YouTube2Text

"현실은 미친 듯이 어렵다" 일론 머스크가 말하는 AI 버블의 진실 | 2026 바론 투자 컨퍼런스 — Transcript

by 머니네버슬립 · 3,970 words · 568 segments · language en · Watch on YouTube

Full transcript

  1. 0:00I think there's a lot of speculators out
  2. 0:02there who think you can simply order
  3. 0:04some GPUs and just uh rake in cash.
  4. 0:07Um and the reality is it's insanely hard
  5. 0:10will exceed 500 billion. Now um this is
  6. 0:13just my this is this this is not a
  7. 0:15promise. This is just this is what I
  8. 0:17think will happen. Um if it if it if it
  9. 0:20fails it'll be crazy and if it succeeds
  10. 0:22it'll be enormous.
  11. 0:24terrifyingly large.
  12. 0:26You really need to go to space because
  13. 0:28just there isn't enough uh power on the
  14. 0:30ground.
  15. 0:32We could just acquire that legacy telco
  16. 0:34and use those spectra. There's literally
  17. 0:36the opposite of what I sometimes I'm a
  18. 0:38little optimistic on the time frame but
  19. 0:39I always deliver in the end but it's
  20. 0:41overwhelmingly built in Asia. Um the
  21. 0:47so so my my concern really is that there
  22. 0:50is always like some geopolitical risk in
  23. 0:53that region.
  24. 0:54Um and I think that that risk may come
  25. 0:57to fruition at some point in which case
  26. 1:00uh if we don't have all the pieces
  27. 1:03necessary to build chips at scale in
  28. 1:06America then um we may not have any
  29. 1:08chips at all or very few, the 10%number
  30. 1:11type of thing.
  31. 1:13Um and then there's the the maybe
  32. 1:16looking ahead you know past any sort of
  33. 1:20medium term geopolitical bumps in the
  34. 1:24road.
  35. 1:25Uh we we're simply even if you exclude
  36. 1:28geopolitical crisis issues, the rate at
  37. 1:31which we're making chips is simply too
  38. 1:35slow.
  39. 1:37So if you if you want to get to say a
  40. 1:39terawatt of chip production which is our
  41. 1:42goal that's the reason it's called the
  42. 1:44terawatt it's people when I say terawatt
  43. 1:46there's you know three possible meanings
  44. 1:49to terawatt one is terror t e r r o r
  45. 1:54which I think sometimes people like why
  46. 1:57do we want a terawatt um oh and then
  47. 2:00there's t e r r a which just means earth
  48. 2:03um but it's actually t e r a uh as in
  49. 2:06terawatt um so but arguably all three
  50. 2:09apply. So
  51. 2:15it's terrifyingly large. Um so
  52. 2:19so the rest of the world is making we'll
  53. 2:21make this year around 20 gigawatts of AI
  54. 2:24chips maybe 25.
  55. 2:27Um, and then next year depending on
  56. 2:29whose projections you believe, it's
  57. 2:32maybe around 35 to 45 gigawatts of of AI
  58. 2:34chips next year.
  59. 2:38Um, and um
  60. 2:42but if you want to do a terawatt, that's
  61. 2:45now you need you need 30 times what
  62. 2:47current global production will be
  63. 2:49roughly next year for just us.
  64. 2:54We're building that for us.
  65. 2:57Well, yes, we're building I mean we
  66. 2:58could sell it to other people too, but
  67. 3:00basically we're building 30 times what
  68. 3:02the annual production is in one plan.
  69. 3:06Do you think it will have a production
  70. 3:08value as large as the GDP of the 20th
  71. 3:11largest country in the world?
  72. 3:13One, one plan. Yeah, it might even make
  73. 3:16the top 10.
  74. 3:18So, um, no, obviously, you know, a plan
  75. 3:22like this, uh...I think, if it fails,
  76. 3:25it'll be crazy, and if it succeeds,
  77. 3:28it'll be audacious.
  78. 3:30So, it sort of depends on the outcome. I
  79. 3:33think it will succeed at least to some
  80. 3:34degree. Um, hope so. So,
  81. 3:38yeah, I mean it'll cost a trillion.
  82. 3:46Um,
  83. 3:50but we won't need that money immediately
  84. 3:51. It'll be in phases.
  85. 3:55The reason they don't want to try
  86. 3:57anything adventurous with chips is that
  87. 4:00these things require a trillion in
  88. 4:02capital, and if they have a recipe that
  89. 4:05works, they simply don't want to risk
  90. 4:08something new.
  91. 4:11So, generally, the preference in the
  92. 4:14chip industry is to repeat what they did
  93. 4:17last time with minor changes and not
  94. 4:20take new risks.
  95. 4:24So we're building a research fab in
  96. 4:29Texas, right in Austin, at Giga Texas.
  97. 4:35That research fab is going to try many
  98. 4:37different approaches to chip fabrication
  99. 4:39.
  100. 4:41Um, and many of those approaches won't
  101. 4:43work, but some of them, I'm sure, will
  102. 4:46work, and then we will translate those
  103. 4:49improvements to the terafab itself, the
  104. 4:52production terafab, which is in Grimes
  105. 4:54County, Texas.
  106. 4:58Um, and that's where we'll scale up
  107. 5:01production of our AI chips.
  108. 5:06I think one of the things that will be
  109. 5:09interesting about the research in Austin
  110. 5:12is that I wonder if we will have the
  111. 5:15ability to do logic, memory, packaging,
  112. 5:18and lithography.
  113. 5:23So that we can design new chips and,
  114. 5:27under one roof, make the whole chip,
  115. 5:30test it, and iterate very quickly.
  116. 5:37Something that would normally take a
  117. 5:39year, I think we may be able to reduce
  118. 5:41to a month.
  119. 5:43So,
  120. 5:45for any given technology endeavor, the
  121. 5:48pace of progress is determined by how
  122. 5:50many iterations you can do and how much
  123. 5:53progress you can make between those
  124. 5:55iterations.
  125. 5:57음, 그러니까, 만약 반복 의 속도 와 단위 시간당 얼마나 많은 반복 을
  126. 6:02할 수 있는지, 그리고 특정 기술 의 각 반복 사이 의 진전 을 개선 할
  127. 6:08수 있다면, 여러분 은 다른 누구 보다 훨씬 더 빠른 속도 로 발전 을
  128. 6:13가속화 할 수 있을 것 입니다. 만약 여러분 이, 아시 다시피, 한
  129. 6:17자릿수 더 빠르게 반복 할 수 있다면, 일차 근사치 로서 여러분 의 기술
  130. 6:22도 한 자릿수 더 빠르게 발전 할 것 입니다.
  131. 6:25그래서, 저 에게 흥미로운 당신 의 문화 중 하나 는 저 에게 핵심 인물
  132. 6:31을 소개해 준다는 점 입니다.
  133. 6:36관심 을 가질 만한 사람 이라서 그 와 시간 을 보냈고, 그가 저 를
  134. 6:42방문 했을 때 "일론 과 함께 일하는 것은 어떤가요?"
  135. 6:47라고 물었 습니다.
  136. 6:49그는 "다른 사람들 은 오리건 에 있는 제 시설 에 와서 한 시간 반
  137. 6:54정도 둘러 보고 떠 납니다." 라고 말했죠. 그런데 일론 은 와서 서너
  138. 7:00시간 동안 앉아 있었습니다.
  139. 7:02그리고 그는 엔지니어 들을 한 명씩 차례 대로 인터뷰 하며 모든 개념 에
  140. 7:06대해 끊임없이 질문 했습니다. 그래서 그가 하는 모든 일은 속도 와 관련
  141. 7:10이 있습니다. 이제 뉴스 에 자주 나오고 사람들 이 많이 이야기 하는
  142. 7:14데이터 센터 에 대해 조금 이야기 해 보죠.
  143. 7:19'어떻게 친구 를 만들고 사람들 을 설득 할 것인가' 처럼, 데이터 센터
  144. 7:24에 대한 여론 의 반응 은 꽤 날카 롭습니다.
  145. 7:28네, 그렇다고 생각 합니다.
  146. 7:30음, 그래서
  147. 7:33하지만 네, 제가 OpenAI 를 오픈 소스 비영리 단체 로 시작 했는데
  148. 7:37지금 은 최대 이익 을 추구 하는 폐쇄 형 영리 법인 이 되었죠. 제가
  149. 7:43시작한 것과는 정반대 가 된 셈 입니다. 그리고 앤 스로 픽 은
  150. 7:47OpenAI 를 떠난 7 명이 설립 했는데, 그들이 샘 올 트먼 을 신뢰
  151. 7:52하지 않았기 때문 이고, 그럴 만한 이유 가 있었습니다. 음, 그들은
  152. 7:57이를 공개적 으로 밝혔고 비밀 도 아닙니다. 그래서 사실상 앤 스로 픽
  153. 8:01은 OpenAI 의 파생 기업 입니다.
  154. 8:03음, 그래서
  155. 8:07네, 그냥 OpenAI 를 비영리 가 아니라 일반적인 C-Corp (
  156. 8:11일반 법인) 으로 시작 했어야 했다는 생각 이 드네요.
  157. 8:16음, 하지만 저는 이것이 세상 을 위한 자선 적인 일 이 될 것이라고
  158. 8:20생각 했습니다.
  159. 8:22언젠가 이것이 폐쇄 형 영리 기업 으로 변할 줄 은 예상 하지 못했습니다
  160. 8:28.
  161. 8:29그리고
  162. 8:31심각한
  163. 8:33소유권 에 대해 당신 은 아니라고 말했고, 당신 은
  164. 8:38Well, I was confused as to how you could
  165. 8:40have a commercial interest in a
  166. 8:42non-profit; that just doesn't seem to
  167. 8:44make sense to me.
  168. 8:45I provided almost all the capital in the
  169. 8:48beginning for the company. It was my
  170. 8:50idea, and I named it, recruited the key
  171. 8:52people, and taught them everything I
  172. 8:55know—but apart from that...
  173. 9:00So, to go back to data centers. So you
  174. 9:02thought the idea about putting all this
  175. 9:05AI into operation was that we need data
  176. 9:07centers, and people want to build them.
  177. 9:10It ordinarily takes them about two and a
  178. 9:13half years to build one, and it costs
  179. 9:15maybe $ 10 billion to build it—$ 50 or $
  180. 9:1860 billion to build a "gigawatt."
  181. 9:21Then, the people who build them for
  182. 9:23hyperscalers and rent them out, in order
  183. 9:26to get the capital, they need to borrow
  184. 9:28it or get it in some way where they have
  185. 9:30to have long-term leases—years at least
  186. 9:33—to secure them.
  187. 9:35So, the rents people pay on five-year
  188. 9:37leases are maybe $ 1 billion a year for
  189. 9:40a $ 10 billion investment.
  190. 9:43What we do instead is we build them, and
  191. 9:46we build them because we have a use for
  192. 9:49them, because we’re vertically
  193. 9:51integrated for our cars, for our robots
  194. 9:54, for our rockets, for everything.
  195. 9:58And because we need them, we can take
  196. 10:01the risk of building them without
  197. 10:02knowing we have contracts for other
  198. 10:05people.
  199. 10:06What happens, though, is we have them
  200. 10:09built—and it takes us somewhere that
  201. 10:12used to take four months, now it takes
  202. 10:14us eight or nine months to build.
  203. 10:18It takes them two and a half years, it
  204. 10:20takes us eight months.
  205. 10:22It costs them $ 60 billion; it costs us
  206. 10:24$ 3.5 billion.
  207. 10:26And the demand for compute is such that
  208. 10:28people's centers are dark because they
  209. 10:31don't have the power to turn them on.
  210. 10:35We rent them out for compute until we
  211. 10:37need them for ourselves for Grok, and we
  212. 10:39rent it out for somewhere around $ 500
  213. 10:42million a year.
  214. 10:44So, we're getting $ 500 million a year
  215. 10:46on a $ 3.5 billion investment. The
  216. 10:49revenues for xAI are about to explode,
  217. 10:52going up—yes, they are exploding. I
  218. 10:55think, yeah, our Q2 to Q3 revenue
  219. 10:57increase was, I think, over 80%this year
  220. 11:00.
  221. 11:01I'm not sure what the exact numbers will
  222. 11:05be Q3 to Q4, but probably something like
  223. 11:0870%Q3 to Q4.
  224. 11:11And I would say maybe it's something
  225. 11:13similar Q4 to Q1.
  226. 11:16It is extraordinarily difficult to bring
  227. 11:19on compute at scale.
  228. 11:24I think there are a lot of speculators
  229. 11:27out there who think you can simply order
  230. 11:29some GPUs, hand a purchase order to some
  231. 11:32company to build the data center, and
  232. 11:34just rake in cash.
  233. 11:37The reality is, it's insanely hard.
  234. 11:40음, 제가 꽤 재미 있다고 생각 하는 '호세 밈 (Jose meme)'
  235. 11:44이라는 게 있는데, 아마 보신 적 있을 겁니다. 도로 에서 한 사람 이
  236. 11:48구덩이 를 파고 있고, 여덟 명 정도 가 그 사람 이 일하는 걸 구경 만
  237. 11:53하는 장면 이죠.
  238. 11:55그러니까 그 밈 에서는 한 명만 실제로 일 을 하고 있고, 나머지 여덟
  239. 12:00명은 그 사람 이 일하는 걸 지켜 보고 만 있는 겁니다.
  240. 12:04이제, 나라 안의 '호세' 들 수 를 늘리지 않는 한, 실제로 수행 되는
  241. 12:09일 의 양은 늘어나지 않습니다.
  242. 12:13그저 호세 에게 구매 주문서 만 발송 하는 사람들 을 더 늘린다 고 해서
  243. 12:17더 많은 일 이 일어나지 는 않습니다. 오직 호세 가 더 많아 져야만 일
  244. 12:21이 더 많이 진행 되죠. 그래서 저는 호세 가 되기를 열망 합니다.
  245. 12:25그리고 스페이스 X 도 호세 가 되기를 원합니다. 우리 는 실제로 일 을
  246. 12:29합니다. 이건 정말로 아주 중요한 문제 입니다.
  247. 12:34미국 에서 건설 부문 이 엄청나게 확장 되고 있지만, 솔직히 말해서 현장
  248. 12:38에서 물리적 인 노동 을 하는 사람들 이 크게 늘어난 것은 아닙니다.
  249. 12:43그저 구매 주문서 를 발행 하는 사람들 만 엄청나게 늘어 났을 뿐이죠.
  250. 12:48그렇기 때문에 실제로 일 을 하는 사람들 은 극도 로 바 빠질 수밖에
  251. 12:51없습니다.
  252. 12:52그래서 전국적 으로 보면 미완성 인 데이터 센터 프로젝트 들이 아주 많이
  253. 12:57쌓여 있습니다.
  254. 12:59왜냐하면 이 시점 에서는 땅콩 버터 를 너무 얇게 펴 발라서 맛 조차
  255. 13:02거의 느껴 지지 않는 수준 이 거든요.
  256. 13:06차이점 이라면 스페이스 X 는 거대한 로켓 을 제작 하고 광활한 발사
  257. 13:11단지 를 건설 하는 등, 극도 로 도전 적인 물리적 작업 을 수행 하는
  258. 13:15데 익숙 하다는 점 입니다.
  259. 13:19테슬라 또한 대규모 건설 에 매우 능숙 합니다. 아마도 중국 을 제외
  260. 13:24하고 지구상 에서 물리적 인 하드웨어 를 다루는 가장 뛰어난 두 기업 이
  261. 13:28바로 스페이스 X 와 테슬라 일 겁니다.
  262. 13:30테슬라 모델 Y 가 베스트셀러 라는 점도 주목할 만하 죠.
  263. 13:35중국 에서 말이죠. 중국 은 전통적 으로 사업 하기 매우 어려운 곳 으로
  264. 13:40알려져 있는데, 그럼에도 우리 는 그 모델 의 차량 을 판매 하고
  265. 13:45있습니다. 잠시 생각 해보세요. 그리고 말씀 드리 자면 중국 자동차 회사
  266. 13:51들은 정말 뛰어 납니다. 포드 에서 온 사람 도 중국 차 들의 품질 에
  267. 13:56완전히 압도 되었다고 하더군요.
  268. 14:01중국 은 대단 합니다. 그런데 저는 중국 을 정말 좋아 합니다.
  269. 14:05그곳에 가면 영감 을 받게 됩니다.
  270. 14:07보통 중국 에서 건물 을 짓는 사람들 은 사업 운영 의 상당 부분 을
  271. 14:11중국 측 이 소유 하게 해야 해서 다들 매우 불안해 했죠.
  272. 14:16일반적 으로 중국 에서는 외국인 이 최대 49%까지만 소유 할 수
  273. 14:19있습니다. 하지만 테슬라 의 경우, 중국 측 에서 우리 가 그곳에 있기
  274. 14:23를 너무나 간절히 원했기 때문에 100%소유권 을 허용 해 주었습니다.
  275. 14:27네.
  276. 14:28음, 중국 관리 들을 만나는 데 꽤 여러 해 가 걸렸 습니다. 정부
  277. 14:32관계자 중 일부 는 정말 저 에게 소리 를 지르기도 했죠. 침 을 튀기
  278. 14:36면서 말이죠. 우리 가 중국 파트너 와 일종의 '결혼' 을 해야 한다고
  279. 14:40강요 했는데, 저는 우리 는 이제 막 걸음마 를 뗀 회사 일 뿐이 라
  280. 14:44아기 가 결혼 하는 건 말이 안 된다고 했습니다. 제가 아기 를 안고
  281. 14:48식장 에 들어가는 모습 이 우스꽝 스러울 테니까요. 그랬더니 그들은 우리
  282. 14:52가 아기 가 아니라고 하더군요. 그래서 저는 아니라고, 우리 는 아기
  283. 14:56라고 말했죠.
  284. 14:57그리고 아기 는 아직 너무 어려서 결혼 하면 안 된다고 했습니다.
  285. 15:01결국 그들은 알 겠다고 했고, 법 을 바꿔서 저희 가 공장 을 소유 할
  286. 15:04수 있게 해주었 습니다.
  287. 15:07그리고 스페이스 X 의 매출 은 크게 증가 하고 있습니다.
  288. 15:14작년 매출 은 150 억 ~ 160 억 달러 정도 였고, 올해 는 연간
  289. 15:19매출 기준 으로 260 억 달러 정도 에 이를 것으로 예상 합니다.
  290. 15:26내년 에는 5,000 억 달러 에서 1 조 달러 사이 가 될 것으로 보고
  291. 15:31있습니다.
  292. 15:33그러니까 150 억 달러 에서 연간 1,000 억 달러 수준 으로,
  293. 15:39그리고 다시 5,000 억 달러 로 성장한 셈 이죠.
  294. 15:45몇 가지 언급 하자면, 12 월 에 연간 매출액 이 상당한 규모 에 도달
  295. 15:51할 것으로 보입니다.
  296. 15:5512 월 이 얼마 남지 않았 으니 천재 지변 같은 일 이 없다면 충분히
  297. 16:01달성 할 수 있을 겁니다.
  298. 16:06그리고 내년 계획 을 고려 하면, 내년 12 월 의 연간 반복 매출 이
  299. 16:135,000 억 달러 를 넘어 설 가능성 이 매우 높습니다.
  300. 16:20이것은 약속 이 아닙니다. 단지 제가 생각 하는 미래 의 모습 일
  301. 16:23뿐입니다. 제 최선 의 추측 이죠.
  302. 16:26무언가 를 하겠다는 약속 과 제가 생각 하는 예상 을 구분 하고 싶습니다
  303. 16:33.
  304. 16:34내년 12 월 ARR 이 5,000 억 달러 에 이를 것이라 생각 하며
  305. 16:39, 1 조 달러 도 불가능한 수치 는 아니라고 말씀 드리는 겁니다.
  306. 16:46즉, 이전 에는 불가능 했던 일 이지만 이제는 불가능한 일 이 아니라는
  307. 16:49뜻 입니다.
  308. 16:51확률 로 치면 2%정도 일지 모르지만, 중요한 건 그게 0%가 아니라는
  309. 16:56점 입니다.
  310. 16:58그게 흥미로운 부분 이죠.
  311. 17:01시간 이 지날 수록 확률 은 점점 높아지고 있습니다.
  312. 17:08My biological neural net predicts that
  313. 17:11digital superintelligence will have an
  314. 17:14effect on the global economy of about a
  315. 17:1820 to 30%increase. This is without even
  316. 17:21taking robotics into account.
  317. 17:24I think a 20 to 30%increase, which would
  318. 17:27be roughly, let's call it a $ 15
  319. 17:30trillion increase in the global economy
  320. 17:33, and this will happen very fast.
  321. 17:38Trillion dollar global economy, and you
  322. 17:41think digital AI is going to increase
  323. 17:43that by 25-30%?
  324. 17:4520 to 30%, uh, yes, that is...
  325. 17:50That is my guess.
  326. 17:51What is your guess? By how much would it
  327. 17:53increase the economy?
  328. 17:54I think probably 20-30%, um, I mean, I
  329. 17:56think it could even be higher than that
  330. 17:58.
  331. 17:59Um, but it probably doubles the
  332. 18:01productivity of white-collar work.
  333. 18:04So that is just digital. That is not
  334. 18:06physical, that is digital. Digital AI.
  335. 18:09Yes.
  336. 18:10Well, the way to think about physical AI
  337. 18:13in the form of robotics is that, um, any
  338. 18:15given economy is the number of people in
  339. 18:18it times the average productivity per
  340. 18:21person.
  341. 18:23So if you have a humanoid robot that is
  342. 18:25extremely productive and can work 24/7,
  343. 18:27so it is working 168 hours a week...um,
  344. 18:30it doesn't take vacations or bathroom
  345. 18:32breaks or need to eat food or anything
  346. 18:34like that.
  347. 18:36And um, the speed at which it can
  348. 18:39operate is faster than human speed.
  349. 18:43You can expect roughly a five, you know
  350. 18:46, a productivity per week that is, uh,
  351. 18:48on average 5 times better than a person
  352. 18:51, roughly, because a person is maybe
  353. 18:53working 40 hours a week and doesn't need
  354. 18:56to take breaks, obviously, um, and
  355. 18:58doesn't move as fast as a robot. So, I
  356. 19:01would say a minimum productivity
  357. 19:02expectation for a humanoid robot would
  358. 19:04be five times that of a human.
  359. 19:07Then, um, you know, take the population
  360. 19:09of the United States; say, how many
  361. 19:11people are either in school or retired?
  362. 19:13Well, that is about half the people,
  363. 19:16about half of people are either in
  364. 19:18school or retired, and then of those who
  365. 19:20are working age, not everyone works, um
  366. 19:23, or is engaged in highly productive
  367. 19:25employment. Um, so then you are left
  368. 19:28with, you know, perhaps, um, around, I
  369. 19:31don't know, 120 million people in
  370. 19:33America that have a net, uh, positive
  371. 19:36output of goods and services.
  372. 19:39Um, so then you say, well, how many
  373. 19:41robots then would be needed to match
  374. 19:43that productivity? Well, I guess you
  375. 19:46would need, um, you know, about 20 to 30
  376. 19:49billion robots to match the goods and
  377. 19:52services output of the United States.
  378. 19:59So then, moving them to data centers and
  379. 20:01space is the obvious solution.
  380. 20:04So people don't have to worry about them
  381. 20:05in their communities anymore, if they
  382. 20:07want, or if they don't want to.
  383. 20:10And for us, it's less expensive to put
  384. 20:12them in space, and every time we launch
  385. 20:14new generations, they become more and
  386. 20:17more powerful than they were before.
  387. 20:20Isn't that true?
  388. 20:21Everyone says it’s 10 years away, but
  389. 20:23we’re already starting.
  390. 20:26Google has already tested this with a
  391. 20:28satellite we operate for them; it’s been
  392. 20:30effective, and we’re going to start
  393. 20:32launching in '27 or '28.
  394. 20:35How long until it makes an impact?
  395. 20:39Yes, currently people have tested Nvidia
  396. 20:43GPUs on a satellite we launched, and
  397. 20:46there’s a satellite we recently launched
  398. 20:51with a Google TPU, and these are
  399. 20:54operating and working just fine in space
  400. 20:58.
  401. 20:59We expect to start launching the first
  402. 21:01Starlink computers, which we’ve designed
  403. 21:04with Nvidia, probably—I’m just
  404. 21:07estimating here—in the third quarter of
  405. 21:10next year.
  406. 21:11We’ll start off slow, but then we’ll
  407. 21:14ramp up to significant volume, probably
  408. 21:16in '28, and then very, very large volume
  409. 21:19in '29.
  410. 21:21And to really reach the
  411. 21:23terawatt-per-year scale, or even say the
  412. 21:27200-gigawatt-per-year scale, you really
  413. 21:30need to go to space because there just
  414. 21:33isn't enough power on the ground to
  415. 21:36realistically reach those levels every
  416. 21:39year.
  417. 21:42Just as a frame of reference, the
  418. 21:45average power consumption in the United
  419. 21:47States is just over 500 gigawatts.
  420. 21:52So if you look at it, the peak power
  421. 21:55capacity is roughly 1.2 terawatts, but
  422. 21:58average power consumption is just over
  423. 22:01500 gigawatts. You can really see that
  424. 22:04if you're doing, say, 200 gigawatts a
  425. 22:07year of AI, that would be like a 40%
  426. 22:10increase in the entire US power grid
  427. 22:13every year.
  428. 22:16Which is simply not feasible.
  429. 22:20When you take solar panels and put them
  430. 22:23in space, because it is essentially
  431. 22:26always sunny, the solar panels can
  432. 22:29always face the sun directly, not at an
  433. 22:32off-angle.
  434. 22:37Also, you don't have atmospheric
  435. 22:39attenuation of solar power, so you get
  436. 22:4130%more solar power in space.
  437. 22:45It's sunny all the time, there’s no
  438. 22:47seasonality, and you don’t really need a
  439. 22:50large battery backup because you don't
  440. 22:53need to last through the night.
  441. 22:57The cost of intelligence in space will
  442. 23:00be dramatically lower than on the ground
  443. 23:03once you have a low cost per ton;
  444. 23:05Starship, being the first reusable
  445. 23:08rocket, will lower the cost of access to
  446. 23:11space by a factor of 100 or more.
  447. 23:16We want to be highly confident in the
  448. 23:19reliability of the ship, and I think we
  449. 23:22are finally there where we feel it is
  450. 23:25worth it.
  451. 23:28with ship 42, um, which hopefully is
  452. 23:30prophetic because you know 42 is kind of
  453. 23:33a special number, uh, that we will, we
  454. 23:36can try to catch it with the tower. So
  455. 23:38that's what we're going to do in about a
  456. 23:41week from now.
  457. 23:44Rooting for it.
  458. 23:47Yeah. I mean Starlink revenue is growing
  459. 23:49very rapidly by any sort of normal
  460. 23:50business standards. You know, like
  461. 23:52you're roughly doubling revenues per
  462. 23:55year, um, at large scale, and you know,
  463. 23:57I remember when that used to be a big
  464. 24:00deal.
  465. 24:01Um, it wasn't that long ago, uh, that,
  466. 24:03you know, doubling revenues with high
  467. 24:06gross margin and positive cash flow
  468. 24:08every year would normally be an
  469. 24:10impressive thing.
  470. 24:12Um, although I have to say with AI it
  471. 24:15does exceed that.
  472. 24:18So, uh, with AI you have order of
  473. 24:20magnitude per year growth potential.
  474. 24:23So,
  475. 24:24um,
  476. 24:28and, uh, I mean, if you think about,
  477. 24:30like, things I just said, 'cause I say a
  478. 24:32bunch of things I've said sound
  479. 24:34extremely crazy, uh, but I think we'll
  480. 24:36get there.
  481. 24:37You know, sometimes I'm a little
  482. 24:39optimistic on the time frame, but I
  483. 24:40always deliver in the end.
  484. 24:42Um, so,
  485. 24:46you know, we have 11,000 satellites up
  486. 24:49there right now. Um, primarily the
  487. 24:52version 2 satellite. We've just started
  488. 24:54launching the version 3 satellite.
  489. 24:55Version 3 satellite has over 10 times
  490. 24:57the capability of version 2 satellite
  491. 24:59and we'll be launching 10 times as many
  492. 25:01of them. So the amount of bandwidth that
  493. 25:04Starlink will produce, um, in I don't
  494. 25:06know, two, two or three years, will be
  495. 25:08100 times more bandwidth than it
  496. 25:10currently offers.
  497. 25:12And, um,
  498. 25:14yeah, so I, I think you said, like,
  499. 25:16what's the revenue per, you know,
  500. 25:19gigabit per second type of thing?
  501. 25:21It's going to be, it's going to be high
  502. 25:23. I mean, and then you say, how many
  503. 25:25gigabits are you transferring? It's a
  504. 25:27very big number.
  505. 25:28So, and, and with massive demand, you
  506. 25:31know, by, by AI and robots, far, far
  507. 25:33exceed human demand by, like, literally
  508. 25:361,000. Well,
  509. 25:39um, should we pass that question?
  510. 25:46I think SpaceX will be able to get very
  511. 25:47far with the spectrum that it's acquired
  512. 25:49.
  513. 25:50Uh, but, you know, at a certain point,
  514. 25:52if we're able to make much better use of
  515. 25:55the spectrum than, um, than say a legacy
  516. 25:57telco is, we could just acquire that
  517. 25:59legacy telco and use that spectrum, you
  518. 26:02know, 10 times more effectively.
  519. 26:08Um,
  520. 26:11but, um,
  521. 26:14yeah, I, I think there's, in terms of,
  522. 26:17uh, being of competing with China, uh,
  523. 26:19you know, from like an electricity
  524. 26:22generation standpoint and, um, just
  525. 26:25overall sort of how much digital
  526. 26:27intelligence have you harnessed for
  527. 26:30economic activity, I think the only way
  528. 26:33to be competitive with China is to take
  529. 26:36the intelligence to space.
  530. 26:39Um, so there we can sort of bring on a
  531. 26:42few hundred gigawatts a year, um, and
  532. 26:46potentially, you know, a terawatt a year
  533. 26:49.
  534. 26:50Um, so if if we are able to do a
  535. 26:53terawatt a year of intelligence to space
  536. 26:55. So, two more questions.
  537. 26:58Uh, why do we need so much power? Why do
  538. 27:01we need so many chips? Why? What is the
  539. 27:04demand that’s going to come for that
  540. 27:08intelligence, for that power, and for
  541. 27:12those chips?
  542. 27:16The demand will be...
  543. 27:19This again sounds somewhat circular, but
  544. 27:21the demand for intelligence, for digital
  545. 27:24superintelligence, will come from
  546. 27:26digital superintelligence.
  547. 27:30Which is for medical purposes—I mean,
  548. 27:33what does that mean?
  549. 27:38Um, it means that superintelligence will
  550. 27:41saturate all human needs by far; uh,
  551. 27:43there will be universal high income,
  552. 27:45universal extremely high income.
  553. 27:49Um, universal medical care that is
  554. 27:51better than the best medical care
  555. 27:52available on Earth today.
  556. 27:58Where you could wish for a castle or
  557. 28:02whatever you want, basically, and the
  558. 28:05digital superintelligence can make it
  559. 28:07for you as a side project.
  560. 28:10Um, so well, I guess I’d like—I would
  561. 28:13like—there’s perception and reality; I
  562. 28:16mean, I would like the reality, whatever
  563. 28:19the perception is, but like the reality
  564. 28:22to be that I made...
  565. 28:26that...
  566. 28:34If you enjoyed this content, please
  567. 28:36subscribe to our YouTube channel and
  568. 28:37newsletter right now. Thank you.

About this transcript

This page contains the full transcript of "현실은 미친 듯이 어렵다" 일론 머스크가 말하는 AI 버블의 진실 | 2026 바론 투자 컨퍼런스 by 머니네버슬립, generated from the public captions YouTube serves with the video. The transcript has 3,970 words across 568 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.