YouTube2Text

젠슨 황 퇴진, AI 패권 전쟁의 서막! 엔비디아가 주인이 아니었다… 앤트로픽·스페이스X가 준비하는 새로운 AI 시대 — Transcript

by 엔지니어TV · 2,537 words · 454 segments · language en · Watch on YouTube

Full transcript

  1. 0:00There are two companies that NVIDIA TV
  2. 0:02has been focusing on intensively lately
  3. 0:04. They are Anthropic and SpaceX. I
  4. 0:07believed they were companies creating AI
  5. 0:16models and showing us where we are
  6. 0:23headed. And at this very moment, a
  7. 0:26surprising event occurred. Jensen Huang
  8. 0:28has suddenly stepped down from the CEO
  9. 0:30position. NVIDIA appointed its CFO as
  10. 0:33interim CEO and began searching for new
  11. 0:35leadership.
  12. 0:36Now, Jensen Huang was not just a typical
  13. 0:38CEO of a semiconductor company.
  14. 0:41For the past few years, NVIDIA has been
  15. 0:43at the center of the AI industry, and
  16. 0:46Jensen Huang was at its core. Whether
  17. 0:48OpenAI announced a new model or Google
  18. 0:50expanded its AI data centers, NVIDIA's
  19. 0:52GPUs were essential, and NVIDIA remained
  20. 0:55in the spotlight.
  21. 0:56Jensen Huang appeared every time the
  22. 0:58future of the AI industry was discussed
  23. 1:00, and he recently went around personally
  24. 1:01explaining this future in detail.
  25. 1:03It was as if he were the leader guiding
  26. 1:06the entire AI industry.
  27. 1:07But suddenly, that figure has
  28. 1:09disappeared. Now, will he be able to
  29. 1:10return? Of course, it’s not that
  30. 1:12NVIDIA’s GPU technology has vanished.
  31. 1:14CUDA is still there, and their
  32. 1:15world-class semiconductor development
  33. 1:17talent remains in place. However,
  34. 1:19watching this event, one question came
  35. 1:21to mind. Was NVIDIA truly the master of
  36. 1:24the AI industry? And will AI companies
  37. 1:26have no choice but to continue relying
  38. 1:28on NVIDIA in the future? Let’s examine
  39. 1:30this issue from a broader perspective.
  40. 1:33Let's look at how the AI industry is
  41. 1:35actually moving at the moment. They say
  42. 1:38the US and China are currently engaged
  43. 1:39in a war for AI supremacy.
  44. 1:41In the US, world-class AI models are
  45. 1:43being developed, creating new markets
  46. 1:45through AI services.
  47. 1:46There are the companies creating those
  48. 1:49AI models, and NVIDIA sits at the
  49. 1:51hardware core, the GPU. They supply the
  50. 1:53GPUs, networking, and software ecosystem
  51. 1:56needed to train and run AI models.
  52. 1:58But who actually makes NVIDIA's chips?
  53. 2:02It is Taiwan’s TSMC. They produce
  54. 2:04advanced semiconductors and even handle
  55. 2:06high-performance packaging.
  56. 2:08And the places that assemble these
  57. 2:10semiconductors into actual AI servers,
  58. 2:12like Foxconn and Quanta, play a vital
  59. 2:14role.
  60. 2:16And recently, the importance of Korean
  61. 2:17companies has also grown. It is the HBM
  62. 2:19supplied by SK Hynix and Samsung
  63. 2:21Electronics.
  64. 2:23There is even talk that even if a GPU is
  65. 2:25fast, if it cannot receive data quickly
  66. 2:27enough, HBM is what truly determines the
  67. 2:29actual performance.
  68. 2:31It has become a core component.
  69. 2:33Now, while the US is leading in AI
  70. 2:35models and semiconductor design, the
  71. 2:37hardware that actually runs the AI
  72. 2:39depends on a global manufacturing supply
  73. 2:41chain, including Taiwan and South Korea
  74. 2:44.
  75. 2:44Now, on the other side of this, there is
  76. 2:46China.
  77. 2:48China is developing its own AI
  78. 2:49semiconductors centered around Huawei,
  79. 2:51while expanding servers, data centers,
  80. 2:52and power infrastructure.
  81. 2:54China is responding by combining its
  82. 2:56proprietary technology, manufacturing
  83. 2:58capabilities, and low infrastructure
  84. 3:00costs.
  85. 3:01However, Jensen Huang has been a very
  86. 3:03special figure in this complex supply
  87. 3:04chain.
  88. 3:05This is because he has long connected
  89. 3:08U.S. AI companies with semiconductor and
  90. 3:10server manufacturers in Taiwan. The
  91. 3:12relationship with TSMC was not built
  92. 3:14over a short period, either. For decades
  93. 3:16, they have developed technology
  94. 3:18together, secured production capacity,
  95. 3:20and pioneered new markets. Such
  96. 3:22relationships are vital, especially in
  97. 3:24markets like the AI industry where
  98. 3:26demand is surging rapidly.
  99. 3:29This is because having just a GPU does
  100. 3:31not mean you can build an AI data center
  101. 3:32. You must secure semiconductor
  102. 3:34production, packaging, memory, servers,
  103. 3:36and even network equipment on time. The
  104. 3:38ability to coordinate all of these is
  105. 3:40necessary. Jensen Huang has been
  106. 3:42performing that very role. But now,
  107. 3:44things are set to change. Of course,
  108. 3:46technical agreements and cooperation
  109. 3:48will continue. The question is how
  110. 3:50naturally the trust and network built by
  111. 3:53Jensen Huang personally will continue
  112. 3:56into the future. Now, there will clearly
  113. 3:58be some changes. This means it is a more
  114. 4:01significant issue than a simple CEO
  115. 4:03replacement.
  116. 4:04From the perspective of AI supremacy,
  117. 4:06Nvidia was not just a company selling
  118. 4:09GPUs, but a company that connected the
  119. 4:11global AI ecosystem into a single system
  120. 4:14.
  121. 4:15And at the core of this, there is Jensen
  122. 4:17Huang.
  123. 4:18Now, interestingly, a new company has
  124. 4:20suddenly appeared here.
  125. 4:22It is Anthropic. Most AI companies have
  126. 4:25competed over how many of the latest
  127. 4:27Nvidia GPUs they could secure. However,
  128. 4:30Anthropic's strategy was a little
  129. 4:31different. They also use Amazon's
  130. 4:34Trainium and Google's TPU. Of course,
  131. 4:36they use Nvidia GPUs as well. But on top
  132. 4:39of that, they added AMD's AI
  133. 4:41accelerators. In particular, they
  134. 4:43decided to adopt 4 th-generation AI
  135. 4:45computing infrastructure from AMD on a
  136. 4:47scale of up to 2 GW. 2 GW is not a small
  137. 4:50scale. It is a power capacity level
  138. 4:53equivalent to several large power plants
  139. 4:55. Anthropic did not say they would stop
  140. 4:57using Nvidia. But they did say they
  141. 4:59would break free from situations where
  142. 5:01they*had*to use it. Why is this
  143. 5:03important? Until now, one of Nvidia's
  144. 5:05strongest competitive edges was CUDA,
  145. 5:08which is software. This is because the
  146. 5:10software ecosystem for developing and
  147. 5:13running AI models was built around
  148. 5:15Nvidia GPUs.
  149. 5:17But recently, AI itself has begun to
  150. 5:19help with the task of running and
  151. 5:22optimizing software on other
  152. 5:23semiconductors.
  153. 5:26Anthropic is also utilizing the cloud to
  154. 5:28push for software development and
  155. 5:30optimization for AMD GPUs.
  156. 5:32They are using AI to develop software
  157. 5:34for AI semiconductors. Now, the power of
  158. 5:37this AI is emerging one by one through
  159. 5:40agents, as I said.
  160. 5:43All software is becoming a target, and
  161. 5:45especially the software that causes
  162. 5:47bottlenecks and critical software will
  163. 5:49all be touched by AI.
  164. 5:51As these technologies advance, the cost
  165. 5:53and time to switch semiconductor
  166. 5:55suppliers are likely to decrease.
  167. 5:57Of course, this alone does not mean
  168. 5:59NVIDIA's technological edge will vanish
  169. 6:01. NVIDIA won't just sit still either.
  170. 6:03But from the perspective of AI companies
  171. 6:05, their options are increasing. There
  172. 6:06are many bottlenecks right now.
  173. 6:09For AI firms, hardware is difficult, but
  174. 6:10they are actually starting to enter that
  175. 6:12space, and all software-related things
  176. 6:14are becoming targets.
  177. 6:17Tasks favorable to Amazon's chips can
  178. 6:18run on Amazon EC2, and tasks efficient
  179. 6:20on Google TPUs can run on Google.
  180. 6:23They should also be able to select
  181. 6:24between tasks where AMD or NVIDIA have
  182. 6:26the advantage. If this happens, the
  183. 6:28negotiation structure between
  184. 6:29semiconductor suppliers and AI model
  185. 6:31firms could change.
  186. 6:33Here, the figure Dario Amodei is also
  187. 6:35important. Dario previously worked at
  188. 6:38the U.S. AI lab of China's Baidu.
  189. 6:41At the time, Baidu led an incredible
  190. 6:43group of AI talent, but they have faded
  191. 6:45a bit now.
  192. 6:46From the early days of AI, China was
  193. 6:48investing significantly. He studied
  194. 6:51speech recognition in a team led by
  195. 6:53Andrew Ng, and after passing through
  196. 6:55Google and OpenAI, he founded Anthropic
  197. 6:58. For Dario, the sibling structure is
  198. 7:00also important. Essentially, he is
  199. 7:02someone who has directly experienced
  200. 7:04China's AI research ecosystem. So, the
  201. 7:06AI industry has a quite complex
  202. 7:07structure, or should I say, a limited
  203. 7:10cast of characters?
  204. 7:11Now, Dario has shown a strong stance on
  205. 7:14controlling exports of advanced AI chips
  206. 7:16to China.
  207. 7:18On the other hand, Jensen Huang has
  208. 7:19emphasized the global spread of U.S. AI
  209. 7:21technology and the importance of access
  210. 7:22to the Chinese market. It is a somewhat
  211. 7:24opposing stance. This difference cannot
  212. 7:26be explained solely by political
  213. 7:28background, as the business structures
  214. 7:30of the two companies are different.
  215. 7:32It is the difference between
  216. 7:33semiconductors and AI models.
  217. 7:36NVIDIA may want a wider semiconductor
  218. 7:38market, while Anthropic may judge it
  219. 7:41strategically advantageous to limit
  220. 7:43rival nations 'access to advanced
  221. 7:45computing.
  222. 7:46Here, a new face of the AI supremacy war
  223. 7:48is being revealed.
  224. 7:50AI companies are no longer just
  225. 7:52customers buying NVIDIA GPUs, but are
  226. 7:54becoming entities that choose
  227. 7:55semiconductor supply chains, influence
  228. 7:57tech policy, and build their own
  229. 7:59computing strategies.
  230. 8:02So then, how is China responding? One of
  231. 8:05China's most powerful weapons is cost.
  232. 8:09We know that the recent cost of building
  233. 8:11AI data centers in China is much lower
  234. 8:13than in the U.S., but just how much
  235. 8:15lower?
  236. 8:16The recent comparison region is Inner
  237. 8:18Mongolia. It is a lower-cost area in
  238. 8:21China, and they say it's about 1/5th the
  239. 8:23cost. Of course, that's excluding GPUs.
  240. 8:27Since it mainly compares construction
  241. 8:28and infrastructure costs, China’s
  242. 8:30ability to leverage cheaper construction
  243. 8:32, power infrastructure, and supply
  244. 8:33chains is a key competitive edge.
  245. 8:36However, there is a decisive problem
  246. 8:37here.
  247. 8:39No matter how cheaply you build data
  248. 8:41centers, you can't run state-of-the-art
  249. 8:43AI models efficiently without the latest
  250. 8:45AI semiconductors.
  251. 8:47Now, evidence of this is emerging from
  252. 8:50various sources. Jensen Huang's recent
  253. 8:52issue is also due to high-end GPUs
  254. 8:54leaking into China. If Huawei chips were
  255. 8:56superior, they wouldn't need Nvidia GPUs
  256. 8:58, right? From China's perspective. But
  257. 9:00that is not the case.
  258. 9:02That’s why Chinese companies have been
  259. 9:04keen on Nvidia's latest GPU computing
  260. 9:06resources even while developing their
  261. 9:07own chips.
  262. 9:09That is one of the reasons why regions
  263. 9:11like Malaysia and Singapore have become
  264. 9:13important recently.
  265. 9:15Now, what happens if the U.S. starts
  266. 9:17restricting Chinese companies' use of
  267. 9:19overseas cloud services?
  268. 9:20Chinese firms may find it difficult to
  269. 9:22access GPUs secured abroad. The bigger
  270. 9:25problem lies with the operators who
  271. 9:26built the data centers.
  272. 9:28While Southeast Asian data centers
  273. 9:30recently gained attention, sentiment is
  274. 9:32shifting rapidly toward them being at
  275. 9:34risk due to this event.
  276. 9:36The situation will depend on reliance on
  277. 9:38Chinese clients, contract terms, and the
  278. 9:40ability to secure other customers. This
  279. 9:42shows that the competitiveness of an AI
  280. 9:44data center isn't decided solely by
  281. 9:45construction costs. Access to
  282. 9:47semiconductors, power prices, financial
  283. 9:49structures, client mix, and
  284. 9:51international regulations all play a
  285. 9:53part.
  286. 9:54This is exactly why we need to examine
  287. 9:56Elon Musk’s SpaceX strategy.
  288. 9:59Musk is developing AI models, but he is
  289. 10:02even more focused on building
  290. 10:03large-scale data centers.
  291. 10:05Tesla is developing its own AI
  292. 10:07semiconductors for autonomous driving
  293. 10:09and robotics. While these are separate
  294. 10:11businesses, they share a common
  295. 10:12direction.
  296. 10:13They are trying to secure the necessary
  297. 10:16computing, energy, networks, and
  298. 10:18hardware for AI more directly.
  299. 10:21This explains why Elon Musk is obsessed
  300. 10:24with building data centers quickly,
  301. 10:26securing power, and investing in chip
  302. 10:28design and manufacturing.
  303. 10:31Improving AI model performance alone is
  304. 10:33not enough. The cost of training and
  305. 10:35running those models is what matters.
  306. 10:38And in the future, AI will operate in
  307. 10:40cars, robots, and industrial facilities
  308. 10:43—meaning it will be integrated with
  309. 10:45physical hardware.
  310. 10:47The AI race no longer ends with who
  311. 10:49builds the smarter model.
  312. 10:52It’s becoming about who can operate
  313. 10:54those models more affordably and
  314. 10:55reliably.
  315. 10:57I believe this is exactly why we should
  316. 10:59look at Anthropic and Musk’s companies
  317. 11:01together.
  318. 11:03Anthropic is trying to reduce its
  319. 11:05reliance on suppliers by utilizing
  320. 11:06various semiconductors, while Musk is
  321. 11:08directly securing more infrastructure
  322. 11:10and physical systems to run AI.
  323. 11:13Both strategies stem from the
  324. 11:15realization that AI models alone are not
  325. 11:17enough.
  326. 11:19However, Nvidia has also been moving
  327. 11:20into the territory of AI model companies
  328. 11:22.
  329. 11:23Beyond building faster GPUs, they have
  330. 11:26been specifically emphasizing the
  331. 11:28development of AI platforms for robots.
  332. 11:30Physical AI. They have approached
  333. 11:33software from a more advanced level.
  334. 11:34They have been heavily investing in
  335. 11:36simulation and training environments for
  336. 11:38autonomous driving, known as "world
  337. 11:40models." World models are AI
  338. 11:41technologies that understand and predict
  339. 11:43the movements and physical changes of
  340. 11:44the real world. Autonomous driving is
  341. 11:46also a technology contained within world
  342. 11:48models. These are important parts of
  343. 11:50Physical AI.
  344. 11:51What does it mean that Nvidia has
  345. 11:53developed this technology?
  346. 11:55Up until now, their core role was
  347. 11:57supplying semiconductors, but moving
  348. 11:59forward, they intended to provide
  349. 12:01software platforms that help AI
  350. 12:02understand and operate in the real world
  351. 12:05. This is where their business areas
  352. 12:06begin to overlap with existing AI
  353. 12:08companies.
  354. 12:09Let’s imagine a robot company adopting
  355. 12:13AI.
  356. 12:13They could use Nvidia’s robot AI
  357. 12:15platform, or they could utilize
  358. 12:16technology from other AI model companies
  359. 12:18. Whenever a new AI model is released, a
  360. 12:22major concern in the robotics industry
  361. 12:25is how well it makes the robot operate.
  362. 12:28Unitree’s Wang Xingxing recently proved
  363. 12:31this. Such choices will become even more
  364. 12:33important in the future.
  365. 12:34If that happens, can Nvidia and existing
  366. 12:37AI companies continue to only cooperate
  367. 12:39? That might not be the case. They will
  368. 12:41end up competing with each other.
  369. 12:43Conversely, AI model companies are
  370. 12:45developing their own semiconductors and
  371. 12:47moving toward utilizing various types of
  372. 12:50chips.
  373. 12:51Now, the boundaries between companies
  374. 12:52are breaking down.
  375. 12:53I think this is a more significant
  376. 12:55long-term change than Jensen Huang’s
  377. 12:58eventual retirement.
  378. 13:00Why have I been keeping a close eye on
  379. 13:02Anthropic and SpaceX recently?
  380. 13:04It is the strengths in software and AI
  381. 13:06models versus hardware between Anthropic
  382. 13:08and SpaceX, as well as the respective
  383. 13:10strengths of China and the US.
  384. 13:13The common thread among all these
  385. 13:15companies and nations is that they are
  386. 13:17all transcending their existing business
  387. 13:19domains.
  388. 13:21When China steals Anthropic’s latest
  389. 13:23models to build their own, it is
  390. 13:25ethically problematic, to say the least
  391. 13:27.
  392. 13:29They are essentially taking something
  393. 13:30that required massive investment and
  394. 13:31getting it for free.
  395. 13:33However, the reason behind these actions
  396. 13:36is hegemony; while there is a war for
  397. 13:38hegemony between nations, the core
  398. 13:41technology here is AI, and as this AI
  399. 13:43becomes physical AI, it begins to
  400. 13:45influence hardware and actual physical
  401. 13:48situations.
  402. 13:50Things are no longer happening just
  403. 13:52within networking. In the future, it
  404. 13:54will be possible to have a direct
  405. 13:56physical impact or influence.
  406. 13:58Manufacturing companies are trying to
  407. 14:00change actual industries using AI. This
  408. 14:02means that AI competition is going to
  409. 14:05change. Even if you secure many GPUs, if
  410. 14:07power costs are too high,
  411. 14:08competitiveness drops, and even if an AI
  412. 14:10model has excellent performance, it
  413. 14:12could be limited if it is difficult to
  414. 14:13apply to actual industries. It might not
  415. 14:16be chosen by the market.
  416. 14:18Conversely, if you combine appropriate
  417. 14:19semiconductors, affordable power, and
  418. 14:21efficient AI models, you can provide
  419. 14:23services at a much lower cost.
  420. 14:26And here, in the era of physical AI,
  421. 14:28actual hardware like robots, automobiles
  422. 14:30, and factories is being added.
  423. 14:32Now, AI is evolving from a technology
  424. 14:34that simply answers questions into a
  425. 14:36technology that moves the real world. In
  426. 14:39that process, the integration of brains
  427. 14:41and hardware becomes even more important
  428. 14:43. Now, AI and hardware exist separately
  429. 14:46, but how do we combine them?
  430. 14:50I expect that cooperation between
  431. 14:51companies that develop chips and
  432. 14:53companies that develop AI models will
  433. 14:55increase.
  434. 14:57Now, Jensen Huang stepping down doesn't
  435. 14:59mean Nvidia is over, and he could even
  436. 15:01return.
  437. 15:02But the point is that there is a reason
  438. 15:04why this situation has arisen.
  439. 15:06It shows the current AI industry
  440. 15:08evolving from a single center called
  441. 15:09Nvidia into a more complex and fierce
  442. 15:11competitive structure.
  443. 15:14Even if Jensen Huang's retirement did
  444. 15:16not directly cause this change, it is
  445. 15:18clearly an important event that makes us
  446. 15:20rethink the existing order.
  447. 15:23Nvidia did, in fact, open the door to
  448. 15:25the AI era, but the master of that era
  449. 15:28does not necessarily have to be Nvidia.
  450. 15:31I think this is one of the most
  451. 15:33important changes that will take place
  452. 15:34in the AI industry going forward.
  453. 15:37The real AI hegemony war is just
  454. 15:39beginning now.

About this transcript

This page contains the full transcript of 젠슨 황 퇴진, AI 패권 전쟁의 서막! 엔비디아가 주인이 아니었다… 앤트로픽·스페이스X가 준비하는 새로운 AI 시대 by 엔지니어TV, generated from the public captions YouTube serves with the video. The transcript has 2,537 words across 454 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.