YouTube2Text

广达鸿海斗了二十年,为什么就只服黄仁勋?黄老板被扫地出门,AI 产业链还稳吗?兆元宴合影藏着整个台湾代工江湖|艾森 Essen — Transcript

by 艾森 Essen · 5,183 words · 900 segments · language en · Watch on YouTube

Full transcript

  1. 0:00On the evening of January 31st this year
  2. 0:02, in front of a Taiwanese restaurant
  3. 0:05called Zhuan Yao Gu Zao Wei in Taipei,
  4. 0:07about 40 people crowded together for a
  5. 0:09group photo. The combined net worth of
  6. 0:12the chairmen and CEOs in this photo
  7. 0:14exceeds one trillion New Taiwan Dollars
  8. 0:16.
  9. 0:17The person who hosted this meal, dubbed
  10. 0:19by the outside world as the "
  11. 0:21Trillion-Dollar Banquet," is Jensen
  12. 0:23Huang.
  13. 0:24He sat in the center of the front row,
  14. 0:27with TSMC Chairman C.C. Wei on one side
  15. 0:29and Quanta Chairman Barry Lam on the
  16. 0:31other.
  17. 0:33However, if you look closely at the
  18. 0:35second row, you will notice a very
  19. 0:37strange positioning.
  20. 0:39Foxconn Chairman Young Liu stood
  21. 0:41directly behind Jensen Huang.
  22. 0:44Foxconn's revenue last year was 8.1
  23. 0:47trillion NTD, nearly four times that of
  24. 0:49Quanta; it is the largest company in
  25. 0:52Taiwan and Nvidia's top AI server
  26. 0:54contract manufacturer.
  27. 0:57The one with the largest market share is
  28. 0:59standing, while the one with the
  29. 1:00second-largest is sitting. Seniority is
  30. 1:03certainly one reason, as Barry Lam is a
  31. 1:05veteran in the industry. But everyone in
  32. 1:07Taiwan's tech circles knows that these
  33. 1:09two companies have been bitter rivals
  34. 1:11for 20 years.
  35. 1:13These Taiwanese bosses, who fought tooth
  36. 1:15and nail under Apple years ago, stood
  37. 1:18obediently in the same photo today.
  38. 1:21Among those present was even the only
  39. 1:23mainland Chinese boss in the room.
  40. 1:25Four months later, at the same
  41. 1:27restaurant, against the same wall, the
  42. 1:29same group of people took another photo
  43. 1:31, with the seating in the first two rows
  44. 1:33almost unchanged. Half a year later, on
  45. 1:36November 12, Nvidia announced that the
  46. 1:38person who had been sitting in the
  47. 1:39middle had resigned.
  48. 1:43This video tells a story about how
  49. 1:45commercial interests triumph over
  50. 1:46personal friendships and how a new order
  51. 1:49is established.
  52. 1:52We want to answer what it was that
  53. 1:53filled this table in the first place.
  54. 1:56Now that the person sitting in the
  55. 1:58middle has left, can they still sit at
  56. 2:00that table?
  57. 2:01To answer this, we must go back to a
  58. 2:03plate of dumplings. Years ago, Foxconn
  59. 2:06started by making connectors.
  60. 2:08Terry Gou's wife, Lin Shu-ju, wanting to
  61. 2:11help win business from Quanta, found out
  62. 2:13Barry Lam loved dumplings, so she
  63. 2:15personally cooked a plate for him. This
  64. 2:18plate of dumplings earned the friendship
  65. 2:20of the two families, and Quanta
  66. 2:22subsequently became Foxconn's largest
  67. 2:24client for laptop connectors.
  68. 2:27At the end of 2004, at the retirement
  69. 2:29dinner for the founder of a red-flag
  70. 2:32company, according to Taiwanese
  71. 2:34magazines, Terry Gou hugged Barry Lam
  72. 2:36from behind, called him "big brother" in
  73. 2:38front of the media, and promised that as
  74. 2:41long as Quanta was around, Foxconn would
  75. 2:43not touch laptop manufacturing.
  76. 2:46This promise only lasted three years. In
  77. 2:492007, the global laptop market exploded
  78. 2:51, and the subsequent financial crisis
  79. 2:54pushed Foxconn's revenue toward a
  80. 2:56recession.
  81. 2:57Terry Gou needed growth, and the biggest
  82. 3:00prize right in front of him was laptops
  83. 3:02.
  84. 3:04In March 2009, a new face appeared at
  85. 3:06Foxconn’s Neihu office in Taipei: a
  86. 3:08senior vice president from Quanta whom
  87. 3:11colleagues called Max.
  88. 3:15He managed the entire Dell product line
  89. 3:17, and Dell had even declared that if Max
  90. 3:19wasn’t the one handling their account,
  91. 3:21they wouldn't give the business to
  92. 3:23Quanta.
  93. 3:24Terry Gou didn’t just poach a vice
  94. 3:26president; he poached a client
  95. 3:28relationship, along with a whole team of
  96. 3:30people who would follow him.
  97. 3:33Nearly 60 engineers followed him to
  98. 3:35Foxconn.
  99. 3:38Barry Lam’s reaction was brutal; the
  100. 3:41vice president had submitted his
  101. 3:42resignation through proper channels.
  102. 3:47After learning he had been poached by
  103. 3:49Foxconn, Lam changed the status to "
  104. 3:52dismissal" and ordered that Quanta never
  105. 3:54hire him again.
  106. 3:56"Resignation" means leaving on your own
  107. 3:59; "dismissal" means being kicked out.
  108. 4:02This was a message for all the remaining
  109. 4:04Quanta engineers.
  110. 4:07Three months later, Quanta forcefully
  111. 4:09snatched back an Apple MacBook order
  112. 4:11worth billions of NT dollars that
  113. 4:13Foxconn had already secured.
  114. 4:16At the investor conference in August,
  115. 4:18someone asked about the group that had
  116. 4:20been poached. Barry Lam simply said, "
  117. 4:23They weren't our best employees anyway,
  118. 4:25and they haven't achieved anything since
  119. 4:27moving over."
  120. 4:30He didn’t say a single word against
  121. 4:32Terry Gou; he attacked the 60 engineers
  122. 4:34, essentially saying: "You only poached
  123. 4:36the people I didn't want."
  124. 4:40Many people think this feud was resolved
  125. 4:42by Jensen Huang, but in reality, he had
  126. 4:45nothing to do with it.
  127. 4:47On the evening of November 25, 2009, the
  128. 4:50two attended the same banquet and went
  129. 4:52to the restroom together.
  130. 4:55When they came out, they were holding
  131. 4:57hands, fingers interlaced.
  132. 5:00Terry Gou held their joined hands up for
  133. 5:02the reporters to see, saying, "Look, our
  134. 5:05fingers are locked.
  135. 5:07He is my big brother, my big brother, my
  136. 5:10big brother forever."
  137. 5:12That year, NVIDIA was still just a
  138. 5:14company making graphics cards for gaming
  139. 5:16.
  140. 5:18Yet, even after the handshake, the
  141. 5:20battle continued.
  142. 5:22In 2015, Apple’s iPad Pro was originally
  143. 5:24slated for Quanta, but Terry Gou
  144. 5:27snatched it back at the last minute with
  145. 5:29a better price.
  146. 5:32By 2017, Foxconn had replaced Quanta as
  147. 5:35the largest manufacturer of MacBooks.
  148. 5:39People often compare these two scenes.
  149. 5:41In 2004, Terry Gou called Barry Lam "
  150. 5:43big brother" in front of hundreds of
  151. 5:46people, but three years later, for the
  152. 5:48sake of company growth, he tore up that
  153. 5:50promise. In 2009, the two interlaced
  154. 5:53their fingers before reporters, yet six
  155. 5:55years later, they were still fighting
  156. 5:58tooth and nail over Apple orders.
  157. 6:00It’s not that these two men didn't value
  158. 6:03their friendship; it’s just that they
  159. 6:05knew all too well that in the Apple
  160. 6:07supply chain, friendship cannot secure
  161. 6:09orders, and promises cannot stop
  162. 6:10business.
  163. 6:12Why is that?
  164. 6:14The reason they were forced to stab each
  165. 6:16other in the back was their shared major
  166. 6:18client, Apple.
  167. 6:21Apple's method for managing suppliers is
  168. 6:23simple: for many key products, they
  169. 6:26insist on cultivating more than one
  170. 6:28supplier.
  171. 6:30iPhone assembly was originally dominated
  172. 6:32by Foxconn, so Apple propped up Pegatron
  173. 6:35as a second supplier. In 2014, for the
  174. 6:38iPhone 6, Foxconn took 70%and Pegatron
  175. 6:41took 30%.
  176. 6:43This approach means you can be replaced
  177. 6:45at any time, so no company dares refuse
  178. 6:47to lower their prices.
  179. 6:50This is clearly visible on the balance
  180. 6:52sheet.
  181. 6:53In 2017, the gross margins of all four
  182. 6:56companies—Quanta, Foxconn, Compal, and
  183. 7:00Pegatron—dropped together.
  184. 7:03The worst hit was Pegatron, the second
  185. 7:06supplier specifically propped up to
  186. 7:09compete with Foxconn; their gross margin
  187. 7:12was slashed in half from 6.2%in 2015 to
  188. 7:153%in 2018.
  189. 7:16In the Taiwanese industry, they call
  190. 7:18this "gross margin of three to four"—
  191. 7:20working hard for a year, earning only
  192. 7:22three or four dollars in profit for
  193. 7:24every hundred dollars of revenue. Since
  194. 7:26all four dropped together, it shows the
  195. 7:28problem wasn't the management of any
  196. 7:30single company.
  197. 7:32Rather, the same client tightened its
  198. 7:34grip on the entire supply chain during
  199. 7:36those two years.
  200. 7:39That’s how Quanta’s Apple Watch business
  201. 7:41went. When the first generation launched
  202. 7:44in 2015, Quanta was the exclusive
  203. 7:46assembler; they didn't make a profit for
  204. 7:49two years, only seeing returns by 2017.
  205. 7:51In 2019, Quanta decided not to take on
  206. 7:54new models. Barry Lam reasoned that they
  207. 7:57would only do what was profitable, as
  208. 7:59profit is paramount. The new orders
  209. 8:01eventually went to Foxconn and Compal.
  210. 8:05In the Apple supply chain, what everyone
  211. 8:07fights over is often a money-losing
  212. 8:09contract.
  213. 8:12That is the rule of the game: you can be
  214. 8:14replaced at any time, so no one’s
  215. 8:16promises are worth anything. Back then,
  216. 8:18no one could have imagined that ten
  217. 8:20years later, someone would appear who
  218. 8:22could make that same group of incredibly
  219. 8:24shrewd Taiwanese businessmen willing to
  220. 8:26bet billions of New Taiwan dollars on a
  221. 8:28long-term promise.
  222. 8:30Investing tens of billions of NT dollars
  223. 8:32.
  224. 8:34The other half of this story goes back
  225. 8:36to Quanta’s annual year-end party in
  226. 8:382010. That year, Barry Lam, who had just
  227. 8:41lost 60 people to Foxconn, took the
  228. 8:43stage dressed as the Monkey King in
  229. 8:45front of thousands of employees. He said
  230. 8:47, "I won't go to the 'Blue Ocean,'
  231. 8:49because one day it will turn into a '
  232. 8:51Red Ocean'—I’m going to the cloud."
  233. 8:53Everyone in the audience understood. "
  234. 8:55Red Ocean" referred to Foxconn. Barry
  235. 8:57Lam meant: "If you want to fight for
  236. 9:00this laptop business, go ahead; I’m done
  237. 9:02playing with you. I’m going to make
  238. 9:05servers." At that time, Quanta’s primary
  239. 9:07business was still laptops. He didn't
  240. 9:10even know if the cloud could support the
  241. 9:12company, or how many years it would take
  242. 9:14.
  243. 9:17Across the Pacific, someone else was
  244. 9:19doing something even harder to fathom:
  245. 9:21he was pushing a company that made money
  246. 9:23on gaming graphics cards into a
  247. 9:25computing market that almost no one was
  248. 9:27willing to pay for at the time.
  249. 9:32On October 28, 2009, in Beijing, NVIDIA
  250. 9:34presented a plaque for a "CUDA Center of
  251. 9:37Excellence" to Tsinghua University.
  252. 9:41The person who personally handed the
  253. 9:43plaque to the director of Tsinghua’s
  254. 9:45Computer Department was Jensen Huang
  255. 9:47himself.
  256. 9:48What exactly was it? To put it simply,
  257. 9:50it was a set of tools that allowed
  258. 9:52graphics cards to be used not just for
  259. 9:54gaming, but also for scientific
  260. 9:55computing. Back then, people doing
  261. 9:58scientific computing only used CPUs. If
  262. 10:00you applied to buy a graphics card in
  263. 10:02the lab, your advisor would suspect you
  264. 10:04just wanted to play games on the sly. A
  265. 10:06year later, the National University of
  266. 10:08Defense Technology’s Tianhe-1A was
  267. 10:11equipped with 7,168
  268. 10:13NVIDIA graphics cards, reaching the
  269. 10:15number one spot on the TOP500 list of
  270. 10:18the world's supercomputers.
  271. 10:23Putting CUDA on every GPU, that was a
  272. 10:25strategic decision that was very, very
  273. 10:27hard to do because the company consumed
  274. 10:29enormous amounts of our profits and we
  275. 10:32couldn't afford it at the time.
  276. 10:35For 10 years, that made NVIDIA what it
  277. 10:38is today.
  278. 10:40>> However, to Wall Street, these were all
  279. 10:43losing propositions.
  280. 10:47Huang later recalled that putting CUDA
  281. 10:49into every gaming graphics card was an
  282. 10:51incredibly difficult decision that
  283. 10:53drained huge profits from the company—
  284. 10:55something NVIDIA couldn't afford at the
  285. 10:57time.
  286. 11:01He said at Stanford that NVIDIA spent
  287. 11:03about ten years investing in a future
  288. 11:05that didn't even exist yet.
  289. 11:09For 10 to 15 years, the markets that
  290. 11:12sustain NVIDIA today didn't exist at all
  291. 11:14.
  292. 11:17When asked how he reacted when the
  293. 11:19market cap dropped by 80%, he said his
  294. 11:22pulse was exactly the same as usual.
  295. 11:26The stock price changed, but did the
  296. 11:28laws of physics change?
  297. 11:30Did gravity change? Nothing changed, so
  298. 11:33we just kept moving forward.
  299. 11:35One was waiting for the cloud, and the
  300. 11:37other was building the computing tools
  301. 11:39needed for the cloud.
  302. 11:42At the time, neither could guarantee
  303. 11:44they would win; their paths converged on
  304. 11:47the AI server.
  305. 11:51According to Quanta Vice Chairman C.C.
  306. 11:53Leung, the two companies’ AI
  307. 11:55collaboration likely began around 2015
  308. 11:58or 2016.
  309. 12:00It was still a small business back then
  310. 12:02.
  311. 12:03Seven or eight years later, AI really
  312. 12:05took off.
  313. 12:07"Nice to see you, ladies and gentlemen.
  314. 12:09Barry, my hero."
  315. 12:15>> On the evening of June 5, 2024, it was
  316. 12:18raining heavily in Neihu, Taipei. Barry
  317. 12:21Lam hosted a dinner for Jensen Huang at
  318. 12:23a restaurant called Shih Hsiang Fong. He
  319. 12:26stood in the rain waiting for Jensen
  320. 12:28Huang to get out of the car, first
  321. 12:30apologizing in English, then adding in
  322. 12:33Taiwanese, "phái-sè, phái-sè," and
  323. 12:35following up in Mandarin, "sorry." Barry
  324. 12:38Lam replied in Taiwanese, "bô-iàu-kín"(
  325. 12:40it doesn't matter). Then Jensen Huang
  326. 12:42turned to the camera and said, "This is
  327. 12:45my hero.
  328. 12:46Thank you for waiting for me in the rain
  329. 12:48."
  330. 12:49Barry Lam gave a thumbs up and replied
  331. 12:51in English, "Now you are our King."
  332. 12:55Fourteen years ago, the man who was like
  333. 12:57the Monkey King dodging the Red Sea to
  334. 12:59reach the cloud is now standing in the
  335. 13:02rain ten years later, calling another
  336. 13:04gambler "King."
  337. 13:06However, Barry Lam doesn't follow Jensen
  338. 13:08Huang blindly on everything.
  339. 13:10That very week, 12 days after Jensen
  340. 13:13Huang declared at the NTU Sports Center
  341. 13:15that the era of humanoid robots was
  342. 13:17coming, Barry Lam said at the Quanta
  343. 13:19shareholder meeting that humanoid robots
  344. 13:21are expensive and have no customers, and
  345. 13:24their only advantage might be that they
  346. 13:26don't complain or go on strike.
  347. 13:30The person sitting at Jensen Huang's
  348. 13:32side is not his follower, but his
  349. 13:34partner.
  350. 13:36From May 2024 to May of this year,
  351. 13:39Jensen Huang hosted five dinners at that
  352. 13:41same Taiwanese restaurant; as Pegatron's
  353. 13:44Jason Chang entered, he remarked that he
  354. 13:46was here to attend the year-end party
  355. 13:48for "On" and everyone else.
  356. 13:52In January 2025, to squeeze all 35 CEOs
  357. 13:55into the frame, there was quite a
  358. 13:57scramble over who would occupy the front
  359. 14:00row.
  360. 14:01In the end, Jensen Huang himself
  361. 14:03crouched down in the front row.
  362. 14:06He joked that those crouching in the
  363. 14:08front were the ones who exercised, while
  364. 14:11those in the back did not, but don't
  365. 14:12forget, these people are still fighting
  366. 14:15over orders today, just as fiercely as
  367. 14:17they fought for Apple's business back
  368. 14:19then.
  369. 14:21Although they are still competitors, in
  370. 14:23the Apple era, bosses treated each other
  371. 14:25like enemies, but in the NVIDIA era,
  372. 14:27they can sit at the same table and toast
  373. 14:30one another.
  374. 14:31Jensen Huang being more charismatic than
  375. 14:33Tim Cook is only a small part of the
  376. 14:35reason; the bigger answer is written in
  377. 14:37the financial statements of these
  378. 14:39companies.
  379. 14:42Who could have imagined that Barry Lam,
  380. 14:44who had 60 employees poached and hid
  381. 14:46away in the cloud like Sun Wukong, would
  382. 14:49trade places with Terry Gou on Taiwan's
  383. 14:51rich list over a decade later?
  384. 14:54On the 2018 Forbes Taiwan Rich List,
  385. 14:57Terry Gou ranked first, with a net worth
  386. 15:00twice that of Barry Lam.
  387. 15:03By 2025, Barry Lam had overtaken Terry
  388. 15:06Gou, but the more interesting part of
  389. 15:08this story is yet to come.
  390. 15:12The path Barry Lam chose back then to
  391. 15:14avoid Terry Gou eventually became the
  392. 15:16new battlefield where both of them are
  393. 15:19now fighting tooth and nail.
  394. 15:21Foxconn got its start by assembling
  395. 15:23iPhones for Apple. By the second quarter
  396. 15:26of 2025, Foxconn's cloud business,
  397. 15:28driven primarily by AI servers,
  398. 15:30surpassed its consumer electronics
  399. 15:32business—including the iPhone—in revenue
  400. 15:35for the first time.
  401. 15:39The world's largest Apple contract
  402. 15:41manufacturer has seen its biggest
  403. 15:43business change hands; Foxconn remains a
  404. 15:45giant in this new battlefield, still
  405. 15:47holding the number one market share.
  406. 15:54On this year's list, Terry Gou is ranked
  407. 15:56ahead of Barry Lam again, with only a
  408. 15:58slight difference between them; Barry
  409. 16:00Lam didn't defeat Terry Gou, but the
  410. 16:02battlefield and the price the capital
  411. 16:05market puts on this war have changed.
  412. 16:11Many might assume the high valuations
  413. 16:13are because Jensen Huang pays
  414. 16:15manufacturers better than Apple,
  415. 16:16allowing them to earn higher gross
  416. 16:18margins.
  417. 16:20But the numbers don't say that; Quanta's
  418. 16:23gross margin in 2018 was just over 4%,
  419. 16:26which is a standard report card for
  420. 16:28laptop contract manufacturing.
  421. 16:31After AI servers took off, it once
  422. 16:33climbed to nearly 8%in 2024.
  423. 16:36However, Quanta's revenue in the first
  424. 16:38few months of this year is already more
  425. 16:41than double that of the entire year of
  426. 16:432024.
  427. 16:44Yet the margin has fallen back to under
  428. 16:475%.
  429. 16:48Similar to when they made laptops for
  430. 16:51Apple back then, some might say that if
  431. 16:53the gross margin has dropped below 5%,
  432. 16:55isn't that just being exploited by
  433. 16:57NVIDIA?
  434. 17:01Quanta's CFO explained at the May
  435. 17:03earnings call this year that the decline
  436. 17:05in gross margin is mainly because the
  437. 17:07product has become a $ 3 million AI
  438. 17:09server rack, where chips and memory
  439. 17:12account for over 80%of the cost.
  440. 17:17The contract manufacturer only earns
  441. 17:19money from assembly and testing; revenue
  442. 17:21is inflated by expensive components, so
  443. 17:23the gross margin percentage is naturally
  444. 17:26diluted.
  445. 17:28Profit margins are falling, but profits
  446. 17:30are rising.
  447. 17:32Quanta's operating profit has grown
  448. 17:35sixfold from 2018 to 2025, but what has
  449. 17:38truly turned things upside down is the
  450. 17:40market cap.
  451. 17:43From the end of 2018 to November 12 of
  452. 17:46this year, Quanta's market cap has grown
  453. 17:495.6 times.
  454. 17:50Wistron and Foxconn have also multiplied
  455. 17:53several times over.
  456. 17:54Of course, this money didn't all come
  457. 17:56from Jensen Huang alone.
  458. 17:59No one can stop this wave of AI.
  459. 18:03But Jensen Huang knows exactly what he
  460. 18:05has given these people.
  461. 18:09After his fifth dinner party in May, he
  462. 18:11told reporters that the stock prices of
  463. 18:13all companies partnering with NVIDIA
  464. 18:15have tripled within a year.
  465. 18:20I am very happy for them; they deserve
  466. 18:22it.
  467. 18:25He didn't say how much profit he shared
  468. 18:27with them; he said your stock prices
  469. 18:29have gone up.
  470. 18:31In the Apple era, what contract
  471. 18:33manufacturers cared about most was how
  472. 18:35much money an order could make.
  473. 18:38In the NVIDIA era, they have to
  474. 18:39calculate something else: will there
  475. 18:41still be a place for them in the next
  476. 18:43generation?
  477. 18:45As long as the capital market believes
  478. 18:48you will continue to produce generation
  479. 18:50after generation, even if you only make
  480. 18:52a few dollars on every 100, the stock
  481. 18:54will be revalued based on growth over
  482. 18:57the next decade.
  483. 19:00I think the most valuable thing Jensen
  484. 19:03Huang has given this industry is not
  485. 19:05orders or profits, but a future that can
  486. 19:07be valued in advance by the capital
  487. 19:09market.
  488. 19:11But a future is only valuable if someone
  489. 19:13believes in it.
  490. 19:15Why should Taiwanese bosses believe him
  491. 19:17?
  492. 19:18The answer lies in the balance sheets of
  493. 19:20both companies.
  494. 19:23According to Apple's latest annual
  495. 19:25report, after suppliers deliver goods to
  496. 19:27Apple, they wait about four months on
  497. 19:29average to get paid.
  498. 19:32Yet Apple's own inventory turns over
  499. 19:34extremely fast.
  500. 19:37Apple has taken cash turnover to the
  501. 19:39extreme with a cash conversion cycle of
  502. 19:41negative 67 days, meaning Apple always
  503. 19:43has over two months of cash on hand that
  504. 19:46suppliers are essentially carrying for
  505. 19:48them.
  506. 19:50NVIDIA is the exact opposite. According
  507. 19:53to their annual report, they hold a
  508. 19:55large amount of inventory themselves and
  509. 19:57pay suppliers quickly, resulting in a
  510. 20:00cash conversion cycle of positive 133
  511. 20:02days.
  512. 20:04Even more staggering are Nvidia's
  513. 20:06procurement and capacity commitments,
  514. 20:09which hit $ 279 billion by the end of
  515. 20:11July—nearly tripling in just six months
  516. 20:14.
  517. 20:16The bulk of this increase comes from
  518. 20:18prepurchasing memory.
  519. 20:22Most of this money hasn't been paid yet
  520. 20:24, and some can still be adjusted before
  521. 20:26formal orders, but the intent is clear:
  522. 20:29Nvidia is locking in future capacity for
  523. 20:31the entire supply chain.
  524. 20:37One makes suppliers bear the cost, while
  525. 20:39the other steps up to secure capacity;
  526. 20:41both are dominant, but they are at
  527. 20:43different stages.
  528. 20:48Apple faces a mature market with a fixed
  529. 20:50pie, so it pushes to minimize every cost
  530. 20:53.
  531. 20:56AI is a rapidly growing market where
  532. 20:58chips and memory are both scarce; Nvidia
  533. 21:00needs to snatch up tomorrow's supply
  534. 21:03today.
  535. 21:07Apple forces suppliers to fight for a
  536. 21:09visible market, while Nvidia forces them
  537. 21:11to compete for a future not yet realized
  538. 21:13.
  539. 21:16Apple relies on the fear that you are
  540. 21:18replaceable.
  541. 21:20Jensen Huang relies on the expectation
  542. 21:22that you are needed for the next
  543. 21:24generation.
  544. 21:27An order can get a factory running, but
  545. 21:30to make a boss sink billions into a line
  546. 21:32only for the next generation, orders
  547. 21:34aren't enough; they must believe that
  548. 21:36line will still be busy in two years.
  549. 21:40Fear can force price cuts, but capacity
  550. 21:43expansion relies on expectations.
  551. 21:46This brings up something that doesn't
  552. 21:48quite add up.
  553. 21:49At the 2024 Computex, Jensen Huang
  554. 21:52called Barry Lam a hero, and Barry Lam
  555. 21:55called Jensen Huang a king.
  556. 21:59Yet, less than two weeks after Computex
  557. 22:02, Taiwanese media reported that for the
  558. 22:05first batch of Nvidia's next-gen AI
  559. 22:08GB200, Foxconn took 40%and Quanta took
  560. 22:1130%.
  561. 22:16Back in March of that year, analyst
  562. 22:18Ming-Chi Kuo warned the market that
  563. 22:20Nvidia was putting more effort into
  564. 22:22building its own systems to reduce
  565. 22:24reliance on Foxconn.
  566. 22:28If Jensen Huang and Barry Lam are so
  567. 22:30close, why isn't Quanta the biggest
  568. 22:33winner? And if Foxconn's reliance was
  569. 22:35being reduced, why did they still take
  570. 22:37the top spot? Because Jensen Huang knows
  571. 22:39better than anyone how Apple managed
  572. 22:41these people in the past.
  573. 22:44He can make everyone believe the future
  574. 22:46pie is huge, but he will never let any
  575. 22:48single one of them hog the whole thing.
  576. 22:52Friendship is for the dinner table, but
  577. 22:54orders are business; exactly how orders
  578. 22:56are allocated isn't decided over dinner
  579. 22:58.
  580. 22:59Once a batch of the tightest parts
  581. 23:01arrives, which factory should they be
  582. 23:03sent to?
  583. 23:06Nvidia wrote in their own tech blog that
  584. 23:08they built a system with an American
  585. 23:10data analytics firm, Perplexity, to
  586. 23:12recalculate every week, with the
  587. 23:14resulting plan handed to supply chain
  588. 23:16planners to finalize.
  589. 23:20Clients also choose which foundry gets
  590. 23:22their orders.
  591. 23:25Industry reports state that Wistron has
  592. 23:27been the exclusive foundry for Dell from
  593. 23:30the B200 to the GB200.
  594. 23:33Quanta lost a portion of Meta's orders
  595. 23:35over the past two years; it was Meta
  596. 23:37that changed its supplier.
  597. 23:39So, what does Jensen Huang hold in his
  598. 23:42hands? At the end of 2024, Taiwanese
  599. 23:44media reported that NVIDIA had finalized
  600. 23:47the preliminary order allocation for the
  601. 23:49GB200, with Foxconn as the largest
  602. 23:51supplier and Quanta as the second.
  603. 23:55Note the term "finalized," because
  604. 23:57NVIDIA’s headquarters sets the framework
  605. 24:00: whose technology to use, who to
  606. 24:02certify, who receives platform support,
  607. 24:04who is eligible to sit at the table this
  608. 24:07generation, and the approximate capacity
  609. 24:09each company will handle.
  610. 24:12Simply put, Jensen Huang issues licenses
  611. 24:15, not orders.
  612. 24:18Getting a license doesn't guarantee your
  613. 24:20share, but without it, you don't even
  614. 24:21have the qualification to sit at the
  615. 24:23table.
  616. 24:25How do you get this license?
  617. 24:27It depends on whether you can do what
  618. 24:29others cannot; personal relationships
  619. 24:31come second. High-bandwidth memory from
  620. 24:34South Korea is exactly like this. This
  621. 24:36is a type of high-speed memory
  622. 24:38specifically used alongside AI chips. SK
  623. 24:40Hynix, Micron, and Samsung have all
  624. 24:43received NVIDIA's certification. Samsung
  625. 24:46once failed certification, causing its
  626. 24:48share to drop from 40%to less than 20%,
  627. 24:51but it was pulled back up after passing
  628. 24:54.
  629. 24:54The only entrepreneur from mainland
  630. 24:57China present that day was Chen Tao,
  631. 24:59chairman of Huizhou Shenghong Technology
  632. 25:01.
  633. 25:02His factory is one of the few in the
  634. 25:04world capable of mass-producing the
  635. 25:06high-end multi-layer circuit boards
  636. 25:08NVIDIA requires.
  637. 25:11Taiwanese experts say without this
  638. 25:12capability, he wouldn't be able to sit
  639. 25:14at this table.
  640. 25:16The company with the highest gross
  641. 25:18margin in this supply chain is actually
  642. 25:20Chuan Hsin, a company from Kaohsiung
  643. 25:22that started out making cabinets.
  644. 25:24NVIDIA's new platform's server rails,
  645. 25:26those with seven layers or more, are
  646. 25:29made by Chuan Hsin, and their gross
  647. 25:31margin is even higher than NVIDIA's. So
  648. 25:34, do you understand this group photo now
  649. 25:36?
  650. 25:37Whether one is sitting or standing isn't
  651. 25:40that important; what matters is who is
  652. 25:42qualified to be in the photo at all.
  653. 25:46Jensen Huang didn't abolish Apple's rule
  654. 25:48of having suppliers compete against each
  655. 25:51other; he simply moved that competition
  656. 25:53into a high-growth market.
  657. 25:59Facing a cake that keeps getting bigger
  658. 26:01, everyone still fights for it, but even
  659. 26:03those who lose the fight get to eat, so
  660. 26:05no one flips the table.
  661. 26:10Jensen Huang didn't make the Taiwanese
  662. 26:12bosses stop competing; he just made them
  663. 26:15believe that even while competing,
  664. 26:16everyone can get rich together.
  665. 26:21And this belief relies on one person's
  666. 26:23credibility.
  667. 26:24On November 11, NVIDIA's board accepted
  668. 26:27Jensen Huang's resignation. The next day
  669. 26:30, it was announced that he would step
  670. 26:33down as President, CEO, and Director,
  671. 26:35with CFO Colette Kress serving as acting
  672. 26:38CEO. For that entire week, almost none
  673. 26:41of these Taiwanese bosses said a word.
  674. 26:43On November 10th, Barry Lam attended an
  675. 26:45awards ceremony.
  676. 26:47Reporters hounded him about whether his
  677. 26:49supply chain would be affected, but he
  678. 26:51wouldn't comment.
  679. 26:52When asked if his mood was affected, he
  680. 26:55said "not bad" twice.
  681. 26:57Executives like Terry Gou and Jason
  682. 26:59Chang remained silent that week.
  683. 27:04The only person to stand up for Jensen
  684. 27:06Huang was Quanta’s number two, C.C.
  685. 27:08Leung.
  686. 27:10He stated that Jensen Huang would
  687. 27:12certainly not betray anyone, and added
  688. 27:15that everyone still deeply needs AI.
  689. 27:18Current orders remain untouched.
  690. 27:21On November 13th, Taiwan's Central News
  691. 27:23Agency interviewed four semiconductor
  692. 27:26supply chain executives; they said
  693. 27:28orders through 2027 are unchanged, and
  694. 27:30Nvidia's technical roadmap has not been
  695. 27:33adjusted.
  696. 27:35One testing executive noted that
  697. 27:37Nvidia’s platform barriers are getting
  698. 27:39higher, so it isn't easy to just switch
  699. 27:41.
  700. 27:44Nvidia’s new chairman also stated in an
  701. 27:46announcement that company operations,
  702. 27:48customer commitments, and product
  703. 27:50roadmaps would not be interrupted.
  704. 27:54However, behind these public statements
  705. 27:56, some have already started calculating
  706. 27:58a different bottom line.
  707. 28:01The Koreans have been the most blunt
  708. 28:02about it. On November 13th, both Samsung
  709. 28:05and SK Hynix refused to comment, but a
  710. 28:08vice chairman of the Korea Semiconductor
  711. 28:10Industry Association spoke up.
  712. 28:13He said this might not necessarily be a
  713. 28:15bad thing for Korea.
  714. 28:17Even if other companies fill the market
  715. 28:20share left by Nvidia, those companies
  716. 28:22will still require Korea's
  717. 28:23high-bandwidth memory.
  718. 28:27He added that they don't need to
  719. 28:29concentrate production on a single
  720. 28:31client; they can diversify among
  721. 28:33multiple clients, which is actually more
  722. 28:35favorable for them.
  723. 28:38If Nvidia wants to change suppliers, he
  724. 28:41countered: where are the alternative
  725. 28:43suppliers?
  726. 28:45The Koreans have the confidence to say
  727. 28:47this because only a few companies
  728. 28:49worldwide can supply advanced
  729. 28:51high-bandwidth memory at scale.
  730. 28:54Today, it isn't just that the Koreans
  731. 28:56can't live without Nvidia; Nvidia cannot
  732. 28:59live without Korea.
  733. 29:01A U.S. semiconductor research firm
  734. 29:03published an article on the day of
  735. 29:04Jensen Huang's resignation.
  736. 29:09It estimated that in 2027, the price per
  737. 29:11unit of SK Hynix's next-gen
  738. 29:13high-bandwidth memory for Nvidia could
  739. 29:16be nearly 20%cheaper than for clients
  740. 29:19like AMD.
  741. 29:22How does Nvidia secure such terms?
  742. 29:25Because Nvidia is one of their largest
  743. 29:28clients, the two parties have
  744. 29:29collaborated for many years, possibly
  745. 29:32even exchanging benefits like priority
  746. 29:34GPU supply.
  747. 29:36How many of these terms are in contracts
  748. 29:39versus based on mutual trust for future
  749. 29:41cooperation is unclear to outsiders.
  750. 29:45What is certain is that after the new
  751. 29:47management takes over, holding only past
  752. 29:49contracts won't be enough when sitting
  753. 29:51down to negotiate prices and capacity.
  754. 29:56They still have to convince the Koreans
  755. 29:58that Nvidia remains the client most
  756. 30:00worthy of priority.
  757. 30:03When Jensen Huang was there, everyone
  758. 30:05knew how these interest exchanges were
  759. 30:07coordinated.
  760. 30:10If Google and Amazon come knocking with
  761. 30:12major orders in the future, why should
  762. 30:14Nvidia remain at the front of the line?
  763. 30:18This isn't a problem that can be solved
  764. 30:20with a meal or a photo op.
  765. 30:23More troublingly, even if Jensen Huang
  766. 30:26hadn't resigned today, Nvidia's
  767. 30:27competitors have long been carving into
  768. 30:30their market share.
  769. 30:33Google has its own AI chips, Amazon does
  770. 30:36too, and AMD has been catching up
  771. 30:38steadily.
  772. 30:42Morgan Stanley predicts that by 2027,
  773. 30:45Nvidia's share of global high-bandwidth
  774. 30:48memory demand will drop to around 37%,
  775. 30:50while Google's self-developed chips will
  776. 30:53account for 36%, putting them neck and
  777. 30:56neck.
  778. 31:00In that CNA report, a high-level
  779. 31:02executive at a substrate manufacturer
  780. 31:04also mentioned that the landscape
  781. 31:06between GPUs and application-specific
  782. 31:09chips might shift toward the latter.
  783. 31:13So the competition itself began long ago
  784. 31:15, regardless of whether Jensen Huang
  785. 31:17leaves or not.
  786. 31:20The real issue is that at the peak of
  787. 31:22this intense competition, Nvidia
  788. 31:24suddenly lacks the person who was best
  789. 31:26at binding technology suppliers and
  790. 31:28capital together.
  791. 31:30At a time like this, will others seize
  792. 31:32the opportunity to snatch away the next
  793. 31:34orders that originally belonged to
  794. 31:36Nvidia?
  795. 31:38A signed order won't become void just
  796. 31:40because one person resigns.
  797. 31:43But what about the next unsigned order,
  798. 31:45the next factory that hasn't broken
  799. 31:47ground, or the next year’s capital
  800. 31:49expenditure yet to be approved?
  801. 31:52A former Nvidia Korea representative
  802. 31:54stated clearly that the contracts signed
  803. 31:57by Samsung and SK Hynix with Nvidia
  804. 31:59already cover next year, so there won't
  805. 32:01be much impact in the short term.
  806. 32:05However, variables may emerge after 2028
  807. 32:08.
  808. 32:10Looking at it from another angle, if
  809. 32:12this supply chain really cannot function
  810. 32:14without Jensen Huang, it would suggest
  811. 32:16that in his 30-plus years of running
  812. 32:18Nvidia, he failed to turn his most
  813. 32:20important capabilities into company
  814. 32:22systems.
  815. 32:23But Nvidia's technical standards,
  816. 32:25customer network, and supply chain
  817. 32:27systems are all still in place. TSMC’s
  818. 32:30production capacity won't simply vanish
  819. 32:32into thin air just because he resigned.
  820. 32:35Can Nvidia keep spinning without Jensen
  821. 32:37Huang?
  822. 32:38The answer is already clear.
  823. 32:41It certainly can.
  824. 32:44The real test of this transition is
  825. 32:46whether Nvidia can continue to convince
  826. 32:49the entire industry chain that, even
  827. 32:51without Jensen Huang, betting the next
  828. 32:53round of investment here remains the
  829. 32:55best business move.
  830. 32:59TSMC has already provided a hint through
  831. 33:01its actions.
  832. 33:05On November 10, the day before Nvidia's
  833. 33:07board accepted his resignation, TSMC's
  834. 33:10board approved three investment quotas
  835. 33:13of up to $ 500 million each: subscribing
  836. 33:15to shares of AI company Anthropic upon
  837. 33:18its IPO, buying OpenAI preferred stock,
  838. 33:20and establishing a strategic investment
  839. 33:23firm.
  840. 33:28This certainly doesn't prove that C.C.
  841. 33:30Wei knew anything in advance, but it at
  842. 33:32least shows that TSMC has never bet on
  843. 33:35any one individual at NVIDIA, but rather
  844. 33:37on the future demand for AI as a whole.
  845. 33:42Some of these companies are now making
  846. 33:44their own chips.
  847. 33:46As mentioned in my video a couple of
  848. 33:48days ago, production lines have inertia
  849. 33:51, but bridges don't.
  850. 33:53To be more specific now: orders have
  851. 33:55inertia, but trust does not.
  852. 33:59When the next round of discussions
  853. 34:01regarding pricing, capacity, and
  854. 34:02next-gen platforms begins, the person
  855. 34:05sitting across the table will no longer
  856. 34:07be the one known in Taipei’s industry
  857. 34:09circles as "The King."
  858. 34:12The November 17th earnings call will be
  859. 34:14Kress's first time addressing the market
  860. 34:17as acting CEO.
  861. 34:19As of November 13th, those bosses in
  862. 34:21Taiwan and South Korea are still waiting
  863. 34:23for him to speak.
  864. 34:25This year, on May 28th, was the fifth "
  865. 34:28AI Banquet."
  866. 34:30That day, Jensen Huang's son Spencer and
  867. 34:32daughter Madison were also there.
  868. 34:35C.C. Wei was still on one side of him,
  869. 34:38Barry Lam on the other, and Young Liu
  870. 34:40was still standing in the second row.
  871. 34:44That was the last banquet he hosted in
  872. 34:47his capacity as NVIDIA's CEO.
  873. 34:51In 2009, Terry Gou and Barry Lam locked
  874. 34:53fingers in a restroom, but once they
  875. 34:55came out, they continued their rivalry
  876. 34:58just the same.
  877. 34:59A few years later, this group finally
  878. 35:01managed to sit at the same table.
  879. 35:05In truth, they never stopped competing
  880. 35:07for a single day, but someone appeared
  881. 35:09in their midst who made everyone believe
  882. 35:11the future pie would keep growing—large
  883. 35:14enough for everyone to get a slice.
  884. 35:17Now that person is gone. By this time
  885. 35:19next year, there might still be a group
  886. 35:21photo taken in front of that Taiwanese
  887. 35:23restaurant.
  888. 35:25Who sits in the first row or stands in
  889. 35:27the second row really doesn't matter as
  890. 35:29much anymore.
  891. 35:31What truly matters is whether these
  892. 35:33people are still willing to bet billions
  893. 35:35upon billions of capital on the next
  894. 35:37generation of chips once they go back.
  895. 35:46A photo can keep everyone in the same
  896. 35:48frame, but only a future that everyone
  897. 35:50believes in can keep them sitting at the
  898. 35:52same table.
  899. 35:59Thank you for watching today's video, we
  900. 36:01...

About this transcript

This page contains the full transcript of 广达鸿海斗了二十年,为什么就只服黄仁勋?黄老板被扫地出门,AI 产业链还稳吗?兆元宴合影藏着整个台湾代工江湖|艾森 Essen by 艾森 Essen, generated from the public captions YouTube serves with the video. The transcript has 5,183 words across 900 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.