1007. 32:10description, or just point your phone
  1008. 32:12camera at the QR code on the screen.
  1009. 32:13Whoever joins the list finds out as soon
  1010. 32:15as it opens, okay? And that's it,
  1011. 32:16cousins. I hope it's become clear what
  1012. 32:18is happening in Korea, why this memory
  1013. 32:20has become the most sought-after piece
  1014. 32:21in artificial intelligence, and why this
  1015. 32:23market will still give us much to talk
  1016. 32:25about. If this video helped you
  1017. 32:26understand this landscape a little
  1018. 32:28better, leave a like and send it to that
  1019. 32:29friend who is always talking about
  1020. 32:31artificial intelligence and the stock
  1021. 32:32market. And tell me here in the comments
  1022. 32:33what you thought, I'll be reading them.
  1023. 32:35Best regards, see you next time, and bye
  1024. 32:36.

About this transcript

This page contains the full transcript of A BOLHA DA COREIA DO SUL ACABOU DE ESTOURAR (explicado) by O Primo Rico, generated from the public captions YouTube serves with the video. The transcript has 6,542 words across 1,024 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.