YouTube2Text

Sam Altman: ‘It’s Hard To Imagine the AI Revolution Without Jensen’ | Silicon Valley & The World — Transcript

by Semafor · 5,324 words · 766 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Um, great room. Sam, thank you for
  2. 0:02coming.
  3. 0:02>> Thank you for having me.
  4. 0:03>> Um, so just curious, did you catch
  5. 0:05Saturday Night Live this weekend?
  6. 0:06>> I just did.
  7. 0:08What did you think?
  8. 0:09>> Uh, it like a little funny. I love
  9. 0:13Saturday Night Live.
  10. 0:14>> A little funny. Okay. Well, so that was
  11. 0:17about the Jeden Wong uh scandal. Um, you
  12. 0:20a few years ago, almost exactly three
  13. 0:22years ago, uh, had a situation where you
  14. 0:24were also ousted by the board. Uh there
  15. 0:27was a letter going around about Jensen
  16. 0:29that was I think copy pasted from the
  17. 0:30one that your supporters drafted a few
  18. 0:33years ago. Um so I'm curious, do you
  19. 0:36think he should do you think Jensen
  20. 0:38should come back?
  21. 0:40>> Um
  22. 0:42I don't think I know enough to answer
  23. 0:44that question. I Jensen has long been a
  24. 0:47hero of mine. Heroes have flaws. I am
  25. 0:50grateful for the company he's built. Um
  26. 0:53I
  27. 0:55I do have like personal strong empathy
  28. 0:58to a situation like this. I remember uh
  29. 1:02what it was like although it feels like
  30. 1:03forever ago. Um I
  31. 1:06but it's hard to imagine the AI
  32. 1:08revolution without Jensen. Um and the
  33. 1:12company he has built took an early bet
  34. 1:15on us. Uh we took an early bet on him. I
  35. 1:17have extreme extreme gratitude for that.
  36. 1:21They also guaranteed I think $ 105
  37. 1:24billion for your data center in Ohio.
  38. 1:27They're also, you know, they're the
  39. 1:28bank, they're sort of backing the entire
  40. 1:30industry. Do you think that guarantee is
  41. 1:32worth as much today as it was, you know,
  42. 1:35a couple weeks ago.
  43. 1:37Um,
  44. 1:39I mean, Nvidia is like a behemoth of a
  45. 1:42company. I think people still trust
  46. 1:44their guarantees very much. But but
  47. 1:45yeah, I mean Nvidia without Jensen like
  48. 1:47Jensen and Nvidia clearly very deeply
  49. 1:49associated uh and
  50. 1:54I I think the ambition he has brought
  51. 1:56and the scale has has been quite
  52. 1:58important to the industry.
  53. 1:59>> So you think there there will be a
  54. 2:01ripple effect there'll be an impact on
  55. 2:03the industry. Um,
  56. 2:05look at this point I think there's so
  57. 2:07much demand in and it's so clear that
  58. 2:11like AI is real and people want to use
  59. 2:13it and the revenues and demand are
  60. 2:14growing so strongly uh that I think the
  61. 2:18industry will be fine uh but again I
  62. 2:21think Jensen has played this
  63. 2:24extremely special and valuable role as
  64. 2:27someone to really drive things forward
  65. 2:28with great ambition. So anthropic going
  66. 2:32public this week. Um in a time when I
  67. 2:34think there's there's a lot of
  68. 2:35skepticism from investors. Um curious,
  69. 2:39do you actually want to go public? Like
  70. 2:42if you could just avoid it, would you?
  71. 2:45>> I think there's a lot of skepticism from
  72. 2:46journalists. Um investors seem to be
  73. 2:49pretty bullish on AI. I uh things will
  74. 2:53go up and down. Um prices will go up and
  75. 2:55down. You know, there will be periods of
  76. 2:57overbuild and underbuild. Um there will
  77. 3:00be periods where something goes wrong or
  78. 3:03something goes really right. But the the
  79. 3:05overall trend is that huge value is
  80. 3:09being created. Um huge value is being
  81. 3:12disrupted too. And I think investors
  82. 3:15seem I just as we were backstage
  83. 3:17listening to the amount of capital being
  84. 3:19invested. Investors seem to realize
  85. 3:21that. Um do I want to go public? We we
  86. 3:24will go public at some point. I
  87. 3:27I think there are real downsides to
  88. 3:29being public. Uh
  89. 3:33and it's a very volatile time in our
  90. 3:35industry with very high stakes
  91. 3:36consequential decisions to make around
  92. 3:40safety and how society's going to
  93. 3:41interact with this technology. Um and
  94. 3:43there are doubtless things that are
  95. 3:44easier as a private company. On the
  96. 3:47other hand, uh I think it's not just
  97. 3:49like investors that h happen to have
  98. 3:51access to private markets that want to
  99. 3:54invest. the world as a whole would like
  100. 3:56exposure to this and I think they're
  101. 3:57that's you know directionally a good
  102. 3:59thing to do.
  103. 3:59>> Yeah. Do you think if so Anthropic is
  104. 4:02going to go first? Uh that's clear. Once
  105. 4:05you go public, is there enough, you
  106. 4:07know, do you think there's a just enough
  107. 4:09investor demand left or are you going to
  108. 4:11be sort of the lift to their Uber, do
  109. 4:12you think? Is that a worry?
  110. 4:14>> That is not a worry.
  111. 4:17And again, I I get that you have to like
  112. 4:18write something every day and there's
  113. 4:20got to be like, oh, this, you know, this
  114. 4:22fight and that one and is this winner
  115. 4:24and this loser and what's happening here
  116. 4:25and there, but like the macro trend and
  117. 4:28the value being created. Um, I am not I
  118. 4:32am not at all worried about there not
  119. 4:33being a huge amount. In fact, you know,
  120. 4:36one way you could look at the last year
  121. 4:37is there's been way too much investor
  122. 4:40demand and not enough supply. Um but you
  123. 4:43know I I think we'll be able to continue
  124. 4:44to build we the whole industry just
  125. 4:47incredible incredible value.
  126. 4:49>> Yeah. Do you So you made these these
  127. 4:51comments behind closed doors it was
  128. 4:53reported to Samsung executives where you
  129. 4:55said there's been some double counting
  130. 4:57in the industry. What did you mean by
  131. 4:59that? Can you explain?
  132. 5:00>> Sure. uh when
  133. 5:04for a long time uh there was basically
  134. 5:07one chip provider in the industry and uh
  135. 5:11so when people were counting wafers they
  136. 5:14could ask companies like ours what
  137. 5:16demand was going to be um and they could
  138. 5:18I think get a sense of where end user
  139. 5:21demand for the chips they could make
  140. 5:23would be um there are now many people
  141. 5:25making chips and many of those people
  142. 5:28think they're going to sell most of the
  143. 5:30volume to us anthropic and others. Um,
  144. 5:33and we have, you know, some limited
  145. 5:35amount of demand and data centers and
  146. 5:38everything else. And I think there are
  147. 5:39many people making chips all talking to
  148. 5:42the silicon supply chain. Uh, assuming
  149. 5:44that we're going to buy all of their
  150. 5:46chips. And, uh, I think the demand, the
  151. 5:51data center buildout, that stuff is
  152. 5:53solid. There are,
  153. 5:55but we're not we're not all going to be
  154. 5:57able to or need to buy every, you know,
  155. 5:59all of our supply from every chip
  156. 6:00vendor.
  157. 6:01>> What does that mean for the industry?
  158. 6:02Are people is it over, you know, are
  159. 6:05things overpriced? What do you think?
  160. 6:07>> Um,
  161. 6:09again, I'm sure there will be ups and
  162. 6:10downs. The long-term demand for huge
  163. 6:14amounts of compute. Um, and and the
  164. 6:17value that that unlocks, just the
  165. 6:18incredible stuff that people are now
  166. 6:20doing at at real scale with this
  167. 6:22technology, I think that will keep
  168. 6:24going. Um and you know like any other
  169. 6:29like any other tech revolution um the
  170. 6:31long-term trend is clear the forecast
  171. 6:33any quarter to quarter and year to year
  172. 6:36of supply and demand exactly matching.
  173. 6:37Harder to say it's interesting you say
  174. 6:39you know they could count on us buying
  175. 6:41all the chips. I thought you did want to
  176. 6:42buy all the chips. Um but do you think
  177. 6:45like one one way of looking at this
  178. 6:46whole Nvidia thing is now you can.
  179. 6:49They're not shipping to China anymore.
  180. 6:51that seems to be
  181. 6:51>> um
  182. 6:52>> the supply chain is very complicated.
  183. 6:55We've been constrained by different
  184. 6:56things at different times. Sometimes
  185. 6:58it's been memory, sometimes it's been
  186. 6:59rack assembly, sometimes it's been, you
  187. 7:01know, data center capacity. Um sometimes
  188. 7:06it's been some random component like
  189. 7:08you're like really we can't make you
  190. 7:10know we can't sell more services because
  191. 7:12the world is out of like PCB laminate or
  192. 7:14whatever. Um, so it's it's just a very
  193. 7:17complex supply chain um at now at huge
  194. 7:20huge scale.
  195. 7:21>> Yeah, we also just had this blue wave
  196. 7:23that was the other news before the the
  197. 7:25Wong stuff. Um I think it was all but
  198. 7:28one of these data center referendum
  199. 7:30bills passed. In other words, people
  200. 7:33hate the data centers. Um a lot of voter
  201. 7:36sentiment on both sides of the aisle
  202. 7:37against AI. I'm just curious like how do
  203. 7:39you think the industry screwed this up
  204. 7:41so badly in telling the story about this
  205. 7:43technology?
  206. 7:46I I think if the sales pitch is like we
  207. 7:48are going to offer you a cure to cancer
  208. 7:51and cheap stuff and the price of that is
  209. 7:54you are going to lose all agency and
  210. 7:56control and voice in the future. You are
  211. 7:58going to live like a Labrador. Maybe the
  212. 8:01machines will try to make you happy, but
  213. 8:02you certainly won't have any ability to
  214. 8:04do that. and we're going to like rush to
  215. 8:05hand over um you know control of the
  216. 8:09future to AI and we're going to like
  217. 8:12make our make our AI really nice uh and
  218. 8:14he's going to take good care of you but
  219. 8:15that's going to be that. I think it's a
  220. 8:17terrible sales pitch. I don't want that
  221. 8:19world. Um I and I don't think most
  222. 8:22people want that world. There's other
  223. 8:24big questions about the economics too.
  224. 8:26But I I think fundamentally if you talk
  225. 8:29to people who really don't like AI and
  226. 8:31you get past the initial things. I think
  227. 8:34data centers are easy to point to
  228. 8:35because they're like a big physical
  229. 8:37instantiation of this broader thing. I
  230. 8:39think the job stuff it's it's a real
  231. 8:42immediate term worry that people feel
  232. 8:44very deeply. But there is this like
  233. 8:46long-term question about what are
  234. 8:47people's role going to be in society and
  235. 8:51what you know what is the balance of
  236. 8:53power going to look like either between
  237. 8:55people and AI or between people that
  238. 8:57have a lot of AI and a lot of computing
  239. 8:59power to use AI versus people who don't.
  240. 9:02Um and I think this is a place where the
  241. 9:05AI industry needs to take a very strong
  242. 9:06stance which is we are
  243. 9:09aggressively in favor of and committed
  244. 9:11to people staying in control of the
  245. 9:13future. um we should not want uh to have
  246. 9:17a handover, I think that's the current
  247. 9:19term people are using in Silicon Valley.
  248. 9:21Uh either gradual or sudden to AI uh you
  249. 9:23know making decisions about our future.
  250. 9:25This should be a tool that people get to
  251. 9:26use and
  252. 9:29I think people can kind of tell that not
  253. 9:30everybody in the AI industry thinks that
  254. 9:32way and are appropriately terrified. uh
  255. 9:34and that combined with the economic
  256. 9:36uncertainty and you know various ways
  257. 9:38that AI has you know caused cyber issues
  258. 9:41or whatever um has people against data
  259. 9:45centers but I think from the
  260. 9:48conversations I've had that what they're
  261. 9:49really against is this feeling that man
  262. 9:52are the people building AI actually
  263. 9:55really truly on team human. You said
  264. 9:57back in uh July 2025 you made this
  265. 10:00comment about Democrats uh you're
  266. 10:02politically homeless. Democrats have
  267. 10:04lost the culture of innovation. Do you
  268. 10:06still feel that way?
  269. 10:09>> Um
  270. 10:11roughly speaking, you know.
  271. 10:13>> How does that affect your view on the
  272. 10:15the the results, you know, what we're
  273. 10:18seeing now? I mean, obviously I will I
  274. 10:21want to work with whoever our elected
  275. 10:23leaders are. Do I still think there is
  276. 10:24room for like a party that more closely
  277. 10:28matches
  278. 10:30the ideals that I think have worked for
  279. 10:32America and the spirit of innovation for
  280. 10:34a long time. Um, I think that'd be
  281. 10:37great. I hope that both parties will
  282. 10:39move in that direction and I think there
  283. 10:40are like positive signs in both
  284. 10:43directions. Um,
  285. 10:46but yeah, you you said you want to work
  286. 10:48with the administration. you've, you
  287. 10:50know, before. I mean, you could call up
  288. 10:51David Saxs, you could call up Susie
  289. 10:53Wilds, can you call up Steven Miller?
  290. 10:56>> I don't know. I'll try.
  291. 10:59Have you, Do you still, you know, just
  292. 11:01curious like when was the last time you
  293. 11:03talked with Trump?
  294. 11:05>> Um,
  295. 11:08the China State Dinner.
  296. 11:10>> Yeah. And And what was the discussion
  297. 11:13then? What did you guys talk about? I
  298. 11:14>> mean, you don't get to have like a deep
  299. 11:16in-depth conversation. That's a lot of
  300. 11:19stuff. But yeah.
  301. 11:20>> So, you know, I can't help but think
  302. 11:22about you had your own spy drama, right,
  303. 11:24with this, you know, person putting a
  304. 11:26hidden uh a hidden microphone in your
  305. 11:29office. I think you've talked to
  306. 11:31employees about the the danger of, you
  307. 11:33know, people pretending to hit on them
  308. 11:36and potentially being agents of a
  309. 11:38foreign government. Um, you know, I'm
  310. 11:40just curious like has this changed how
  311. 11:42you interact with you have a lot of
  312. 11:44investors from the Gulf. Do you view
  313. 11:46them differently because of all this
  314. 11:48stuff?
  315. 11:50I
  316. 11:52we have now had I think more practice
  317. 11:55than is maybe obvious from the outside
  318. 11:57about living in this world where it is
  319. 12:00clear that AI is an extremely important
  320. 12:03thing happening and that people are
  321. 12:05looking for an edge on in all sorts of
  322. 12:08ways whether it's you know putting a
  323. 12:11camera in one of our conference rooms or
  324. 12:13trying to take our engineers out for a
  325. 12:15few drinks and get them to tell
  326. 12:16pre-training secrets um or whatever.
  327. 12:18ever. So,
  328. 12:21I think we have been living in this
  329. 12:24reality that everybody else is now kind
  330. 12:26of very publicly and suddenly coming to
  331. 12:28grips with. Um, and
  332. 12:32I'm sure we've been tricked sometimes
  333. 12:34and I'm sure that our operational
  334. 12:37security and overall security has not
  335. 12:39been perfect, but we have had time to
  336. 12:41adjust to this. the the a thing that we
  337. 12:44talk a lot about uh when it comes to
  338. 12:46safety in which I put this bucket of of
  339. 12:48stuff like the I put security in the
  340. 12:50broader bucket is being pragmatic
  341. 12:52centrists. Um it it totally does not
  342. 12:56work to I think say we're going to just
  343. 12:59build AI in absolute secret in a lab.
  344. 13:02We're not going to ever trust anybody
  345. 13:03else with it. We're not going to bring
  346. 13:04the world along. We're not going to
  347. 13:06iteratively deploy this technology and
  348. 13:07get the world to, you know, be able to
  349. 13:10use it. Um, but also we're not going to
  350. 13:12pretend like this is like any other
  351. 13:14technology. And there are no moments
  352. 13:15where we have to do things maybe
  353. 13:18differently than any kind of company or
  354. 13:19any industry has come before. And that
  355. 13:22can range from things like being very
  356. 13:24careful in our external interactions to
  357. 13:26saying, you know, at some point this
  358. 13:28technology may get so powerful so
  359. 13:30quickly that we really have to change
  360. 13:31all of our our plans uh to be able to
  361. 13:33continue to make the safety guarantees
  362. 13:35we want to make. And um that thing we
  363. 13:38talked about earlier I think this
  364. 13:39principle of people staying always
  365. 13:40standing control of the future uh that's
  366. 13:43going to require some very different it
  367. 13:45may require some very different actions
  368. 13:47as these models get so capable. People
  369. 13:49talk a lot about recursive
  370. 13:50self-improvement. Easy to say those
  371. 13:52words. I don't think most people like
  372. 13:53really have wrapped their head around
  373. 13:55what that would mean and what we might
  374. 13:56have to do. Um so I think for all of our
  375. 14:00interactions we try to like have
  376. 14:04the weight of what is happening and what
  377. 14:06we are doing on our mind in every way
  378. 14:09and then act as these pragmatic
  379. 14:11centrists where you have to kind of like
  380. 14:14engage with the messy reality of the
  381. 14:15world.
  382. 14:16>> Since you mentioned RSI are you there
  383. 14:17yet? Is that happening? Um
  384. 14:22I think like many other things including
  385. 14:24the term AGI it is like very hard to
  386. 14:27have a clean demarcation and say now we
  387. 14:29didn't have AGI now we do. Um are we
  388. 14:31developing
  389. 14:34new models faster than we would without
  390. 14:36using our technology? Yes. Are we
  391. 14:39pushing a button and just saying hey
  392. 14:40model like
  393. 14:43design the next version and rapidly
  394. 14:45self-improve with no human oversight?
  395. 14:46Definitely not. um we could be pushing
  396. 14:49more towards the RSI spectrum than we
  397. 14:51are, but I don't think we could safely
  398. 14:52do that right now. And again, I'm not
  399. 14:58I I think this like question of if we
  400. 15:00make all decisions
  401. 15:02first from a principle of saying people
  402. 15:04need to be in control of the future
  403. 15:06forever. There are going to be a lot of
  404. 15:08follow-on implications from that for
  405. 15:10things like how quickly someone should
  406. 15:12be willing to push onto the R.
  407. 15:14>> So, you're going to hold back on RSI a
  408. 15:15little bit because of that. I I think
  409. 15:18that holding back on RSI is a tactic. Uh
  410. 15:22I think RSI can be a good thing.
  411. 15:23Certainly using this research, using
  412. 15:25this technology to do better and faster
  413. 15:27alignment research is a great thing. Um
  414. 15:30but I think you should model us as we
  415. 15:31will make our decisions um from a place
  416. 15:34of being unwavering on people needing to
  417. 15:38stay in control of the future and
  418. 15:41sort of this pragmatic centrist view of
  419. 15:43safety and interaction with the world.
  420. 15:45Meter uh had this report I'm sure you
  421. 15:47saw about Anthropic saying that they
  422. 15:49they've doubled the speed of their
  423. 15:51capability development on specific tasks
  424. 15:53compared to like a couple years ago. Are
  425. 15:55you also seeing that rate of increase?
  426. 15:59Um
  427. 16:00I don't know of a measurement like quite
  428. 16:02like that. Uh and and I think it'd be
  429. 16:04like it it is hard this these things are
  430. 16:08still jagged enough that it's hard to
  431. 16:09say like overall we're doing this whole
  432. 16:11class of things two times as fast. I'm
  433. 16:13sure you could find some things we're
  434. 16:14doing much faster than that, some things
  435. 16:16that are much slower. The blended
  436. 16:17numbers I think are always harder than
  437. 16:19they seem. Um, we are speaking of Meter,
  438. 16:22uh, we are working with embedded
  439. 16:24evaluators now, uh, including meter. And
  440. 16:28I think that this principle of having
  441. 16:31people
  442. 16:33getting like an outside opinion on the
  443. 16:35safety work and the sort of robustness
  444. 16:37of these systems is going to be
  445. 16:40increasingly important and valuable.
  446. 16:42So, let's talk about products. I use
  447. 16:44dots. I find it incredibly useful. Um, I
  448. 16:48was looking at the terms of service and
  449. 16:49I was sort of wondering like where's
  450. 16:51where's my data? Like where's this
  451. 16:52memory being stored and it's somewhere
  452. 16:54on the cloud. Um, currently I can't
  453. 16:57change it or you know edit it in any
  454. 17:00way. Um, and I'm just curious like it
  455. 17:03worries a little bit worries me a little
  456. 17:05bit about like lock in. would you allow
  457. 17:07people to sort of have a mechanism to
  458. 17:09export their dots uh memory to other
  459. 17:13>> so we do for chipt memory um and that's
  460. 17:16like an easier structure the the
  461. 17:20I think in principle yes people should
  462. 17:23be able to take their data with them the
  463. 17:25dots version of this is harder than the
  464. 17:30chatebt version of this like dots really
  465. 17:33can build up like a quite sophisticated
  466. 17:36understanding
  467. 17:37of your life. Um, so I don't have like a
  468. 17:40ready-made answer. I'll say I I agree
  469. 17:42with the principle in general of like
  470. 17:43people people should have their data and
  471. 17:45be able to do what they want with it.
  472. 17:46But it do you think people will be sort
  473. 17:48of worried about that? It's like a when
  474. 17:50you talk about this this thing living in
  475. 17:52the cloud that's like got a full
  476. 17:53understanding of my life. It just seems
  477. 17:55like I want to sort of own that. I don't
  478. 17:58know. I don't want that to be out in the
  479. 18:00cloud and I have no control over it. I
  480. 18:02can't even see it. Well, I think you
  481. 18:03want some way to be able to see it, to
  482. 18:06modify it, and to take it if you want to
  483. 18:08take it. Uh, exactly what that look
  484. 18:09like. It's easy to think about, you
  485. 18:12know, and someone said to me recently, I
  486. 18:14thought it was a good insight that the
  487. 18:15biggest lock in the entire tech industry
  488. 18:17is your um your iPhone camera roll and
  489. 18:20your eye messages. you just like that is
  490. 18:23over now 20 years of our lives and it's
  491. 18:25there and you really like you don't
  492. 18:26think about it because you can kind of
  493. 18:28sit there and look at it and probably
  494. 18:30theoretically you could export it but
  495. 18:31you're not going to. Um and
  496. 18:35but that's at least like you can just
  497. 18:36say like I have 100,000 photos I'm going
  498. 18:38to move them somewhere else feels
  499. 18:40theoretically plausible. Um, if there is
  500. 18:42just this like
  501. 18:44subtle understanding in the same way
  502. 18:46that if you have like a really great
  503. 18:48assistant that understands your life and
  504. 18:50your preferences and that like oh he
  505. 18:52normally wants this but in this weird
  506. 18:53circumstance he wants this. Those are
  507. 18:54like hard rules to write down. I mean
  508. 18:57maybe maybe dots can do it in a way that
  509. 18:59person that people can't. Um and again I
  510. 19:02agree with you on the principle. I just
  511. 19:04like haven't thought before this about
  512. 19:05like what the actual button in the
  513. 19:07product version of this looks like.
  514. 19:08>> Well, you never want to fire that
  515. 19:09assistant. That's for sure. Right. So
  516. 19:11that's yeah
  517. 19:13>> good good customer lock in there. Um
  518. 19:15okay so speaking of consumers and
  519. 19:18everything you got in some hot water
  520. 19:20talking about computer RAM uh the other
  521. 19:22day. Um you know you said I think it's a
  522. 19:26bad allocation of resources for people
  523. 19:27to have all this RAM on their home
  524. 19:29computers. Um you know people feel like
  525. 19:32I should be able to tinker with this
  526. 19:33stuff. I should be able to use it on my
  527. 19:36you know run models on locally etc. um
  528. 19:38that was a big part of like the PC
  529. 19:41internet revolution was like these home
  530. 19:44tinkerers who do cool stuff in their
  531. 19:46home computers. Do you think that we're
  532. 19:48like we're missing that in this era of
  533. 19:50cloud everything in the cloud? I
  534. 19:51>> I want to be clear. I want people to
  535. 19:53have a lot of RAM. I think everybody
  536. 19:55should buy a lot of RAM, be able to buy
  537. 19:56a lot of RAM for their own computer. I
  538. 19:58think this is good. I think the home
  539. 20:00tinkering is great. Um unfortunately a
  540. 20:03few years ago we started saying like the
  541. 20:05world needs to build a lot more
  542. 20:08memory supply chain in particular. We
  543. 20:10said the whole chip supply chain, logic,
  544. 20:12memory, packaging, whole thing. We said
  545. 20:14in particular there was going to be a
  546. 20:15huge squeeze on memory coming and you
  547. 20:18know like someone quoted me as saying
  548. 20:21the world needs to invest $7 trillion in
  549. 20:23capex and for like I was in much more
  550. 20:27hot water for that than the RAM comment.
  551. 20:29Everyone was saying this is ridiculous.
  552. 20:31This is impossible. This is insane. The
  553. 20:32world will never do this. This is
  554. 20:34ludicrous. this is, you know, there's no
  555. 20:36there's no way there's this much demand
  556. 20:38in the world and the world did not build
  557. 20:40out that much and now we have seen DRAM
  558. 20:43prices. I don't know exactly but maybe
  559. 20:45they went up like 5x 7x something like
  560. 20:47that. Um I think it would be much better
  561. 20:49if the world just had a lot of memory.
  562. 20:54um that comment which I think got
  563. 20:57misinterpreted but there what there is
  564. 21:00like a true part of it which I will
  565. 21:01stand by is that if we're going to be in
  566. 21:03this memory supply shortage then in some
  567. 21:06sense whether or not we'll decide to do
  568. 21:08it because I think there are other
  569. 21:09trade-offs like people having their own
  570. 21:11you know at home inference the optimal
  571. 21:14thing would be can we share this memory
  572. 21:16and have it working 100% of the time
  573. 21:18rather than like sometimes off in
  574. 21:20someone's edge device uh again optimal
  575. 21:22use of this limited supply maybe is not
  576. 21:25the reason to make the decision but
  577. 21:27that's what I was trying to say
  578. 21:28>> what what's the argument I mean we have
  579. 21:29like you know the CEO of ASML here right
  580. 21:31like what's your argument when you meet
  581. 21:33with these people like because they're
  582. 21:34saying I can't do I can't increase
  583. 21:37production really
  584. 21:38>> uh no I think now people are increasing
  585. 21:41production like I I think we will see
  586. 21:43production get ramped up a lot it just
  587. 21:47it would have been good to do it a few
  588. 21:48years earlier
  589. 21:49>> do you think they did sort of drag their
  590. 21:50heels on that they didn't they didn't
  591. 21:52believe that the this was going to
  592. 21:54happen. The way that someone explained
  593. 21:56this to me is that being in the memory
  594. 21:58business has been approximately the
  595. 22:01following story for the last several
  596. 22:02decades, which is you invest $20 billion
  597. 22:05and you make $19 billion
  598. 22:08and that's it.
  599. 22:12And so and like you know there's booms
  600. 22:14and busts and some years you make money
  601. 22:16but on the whole um it has not been a
  602. 22:19great business and now all of a sudden
  603. 22:21it is. But uh I I like the way that I've
  604. 22:26explained it is the way that it's been
  605. 22:28explained to me is that the industry has
  606. 22:30just been through too many booms and
  607. 22:31busts and so people are nervous to or
  608. 22:35they were nervous to really make a very
  609. 22:37bold capex commitment.
  610. 22:39So on the data center point, do you
  611. 22:42think there is all this opposition? Will
  612. 22:44the data centers still get built or is
  613. 22:46this going to really hamper your ability
  614. 22:48to to do the buildout?
  615. 22:51Um I think data centers will get built.
  616. 22:53Exactly where exactly how you know what
  617. 22:56what the things that make sense for the
  618. 22:57communities they're in or whether we
  619. 22:58kind of put them further out away from
  620. 23:00communities,
  621. 23:03that stuff I I think wins and public
  622. 23:06sentiment will change over time. Um it
  623. 23:09wasn't that long ago that everybody
  624. 23:10wanted a data center in their community
  625. 23:13for the jobs and the kind of industry
  626. 23:14and you know like it was now clearly
  627. 23:17they don't people like put them way out
  628. 23:19in the desert. Uh maybe people want them
  629. 23:21in their communities again. Um I I
  630. 23:24personally have like a lot of sympathy
  631. 23:26to not wanting to
  632. 23:28have a data center in a community but I
  633. 23:30also think there's a lot we have a giant
  634. 23:32country. We have a lot of space we can
  635. 23:33do this. And I think people really want
  636. 23:35to use AI. People also want a lot of the
  637. 23:37other things that data centers provide.
  638. 23:39People like the internet and like
  639. 23:40watching videos on their phones or
  640. 23:42whatever. Um, so I think data centers
  641. 23:44will get built, but clearly the current
  642. 23:47way they're working being built and the
  643. 23:50engagement with communities is not
  644. 23:52working for people and you know the
  645. 23:54ecosystem will adapt. Do you think this
  646. 23:55it's a concern? I people talk about this
  647. 23:58buildout being sort of funded by debt
  648. 24:01and that it's all a house of cards. Do
  649. 24:03you do you think the debt is a problem?
  650. 24:06Um
  651. 24:08I have never seen the actual revenue
  652. 24:12growth
  653. 24:14including great margins show up for any
  654. 24:18revolution in tech that I have witnessed
  655. 24:20in my career. If it were not for that uh
  656. 24:23then I think it would be a problem. And
  657. 24:24I also think that you know these are
  658. 24:26like smart people on the whole issuing
  659. 24:28the debt. Um there's going to be like in
  660. 24:31any moon, there's going to be a lot of
  661. 24:33foam on the top that maybe needs to come
  662. 24:35off. Um there will be dumb things here
  663. 24:37and there, but but the fundamentals of
  664. 24:43not only transformative
  665. 24:45things happening in business today, but
  666. 24:48business is now coming to us and saying
  667. 24:49like I have seen the light. I see what's
  668. 24:51going to happen here. I can see what
  669. 24:52this does to the cap products and
  670. 24:55services I can offer in the future. um
  671. 24:59that just seems
  672. 25:02extremely clear and compelling. Now, is
  673. 25:06it possible that the buildout can get
  674. 25:09ahead of how quickly the economy can
  675. 25:11digest it? Yes, that could happen. Um I
  676. 25:13don't see that level of recklessness
  677. 25:16happening right now. Um the and I also
  678. 25:21think we're just we're underpricing
  679. 25:25we're underweaiting what can happen with
  680. 25:26the next generation of products. You
  681. 25:27know we had the chat bots then we had
  682. 25:30the coding agents. I think in the last
  683. 25:32few months we started to see the
  684. 25:34beginning of AI for science and what can
  685. 25:36happen with
  686. 25:38using a lot of compute to discover new
  687. 25:41math new physics. Um, I think we're
  688. 25:44going to start to see what doing a lot
  689. 25:46of compute can do for chip design and
  690. 25:49other some of the most valuable kind of
  691. 25:51engineering projects we have. I think
  692. 25:53we're going to see what a lot of compute
  693. 25:54can do to do a lot of the kind of wrote
  694. 25:57work inside an organization. So, and all
  695. 26:00of that has great value and I expect it
  696. 26:03will keep going.
  697. 26:03>> I mean, you announced actually just this
  698. 26:05morning that you're going to expand free
  699. 26:07compute for academic researchers. Um,
  700. 26:10you mentioned these math experiments.
  701. 26:12You've you've solved these historically
  702. 26:14difficult math theorems. Um, I'm curious
  703. 26:17what you think the impact of those, you
  704. 26:20know, particularly the math discoveries.
  705. 26:22Like you're you're invested, for
  706. 26:23instance, in nuclear fusion companies
  707. 26:25in, you know, companies that are
  708. 26:27partially reprogramming cells, like
  709. 26:29really cutting edge science. Do you see
  710. 26:32immediate impacts or will it be more of
  711. 26:34a long-term thing?
  712. 26:35>> Some of both. Uh generally I'm I'm I'm a
  713. 26:38believer that
  714. 26:40maybe the most important driver of
  715. 26:42long-term human progress comes from
  716. 26:44better scientific understanding. And
  717. 26:46sometimes that happens that very quickly
  718. 26:50changes uh society and the economy like
  719. 26:55you know we figured out how to make
  720. 26:56language models and very quickly the
  721. 26:57economy changed. Um, and sometimes it's
  722. 27:00like a fundamental discovery in physics
  723. 27:02and it's going to apply in a super
  724. 27:05important way to a new material or
  725. 27:06whatever, but that will take decades
  726. 27:07longer. So, I think it will be a mix.
  727. 27:09But what we're hearing from scientists
  728. 27:11is so compelling now about how much more
  729. 27:13productive they are. And not just that
  730. 27:14they can work faster, but they can
  731. 27:16discover things that they just weren't
  732. 27:18going to be able to figure out
  733. 27:19otherwise. We heard this a lot from
  734. 27:20mathematicians. We're now hearing it
  735. 27:22from physicists, chemists, material
  736. 27:24scientists. um and a world where we can
  737. 27:28get
  738. 27:2910 years or 100 years of science done in
  739. 27:32a year. Um and then the ability to
  740. 27:36compound that into valuable products and
  741. 27:39services and technologies we all use. Um
  742. 27:42I think the world can get so much better
  743. 27:44if we think backwards
  744. 27:46100 or 500 years and what it was like to
  745. 27:48live then. I think we are very grateful
  746. 27:51for the scientific discoveries and the
  747. 27:53engine to get them usable in people's
  748. 27:55hands we have had since um and I think
  749. 27:58in the notistant future I hope people
  750. 28:00will look back at 2026 and be like they
  751. 28:02had it really rough and the main reason
  752. 28:04will be the scientific discovery and the
  753. 28:06compute and the AI that is being used to
  754. 28:08and the robots whatever to make our
  755. 28:10lives uh just much more enjoyable higher
  756. 28:13quality more fulfilling higher agency to
  757. 28:15live. Um, so we were thrilled to give AI
  758. 28:19uh to the scientists and I think that
  759. 28:21the next couple of generations of models
  760. 28:23are really going to surprise people on
  761. 28:25the upside of how much they can be an
  762. 28:27engine of scientific discovery.
  763. 28:29>> Well, I think we're going to end it on
  764. 28:30that optimistic note. Okay, it's a good
  765. 28:32one. I really appreciate the time. Thank
  766. 28:33you very much.

About this transcript

This page contains the full transcript of Sam Altman: ‘It’s Hard To Imagine the AI Revolution Without Jensen’ | Silicon Valley & The World by Semafor, generated from the public captions YouTube serves with the video. The transcript has 5,324 words across 766 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.