YouTube2Text

Nvidia Stock Crashed — Understand It Better Than 99% of Investors — Transcript

by Drew Cohen · 15,230 words · 2,137 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Nvidia is a story of a company that has
  2. 0:03always looked far into the future. In
  3. 0:062016, they were talking about how their
  4. 0:09GPUs would uniquely benefit from end
  5. 0:13markets such as AI, autonomous vehicles,
  6. 0:16and robotics. Fast forward 10 years, and
  7. 0:20looks like they were pretty correct
  8. 0:22about that AI thing. In fact, their
  9. 0:24revenues when they first made that
  10. 0:26prediction were just $5 billion in total
  11. 0:30with less than a billion dollars of
  12. 0:32profit. Fast forward 10 years till today
  13. 0:36and they have over $300 billion in
  14. 0:39revenues and $200 billion in profits.
  15. 0:43The stock over that period was up 100x.
  16. 0:48And in fact, if we rewind all the way to
  17. 0:50the 1991 IPO of Nvidia, we'll see that
  18. 0:55this is one of the strongest stock
  19. 0:57performances of all time. If you
  20. 1:00invested $1,000 at IPO, congratulations.
  21. 1:04You're now a centmillionaire. That is
  22. 1:06someone worth $10 million, almost as
  23. 1:09much as many Nvidia employees. And while
  24. 1:12no one can question their past success,
  25. 1:15questions about what the future brings
  26. 1:17for Nvidia are very open-ended and have
  27. 1:20gotten much harder to answer, especially
  28. 1:23this week with the announcement that
  29. 1:26Nvidia CEO and founder Jensen Hong would
  30. 1:30resign and step down after not one but
  31. 1:33two scandals in the same week. one
  32. 1:36involving an affair with a potential
  33. 1:38Chinese spy, another one involving an
  34. 1:41affair of Nvidia executive. And so now
  35. 1:45that Jensen Hang, who has led Nvidia for
  36. 1:48the past three plus decades and has seen
  37. 1:51around the corner of all sorts of
  38. 1:53potential risks to Nvidia, many of which
  39. 1:55we'll cover in this video. The question
  40. 1:58now is who will be the new leader of
  41. 2:00Nvidia? And I don't have an answer for
  42. 2:02you. And this leadership vacuum comes
  43. 2:05all at the same time that the risks are
  44. 2:07piling up. Some of Nvidia's biggest
  45. 2:10customers, Microsoft, Amazon, Meta,
  46. 2:15Google, not only spend hundreds of
  47. 2:18billion dollars together on their chips.
  48. 2:21They also are actively trying to cut
  49. 2:23them out of the data centers because
  50. 2:26they want to create their own chips and
  51. 2:28keep that margin for themselves. On top
  52. 2:31of that, the CUDA programming language,
  53. 2:34which was thought to be Nvidia's moat
  54. 2:37that would be impossible to surmount,
  55. 2:39has started to see some cracks with real
  56. 2:42competition coming up just in the past
  57. 2:45couple years and getting real adoption.
  58. 2:47And lastly, what about all of these
  59. 2:50circular financing arrangements and the
  60. 2:52fact that Nvidia is becoming a sort of
  61. 2:54credit back stop for over a hundred
  62. 2:57billion dollar in various guarantees? Is
  63. 3:00that concerning? We will cover all of
  64. 3:03this in much more in this deep dive
  65. 3:06video on Nvidia. If this is your first
  66. 3:08time here, what have you been doing? But
  67. 3:12you should know that I am a financial
  68. 3:14adviser at Davidson Con Management where
  69. 3:17we manage over half a billion dollars in
  70. 3:19AUM. I used to work at Goldman Sachs and
  71. 3:22Capital Group. On this channel, you'll
  72. 3:25find deep dives on a lot of different
  73. 3:27businesses as well as personal finance
  74. 3:30content that will help you make better
  75. 3:32investment and better money decisions.
  76. 3:34And so, if that is your thing and you
  77. 3:35want to get better with investing, go
  78. 3:36ahead and hit subscribe. Now, let us get
  79. 3:39into it. So, uh the way I structured
  80. 3:42this video is a little different than
  81. 3:43past ones because I feel like there was
  82. 3:46just so much to talk about and I wasn't
  83. 3:48sure how to fit it all in. So we are
  84. 3:51going to still start with the business
  85. 3:52section. I'll still touch on you know
  86. 3:55the overviews of the business, get into
  87. 3:57some of the products, a little bit of
  88. 3:58history and and context will kickstart
  89. 4:01us there. Then the next section I have
  90. 4:04I'm just calling six considerations.
  91. 4:07There's these sort of six different
  92. 4:08things that I came up with that are are
  93. 4:10kind of like key debate points uh things
  94. 4:12going on with Nvidia, things to
  95. 4:14consider. Some of them are risk uh key
  96. 4:17risk. Others are just kind of things
  97. 4:19that are kind of interesting to think
  98. 4:20about dynamics that are going on. Not
  99. 4:22clear whether or not it's going to be a
  100. 4:24beneficiary or you potential headwind or
  101. 4:27nothing for them. So that's going to be
  102. 4:29the next section. Then we're going to
  103. 4:31get into a bunch of other risk and then
  104. 4:33finally conclude with valuation where
  105. 4:36spoiler alert I'm going to disappoint
  106. 4:38you a little bit because as we're going
  107. 4:40to talk about uh valuation of a cyclical
  108. 4:43business is very very tricky. So I'm
  109. 4:46going to give you a little bit of a
  110. 4:47framework on how to do that. But there
  111. 4:49is a reason why a lot of investors have
  112. 4:52a stewed uh buying cyclical businesses
  113. 4:55and that includes of course Warren
  114. 4:57Buffett himself. uh if he does, he's
  115. 4:59just going to go ahead and buy him at a
  116. 5:00trough multiple and call it a day. Not
  117. 5:03going to go about with trying to
  118. 5:04estimate out uh all sorts of different
  119. 5:06earnings, but we will take a stab at
  120. 5:08doing that too. Aman, by the way, is
  121. 5:10another one who also just avoids
  122. 5:12cyclical businesses. A lot of investors
  123. 5:14do. So, one of the key questions though
  124. 5:15with Nvidia is going to be whether or
  125. 5:18not this is actually a cyclical business
  126. 5:20because there is a lot of people that
  127. 5:22have the opinion that the AI demand is
  128. 5:25so real and so high that this is going
  129. 5:28to be they might call it a Super Bowl
  130. 5:29cycle, but it's equivalent to saying
  131. 5:31that there's going to be sufficient
  132. 5:33demand for decades or longer that you
  133. 5:36don't need to think about this as being
  134. 5:38a cyclical business anymore. Going to be
  135. 5:40something we're going to discuss in this
  136. 5:42video. So, let us get started with what
  137. 5:45is the business. Now, the first thing
  138. 5:47you should know about Nvidia is that
  139. 5:49they're on a really weird fiscal year.
  140. 5:51Okay, that's definitely not the first
  141. 5:52thing you need to know about them, but
  142. 5:54it's worth mentioning because right now,
  143. 5:57you know, coming up in the next week,
  144. 5:59they're going to have 3Q earnings.
  145. 6:00They're calling that 3Q 2027 earnings.
  146. 6:04Obviously, the year is still 2026, but
  147. 6:07for some reason, I honestly don't even
  148. 6:09actually have a great answer for this.
  149. 6:10Some businesses, they have very weird
  150. 6:12fiscal years. And so, uh, their fiscal
  151. 6:14year, uh, the last one ended in January
  152. 6:172026, which really is basically 2025.
  153. 6:20It's 11 months of 2025, one month of
  154. 6:232026. Okay, if that was confusing at
  155. 6:26all, I'm sorry for taking up any of your
  156. 6:28bandwidth already because there's so
  157. 6:30much more we need to talk about. So, uh,
  158. 6:33I'm probably gonna flip between the two
  159. 6:35throughout the video. I'm going to try
  160. 6:36to clarify and just do calendar years
  161. 6:38cuz I feel like that's simpler, but I
  162. 6:40feel like someone's going to call me out
  163. 6:41and make a mistake on that at some
  164. 6:43point. And so, the first thing though is
  165. 6:44going to be LTM. They generated $32
  166. 6:48billion in revenue. So, that's the last
  167. 6:5012 months. Uh, and then in the last
  168. 6:52quarter alone, they did $105 billion in
  169. 6:55revenue with 106% year-over-year growth.
  170. 6:58And so that growth has actually been
  171. 7:01accelerating over the past couple
  172. 7:03quarters, which is just crazy when
  173. 7:04you're talking about a business that's
  174. 7:06doing hundred billion, growing 100%. Um,
  175. 7:09and this is kind of getting ahead of
  176. 7:11ourselves, but Jensen Hang talked about
  177. 7:13on the last couple earnings calls, how
  178. 7:16uh for the following year, which is
  179. 7:18going to be 2027,
  180. 7:20uh, he believes that there's going to be
  181. 7:22at least 70% capacity growth, but
  182. 7:25there's actually 100% demand growth as
  183. 7:27well there. That's kind of the
  184. 7:28difference between what they believe
  185. 7:29they could deliver versus the demand
  186. 7:31that's there. The point being it's still
  187. 7:33growing very quickly, thought to be
  188. 7:35sustained for at least another year or
  189. 7:37so. We'll talk more about that when we
  190. 7:39get to valuation. Now, gross margins
  191. 7:4175%. Those are up 12 points in the past
  192. 7:455 years. So, as they are getting bigger,
  193. 7:47you are seeing that they are getting
  194. 7:49some gross margin leverage there as
  195. 7:51well. Uh some pricing power too. some of
  196. 7:54that is coming through and so gross
  197. 7:56margins expanded to 75%. What's more
  198. 7:59impressive though is what happened on
  199. 8:00the operating income line. Uh on the
  200. 8:03operating income line we actually see
  201. 8:05they now in the last 12 months are doing
  202. 8:0765% margins and that's up 35 points in
  203. 8:12about the past 5 years. So margins went
  204. 8:14from you know 30-ish% range to now 65%.
  205. 8:18And if you're looking at where that
  206. 8:20operating leverage actually came from,
  207. 8:21it came pretty evenly from uh SGNA as
  208. 8:24well as R&D. And so if you're looking at
  209. 8:26just expenses as a percent of revenue,
  210. 8:29we see that uh SNM has gone from around
  211. 8:3110% to 2%. So very low. And uh R&D has
  212. 8:35gone from 20% to 8%. You know, it's
  213. 8:38still a high number in absolute dollars
  214. 8:40because they're such a massive business,
  215. 8:41but you're seeing a lot of that
  216. 8:43operating leverage uh flowing through.
  217. 8:45And that's really what has allowed them
  218. 8:47to now have, you know, that $200 billion
  219. 8:48in LTM profit. So, uh, before we get
  220. 8:51into what their products are, what the
  221. 8:54business is, you know, the advantage of
  222. 8:56the products, all of that, let's give a
  223. 8:58little bit of context and history about
  224. 9:00this business. I think it's very
  225. 9:02important to have some idea of where
  226. 9:04they started, some of the things they've
  227. 9:06been through to get a better
  228. 9:07understanding of a business. So, we're
  229. 9:09going to start all the way in 1993.
  230. 9:11Don't worry, the history part will be
  231. 9:14quick. But uh Nvidia was founded that
  232. 9:16year uh by Jensen Hang and two other
  233. 9:19co-founders. In 1999, they rolled out
  234. 9:22the first uh GPU that they had, the
  235. 9:25GeForce. They didn't invent the GPU
  236. 9:27exactly, but they did popularize it. And
  237. 9:29so that was a gaming focused GPU because
  238. 9:32that's where they started. That's where
  239. 9:34their roots were. They were in gaming.
  240. 9:36Now, one thing that's interesting about
  241. 9:38a GPU, if you're wondering what's the
  242. 9:40big deal about, you know, the GPU, why
  243. 9:42are these all the rage now instead of
  244. 9:44CPUs? Uh, right. So, the thing that is
  245. 9:47different is if you're thinking of a
  246. 9:49traditional CPU, the way it computes is
  247. 9:52it goes kind of line by line in order,
  248. 9:55whereas a GPU is able to do processing
  249. 9:58in parallel. So, if you think about a
  250. 10:00graphics card, uh, a GPU by the way, it
  251. 10:03just stands for graphics processing
  252. 10:04unit. If you think about this and you
  253. 10:07let's say are playing a video game, the
  254. 10:10pixels you're looking at on the video
  255. 10:11game, let's say you're thinking about
  256. 10:13the pixels up here versus the ones over
  257. 10:15there, they could process at the same
  258. 10:17time. They don't have any uh computation
  259. 10:19that relies on one another, right? So
  260. 10:22the graphics card could simultaneously
  261. 10:24or in parallel as it's usually said uh
  262. 10:27produce you know the pixel over here and
  263. 10:29the mathematical equation that produces
  264. 10:31the pixel over there. So it could run
  265. 10:33both of these equations simultaneously.
  266. 10:35That's what a GPU is. And so this
  267. 10:38started in graphics, but then we fast
  268. 10:40forward to around 2006 and Jensen Hong
  269. 10:44started having this idea that this sort
  270. 10:46of compute could actually be useful for
  271. 10:48a lot of other things and that is why he
  272. 10:52created the CUDA platform. So this was
  273. 10:54for programmers to help program GPUs to
  274. 10:57do a variety of different things to be
  275. 11:00more general purpose than just graphics
  276. 11:02and take advantage of this unique
  277. 11:04functionality of the GPU. Now we're
  278. 11:06going to fast forward a little bit more
  279. 11:07to 2012 and there was this AI
  280. 11:10breakthrough uh something called the
  281. 11:12AlexNet which was a neural network that
  282. 11:14dramatically improved uh image
  283. 11:17recognition and it was trained on Nvidia
  284. 11:20GPUs. So, this was uh one of the first
  285. 11:23times there was this real idea that hey,
  286. 11:26wait a minute. These GPUs are kind of
  287. 11:28uniquely good for this thing called AI.
  288. 11:31Maybe that'll be interesting one day.
  289. 11:33And you fast forward a couple years to
  290. 11:35the 2016 investor day uh that Nvidia
  291. 11:38had. And Jensen Hong is already kind of
  292. 11:40going allin on AI. They're talking about
  293. 11:42in their investor day in 2016, they're
  294. 11:45talking about how big AI is going to be.
  295. 11:47They're talking about autonomous
  296. 11:48vehicles. They're talking about
  297. 11:50robotics. These are big use cases or end
  298. 11:53markets that they really believe are
  299. 11:55going to create a lot of demand in the
  300. 11:57future and it's kind of just admirable
  301. 11:59if you will. Imagine seeing something
  302. 12:01coming and getting ready for it for like
  303. 12:04a decade. But that's that's kind of what
  304. 12:06Nvidia did. So that was back in 2016.
  305. 12:09But uh it should be noted that their
  306. 12:11their entire business at this time it's
  307. 12:13basically just gaming graphics cards,
  308. 12:16right? you know, selling them to desktop
  309. 12:18gamers, you know, they they'll sell them
  310. 12:20sometimes to console makers, to laptops,
  311. 12:22whatever. It's just these PC end markets
  312. 12:25is really what their business is. And,
  313. 12:27you know, they're thought to make good
  314. 12:29graphics card, but as you can imagine,
  315. 12:31it's a very cyclical business because,
  316. 12:33uh, consumer demand for computers, well,
  317. 12:35it fluctuates with the economy a lot.
  318. 12:38And so, in in 2008, revenues actually
  319. 12:40peaked at the time, hit $4.1 billion.
  320. 12:44and they wouldn't surpass that level
  321. 12:46again for 5 years until 2013. So the
  322. 12:49revenues they made in 2008 were not
  323. 12:51surpassed until 5 years later. Uh and
  324. 12:55profit margins over this time, they're
  325. 12:56bobbing around the mid- teens to 20%.
  326. 12:59Then the financial crisis, they're
  327. 13:00actually negative. There was another
  328. 13:02year in there, they also were pretty
  329. 13:03weak. Um so you could really see all of
  330. 13:06kind of the cyclicality in this business
  331. 13:08at the time. And then the crypto boom
  332. 13:09happens, you know, starting in 2017,
  333. 13:11really starts to take off in 2018. uh by
  334. 13:142019 they're now doing 11 billion in
  335. 13:16revenues, 32% margins and this just
  336. 13:19extends even through co there was a
  337. 13:21little bit of a lull in 2020 but uh then
  338. 13:24you know co it comes back with a
  339. 13:25vengeance but then there's the aftermath
  340. 13:27of this and by 2022 revenues are flat
  341. 13:31but they have all of these uh GPUs that
  342. 13:33they have to get rid of. They have to
  343. 13:34discount them basically to get rid of
  344. 13:36them. Can you imagine that? Just 4 years
  345. 13:38ago, this is in 2022, they were
  346. 13:41discounting GPUs to get rid of them. And
  347. 13:43uh their operating margins fell from 37%
  348. 13:46to 16%. And so this is just such a a
  349. 13:49crazy fact. If we rewind just about four
  350. 13:52years ago, their revenues were 26
  351. 13:54billion, their operating margins were
  352. 13:5816%. I'm doing dramatic pauses so it can
  353. 14:01sink in. Um, four years ago, 16%
  354. 14:04margins, 26 billion uh, revenue company.
  355. 14:07Today, 300 billion, 65%
  356. 14:12margins. That was a 4-year transition. I
  357. 14:14know we're all looking at like
  358. 14:15Enthropics ARR and how much that's
  359. 14:17growing. And, you know, it's fast. So,
  360. 14:19now it's like made us blasze about every
  361. 14:22other company that's ever grown before,
  362. 14:23but that is a crazy uh, level of growth
  363. 14:26for this business to have. And so uh
  364. 14:28this is of course the AI boom you know
  365. 14:30kick off in 2023 about with uh chat GPT
  366. 14:34and so their their decade long wait of
  367. 14:36waiting for AI to finally come through
  368. 14:39paid off for them and it paid off big
  369. 14:41time. So I whenever I'm looking at a
  370. 14:43business I love looking at a bit of
  371. 14:45history to understand where they're
  372. 14:46coming from. gives you a better sense of
  373. 14:48where they're going too because I think
  374. 14:50a lot of times if you're just popping in
  375. 14:52at the last quarter, the last year and
  376. 14:55then you're just projecting what
  377. 14:56happened in the last year out forward,
  378. 14:58you're going to miss a lot in the
  379. 15:00history, right? And so understanding
  380. 15:02that this was a pretty cyclical business
  381. 15:04and their success, you know, this level
  382. 15:07of success certainly is like a very new
  383. 15:09phenomena, I think is kind of helpful
  384. 15:11knowledge when you're going to try to
  385. 15:12predict out, you know, the next decade
  386. 15:14and you're real realizing, wait a
  387. 15:16minute, it's only really been a great
  388. 15:17business for the past four years. Um, so
  389. 15:20I think that's something that that's
  390. 15:21kind of worth taking into account. And
  391. 15:23when I'm doing this, I like to look at a
  392. 15:25lot of financial data and I do that on
  393. 15:27Fiscal AI. They are a sponsor of the
  394. 15:29show, but I used them before they
  395. 15:32sponsored the show. And it's nice
  396. 15:33because I can just pull up all of, you
  397. 15:35know, the operating margins going back,
  398. 15:37you know, two decades, two and a half
  399. 15:39decades, and can just see all of this
  400. 15:41data very clearly. And it makes it very
  401. 15:43easy to then pick out something I if
  402. 15:45something looks like it it doesn't
  403. 15:47really belong. And then you could hone
  404. 15:48in on that and ask more questions. So,
  405. 15:50if you want to use fiscal AI to do the
  406. 15:53same sort of analysis I do, uh there's a
  407. 15:55discount code in the show notes below.
  408. 15:57Now, that is enough of a plug. Let's
  409. 15:59move back to an overview of the
  410. 16:01business. So, uh, a year ago, their
  411. 16:04disclosures were a little bit different
  412. 16:06than what they offer now. Uh, I actually
  413. 16:09like some of the new disclosures they
  414. 16:11added. We'll talk about them in a
  415. 16:12second, but uh, on the old disclosures,
  416. 16:15they did have these different market
  417. 16:16segments they would break down, which
  418. 16:18was data center, gaming, professional
  419. 16:21visualization, and automotive. And so,
  420. 16:24we could see the market share stats of
  421. 16:25all of this. And when I tell you these
  422. 16:27numbers, it's going to be clear why they
  423. 16:29kind of didn't bother with these
  424. 16:30disclosures anymore. So, gaming 7% of
  425. 16:33revenues. Professional visualization. Uh
  426. 16:37this is going to be uh like 3D graphics
  427. 16:40and uh doing simulations and models uh
  428. 16:44stuff for professional design. So, this
  429. 16:47this was 1 and a.5% of revenues and then
  430. 16:50automotive was 1% too and data center
  431. 16:53was 90%. So, And since then it looks
  432. 16:57like data center actually grew even
  433. 16:58more. So uh it makes sense why they they
  434. 17:00don't bother doing this breakout
  435. 17:01anymore. Uh when they did do the
  436. 17:04breakout though they also broke down
  437. 17:05data center in terms of compute versus
  438. 17:08networking. So compute that's going to
  439. 17:10be more like the GPUs. Uh they also do
  440. 17:13have CPUs now but u mostly GPUs. And
  441. 17:16then networking. Networking is going to
  442. 17:18be kind of the things involved with
  443. 17:20helping the chips talk to each other.
  444. 17:22And then also the racks of systems, you
  445. 17:25know, the racks and servers talk to each
  446. 17:27other as well. So compute was 83%,
  447. 17:29networking 17%. Networking is actually
  448. 17:32getting more and more important and it's
  449. 17:34actually becoming more their kind of
  450. 17:37moat, if you will, and we're going to
  451. 17:39talk more about that later. So um, in
  452. 17:42terms of kind of their major products,
  453. 17:43like what do they actually sell? So they
  454. 17:45have a bunch of different architects of
  455. 17:47GPUs. you go back uh you know uh to a
  456. 17:50couple architects ago and it was the
  457. 17:52Hopper. So the H100, the H200, then they
  458. 17:55replace that with Blackwell, the B100,
  459. 17:58the B200. Up next is going to be the
  460. 18:00Reuben. Uh they also do have CPUs. Those
  461. 18:03go by the name Grace and Vera. Uh
  462. 18:06there's something called HGX that is
  463. 18:08going to be a server platform. So this
  464. 18:11is going to be basically multiple GPUs
  465. 18:14and then it's connected with their
  466. 18:16networking technology like an NVL link
  467. 18:18and so they'll they'll sell a whole
  468. 18:20server. Then on the next level to that
  469. 18:23though is going to be getting an entire
  470. 18:25rack or like AI factory as Jensen Hang
  471. 18:28likes to talk about that. And so uh DGX
  472. 18:30that's the name of of kind of the the
  473. 18:32leading product here right now from
  474. 18:33them. This is going to be their
  475. 18:35integrated AI system. It's going to give
  476. 18:37you eight of the B200 GPUs, uh, two
  477. 18:41Intel Xeon CPUs, two NV switch chips,
  478. 18:44and then also NVLink all together in the
  479. 18:47same server. And so, you'll note I said
  480. 18:49Intel, so they do still partner with
  481. 18:51Intel for certain CPU chips. looks like
  482. 18:54they're looking to replace those in a
  483. 18:56lot of architectures in the next
  484. 18:57generation, but uh Intel still has the
  485. 19:00x86
  486. 19:01uh CPU platform that they use and
  487. 19:04sometimes that's kind of preferred for
  488. 19:06certain servers. And so generally
  489. 19:08speaking, Nvidia is pretty integrated in
  490. 19:11terms of everything they're doing, but
  491. 19:13they do still become flexible. If it
  492. 19:16looks like there's an advantage uh or a
  493. 19:18reason why what one customer wants to
  494. 19:20like stick with an x86 architect, that's
  495. 19:22fine. They'll accommodate that. You'll
  496. 19:25even see later on this NV link, which is
  497. 19:27their um designed way to have chips talk
  498. 19:30to each other. They even now created
  499. 19:32something that allows chips that are not
  500. 19:34even Nvidia chips to kind of get in on
  501. 19:36that. And so they are you know flexible
  502. 19:39I would say in terms of yes they design
  503. 19:42everything and it works better when it's
  504. 19:43all integrated but if there is something
  505. 19:45that they can sell to uh you know a
  506. 19:48customer they'll they'll still do it and
  507. 19:49try to make sure they still have a
  508. 19:51position in the data center uh or
  509. 19:54conversely if there's a benefit from
  510. 19:56bringing in you know an outside um
  511. 19:59players uh technology into them like
  512. 20:01Intel they'll accommodate that as well.
  513. 20:03Um then coming up next is going to be
  514. 20:05the rack scale AI uh supercomputers. So
  515. 20:09this is like their their biggest thing
  516. 20:10now. One of the names you maybe would
  517. 20:12have heard is the Vera Rubin. That's
  518. 20:15officially going to be called the Vera
  519. 20:16Rubin uh NVL72.
  520. 20:19Very important you remember that. That
  521. 20:21is going to have 72 uh Reuben GPUs. So
  522. 20:24latest GPU architecture there. 36 Vera
  523. 20:28CPUs plus um NVLink plus a bunch of high
  524. 20:31bandwidth memory. All of that together
  525. 20:33is basically set as like a rack that's
  526. 20:36just ready to drop uh into a data
  527. 20:38center. And so uh we're going to talk
  528. 20:40more about this, but over time they've
  529. 20:42been moving from just being a chip to
  530. 20:44becoming an entire AI system. And this
  531. 20:47has actually allowed for them to have
  532. 20:49more of an advantage, a performance
  533. 20:51advantage by taking everything together
  534. 20:53and integrating it and optimizing all of
  535. 20:56it even when they're perchip level sort
  536. 20:59of performance drops. If you're just
  537. 21:00doing comparisons at the chip level, the
  538. 21:03performance level across the entire
  539. 21:05system tends to be maintained. And so
  540. 21:08this has been actually a pretty smart
  541. 21:09thing that uh Nvidia has been doing for
  542. 21:11for the past several years is moving to
  543. 21:13integrate everything together. They
  544. 21:15bought you know networking technology
  545. 21:17six years ago and this has been kind of
  546. 21:19a longcoming effort to really integrate
  547. 21:21an entire platform and Jensen Hong has
  548. 21:25been talking about you know AI factories
  549. 21:27for a long time. This is kind of the
  550. 21:29idea is that this is not about selling,
  551. 21:31you know, just a chip. Now, it's an
  552. 21:32entire AI system. Now, just going to run
  553. 21:34through a few more products quickly and
  554. 21:36then we're going to get into kind of the
  555. 21:38fun stuff, the meat of of this video and
  556. 21:40in the analysis. So, uh networking,
  557. 21:42there's going to be NVLink there and NV
  558. 21:44switch. Uh NVL link basically allows the
  559. 21:47GPUs to talk to each other and then the
  560. 21:49NV switch uh coordinates uh that. So,
  561. 21:52there's a lot of GPUs now that are
  562. 21:54running distinct workloads. And so the
  563. 21:56new thing, especially as models get more
  564. 21:59complex, is figuring out where to send
  565. 22:02these workloads to and orchestrating all
  566. 22:04of them. So having technology to help do
  567. 22:07that, Nvidia also has software for this.
  568. 22:09Uh that's very helpful to actually
  569. 22:11optimize um the the chips to actually
  570. 22:14get the most out of um the performance
  571. 22:17of these. So uh they also have something
  572. 22:19called Infiniband and Spectrum X
  573. 22:21Ethernet. This connects servers and
  574. 22:23racks across the data center. Uh there
  575. 22:26is still, you know, software to all of
  576. 22:27this. CUDA is part of it. There's other
  577. 22:30software though too that we could name
  578. 22:31here, but uh it's going to be just too
  579. 22:33much for right now. So, we'll keep
  580. 22:35going. Uh in gaming, it's still big. You
  581. 22:37know, they still have the GE Force uh
  582. 22:39GPU. Yeah, there's competitor GPUs now
  583. 22:42and all that. Still a good business for
  584. 22:44them, though. Um much more
  585. 22:46inconsequential relative to data center.
  586. 22:48Uh professional visualization and
  587. 22:50automotive, these are also going to be
  588. 22:52smaller businesses for them. Now,
  589. 22:54Automotive does have something called
  590. 22:56Nvidia Drive, which is an autonomous
  591. 22:58vehicle platform to help enable uh all
  592. 23:01these different car makers to have their
  593. 23:02own autonomous vehicle uh uh programs
  594. 23:05and really kind of give them a jump
  595. 23:07start with building their own EV
  596. 23:08technology. And then there's also
  597. 23:10something called Jetson, which is a
  598. 23:12compact AI computer for robots. Not
  599. 23:15going to talk much about that right now,
  600. 23:16but it's worth pointing out that a lot
  601. 23:18of the AI that we've seen in the world
  602. 23:20so far, it it's still, you know, LLMs,
  603. 23:23right? maybe uh agentic AI, but the real
  604. 23:26next lay layer to this afterwards is
  605. 23:29going to be what happens when we combine
  606. 23:31AI with robots and that's going to
  607. 23:33unlock, you know, a whole new wave of
  608. 23:36demand. So, still, you know, a little
  609. 23:38bit far out, at least as of now, but
  610. 23:40kind of just worth mentioning. Um, but
  611. 23:42not kind of in any of the numbers we're
  612. 23:44going to talk about for Nvidia here.
  613. 23:46Now, I mentioned that they added some
  614. 23:48new disclosures, and I like these
  615. 23:51disclosures because I feel like it helps
  616. 23:54allude to a bunch of the conversations
  617. 23:56we're going to be having now. Uh, but it
  618. 23:59also kind of shows maybe a little bit
  619. 24:00their paranoia. So, uh, if you're
  620. 24:03looking at these disclosures, they talk
  621. 24:05about hypers scale revenue versus
  622. 24:08revenue from AI clouds, industrial, and
  623. 24:10enterprise. So hypers scale revenue is
  624. 24:14169 billion in the last 12 months versus
  625. 24:17AI clouds industrial enterprise uh at
  626. 24:21109 billion. So the hyperscalers are
  627. 24:24growing 106% the AI clouds are growing
  628. 24:2768%. So the hyperscalers still very
  629. 24:31material portion of revenue and growing
  630. 24:34quicker right now. And then just before
  631. 24:36we get into kind of the the fun
  632. 24:37conversation all of that I should just
  633. 24:39note they basically have two segments
  634. 24:41now it's data center and then edge
  635. 24:43compute. Edge compute is basically just
  636. 24:45a catch-all bucket of all the you know
  637. 24:48uh compute that's not really data
  638. 24:50center. So it's gaming professional
  639. 24:52visualization automotive. uh that's
  640. 24:54going to be about 25 billion in revenue
  641. 24:56growing 35%. Whereas data centers 277
  642. 25:00billion uh growing 89% and that's going
  643. 25:03to be 92% of the total. Okay, we got a
  644. 25:07lot of background, a lot of numbers out
  645. 25:08of the way. I'm sure you remembered
  646. 25:10everything I already said, so that's
  647. 25:11great. Now, let's get into the fun part.
  648. 25:13So, these are going to be the six
  649. 25:15different considerations. The first one
  650. 25:18is going to be about uh the hyperscalers
  651. 25:21versus the neoclouds. And it kind of uh
  652. 25:25is alluded to in that disclosure change
  653. 25:28that I just mentioned where they're now
  654. 25:30talking about the revenue from the
  655. 25:32hyperscalers versus revenue from other
  656. 25:34sources from the AI clouds. And the
  657. 25:38reason why I think this is very
  658. 25:40important is because and I I do think by
  659. 25:43the way Jensen Wong is very very smart
  660. 25:46regardless of you know what's transpired
  661. 25:49um whatever sort of lapses in judgment
  662. 25:51he may have there's several things we're
  663. 25:54going to talk about here where it's like
  664. 25:56very early on maybe people were doubting
  665. 25:58you know why are they doing these
  666. 25:59arrangements where they're giving money
  667. 26:00to these different NeoClouds and like it
  668. 26:03you you'll kind of start to understand
  669. 26:04why through this conversation. So the
  670. 26:06first thing we should note is that there
  671. 26:08are the hyperscalers right who by and
  672. 26:10large are in the business of selling
  673. 26:13compute now because they buy a lot of
  674. 26:16chips and their business is renting them
  675. 26:18out and they have a lot of scale. They
  676. 26:21have an incentive to optimize their
  677. 26:24workloads to get not just the best
  678. 26:26performance but the p best performance
  679. 26:29on cost. Plus the fact that they have a
  680. 26:31lot of scale and they spend a lot of
  681. 26:32money on chips means that it usually
  682. 26:35makes sense for them to get into
  683. 26:37designing their own chips in order to
  684. 26:39save money. This decision to design your
  685. 26:42own chip in order to save money. It only
  686. 26:45makes sense if you're very very big and
  687. 26:47very massive. So right now we know of
  688. 26:50course Google has the TPU that's the
  689. 26:52name of their chip that they've
  690. 26:54designed. Um Microsoft has uh Maya uh no
  691. 26:59one talks about it as much but they do
  692. 27:01have their own chip they've been
  693. 27:03designing. Amazon uh they have a few but
  694. 27:05the one we'll talk about that's
  695. 27:06competitive to the GPU is Tranium. Uh
  696. 27:09that's their kind of version there. Uh
  697. 27:11and then Meta also has uh their own chip
  698. 27:13too. It's MTIA, however you're supposed
  699. 27:16to say that. And then Oracle is one of
  700. 27:19the only hyperscalers that does not
  701. 27:21actually have their chip. And I know
  702. 27:22Meta is not a hyperscaler, but they're
  703. 27:24in a weird place right now with all the
  704. 27:26money on capex they've been spending.
  705. 27:27They want to be a hyperscaler. Um, and
  706. 27:30so either way though, the point being
  707. 27:31that these are all players that have big
  708. 27:35cloud businesses, Meta the exception
  709. 27:37there, but they buy a lot of Nvidia
  710. 27:38GPUs, right? So they buy a lot of Nvidia
  711. 27:41GPUs and they use them a lot in their
  712. 27:43business. And they realized, hey, why
  713. 27:46are we paying this massive margin to
  714. 27:48Nvidia? Why don't we design our own
  715. 27:49chips and we can basically cut them out?
  716. 27:52We don't need them. Now, that's not the
  717. 27:53easiest thing to do and we're going to
  718. 27:55talk throughout this video why that is,
  719. 27:56why it is hard to actually compete
  720. 27:58against Nvidia. But the thing I want to
  721. 28:00note now is that their scale does enable
  722. 28:03them to make that decision to design
  723. 28:05their own chip. So, if they were small,
  724. 28:09they wouldn't have the resources to be
  725. 28:11able to do that. And so what Nvidia did
  726. 28:14was they started funding a lot of these
  727. 28:17small cloud competitors. And they didn't
  728. 28:20just fund one or two. They funded, you
  729. 28:22know, like half a dozen or more of them
  730. 28:24cuz they just wanted a bunch of them
  731. 28:26because they knew that the more uh small
  732. 28:29cloud competitors there were, they could
  733. 28:32then freely sell their GPU chips to them
  734. 28:35without the fear that they would get big
  735. 28:37enough and get scale to then make the
  736. 28:39decision to design their own chips and
  737. 28:40try to cut Nvidia out. So Nvidia
  738. 28:43invested in a lot of these companies
  739. 28:45including Coreweave, Nebus, Nscale, uh,
  740. 28:48Lambda, Cruso, a bunch of other ones.
  741. 28:51And so they put money into all of these.
  742. 28:54They gave them GPUs instead of giving
  743. 28:56them to Google and Amazon because they
  744. 28:58were trying to basically create their
  745. 29:00own um, competitor chip. So instead,
  746. 29:02Nvidia said, "I'm going to give my chips
  747. 29:04to all these competitors. I'm going to
  748. 29:05help fund them, too, and then this is
  749. 29:08what is going to enable me to be more
  750. 29:10protected because it's very unlikely
  751. 29:13Nebius or Coreweee now is going to go
  752. 29:15and design their own chip." And so a lot
  753. 29:17of their architecture too is now
  754. 29:18dependent on Nvidia. So, uh, it just
  755. 29:21puts them in a much better position. you
  756. 29:24know, yeah, it's possible Google sells
  757. 29:25like TPUs to to Coreweave or something
  758. 29:27like that in the future, but uh the
  759. 29:29Neoclouds are also very scared of
  760. 29:32getting on Nvidia's bad side because
  761. 29:34their entire business is basically we
  762. 29:36were able to get Nvidia GPUs when other
  763. 29:39people weren't. None of these NeoClouds
  764. 29:41would exist if Nvidia didn't want them
  765. 29:43to exist, but they do want them to exist
  766. 29:46uh because it's in their competitive
  767. 29:48interest, too. So, this is the the first
  768. 29:50consideration here. You have this
  769. 29:52dynamic where you have a lot of your
  770. 29:53revenue, majority of revenue coming from
  771. 29:55your customers who are actively trying
  772. 29:58to cut you out while you're
  773. 29:59simultaneously
  774. 30:01funding the competition against them as
  775. 30:04a kind of way to balance out the dynamic
  776. 30:07but also ensure that there are all these
  777. 30:09players buying your chips in case they
  778. 30:11do kind of fully cut you out. Now, I
  779. 30:13don't think that Nvidia is going to ever
  780. 30:15get fully cut out from, you know,
  781. 30:16Google, Amazon, any of them. I think it
  782. 30:19it's more likely that you have some base
  783. 30:22level of compute that is going to be
  784. 30:24Nvidia. There's still reasons to have
  785. 30:26it. It tends to be more flexible.
  786. 30:27There's certain workloads uh things are
  787. 30:29still the best uh performance per dollar
  788. 30:31on. But what's slowly happening is that
  789. 30:33certain workloads are being carved out
  790. 30:36and are going to these custom chips that
  791. 30:38they're making. Um Google in specific
  792. 30:41with the TPU is particularly competitive
  793. 30:43because um it's kind of you know a
  794. 30:46little bit more general purpose uh
  795. 30:49versus some of the other customized
  796. 30:50chips that the hyperscalers are making.
  797. 30:52And so the the one other thing worth
  798. 30:54mentioning here in the same bucket is
  799. 30:55sovereign AI. So this is something that
  800. 30:58Jensen Hong's been talking about for a
  801. 31:00few years. And I was kind of curious.
  802. 31:02I'm like why does he keep mentioning
  803. 31:03this idea? Like who cares that much
  804. 31:06about whether or not you know a country
  805. 31:08has their own you know sovereign AI data
  806. 31:10centers? Well Nvidia cares and do you
  807. 31:14know why? Because if they are able to
  808. 31:16convince a country that they need to
  809. 31:18have sovereign uh their sovereignty over
  810. 31:21their data centers then very likely what
  811. 31:23that means is that they're going to have
  812. 31:25to set it up themselves. And if they set
  813. 31:28it up themselves, the easiest way to do
  814. 31:31that is to go directly to Nvidia and buy
  815. 31:33one of their AI factories. They have
  816. 31:36whole server racks you could just buy
  817. 31:38and plop into a data center. And so
  818. 31:40that's why the AI sovereignty sort of
  819. 31:42idea or theme is actually really
  820. 31:44important to Nvidia because it's just
  821. 31:46it's the same part of the same story
  822. 31:48where you get a more disperate group of
  823. 31:51customers buying your product. So you're
  824. 31:52diversifying out uh where it's going.
  825. 31:55But the bigger part of that too is that
  826. 31:56this is going to be another customer
  827. 31:58who's not going to be a competitor to
  828. 32:00you. Nvidia is sick of having their
  829. 32:02customers try to then design their own
  830. 32:04chips and cut them out. And if you're
  831. 32:06talking about, you know, a country like
  832. 32:09Saudi Arabia setting up their own data
  833. 32:11centers for sovereign AI, they're not
  834. 32:13going to also get into the chip design
  835. 32:15business. They're just going to buy
  836. 32:16Nvidia's chips. So that's why the
  837. 32:18sovereign AI story is also kind of
  838. 32:21another consideration here. So, by the
  839. 32:23way, when I wrap this all up, I I'll
  840. 32:24kind of conclude with Nvidia's moes and
  841. 32:27what I think about them, where I think
  842. 32:28they're strengthening, weakening. But
  843. 32:30for now, these are just kind of
  844. 32:31considerations, things going on. So, the
  845. 32:33second consideration is going to be this
  846. 32:35idea of an individual chip versus an AI
  847. 32:39factory. And so, generally speaking,
  848. 32:41Nvidia's chip is still highly
  849. 32:44performative on a perost basis. So,
  850. 32:47there's lots of different ways you can
  851. 32:49run test. And you know, there's
  852. 32:51different ways you could kind of fudge
  853. 32:53it and get different results and and
  854. 32:55people do that because everyone wants to
  855. 32:56show they have the best chip. Generally
  856. 32:58speaking though, Nvidia's chips are are
  857. 33:00still always up there. However,
  858. 33:03competitors are getting closer to
  859. 33:06sometimes passing uh them on a perch
  860. 33:08chip performance level. So, especially
  861. 33:11if we're thinking about Google's um TPU
  862. 33:14and also AMD's GPU in some cases is
  863. 33:17actually more performant than Nvidia's
  864. 33:20chip. But there's an asterisk on this.
  865. 33:22This is going to be usually compared to
  866. 33:25um their their later generation chip,
  867. 33:28not the latest one like the one that
  868. 33:30they have yet to release. And then also
  869. 33:33uh it's not going to be when you're
  870. 33:34running the test on a system level. So
  871. 33:37what I mean by that is let's say for
  872. 33:39example, this was an analysis uh
  873. 33:41semi-analysis did was they compared
  874. 33:43Google's TPU to Nvidia's uh B300 chip.
  875. 33:48And what they found was the Google TPU
  876. 33:51uh was cheaper on a per token basis by
  877. 33:5319%. So that's just comparing one chip
  878. 33:56to the other chip. However, and this is
  879. 33:59the important part, if you compare it to
  880. 34:02the entire Nvidia system, uh which is
  881. 34:04called, you know, the GB300
  882. 34:07NVL72. I know you remember that from the
  883. 34:10product breakdown. What that is, uh,
  884. 34:12it's one of those AI systems that has,
  885. 34:14uh, 72, uh, GPU chips, a bunch of CPU
  886. 34:18chips, NVLink, NV switch, and if you're
  887. 34:21running it compared to this entire unit,
  888. 34:24then uh, the performance uh, per price
  889. 34:27uh, actually is 30% less than uh, than
  890. 34:30Google's TPU. So, the performance is
  891. 34:33still better on the systems level. And
  892. 34:35so that's kind of the consideration of
  893. 34:38this section is the difference between
  894. 34:39chip performance versus uh system level
  895. 34:42performance. And what's been happening
  896. 34:44is that competitors have caught up by
  897. 34:47and large pretty close maybe you know a
  898. 34:49year behind but pretty close on the chip
  899. 34:51level but not on the systems level and
  900. 34:55the integration of all of that. And it's
  901. 34:57going to be harder for them to do that
  902. 34:58too because philosophically the way
  903. 35:00Nvidia designed uh their chips, their
  904. 35:03hardware is they're tying everything
  905. 35:05together. They're tying together the
  906. 35:07chip, the linking between the chip uh
  907. 35:10then also the linking between the racks
  908. 35:12and then also the software CUDA right on
  909. 35:15top of that. It's all together working
  910. 35:17together uh to to get the best
  911. 35:19performance. And I'm going to talk more
  912. 35:21about CUDA in a different consideration.
  913. 35:23And so the other thing I just want to
  914. 35:24hit on here is that yes, there's lots of
  915. 35:27uh different companies that'll talk
  916. 35:28about how their chip is more performant
  917. 35:31versus, you know, the the regular GPU
  918. 35:33like Amazon. They came out and said
  919. 35:35their their tranium chip is 30% better
  920. 35:38performance per cost versus uh you know
  921. 35:40a general GPU they're talking about
  922. 35:42Nvidia, but that doesn't necessarily
  923. 35:44mean compared to the entire system
  924. 35:46running the entire system. That could
  925. 35:48just be them plopping chips in and out.
  926. 35:50So you have to be careful with that
  927. 35:52distinction and people can run these
  928. 35:54tests and kind of fudge the results a
  929. 35:56little bit if they want to. Um so so
  930. 35:59that's kind of something worth
  931. 36:00considering there. And then the last
  932. 36:01thing I want to say here though is that
  933. 36:03the more real world deployments you
  934. 36:05have, the more there's kind of this
  935. 36:07feedback system of optimizing your
  936. 36:09hardware and software to a real world
  937. 36:12workload in order to continue to improve
  938. 36:14it. That's actually going to be one of
  939. 36:16uh Nvidia's advantages that I'll I'll
  940. 36:18summarize with later. Okay. The third
  941. 36:20consideration is going to be CUDA. So
  942. 36:22this is their programming platform. It
  943. 36:24takes a programming language like
  944. 36:26PyTorch or Jax which themselves
  945. 36:28sometimes call on different CUDA
  946. 36:31libraries which basically means instead
  947. 36:33of you having to like code something out
  948. 36:35and say like move piece of memory from
  949. 36:38here to there, pull it out and then like
  950. 36:40you know save it somewhere else. Instead
  951. 36:42you could just call on the library and
  952. 36:44it just does that all without you
  953. 36:46actually having to code it all out
  954. 36:47yourself. Um, and then it compiles all
  955. 36:50that code and then it gives the actual
  956. 36:53instructions to the GPU. So, the thing
  957. 36:55to know about KOD because it's not
  958. 36:57technically a programming language, uh,
  959. 36:59which I really always want to call it a
  960. 37:01programming language because it just
  961. 37:03sounds like it should be, but it's not.
  962. 37:05Instead, what it is is it's a
  963. 37:07programming platform. And what it allows
  964. 37:10is the programming language to basically
  965. 37:13get translated in a way to give
  966. 37:14instructions to the chip itself. Because
  967. 37:17keep in mind these instructions at the
  968. 37:19chip layer are literally moving
  969. 37:22electricity around transistors, right?
  970. 37:24And a chip may have millions of
  971. 37:26transistors. And so what CUDA is doing
  972. 37:28is it's taking this program and it's
  973. 37:31saying here's how you need to move all
  974. 37:32of the electricity around the chip in
  975. 37:35order to uh execute this program. So the
  976. 37:38feedback loop though that exists between
  977. 37:40design and CUDA is that there's more
  978. 37:43efficient ways to move electricity
  979. 37:46around a chip. basically to do these
  980. 37:48certain functions on a chip. And so over
  981. 37:51time in every iteration they learn, oh
  982. 37:54here's a better way that we can optimize
  983. 37:56the chip to do this function. Let's
  984. 37:58change that in the software. Let's uh
  985. 38:00tell it to to change the path it's
  986. 38:02taking basically on the actual chip
  987. 38:04level. On top of that, they could say,
  988. 38:06oh well, what if we design the chip
  989. 38:08differently? Does that make the software
  990. 38:10more optimized? And so that's the
  991. 38:12feedback loop between the two and why
  992. 38:13this was so so important historically.
  993. 38:16It still is, but this is what's allowed
  994. 38:18um uh different programmers to actually
  995. 38:20get the best performance on the chip.
  996. 38:23You it does require still that software
  997. 38:26layer to squeeze it out. So um the thing
  998. 38:29though that's been happening is you know
  999. 38:31you got these very big companies like
  1000. 38:34Google, like Amazon, uh AMD and all of
  1001. 38:37them have decided that they cannot be
  1002. 38:39beholden to a single proprietary
  1003. 38:42language. So Google has, you know, their
  1004. 38:45whole TPU stack, but they also have
  1005. 38:47something called XLA that's going to be
  1006. 38:49their own kind of version of this that's
  1007. 38:51their own compiler. So it can take uh
  1008. 38:53PyTorch, you could take Jack's code, it
  1009. 38:56can compile it into instructions that
  1010. 38:58then gets sent to a TPU. So very similar
  1011. 39:02uh to CUDA and it could cut CUDA out of
  1012. 39:04the process entirely. Uh Amazon, they
  1013. 39:06have Neuron, so that's their version of
  1014. 39:08the same thing. Uh AMD has Rockm that's
  1015. 39:12their version too and this is uh open
  1016. 39:14source as well. So uh it's shared
  1017. 39:16amongst anyone who wants to use it. Um
  1018. 39:18OpenAI also has their own programming
  1019. 39:21language platform and compiler called
  1020. 39:24Triton. And so you could write GPU
  1021. 39:27programs very similar to how you could
  1022. 39:29write uh Python code and then it could
  1023. 39:31compile for either CUDA or RockM which
  1024. 39:34is AMD's version. So the idea OpenAI had
  1025. 39:37was um let's help get this open-sourced
  1026. 39:40uh uh programming language very popular.
  1027. 39:43So there's a new ecosystem around it.
  1028. 39:45Then well the decision to use either
  1029. 39:47CUDA for Nvidia chips or Rockm which
  1030. 39:49could be AMD chips since it's open
  1031. 39:51source it could also be other chips too.
  1032. 39:53Um the thing to keep in mind though is
  1033. 39:55that just because something can compile
  1034. 39:59for either CUDA or rockm which just
  1035. 40:01again to be clear what I mean by that is
  1036. 40:03you're taking the pietorch the actual
  1037. 40:05language then when you're compiling it
  1038. 40:07you're turning it into actual
  1039. 40:09instructions for uh the chip. So just
  1040. 40:12because you can compile for CUDA or
  1041. 40:14rockm it doesn't mean you're actually
  1042. 40:16optimizing for the best performance. So
  1043. 40:19very often you kind of have to write for
  1044. 40:21one or the other because the chips are
  1045. 40:23different, right? The the language is
  1046. 40:25different, the architecture is
  1047. 40:26different. Now the thing that is really
  1048. 40:28changed in my opinion and I mentioned
  1049. 40:30this in my AMD video, which by the way
  1050. 40:33that if any of this is kind of
  1051. 40:34confusing, I feel like I kind of had a
  1052. 40:36little more patience and explained a lot
  1053. 40:38of things even simpler in that video.
  1054. 40:40This one I knew would be really a long
  1055. 40:42video, so I didn't do the same thing
  1056. 40:44here, but definitely you should watch
  1057. 40:45that one, too. But in that video I
  1058. 40:47talked about one of the key risks I saw
  1059. 40:49to CUDA being AI. And what I mean by
  1060. 40:52that is AI is very good at translation
  1061. 40:56and it's very good at optimization.
  1062. 40:59Whereas you know before if you're
  1063. 41:02translating something you know it's like
  1064. 41:04a dictionary, right? It's like you you
  1065. 41:05look up this word and you see what it
  1066. 41:07translates to. And that's not a very
  1067. 41:08good way to translate things because
  1068. 41:10sometimes things are lost quote unquote
  1069. 41:12in translation as we say because a
  1070. 41:14direct translation doesn't always work.
  1071. 41:16But AI is smart enough to get the gist
  1072. 41:19of what you're saying and then translate
  1073. 41:22to that which is a similar thing uh with
  1074. 41:24chip architecture and programming for
  1075. 41:27chips. It could catch the gist of what
  1076. 41:29you're saying and then try to write to
  1077. 41:31that. On top of that, it could then run,
  1078. 41:34you know, once it compiles the code, it
  1079. 41:36has this instruction set. It could then
  1080. 41:38run something, you know, a thousand
  1081. 41:39times to see what's the best way to
  1082. 41:41actually run it and optimize it, learn
  1083. 41:44that, and then in the future, it knows
  1084. 41:45how best to actually run that that
  1085. 41:47function. And so AI is something that
  1086. 41:50makes the development of these uh
  1087. 41:53programming platforms a lot lot better.
  1088. 41:56On top of that, it could reduce the
  1089. 41:58switching costs for someone who did
  1090. 42:00program entirely in CUDA. Now, maybe
  1091. 42:03it's easier to translate that to a
  1092. 42:05different language. Now, I don't want to
  1093. 42:06overstate this. Translation's still not
  1094. 42:09perfect, and if you're talking about
  1095. 42:10millions of lines of code, it's going to
  1096. 42:12be super annoying trying to figure out
  1097. 42:14where the bugs are, how to optimize it,
  1098. 42:16and all that. But, as AI continues to
  1099. 42:18get better and better and does more and
  1100. 42:20more of this, I I do believe the CUDA
  1101. 42:22moat is shrinking. uh I at least
  1102. 42:25relative to where it was historically, I
  1103. 42:28think that this is a shrinking mode for
  1104. 42:30Nvidia. One other key thing to keep in
  1105. 42:32mind though is that especially when
  1106. 42:34you're dealing with kind of a a chip
  1107. 42:36supply shortage, getting the most
  1108. 42:38performance out of the chip is all the
  1109. 42:41more important, right? To get the best
  1110. 42:43uh cost per performance sort of basis.
  1111. 42:45And so having software optimized for the
  1112. 42:48chip matters even more in these
  1113. 42:50environments. and Nvidia plus CUDA still
  1114. 42:53tends to be the best uh in most
  1115. 42:55circumstances and they also continue to
  1116. 42:57improve every year, right? So they're
  1117. 42:59also not standing still. They also can
  1118. 43:01use AI to optimize this and they're
  1119. 43:03starting from a head start with also a
  1120. 43:05lot of real world AI workloads. So that
  1121. 43:08was the third consideration. The fourth
  1122. 43:10one is going to be training versus
  1123. 43:13inference. So, um, it was generally
  1124. 43:16thought that Nvidia had a better
  1125. 43:19position in training versus inference.
  1126. 43:22And that is because when you train, you
  1127. 43:24need all of your chips to kind of talk
  1128. 43:26to each other. You need to have very low
  1129. 43:28latency because if you're waiting for
  1130. 43:30like one chip to do something before the
  1131. 43:32next chip does it, then your training
  1132. 43:34just takes longer and longer. It becomes
  1133. 43:36more expensive. And so, uh, it was
  1134. 43:38thought that since Nvidia does have very
  1135. 43:40good networking, it could talk to all of
  1136. 43:42its chips, that it was better for
  1137. 43:44training, it had a little bit more of a
  1138. 43:46moat there. Whereas with inference, that
  1139. 43:48would be easier for uh, AS6 um, which is
  1140. 43:52applied um, specialized integrated
  1141. 43:54circuits. It's basically the specialized
  1142. 43:56chips that these companies are making
  1143. 43:59um, would be better for inference. And
  1144. 44:02so you know uh Amazon has tranium that
  1145. 44:05can take some of the inference workloads
  1146. 44:07a lot easier than it would be for them
  1147. 44:08to replace the training. And so with
  1148. 44:10training um Nvidia would be stronger.
  1149. 44:13Now the problem with that is that you
  1150. 44:15know while AI is going to always be
  1151. 44:17training models over time more and more
  1152. 44:19workloads are going to go to inference
  1153. 44:21right and so that would be negative for
  1154. 44:23Nvidia because maybe they keep training
  1155. 44:25but then if inference keeps getting
  1156. 44:27carved out that's a big market that
  1157. 44:28they're losing. Now what's kind of
  1158. 44:31changed more recently and Jensen Huang
  1159. 44:33has been uh clear on on this point is
  1160. 44:36that inference um is actually getting
  1161. 44:38more complex particularly with agentic
  1162. 44:41workflows because with agents maybe
  1163. 44:43you're coordinating a hundred different
  1164. 44:44things all at once and you're waiting
  1165. 44:46for one agent to finish before the next
  1166. 44:48one can step in. That sounds pretty
  1167. 44:50similar to training. And in these cases,
  1168. 44:52again, your um interconnections, your
  1169. 44:55your NVL link, you want the networking
  1170. 44:57uh to really be good so that all of
  1171. 44:59these chips can link and so you're not
  1172. 45:01waiting. And so inference is getting
  1173. 45:02more complex and it's starting to look a
  1174. 45:05little bit more like training in this
  1175. 45:07respect, in which case Nvidia has their
  1176. 45:09advantage there again because of the
  1177. 45:12networking aspect. And so in this
  1178. 45:13consideration, it seemed like this was
  1179. 45:15something that maybe was a risk. I
  1180. 45:17wanted to surface it in case this is an
  1181. 45:19argument that you heard somewhere. It
  1182. 45:21doesn't seem like this is really going
  1183. 45:23to be an issue for Nvidia. So that was
  1184. 45:25the fourth consideration. The fifth
  1185. 45:27consideration is going to be open versus
  1186. 45:29closed source model. So a open- source
  1187. 45:32model is a model that you can download
  1188. 45:34without any restrictions on it generally
  1189. 45:37speaking and you know it could be hosted
  1190. 45:39kind of anywhere. A closed model will be
  1191. 45:42open AI's model or Enthropics model
  1192. 45:44which is going to be hosted by them and
  1193. 45:46they get to decide where it's being run.
  1194. 45:48Now the problem with this is that also
  1195. 45:51means that they get a lot of power uh
  1196. 45:54over deciding where to send uh all of
  1197. 45:57that uh workloads basically. And so they
  1198. 46:01can decide to uh design their own chips
  1199. 46:04for instance and then direct workloads
  1200. 46:07to the chips that they have designed.
  1201. 46:08And this isn't hypothetical. Open AAI is
  1202. 46:11designing or has designed already their
  1203. 46:13own chip and they may be hosting it
  1204. 46:16themselves. Maybe they're still putting
  1205. 46:18it, you know, in partnership with
  1206. 46:19Microsoft. It doesn't really matter.
  1207. 46:21Nvidia is still getting cut out of that.
  1208. 46:23They're also more likely to go to
  1209. 46:24hyperscalers to send a lot of the
  1210. 46:25workload. So uh it's not a huge negative
  1211. 46:29in a world where you know closed models
  1212. 46:31become the most dominant way to access
  1213. 46:33AI but it does present more risk for
  1214. 46:35Nvidia because again they do have this
  1215. 46:37option of cutting them out. Both open
  1216. 46:39AAI and uh Enthropic are designing their
  1217. 46:41own chips. Having said that it doesn't
  1218. 46:44look like the world is really going in
  1219. 46:46that direction where there's just like
  1220. 46:47these couple closed models. Instead it
  1221. 46:49looks like there's a lot of different
  1222. 46:51models. Um, and even within that, even
  1223. 46:54if it is a world with just a lot of
  1224. 46:56closed models, Nvidia still, uh, you
  1225. 46:58know, looks like they'll have a position
  1226. 47:00there. And so, generally thinking, I
  1227. 47:02don't think this is the biggest risk,
  1228. 47:03but the more powerful closed models
  1229. 47:06there are, the more ability they have to
  1230. 47:10basically create their own chips. Um,
  1231. 47:12and also, you know, pressure Nvidia down
  1232. 47:14the line, right? Nvidia doesn't want
  1233. 47:16very concentrated, powerful customers.
  1234. 47:18Instead, they want a bunch of Neoclouds
  1235. 47:20that they can, you know, go ahead and
  1236. 47:22give them chips and they say, "Oh my
  1237. 47:24god, thank you so much. You know, we'll
  1238. 47:25pay whatever." And that's the business
  1239. 47:28that they want. They don't want to sell
  1240. 47:29to customers that then say, "Oh, well,
  1241. 47:31we're going to design our own chip and
  1242. 47:33cut you out." And the result of that is
  1243. 47:34Nvidia has to constantly continue to to
  1244. 47:37move up in performance to justify the
  1245. 47:40premium of their chips. So, that was the
  1246. 47:42fifth consideration. The last
  1247. 47:44consideration has to do with chips
  1248. 47:46obsolescence. So it was generally
  1249. 47:48thought that uh Nvidia's uh GPUs would
  1250. 47:53be more general purpose, more
  1251. 47:55programmable. They could be good for
  1252. 47:57multiple different workloads, training,
  1253. 47:59inference, whatever else. And as a
  1254. 48:01result of that, they would have longer
  1255. 48:03shelf lives and uh better useful lives
  1256. 48:06too as well. And the reason why that's
  1257. 48:08important is because when you're
  1258. 48:09building a data center, you're going to
  1259. 48:11want to put into the data center
  1260. 48:13something that's going to last a long
  1261. 48:14time so you know you're not wasting your
  1262. 48:16money. So for that reason, you may be
  1263. 48:18less likely to buy something like uh you
  1264. 48:20know an Amazon tranium chip if they were
  1265. 48:23selling them or some of these other uh
  1266. 48:25specialized purpose chips that you could
  1267. 48:26get that maybe are cheaper in the short
  1268. 48:28run to run inference but then become
  1269. 48:31more expensive in the long run. So, this
  1270. 48:34was thinking for a while that that uh
  1271. 48:36Nvidia's GPUs would have better useful
  1272. 48:39lives, less obsolescence risk. I think
  1273. 48:41it's generally true, but I think
  1274. 48:43Google's TPU is not that far off, and I
  1275. 48:46think that other chips also won't be
  1276. 48:48that far off in the future, but I do
  1277. 48:50still think this is a little bit of of
  1278. 48:52moderate advantage for Nvidia currently
  1279. 48:54is that their chips are less likely to
  1280. 48:56go obsolete. So people are more
  1281. 48:58comfortable buying them for their data
  1282. 49:00centers knowing that they'll get that
  1283. 49:01full useful life out of them. Um in this
  1284. 49:05environment though with everything so
  1285. 49:07compute constrained it almost doesn't
  1286. 49:09fully matter as much anymore but that is
  1287. 49:12another consideration. Now let us get
  1288. 49:14into Nvidia's moes and so there still is
  1289. 49:17the software the cudas there it's
  1290. 49:20formidable. I did say though I think
  1291. 49:22that moat is starting to erode a little
  1292. 49:24bit but that doesn't mean it's eroded. I
  1293. 49:26just see it not being as strong in
  1294. 49:27several years from now as it is uh you
  1295. 49:30know in the past. But having said that,
  1296. 49:32it still is something that's formidable,
  1297. 49:34hard to overcome because as of now, they
  1298. 49:36still generally do get the best
  1299. 49:38performance out of their chips because
  1300. 49:40of their language and competitors do
  1301. 49:43still have to close that gap. Uh the
  1302. 49:45next thing that's worth noting is going
  1303. 49:46to be the system level. This is kind of
  1304. 49:49the biggest thing worth pointing out is
  1305. 49:51that they still have the best
  1306. 49:53performance on a systems level. that is
  1307. 49:56the cheapest to run. So as long as that
  1308. 49:58continues to be the case, that is the
  1309. 50:00simplest and best selling point for
  1310. 50:02Nvidia. It's very similar to Meta going,
  1311. 50:04you know, we have the highest return on
  1312. 50:06ad spend. Nvidia can say, you know, we
  1313. 50:08have the best performance on a per cost
  1314. 50:11basis, per token basis. And that is
  1315. 50:13enough of a selling point on its own as
  1316. 50:16long as that continues to be the case.
  1317. 50:18It's hard though, right? You have to
  1318. 50:19continue to execute. I'll say more about
  1319. 50:21that in a moment. Um the other advantage
  1320. 50:24is they have a lot of real world AI
  1321. 50:26workloads. This shows up as an advantage
  1322. 50:28for them because that means they're
  1323. 50:30better able to optimize their chips,
  1324. 50:31their software, and then when it
  1325. 50:34actually gets into a data center, the
  1326. 50:35customer actually sees the performance
  1327. 50:38uh actually work the way it's expected
  1328. 50:40to and it continues to improve over
  1329. 50:42time. So that's what allows them to
  1330. 50:44continue to have top performance. Uh the
  1331. 50:46next one's going to be networking and
  1332. 50:47integration. Uh I mentioned a few times
  1333. 50:50this NVL link thing that's pretty
  1334. 50:52important to get all the chips to talk
  1335. 50:53together. NV switch uh helps orchestrate
  1336. 50:56all of them. Uh they also have other um
  1337. 50:59networking technologies so the servers
  1338. 51:01talk to each other. All of this together
  1339. 51:02is what has moved them from just
  1340. 51:04competing on a chip basis to an entire
  1341. 51:06you know AI systems basis and has
  1342. 51:10allowed them to maintain their
  1343. 51:11performance because it's one thing to
  1344. 51:13build a chip. It's another thing to also
  1345. 51:15build networking too on top of that and
  1346. 51:18you know they have to outsource a lot of
  1347. 51:19this to to Broadcom to Arista and then
  1348. 51:22when you're outsourcing it and it's not
  1349. 51:23integrated together it just doesn't
  1350. 51:25quite work as well uh as if it was fully
  1351. 51:28integrated together. So that is going to
  1352. 51:30be another advantage Nvidia has. Uh the
  1353. 51:33next one is that their uh timing of
  1354. 51:35improvements have sped up over time. It
  1355. 51:37used to be every two years they came out
  1356. 51:38with uh new updates on on CUDA and new
  1357. 51:41uh chip architectures. now it's every
  1358. 51:43one year. So, they've sped that up. So,
  1359. 51:45that's been um kind of an operational
  1360. 51:48advantage, just the fact that they've
  1361. 51:49been able to do that um and have an
  1362. 51:52organization that, you know, can have
  1363. 51:54enough R&D spend and and be an R&D
  1364. 51:56organization that's capable of that.
  1365. 51:58That's kind of their their advantage,
  1366. 51:59but it is, you know, execution uh
  1367. 52:02dependent very much so. And then the
  1368. 52:03last one's going to be the developer
  1369. 52:05ecosystem. So this is similar to CUDA
  1370. 52:08but this is the fact that a lot of
  1371. 52:09people are familiar with CUDA know how
  1372. 52:11to use it and will thus prefer to use it
  1373. 52:13and when they use it they'll be you know
  1374. 52:15higher performing. So that that's kind
  1375. 52:18of uh a bunch of the advantages that
  1376. 52:19Nvidia has. The most important one
  1377. 52:21though again is going to be that they
  1378. 52:23have the highest performance on a per
  1379. 52:26cost per token basis. Uh as long as
  1380. 52:29that's the case they're going to be
  1381. 52:30fine. But making sure that's the case
  1382. 52:33far out in the future, that that's a
  1383. 52:34pretty hard thing to do because they
  1384. 52:36have to continue to execute and
  1385. 52:38outperform competition in order to
  1386. 52:41justify their price premium. Whereas a
  1387. 52:43competitor uh like with Google and the
  1388. 52:45TPUs, they could just undercut them on
  1389. 52:47price. And so they could be a little
  1390. 52:49less performant, but since they're
  1391. 52:51cheaper, the performance per cost is the
  1392. 52:54same. So, Nvidia has to continue to do
  1393. 52:57better than everyone else in order to
  1394. 52:58maintain their margins, which is also
  1395. 53:00kind of a key risk as you can imagine.
  1396. 53:03So, uh why don't we actually get into
  1397. 53:05risk? I just gave you one there and some
  1398. 53:07of the other ones we're going to talk
  1399. 53:08about are going to be this circular
  1400. 53:10financing thing that people talk a lot
  1401. 53:12about. Um I'm going to just tell you
  1402. 53:15it's interesting to talk about some of
  1403. 53:18it. It does remind me a little bit of
  1404. 53:20maybe some of the things you hear about
  1405. 53:22during the the you know uh tech uh
  1406. 53:24bubble in the 2000s, but it's nowhere
  1407. 53:27near that level just in terms of dollar
  1408. 53:29amounts. Even though these are large
  1409. 53:30dollars relative to how big Nvidia is
  1410. 53:33and how much cash they generate. I told
  1411. 53:35you they have 200 billion in earnings in
  1412. 53:38just this last year. That is more or
  1413. 53:40basically as much as everything we're
  1414. 53:42talking about here. You're talking about
  1415. 53:43one year of earnings. So circular
  1416. 53:46financing, they've done about hundred
  1417. 53:47billion in equity investments. A lot of
  1418. 53:49these investments that they do do
  1419. 53:51involve them then turning around and
  1420. 53:52buying their chips. The reason why they
  1421. 53:55do that very often is because Nvidia
  1422. 53:57wants to continue to entrench in the
  1423. 53:59ecosystem. So give a couple billion
  1424. 54:01dollars to Coree because they want more
  1425. 54:03of these Neo clouds and then they go and
  1426. 54:05turn around and spend that on chips.
  1427. 54:07Now, Coreweave has a couple billion
  1428. 54:09dollars of Nvidia chips that then get
  1429. 54:11rented out to other um customers that
  1430. 54:14now are stuck using Nvidia chips and and
  1431. 54:17building programs on Nvidia chips. And
  1432. 54:20on top of that, uh Nvidia does still
  1433. 54:22get, you know, an equity stake in this
  1434. 54:24company and then they got a margin when
  1435. 54:26they bought their uh chips on top of
  1436. 54:28that. So, uh it's a little confusing to
  1437. 54:30think about, but I'm generally just not
  1438. 54:33that worried about it because of the
  1439. 54:34size. So, it's been a 100red billion in
  1440. 54:36equity investments. so far. And by the
  1441. 54:38way, a lot of that too is an anthropic
  1442. 54:40and open AAI stake. That's uh those
  1443. 54:42together, my estimates are are probably
  1444. 54:44around like 30 to 40 billion each. Uh so
  1445. 54:47that's like most of it right there. Uh
  1446. 54:49you know, another 25 billion in future
  1447. 54:51commitments. Not the biggest amount. Um
  1448. 54:5336 billion in Neocloud compute um
  1449. 54:56commitments. So this is basically
  1450. 54:57guaranteeing uh compute contracts so
  1451. 55:00that there's going to be demand for it.
  1452. 55:02Nvidia is just so optimistic on how much
  1453. 55:06AI demand there is just generally
  1454. 55:08speaking that this is kind of very
  1455. 55:11little relative to what they expect
  1456. 55:12there to be and especially in this
  1457. 55:14environment where all the compute that
  1458. 55:16exists is basically being rented out and
  1459. 55:18being rented out at very high uh lease
  1460. 55:20rates. I can understand why they'd feel
  1461. 55:23comfortable uh guaranteeing that sort of
  1462. 55:25commitment that they'd be able to lease
  1463. 55:26out that compute. I don't see that as
  1464. 55:28being um very worrisome either. uh there
  1465. 55:31these credit guarantees they're doing.
  1466. 55:33It's an interesting way to kind of
  1467. 55:34finance things and raise money. Those
  1468. 55:36are uh estimated to be around hundred
  1469. 55:38billion. So again, you add all of this
  1470. 55:41up together and you're talking about one
  1471. 55:43year of earnings. So this is a very cash
  1472. 55:47flow generative business. And because of
  1473. 55:49that, it can sustain all of these
  1474. 55:51investments and throwing all of these
  1475. 55:53billions around without it really, you
  1476. 55:55know, being that impactful uh to them.
  1477. 55:58if it gets much bigger, you know, then
  1478. 56:00that can become worrisome as it stands
  1479. 56:02today. I don't personally see it's the
  1480. 56:05biggest risk with Nvidia. Now, I I think
  1481. 56:07one of the biggest risk uh is going to
  1482. 56:10be, as I was kind of alluding to, this
  1483. 56:11is a pretty execution dependent
  1484. 56:13business. They have to continue to stay
  1485. 56:15on top of all of this. And you saw all
  1486. 56:17of these trends that were happening over
  1487. 56:19time that Jensen Wong was pretty good at
  1488. 56:21at catching. you know, the shift from uh
  1489. 56:23just chip to the importance of
  1490. 56:25networking uh to this whole AI factory
  1491. 56:28to AI sovereignty to um funding all of
  1492. 56:31these neo clouds to doing these creative
  1493. 56:34financing arrangements to make sure
  1494. 56:35Nvidia chips continue to get sold. Um
  1495. 56:38all of this has helped really entrench
  1496. 56:41them. uh is, you know, even if some of
  1497. 56:43the financing stuff makes some people
  1498. 56:44nervous, it did get a lot more of their
  1499. 56:47chips sold. And the more chips they get
  1500. 56:48sold now and the more AI kind of
  1501. 56:51continues to get developed on them, the
  1502. 56:53more their ecosystem is kind of
  1503. 56:55entrenched with AI and and the harder it
  1504. 56:58could be to switch off in the future. Uh
  1505. 57:00you want all of these models training on
  1506. 57:02Nvidia chips. you want them running on
  1507. 57:04Nvidia chips because Nvidia then in turn
  1508. 57:06gets better at optimizing its chips for
  1509. 57:09all these AI workloads and it's just
  1510. 57:11harder and harder to catch up to them.
  1511. 57:12So it it makes sense everything he's
  1512. 57:15done. I think he's done an absolute
  1513. 57:16phenomenal job setting this business up
  1514. 57:18um competitively and he's out. He's
  1515. 57:21gone. So the the new CEO is Colette
  1516. 57:23Crest. Uh she's going to be an interim
  1517. 57:25CEO until they find a replacement. She's
  1518. 57:28been the CFO since uh 2012. Uh, but it's
  1519. 57:32going to be very hard to replace Jensen
  1520. 57:34Hong for no less reason other than the
  1521. 57:36fact that he had 60 direct reports. I
  1522. 57:39can't even imagine how confusing uh of a
  1523. 57:42business this will be now. 60 people
  1524. 57:44used to directly report to Jensen Hang
  1525. 57:47and now they're, you know, probably
  1526. 57:50reporting to her. But this is not a
  1527. 57:52structure most people are equipped to
  1528. 57:54handle or would want to handle. And this
  1529. 57:56is not a normal way to run a business.
  1530. 57:58And so I can imagine a ton of confusion
  1531. 58:01and I can also imagine now multiple
  1532. 58:03layers being moved above these 60
  1533. 58:05reports and them resenting that and some
  1534. 58:07of them probably leaving as a result of
  1535. 58:09that. Uh because if you used to directly
  1536. 58:12talk to the CEO to the the founder
  1537. 58:14himself and now you have different
  1538. 58:15people in between you that's
  1539. 58:17demotivating. So that's going to be a
  1540. 58:19problem. That's going to be going to be
  1541. 58:21definitely a lot of you know operational
  1542. 58:24uh issues to figure out within Nvidia.
  1543. 58:26And th this is going to be something
  1544. 58:29that I think is going to eventually lead
  1545. 58:31to the missing things. I don't see how
  1546. 58:33it can't. And you know, you look at all
  1547. 58:35of these very brazen and kind of
  1548. 58:38ambitious things that Nvidia has done.
  1549. 58:40You know, hundred billion dollar equity
  1550. 58:42investments, hundred billion dollar
  1551. 58:44credit financing guarantees. You can't
  1552. 58:46really do these things unless you're
  1553. 58:48like the founder of the company. It's
  1554. 58:50very hard, you know, as an interim uh
  1555. 58:52CEO uh or even whoever comes next to be
  1556. 58:55as brazen as uh Jensen Hang was with all
  1557. 58:59that. So, uh I see him missing things,
  1558. 59:01too. It's also less likely, you know,
  1559. 59:03they're going to be looking at
  1560. 59:04acquisitions, which have been important
  1561. 59:05for them, uh for for the near term. It's
  1562. 59:08kind of like a lot of things just get
  1563. 59:10paused and they need to figure out how
  1564. 59:12to just run this business again without
  1565. 59:14Jensen being a part of it, which is
  1566. 59:17should not be understated how disruptive
  1567. 59:19that can be and how much that could lead
  1568. 59:22to their competitive edge um faltering.
  1569. 59:24And so I think any way you cut it,
  1570. 59:26Jensen Hong not being there does dim the
  1571. 59:29prospects for Nvidia in the future. I I
  1572. 59:32think there's a much higher risk of
  1573. 59:35disruption essentially without someone
  1574. 59:37like him there because again he was
  1575. 59:40really looking over the corners to the
  1576. 59:42future multiple times with the Neoclouds
  1577. 59:45with the financing arrangements moving
  1578. 59:47into networking the AI factories
  1579. 59:49sovereign AI uh the acquisitions with
  1580. 59:52Melanex getting them into networking the
  1581. 59:54Grock licensing getting them a position
  1582. 59:56in some leading uh inference chips uh
  1583. 59:59hugging face getting them a position
  1584. 1:00:00with uh the open models and all of that
  1585. 1:00:02and and kind of seeing where a lot of
  1586. 1:00:04different ways can go and balancing all
  1587. 1:00:06these different uh interest in this
  1588. 1:00:08ecosystem. So definitely a loss uh for
  1589. 1:00:10the business him not being there. also
  1590. 1:00:13just a big risk because now you don't
  1591. 1:00:15know who's running it and execution for
  1592. 1:00:17this business is very very important
  1593. 1:00:19right now because there's a lot of
  1594. 1:00:22competitors who would love to take their
  1595. 1:00:23lunch and you know I talked about how
  1596. 1:00:25now they're trying to move to this
  1597. 1:00:27one-year cycle or they move to a
  1598. 1:00:29one-year cycle of uh new chip design and
  1599. 1:00:32development now this reminds me a little
  1600. 1:00:34bit like Intel who used to have
  1601. 1:00:35something called tick talk where uh tick
  1602. 1:00:39was going to be manufacturing
  1603. 1:00:40improvements then talk was going to be
  1604. 1:00:43architecture uh chip architecture
  1605. 1:00:45improvements and so it went tick tock
  1606. 1:00:48and they did one improvement every two
  1607. 1:00:50years. So there was a full cycle done on
  1608. 1:00:53both of them every two years and they
  1609. 1:00:55did that for 10 years and then it
  1610. 1:00:57stopped working and it's because it's
  1611. 1:00:59very hard to continue to execute at a
  1612. 1:01:02very high level and accomplishing what
  1613. 1:01:04they're trying to accomplish where they
  1614. 1:01:06have to constantly iterate on the last
  1615. 1:01:08design and do better and then do better
  1616. 1:01:11than the competition because they're
  1617. 1:01:12priced at a premium and they need to
  1618. 1:01:15justify that premium by being better
  1619. 1:01:16than the competition. That is the race
  1620. 1:01:19that they are running. And tripping up
  1621. 1:01:21even once one year, they lose that lead
  1622. 1:01:24forever. You never get it back. If a
  1623. 1:01:26competitor catches up to you, it's going
  1624. 1:01:29to be near impossible to ever take that
  1625. 1:01:31lead again. And in fact, it's more
  1626. 1:01:32likely that competitor surpasses you.
  1627. 1:01:34And so that is kind of the key risk here
  1628. 1:01:37because if they do get surpassed and
  1629. 1:01:38they no longer have that best
  1630. 1:01:40performance, they're going to lose their
  1631. 1:01:42pricing power, that that margin that
  1632. 1:01:44they've been able to to maintain, that
  1633. 1:01:4665% operating margin. So that is another
  1634. 1:01:49one of the key risks. And then of course
  1635. 1:01:51I have to just tack on here you know
  1636. 1:01:53general AI potential bubble risk all of
  1637. 1:01:56that. Um it could be data centers stop
  1638. 1:01:59being built because of regulation
  1639. 1:02:00reasons. You know, there's some fears
  1640. 1:02:02that with uh a lot of uh Democratic uh
  1641. 1:02:05senators, governors, uh Congress that
  1642. 1:02:08there could be some legislation that
  1643. 1:02:10either puts a damper on AI development
  1644. 1:02:12or at the individual state level uh more
  1645. 1:02:15of these AI projects uh get stopped uh
  1646. 1:02:18particular fear about that in Texas
  1647. 1:02:20where a lot of these uh builds were
  1648. 1:02:22happening. So, that's just uh something
  1649. 1:02:24else that that could weigh on them
  1650. 1:02:26because if all of a sudden these data
  1651. 1:02:28centers are slowing down, then all of
  1652. 1:02:30this slowly starts to unwind because all
  1653. 1:02:33these uh capex commitments, well, they
  1654. 1:02:35don't come through because they're
  1655. 1:02:37you're not going to build anything and
  1656. 1:02:38so they're not going to buy any chips.
  1657. 1:02:40Um and so that is kind of a big risk
  1658. 1:02:42there. And then even aside from that,
  1659. 1:02:44this whole thing could still be
  1660. 1:02:46cyclical, right? uh it could be the case
  1661. 1:02:48that you can't spend hundreds of
  1662. 1:02:50billions of dollars on on capex and
  1663. 1:02:52continue to see a good return on
  1664. 1:02:54investment and as soon as that starts to
  1665. 1:02:56falter they'll pull back on their capex
  1666. 1:02:58and everything will start to go the
  1667. 1:03:00other way. Um and that's a very hard
  1668. 1:03:03thing to predict. You know right now I
  1669. 1:03:05think there's a lot of people that are
  1670. 1:03:06generally optimistic about the return on
  1671. 1:03:09investment. Um there's certainly some
  1672. 1:03:11evidence if I look at Meta uh some of
  1673. 1:03:13the hyperscalers, Amazon, Google that
  1674. 1:03:15they are getting a good return on
  1675. 1:03:17investment, but uh there's still a lot
  1676. 1:03:20of experimental use and and money being
  1677. 1:03:22spent on AI that maybe when people are
  1678. 1:03:25not experimenting and really demand uh a
  1679. 1:03:28full return on that that no longer uh
  1680. 1:03:31really starts flowing through into those
  1681. 1:03:32hyperscalers. And so ultimately what
  1682. 1:03:34really matters is how much uh economic
  1683. 1:03:37activity sustainable economic activity
  1684. 1:03:40AI is really bringing in and there's
  1685. 1:03:42debate around that. Uh these are hard
  1686. 1:03:45things to actually track. Uh generally
  1687. 1:03:47speaking, my opinion is that I I do
  1688. 1:03:49think it's creating a lot of economic
  1689. 1:03:51activity. But whether or not it is fully
  1690. 1:03:55um enough to sustain, you know, the
  1691. 1:03:57hundreds of billions, you know,
  1692. 1:03:59trillions that are going to be spent on
  1693. 1:04:00this in the coming years, that's a
  1694. 1:04:02harder question to ask. Um, if I look at
  1695. 1:04:05again just like Meta for instance, I
  1696. 1:04:06could see the return on investment
  1697. 1:04:08there, but I don't know about, you know,
  1698. 1:04:11an individual business spending on
  1699. 1:04:12enthropic tokens, whether or not they're
  1700. 1:04:14getting a return on that spend or if
  1701. 1:04:16they find out, wow, we wasting a ton of
  1702. 1:04:18tokens, uh, wasn't really worth it.
  1703. 1:04:20We're overpaying. Hard to say. I
  1704. 1:04:23generally am optimistic about AI, but I
  1705. 1:04:25also just know history and historically
  1706. 1:04:29every time there's ever been a capital
  1707. 1:04:31cycle, we've always overshot. We've
  1708. 1:04:34always overbuilt demand. That's always
  1709. 1:04:36happened. So, is this going to be the
  1710. 1:04:38first time in history where that doesn't
  1711. 1:04:39happen? I don't know. Or maybe this is
  1712. 1:04:42just a very, very long cycle. I also
  1713. 1:04:44don't know. So, uh these are just things
  1714. 1:04:46to consider. I talk more about a lot of
  1715. 1:04:48this though in this video I had on uh
  1716. 1:04:51whether or not we're in an AI bubble.
  1717. 1:04:52So, you can check that one out. But now,
  1718. 1:04:54I'm going to move to the valuation. But
  1719. 1:04:56first, I just want to say if you do have
  1720. 1:04:58questions uh on Nvidia, anything else
  1721. 1:05:00you feel like I didn't do uh a full job
  1722. 1:05:02covering it, uh drop a question below.
  1723. 1:05:05I'm going to have a podcast covering
  1724. 1:05:07Nvidia next week and we could cover
  1725. 1:05:09those questions on that podcast. Uh if
  1726. 1:05:12you don't follow our podcast right now
  1727. 1:05:14though, it's called the synopsis. uh you
  1728. 1:05:16could just throw into your search bar
  1729. 1:05:17the synopsis pod. It'll show up. Hit
  1730. 1:05:20subscribe and then you'll uh make sure
  1731. 1:05:22that you won't miss that video when we
  1732. 1:05:23talk about all your Nvidia questions.
  1733. 1:05:25Okay, so moving on to valuation. And I
  1734. 1:05:29did tell you I was going to disappoint
  1735. 1:05:30you a little bit, but I won't do it
  1736. 1:05:31right now. I'll do it in a moment. Uh so
  1737. 1:05:34at $167 stock price, it's a $4 trillion
  1738. 1:05:37market cap. They have $192 billion in
  1739. 1:05:40trailing earnings. That gets you an LTM
  1740. 1:05:43uh earnings multiple of 21 times. If
  1741. 1:05:45you're looking at forward estimates,
  1742. 1:05:47that's going to be about 14 times
  1743. 1:05:49earnings right now. It trades at that's
  1744. 1:05:50going to be the next 12 months. And then
  1745. 1:05:52if you pull up 2027 estimates, that is
  1746. 1:05:5510 times. So trading at 10 times 2027
  1747. 1:05:58estimates. Now what uh Jensen Hong has
  1748. 1:06:01said is that uh they see 70% capacity
  1749. 1:06:04growth for next year, but they see
  1750. 1:06:06demand growth as much as 100%. So what
  1751. 1:06:09he was alluding to there is if they
  1752. 1:06:10could grow capacity more than 70%, they
  1753. 1:06:12could potentially get up to 100%. Now
  1754. 1:06:15the reason why I said I was going to
  1755. 1:06:16disappoint you with the valuation and I
  1756. 1:06:18am going to do this other valuation
  1757. 1:06:20method in a moment but I think this is
  1758. 1:06:22the more important conversation which is
  1759. 1:06:24that before you invest in Nvidia you
  1760. 1:06:28have to decide and consider how much you
  1761. 1:06:31believe that they are still a cyclical
  1762. 1:06:33chips business or whether or not things
  1763. 1:06:36have really changed from them and
  1764. 1:06:37they're going to be a secular grower
  1765. 1:06:39because if it is a cyclical business it
  1766. 1:06:43becomes very hard to value. And I
  1767. 1:06:45mentioned in the intro, a lot of
  1768. 1:06:46investors assue uh cyclical businesses
  1769. 1:06:49because you can't really value them
  1770. 1:06:52easily. You have to basically just
  1771. 1:06:54underpay a lot and have a large margin
  1772. 1:06:56of safety. And if that's the case, then
  1773. 1:07:00it could work out and you kind of hope
  1774. 1:07:02for the best. But it's it's very hard to
  1775. 1:07:04do because imagine you're doing a DCF of
  1776. 1:07:07a company and every year the earnings
  1777. 1:07:10either go up or down and you don't
  1778. 1:07:12always know whether it's up or down. It
  1779. 1:07:14becomes very hard to have confidence in
  1780. 1:07:16it. Whereas if a company's earnings went
  1781. 1:07:18up every year, you could say, "All
  1782. 1:07:21right, I don't know how much they're
  1783. 1:07:22going to go up, but they're going to go
  1784. 1:07:23up a little bit. Maybe I'll be
  1785. 1:07:25conservative, assume a little bit less."
  1786. 1:07:27And you could get comfortable with a
  1787. 1:07:28number around that, right? And so that
  1788. 1:07:32is why it is so much harder to value a
  1789. 1:07:34cyclical business. And to be honest,
  1790. 1:07:37I've never seen good valuations done for
  1791. 1:07:40cyclical businesses because inherent in
  1792. 1:07:42the idea being cyclical is that you
  1793. 1:07:45don't know exactly where you are in the
  1794. 1:07:47cycle and exactly what trough or peak
  1795. 1:07:51earnings are going to be. These are
  1796. 1:07:52estimates people make and then they put
  1797. 1:07:54a multiple on these estimates. So, if
  1798. 1:07:56you're doing it on uh peak earnings, you
  1799. 1:07:58need to put a pretty low multiple on it,
  1800. 1:08:00or you could put a higher multiple on
  1801. 1:08:02trough earnings. And these multiples
  1802. 1:08:04though in and of themselves have
  1803. 1:08:07different earnings growth assumptions.
  1804. 1:08:09And so, it makes it very confusing.
  1805. 1:08:11Honestly, if you did a DCF of a company
  1806. 1:08:13that's cyclical, you would see some
  1807. 1:08:15years, you know, you'd grow revenues,
  1808. 1:08:17other years you'd shrink revenues. And
  1809. 1:08:18it becomes a very hard thing to do
  1810. 1:08:20because what years are you going to
  1811. 1:08:22decide to start shrinking revenues? what
  1812. 1:08:23years are you going to decide to start
  1813. 1:08:25growing them? These are not things you
  1814. 1:08:27really have any idea how to answer. And
  1815. 1:08:29so it just it's a very tricky thing to
  1816. 1:08:31value a business like that. So generally
  1817. 1:08:34speaking, you just want to assume a very
  1818. 1:08:36very low multiple for a cyclical
  1819. 1:08:38business. And chip stocks, they have
  1820. 1:08:41traded at I think Micron at its low
  1821. 1:08:44traded at a three times earnings
  1822. 1:08:46multiple. And maybe that was too low.
  1823. 1:08:48But uh that just goes to show how low of
  1824. 1:08:51a multiple you could put when a business
  1825. 1:08:53is cyclical and it's at peak earnings.
  1826. 1:08:55That three times was at peak earnings.
  1827. 1:08:57If it's at trough earnings, uh maybe
  1828. 1:08:59it's a high singledigit multiple, maybe
  1829. 1:09:01a low double-digit multiple. And these
  1830. 1:09:04are just the way these things have
  1831. 1:09:05typically traded over time. It's hard to
  1832. 1:09:08ever say that's the right or wrong
  1833. 1:09:09multiple though because you can only do
  1834. 1:09:11it after the fact after you know whether
  1835. 1:09:13or not you know earnings did come back.
  1836. 1:09:16uh through that cycle now for at least
  1837. 1:09:18the next few years it looks like Nvidia
  1838. 1:09:21is in a pretty strong position uh there
  1839. 1:09:23of are the risks that we mentioned but
  1840. 1:09:25generally speaking you know if this AI
  1841. 1:09:27build continues to happen they're going
  1842. 1:09:29to be growing uh revenues still pretty
  1843. 1:09:32quickly and a lot so maybe you look out
  1844. 1:09:34a few years and you see what you think
  1845. 1:09:36they could be earning and a few years
  1846. 1:09:38from now you consider that to be peak
  1847. 1:09:40earnings if you're paying a low multiple
  1848. 1:09:42on that you could decide whether or not
  1849. 1:09:44that makes sense for you so Now, to not
  1850. 1:09:46totally leave you with nothing here, I
  1851. 1:09:48did my own revenue build uh where I with
  1852. 1:09:52my best analyst Claude uh started to try
  1853. 1:09:56to figure out how to estimate gigawatt
  1854. 1:09:59capacity out into 2031.
  1855. 1:10:02And so this is going to be total
  1856. 1:10:04gigawatts built. Then from there, you
  1857. 1:10:06translate to how much of that gigawatt
  1858. 1:10:08power is actually going to go run
  1859. 1:10:10through chips. Then from there you make
  1860. 1:10:13an assumption as to what Nvidia's market
  1861. 1:10:16share of that is going to be. And then
  1862. 1:10:18uh from there it's going to be revenue
  1863. 1:10:21per gigawatt. And I should have backed
  1864. 1:10:23up and said that uh if you've never
  1865. 1:10:24listened to Jensen Huang on an earnings
  1866. 1:10:26call he talks sometimes about uh revenue
  1867. 1:10:30per gigawatt. And it used to be 18
  1868. 1:10:32billion then it was 25 billion with
  1869. 1:10:34Blackwell and then lately it's now 40
  1870. 1:10:36billion he mentioned with Vera. And so
  1871. 1:10:39basically as Nvidia continues to have uh
  1872. 1:10:42more not just chips but networking
  1873. 1:10:44technology, they're able to take over
  1874. 1:10:46more and more of that data center,
  1875. 1:10:48provide more of all of the equipment you
  1876. 1:10:50need to run that data center. Their
  1877. 1:10:52revenue per gigawatt has gone up. So
  1878. 1:10:55that's kind of how he thinks about it.
  1879. 1:10:56So I just jumped on that and I built
  1880. 1:10:58this this revenue model based off of
  1881. 1:11:00that. So I kind of just walked you
  1882. 1:11:01through how that worked. You're going to
  1883. 1:11:03see on the screen now uh kind of the the
  1884. 1:11:05math and the the assumptions how that
  1885. 1:11:07all flows through. I'm going to just
  1886. 1:11:09talk about that base case scenario. So
  1887. 1:11:11uh you could see there it jumps from
  1888. 1:11:14data center gigawatts it converts to how
  1889. 1:11:16much uh is actually going to be on the
  1890. 1:11:18chips. Then it also adds in replacement
  1891. 1:11:20gigawatts. So the idea here is that and
  1892. 1:11:24this is actually very important to
  1893. 1:11:26understand um for Nvidia's potential um
  1894. 1:11:30recurring revenues essentially or how
  1895. 1:11:32durable their revenues will be is how
  1896. 1:11:34much of this gets replaced every four to
  1897. 1:11:37six years and so if chips have a four to
  1898. 1:11:39six year life something like that maybe
  1899. 1:11:41seven whatever it is and they get
  1900. 1:11:43replaced o over that period then you
  1901. 1:11:46have a very different business where you
  1902. 1:11:48do start seeing recurring revenue and it
  1903. 1:11:50reminds me a little bit of when I
  1904. 1:11:51invested in Apple and uh at the time
  1905. 1:11:54people would say oh well people are
  1906. 1:11:55going to buy an iPhone and then you know
  1907. 1:11:58they're not going to want another iPhone
  1908. 1:11:59for many years and I remember thinking
  1909. 1:12:02no like people are going to want you
  1910. 1:12:03know an iPhone every couple years
  1911. 1:12:05they're going to keep wanting the new
  1912. 1:12:06ones and even if maybe that gets a
  1913. 1:12:08little bit delayed over time it's
  1914. 1:12:11generally held true people still just
  1915. 1:12:13want the new the new phones and uh the
  1916. 1:12:15question is whether or not uh data
  1917. 1:12:17center GPUs are going to be similar if
  1918. 1:12:19they could continue to roll out new
  1919. 1:12:21performance uh GPUs that justified the
  1920. 1:12:24data centers replacing them. Um and they
  1921. 1:12:26could continue to get good performance
  1922. 1:12:28uh per uh cost per token on that um per
  1923. 1:12:32gigawatt, whatever metric you want to
  1924. 1:12:34put there on the denominator. Um then
  1925. 1:12:37they'll continue to replace them. And so
  1926. 1:12:39that's what this build has. It has
  1927. 1:12:41replacement gigawatts in there too. This
  1928. 1:12:42is something I've thought a lot about
  1929. 1:12:44though for Nvidia because uh it's kind
  1930. 1:12:46of like we're in this phase where you
  1931. 1:12:47you have this big buildout and
  1932. 1:12:49everything is growing and then there's
  1933. 1:12:51going to be a maintenance phase. But
  1934. 1:12:52when the maintenance phase happens,
  1935. 1:12:55revenues are going to drop a lot, right?
  1936. 1:12:57They're going to drop a lot because
  1937. 1:12:59you're you have a lot less growth and
  1938. 1:13:00then it's just you are replacing what's
  1939. 1:13:02already built. Having said that, if when
  1940. 1:13:05this happens, uh it's already a much
  1941. 1:13:08much bigger base than where they
  1942. 1:13:10currently are at, then even when you get
  1943. 1:13:12to that maintenance level, it could
  1944. 1:13:14potentially potentially uh justify, you
  1945. 1:13:18know, higher revenues than where they
  1946. 1:13:19are today. It seems possible possible
  1947. 1:13:22that they could be at, you know, a
  1948. 1:13:24higher, you know, maintenance uh level
  1949. 1:13:26of revenue than where they are in total
  1950. 1:13:28revenue today. I think that's a pretty
  1951. 1:13:30ambitious assumption. But even if they
  1952. 1:13:33are less, but it's now recurring revenue
  1953. 1:13:36that you have much higher confidence in,
  1954. 1:13:38then maybe it deserves a higher
  1955. 1:13:39multiple. Uh when people see that, okay,
  1956. 1:13:42so this business did, you know, contract
  1957. 1:13:44a little bit, but it didn't contract,
  1958. 1:13:46you know, that much and now it looks
  1959. 1:13:48more stable and now we're coming back
  1960. 1:13:50up. Okay, now we kind of have more of a
  1961. 1:13:52sense of what a cycle looks like. That's
  1962. 1:13:54something that could happen, too. So
  1963. 1:13:56anyway, in this build, we have the
  1964. 1:13:57replacement gigawatts. We assume
  1965. 1:13:59Nvidia's share of it. Um, all these
  1966. 1:14:01assumptions, by the way, this is a
  1967. 1:14:03framework. You could come up with your
  1968. 1:14:04own ideas. You could also literally run
  1969. 1:14:07this math that gets to revenue and then
  1970. 1:14:09you could do it for every year if you
  1971. 1:14:10wanted to and project it out and build
  1972. 1:14:12your own DCF that way. That's that's
  1973. 1:14:14something you could do. Um, I don't know
  1974. 1:14:16if it's going to get you to a better
  1975. 1:14:17result, but if you feel like modeling a
  1976. 1:14:20little bit to make you feel better. Um,
  1977. 1:14:22so that's going to be uh Nvidia share we
  1978. 1:14:25have 55%. This assumes 40 billion
  1979. 1:14:28gawatt. Seems a little high. Um, but
  1980. 1:14:30this is just an assumption. That gets
  1981. 1:14:32you to uh about a trillion in data
  1982. 1:14:35center revenue plus another 50 billion
  1983. 1:14:38in edge revenue. That's going to be a
  1984. 1:14:40$460 billion NOPAT uh figure. So it
  1985. 1:14:43calls this a base case. I think that
  1986. 1:14:45this is still uh seems like a lot to me.
  1987. 1:14:48A billion trillion dollars in revenue.
  1988. 1:14:50Sorry. Um, but I don't know, everyone's
  1989. 1:14:52throwing big numbers out there these
  1990. 1:14:54days. So, if you take this math and then
  1991. 1:14:57you want to put a range of multiples on
  1992. 1:14:59it, let's say you do 15 times multiple,
  1993. 1:15:02that'd get you about a 7 trillion
  1994. 1:15:04valuation. 20 times a multiple would be
  1995. 1:15:06a $9.2 trillion valuation. 25 time
  1996. 1:15:09multiple 11.5 trillion. uh over this
  1997. 1:15:13period from today to uh 2031, this is
  1998. 1:15:16the five years out we're doing, they
  1999. 1:15:18will have collected $1.7 trillion in
  2000. 1:15:21cash. If you assume that that cash is
  2001. 1:15:23being returned to investors as it's
  2002. 1:15:25being generated, a 15 times multiple
  2003. 1:15:27would get you an 18% annualized return.
  2004. 1:15:30A 20 times multiple would get you a 24%
  2005. 1:15:33annualized return and a 25 times
  2006. 1:15:35multiple would get you 29%. Remember
  2007. 1:15:38though, 25 times this is a stock trading
  2008. 1:15:40at you know uh 14 times forward 12 month
  2009. 1:15:44earnings right now. So 15 times though
  2010. 1:15:46that's at 18% right there. The other
  2011. 1:15:48thing I want to note though is that
  2012. 1:15:49these base cases as I was saying they
  2013. 1:15:51are still bullish. It assumes that this
  2014. 1:15:53is a long AI cycle. Uh Nvidia maintains
  2015. 1:15:57their dominance. They maintain their
  2016. 1:15:58margins. They maintain their pricing
  2017. 1:16:00power. Uh the bare case I threw in there
  2018. 1:16:02had earnings uh of $118 billion which is
  2019. 1:16:06you know earnings getting cut almost in
  2020. 1:16:08half. So at that point it would trade
  2021. 1:16:10high30s multiple something like that and
  2022. 1:16:13if it's now proven to see cyclical uh
  2023. 1:16:15probably not going to be able to
  2024. 1:16:16maintain that multiple. So the stock
  2025. 1:16:18would probably get cut um pretty
  2026. 1:16:20severely at that point. Um, the thing is
  2027. 1:16:23that these numbers are very out in the
  2028. 1:16:27air and very hard to predict. And I
  2029. 1:16:29think this is part of the reason why to
  2030. 1:16:31Nvidia trades at the multiple it does,
  2031. 1:16:33right? It's not, you know, a crazy
  2032. 1:16:35multiple to trade at 10 times 2027
  2033. 1:16:38earnings. I think it trades at that
  2034. 1:16:39multiple because a lot of people have a
  2035. 1:16:42lot of questions as to what's going to
  2036. 1:16:43happen in the future, how long this AI
  2037. 1:16:46cycle can really go. And so I think this
  2038. 1:16:49is an opportunity for an investor if
  2039. 1:16:52they do want to take an opinion on that
  2040. 1:16:54to say that I do think that you know
  2041. 1:16:57there's going to be a little bit more
  2042. 1:16:58demand. I think that uh AI is going to
  2043. 1:17:01continue to have very high return on uh
  2044. 1:17:04invested capital for all players and we
  2045. 1:17:06need to build out way more capacity and
  2046. 1:17:08this AI buildout is going to happen for
  2047. 1:17:10many many more years and Nvidia is going
  2048. 1:17:13to continue to be instrumental to that
  2049. 1:17:15buildout. that could be um a situation
  2050. 1:17:18where an investor finds this to to be
  2051. 1:17:20interesting. So, I feel like I kind of
  2052. 1:17:22failed you a little bit with the
  2053. 1:17:23valuation section here, but I think
  2054. 1:17:26that's because you really shouldn't try
  2055. 1:17:29to force too much precision from
  2056. 1:17:33something that is inherently imprecise.
  2057. 1:17:36And that's the issue with cyclical
  2058. 1:17:38businesses because if you try to run a
  2059. 1:17:40DCF on it, it's just going to be
  2060. 1:17:43malarkey, whatever that word Joe Biden
  2061. 1:17:46said was instead of BS. Um
  2062. 1:17:49because whatever your inputs are,
  2063. 1:17:50they're going to be wrong. And you can't
  2064. 1:17:52even make good conservative assumptions
  2065. 1:17:54when trying to do that. That is kind of
  2066. 1:17:56the the issue I have with modeling these
  2067. 1:17:58sorts of things. And so instead what you
  2068. 1:18:01can do is you can look out a couple
  2069. 1:18:04years where you feel comfortable with
  2070. 1:18:06which is maybe you know to 2027 let's
  2071. 1:18:08say and you could see what the earnings
  2072. 1:18:10multiple is on that year that if the
  2073. 1:18:12estimates are correct that's going to be
  2074. 1:18:13a 10 times multiple and then you could
  2075. 1:18:15decide if that makes sense to you or
  2076. 1:18:18not. Now, to give you some context, that
  2077. 1:18:22would be relative to history for chip
  2078. 1:18:25multiples, not as cheap as they get for
  2079. 1:18:28um a trough multiple. So, if you thought
  2080. 1:18:31that those were in earnings peak at
  2081. 1:18:332027, you're going to want to pay a
  2082. 1:18:35lower multiple than that. Uh if you
  2083. 1:18:37thought that there's still going to be
  2084. 1:18:38growth thereafter, then okay, then maybe
  2085. 1:18:41that makes sense. As always, I'm going
  2086. 1:18:43to leave it to you to make that
  2087. 1:18:44decision. But uh a couple things worth
  2088. 1:18:47noting is one, if you're not sure about
  2089. 1:18:49something, there's nothing wrong with
  2090. 1:18:51passing on it. Two, there's also not a
  2091. 1:18:55reason you should need to feel a need to
  2092. 1:18:58force uh certainty from something that
  2093. 1:19:00is uncertain. Uh and three, I think one
  2094. 1:19:04of the biggest issues right now is with
  2095. 1:19:07a company like this, you would really
  2096. 1:19:10want to have comfort in the person
  2097. 1:19:13running it. and knowing that well uh
  2098. 1:19:18even if I can only see out a couple
  2099. 1:19:19years, at least the person in charge can
  2100. 1:19:22see out many more years into the future
  2101. 1:19:24and can run this business and can
  2102. 1:19:26navigate it. And of course, you know
  2103. 1:19:27where I'm going with this. The issue is
  2104. 1:19:29we don't know who that person is. Uh
  2105. 1:19:31because there is not a permanent CEO of
  2106. 1:19:34the business yet. It's up for you though
  2107. 1:19:36to decide whether or not that is more
  2108. 1:19:39than priced in today with the stock down
  2109. 1:19:4130% trading at the multiples it is
  2110. 1:19:44trading at for still a pretty dominant
  2111. 1:19:46franchise in an extremely uh important
  2112. 1:19:50area in the world with a lot of
  2113. 1:19:52tailwinds going for it and growing
  2114. 1:19:54phenomenally quickly. Maybe that can
  2115. 1:19:57make up for them in the short term and
  2116. 1:20:00buy them some time to figure things out
  2117. 1:20:02internally uh and get a new leader. But
  2118. 1:20:05no doubt it makes it much more
  2119. 1:20:07uncomfortable as an investment without
  2120. 1:20:09Jensen Huang there. But maybe that's why
  2121. 1:20:12the opportunity exists. That'll be for
  2122. 1:20:15you to decide. But I will say one last
  2123. 1:20:17thing which is that when I was doing my
  2124. 1:20:20AMD um valuation, I remember at the time
  2125. 1:20:23it was trading at around 10 times 2030
  2126. 1:20:27earnings. And it's possible at the time
  2127. 1:20:29that you know maybe investors really
  2128. 1:20:31thought they were undershooting their
  2129. 1:20:32estimates. So the multiple would have
  2130. 1:20:34been lower at uh for 2030. But that is
  2131. 1:20:37just one data point to give you a little
  2132. 1:20:40bit of context. But if you want much
  2133. 1:20:41more on AMD, go ahead and check this
  2134. 1:20:45video out here. But if you want to know
  2135. 1:20:47what's much more important than picking
  2136. 1:20:49stocks, watch this video here right now.
  2137. 1:20:52Thank you for watching and until next

About this transcript

This page contains the full transcript of Nvidia Stock Crashed — Understand It Better Than 99% of Investors by Drew Cohen, generated from the public captions YouTube serves with the video. The transcript has 15,230 words across 2,137 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.