YouTube2Text

AI Just Got Cheaper. Here's Why That's a Problem for NVIDIA. — Transcript

by AI News & Strategy Daily | Nate B Jones · 3,693 words · 555 segments · language en · Watch on YouTube

Full transcript

  1. 0:00If you've moved work to a cheaper AI
  2. 0:01model this year, you're part of a
  3. 0:03problem Nvidia was already worried about
  4. 0:05back in August. The Wall Street Journal
  5. 0:07reports Jensen Huang told his finance
  6. 0:09team that chip wafers and memory Nvidia
  7. 0:12had committed to were coming at him like
  8. 0:14a quote avalanche. All of that needs
  9. 0:17paying customers, and the customers are
  10. 0:19getting smarter. You found out the
  11. 0:22cheaper model does most of the job, so
  12. 0:24you stop paying for the pricey one. I
  13. 0:26hear that over and over again. Demand
  14. 0:28for AI is very very real. It's
  15. 0:30absolutely enormous. What I want to know
  16. 0:33is whether it is growing at the like 10
  17. 0:35x or more pace that pays the bill for
  18. 0:38the capacity build out. Ramp tracks AI
  19. 0:40spending by businesses. It reported
  20. 0:42virtually no growth in token spending
  21. 0:44across August and September. Two nearly
  22. 0:47flat months followed by a rise of less
  23. 0:49than 2% is still extremely slow growth.
  24. 0:53You're nowhere near the 10-fold increase
  25. 0:55in spending that you want to see to
  26. 0:58sustain the kind of build out that the
  27. 1:01labs are talking about, at least not at
  28. 1:02that pace. Now, I will call out as
  29. 1:04someone who's watched the ramp spending
  30. 1:06numbers that they do bounce around a
  31. 1:09lot. There are some months when you see
  32. 1:1030 40% growth in a month and other
  33. 1:13months where it's completely flat. What
  34. 1:15makes this even more interesting is that
  35. 1:16if you dig into the cross tabs of ramp
  36. 1:18ramp's most recent report, the most
  37. 1:20expensive models actually gained
  38. 1:22spending share in October. So taking a
  39. 1:25larger share is good for the suppliers
  40. 1:28of those models, but it doesn't mean the
  41. 1:30overall market is actually growing
  42. 1:32quickly enough to sustain the capacity
  43. 1:34build out. And that's the problem I am
  44. 1:36interested in. That slow spending growth
  45. 1:38fits a lot of the conversations I'm
  46. 1:39having these days. Companies are
  47. 1:41figuring out that they can move work to
  48. 1:43a cheaper model over and over again.
  49. 1:45They test whether it does the job and as
  50. 1:47models get better, it often does and
  51. 1:49they use it. They don't need the most
  52. 1:51intelligent model for every step of the
  53. 1:52process and they know they don't need to
  54. 1:54pay for it. And that matches what
  55. 1:56businesses have told ramp. They're
  56. 1:57setting defaults that steer employees
  57. 1:59toward less models because those models
  58. 2:02do the work well enough. Here's a really
  59. 2:03simple example. Imagine, you know, a
  60. 2:06customer moving routine work to a
  61. 2:08cheaper model from the same lab. The lab
  62. 2:11is going to keep the customer as a
  63. 2:12paying customer. Token usage is going to
  64. 2:14go way, way up, but the bill either
  65. 2:17doesn't rise or doesn't rise nearly as
  66. 2:20fast. We've covered Anthropic's
  67. 2:22extraordinary growth here. We've covered
  68. 2:24the enormous computing commitments that
  69. 2:26go with that. The question now is where
  70. 2:28the next tenfold increase in spending is
  71. 2:31going to come from if more customers
  72. 2:33learn how to trade down to cheaper
  73. 2:35models. Google has just given customers
  74. 2:37another way to do that in the last week.
  75. 2:39Carbon and hydrogen are two of Google's
  76. 2:42new Gemini models. Carbon handles
  77. 2:44difficult reasoning and hydrogen is
  78. 2:46built for faster execution. Inside
  79. 2:48Google's anti-gravity software, a
  80. 2:49feature called Orchestrator will
  81. 2:51auto-switch between those models during
  82. 2:53tasks. Now, I know that anti-gravity is
  83. 2:56not a super popular software with
  84. 2:58developers right now, but the point is
  85. 3:01that more and more of these kinds of
  86. 3:02models are available. Think of coding
  87. 3:05jobs. You might want a very capable
  88. 3:06model to work out why something is
  89. 3:08broken and maybe plan the repair, but
  90. 3:11once the decision is made, you're going
  91. 3:12to pass a bunch of the rest of the
  92. 3:14coding work to a cheaper, faster model.
  93. 3:16And this is what I see in my own usage,
  94. 3:18and this is what I see in the usage of
  95. 3:20everyone around me who is using AI
  96. 3:22tools, not just at startups, but at
  97. 3:25traditional companies as well.
  98. 3:27In Google's own testing, the
  99. 3:29Orchestrator tool got very close to
  100. 3:32carbon success rate while consuming 24%
  101. 3:35less quota and finishing 19% faster.
  102. 3:38Frankly, some of the numbers I've seen
  103. 3:40from folks who are using open source
  104. 3:42models are even better. They get to 80
  105. 3:4590% cheaper or more, and they can still
  106. 3:48finish faster at an equivalent result.
  107. 3:50So for people who are consuming AI
  108. 3:52tokens right now, this is a phenomenal
  109. 3:54story. It's a phenomenally useful
  110. 3:56result. More work is now fitting inside
  111. 3:58the same allowance, whether you're
  112. 3:59coding or whether you're doing other
  113. 4:01work. Less time is going into waiting,
  114. 4:03the expensive model gets used less, the
  115. 4:05bill either goes down or it's flat. This
  116. 4:09is a direction I see companies moving to
  117. 4:11overwhelmingly. They want AI that works
  118. 4:14and they're getting much more deliberate
  119. 4:16about how much they pay for it.
  120. 4:18And so they don't pay attention to the
  121. 4:20new intelligence scores. Those only
  122. 4:22matter if it lets them do something they
  123. 4:24couldn't do before.
  124. 4:25So for work they already know how to
  125. 4:27complete, cheaper execution is often a
  126. 4:30much much more attractive announcement.
  127. 4:32There's also a clue to all of this in
  128. 4:35RAM's most recent November report about
  129. 4:37where new spending is coming from. Among
  130. 4:40the AI spending RAM can identify by
  131. 4:42department, customer success and support
  132. 4:44appear to have grown from less than 1%
  133. 4:47in June to more than 13% of AI spend in
  134. 4:50October. AI is starting to reach outside
  135. 4:53of engineering into other kinds of work
  136. 4:56and a lot of that work just needs a fast
  137. 4:59economical answer. It does not need a
  138. 5:00frontier answer. If a customer wants to
  139. 5:03know where the shipment is, the company
  140. 5:06just needs the right order information
  141. 5:08and they need to retrieve a correct
  142. 5:09status. That's it. It's not a fancy
  143. 5:11query and paying for deeper reasoning
  144. 5:13just does not make sense on most of
  145. 5:15those requests. And you can grow that
  146. 5:18kind of a business very substantially
  147. 5:20while choosing a cheaper model for most
  148. 5:22of it without changing your overall bill
  149. 5:25footprint and that's exactly the kind of
  150. 5:26pattern that the data reflects. Now look
  151. 5:29at that from the other side of the
  152. 5:30transaction for a minute. The labs have
  153. 5:32ordered computing capacity years and
  154. 5:35years and years and years in advance.
  155. 5:37Their customers are learning how to buy
  156. 5:38intelligence more efficiently at the
  157. 5:40same time. The labs do need enough
  158. 5:43additional business to make up for lower
  159. 5:45prices and then they need to grow far
  160. 5:47beyond that to justify their valuations.
  161. 5:50That's why the new reporting about
  162. 5:52Jensen matters so much. The Wall Street
  163. 5:55Journal describes an August meeting
  164. 5:57about Nvidia's financial support for
  165. 5:59prospective customers. Jensen Huang was
  166. 6:01reportedly worried about commitments
  167. 6:04piling up and major cloud providers
  168. 6:07looking at alternative chips as a
  169. 6:09result. The avalanche remark that I
  170. 6:11quoted earlier in this video came in
  171. 6:13that conversations and concerned supply
  172. 6:16that Nvidia had already committed to
  173. 6:18purchase as part of its own build out
  174. 6:20for chips and capacity to sell to data
  175. 6:23centers. We already know Nvidia has been
  176. 6:25deeply involved in financing the overall
  177. 6:27AI build out. I don't think that's a
  178. 6:29surprise to anyone at this point. What
  179. 6:31this report adds is concern inside
  180. 6:34Nvidia about how those commitments were
  181. 6:36developing and how they might lead to
  182. 6:38exposure for the company long term. None
  183. 6:41of this is a clear evidence that the
  184. 6:44industry has overbuilt so far. Unlike
  185. 6:46.com, we still have skyrocketing token
  186. 6:49demand, but it does give us reason to at
  187. 6:51least examine whether the capacity being
  188. 6:54ordered matches the willingness of
  189. 6:57business to pay more over time. In this
  190. 7:00world, especially as token prices drop,
  191. 7:02demand can still grow astonishingly
  192. 7:05quickly and yet disappoint everyone who
  193. 7:08financed the equipment. Suppose a
  194. 7:10business uses twice or four times as
  195. 7:12much AI but pays a half or a quarter as
  196. 7:14much per unit. That's a world we live
  197. 7:16in.
  198. 7:17Spending hasn't changed overall. The
  199. 7:19provider may serve far more requests,
  200. 7:21but they're not collecting another
  201. 7:22dollar from the customer. Now, that's
  202. 7:24not the only way cost is impacting this
  203. 7:27story. Better hardware and more
  204. 7:29efficient models are lowering provider
  205. 7:30costs as well, and that's part of what
  206. 7:32makes the cheaper service possible at
  207. 7:34all. But whether a particular data
  208. 7:35center earns money back depends on how
  209. 7:37much it can charge, how much business it
  210. 7:39wins, and what it costs to build out in
  211. 7:41the first place. Those are all numbers
  212. 7:43that are one, proprietary, and two,
  213. 7:45different by data center. A cheaper
  214. 7:46model can still be a good business if
  215. 7:48it's inexpensive enough to run, and a
  216. 7:51busy data center can still be a very bad
  217. 7:53investment if the price customers pay
  218. 7:55isn't covering costs. Like that just
  219. 7:57makes sense. At this point, we really
  220. 7:58need to understand on a per data center
  221. 8:01basis where the cost of supplying the
  222. 8:04work and the revenue from selling it add
  223. 8:06up or don't add up. And unfortunately,
  224. 8:08because most of this is, of course, very
  225. 8:10proprietary, we're not in a position to
  226. 8:12know that.
  227. 8:13We don' know for sure what the margins
  228. 8:16are. We don't know for sure where the
  229. 8:18triggers and guarantees are in the
  230. 8:20system, and we don't know for sure
  231. 8:22whether the data centers that are being
  232. 8:24built out are covering their costs as
  233. 8:27they go. And so when we think about the
  234. 8:30future, it's rational to ask ourselves,
  235. 8:32are we in a position where the market is
  236. 8:35growing healthily?
  237. 8:37And when I look at it, I see two
  238. 8:39stories. I see a story of skyrocketing
  239. 8:43token demand, more and more businesses
  240. 8:45leaning into AI. That's all true. But I
  241. 8:48also see a story of dramatically falling
  242. 8:51token costs.
  243. 8:52And the thing that I am looking at
  244. 8:54really carefully is whether frontier
  245. 8:58labs are going to be able to sustain the
  246. 9:02kinds of revenue multiples they're
  247. 9:04looking to grow at over the next year if
  248. 9:07overall token spend remains flat.
  249. 9:11There's still a substantial market for
  250. 9:12the most capable models. I don't
  251. 9:15question that. In fact, we see that in
  252. 9:16RAM's data where the most expensive new
  253. 9:19frontier models make up roughly 70% of
  254. 9:21the spend. The unresolved question is
  255. 9:23how quickly that market grows relative
  256. 9:26to the commitments being made around it.
  257. 9:27I don't have the answer, and anyone who
  258. 9:29says they the answer doesn't. I think
  259. 9:31the combination of customers trading
  260. 9:33down and the reported concern about
  261. 9:34incoming supply makes it a question we
  262. 9:37should all pay attention to. And like it
  263. 9:39or not, RAM's data has become at least
  264. 9:41as important as the Fed in understanding
  265. 9:44one of the key drivers of our economy
  266. 9:46right now. We have to understand AI
  267. 9:49spend to understand where the AI economy
  268. 9:52is going and growing. And just to be
  269. 9:55clear on the macro thesis, I still think
  270. 9:57we are dramatically underallocated to AI
  271. 10:00overall.
  272. 10:01It's just we need to see clear ROI for
  273. 10:04businesses to allocate more. And until
  274. 10:06they see that cash register ring,
  275. 10:09sticking at what, 0.1, 0.2, 0.3, 0.4% of
  276. 10:12overall revenue is something that's
  277. 10:15going to be where most businesses top
  278. 10:16out. And so, as we head into the next
  279. 10:19NVIDIA earnings call, which is right
  280. 10:20around the corner, I would want to hear
  281. 10:22more than just the quarterly results,
  282. 10:24right? And I think investors are going
  283. 10:26to be asking, what about the trajectory
  284. 10:29of NVIDIA as a business? And implicitly,
  285. 10:32they're going to be asking, what about
  286. 10:33the trajectory of the AI build out
  287. 10:35overall? Is there a bigger order book on
  288. 10:37the horizon? Are we seeing secure growth
  289. 10:41in demand and secure growth in financing
  290. 10:43such that we feel comfortable with the
  291. 10:44build out? That's going to be a big
  292. 10:47point to watch in the next week. And
  293. 10:49it's going to be even more important to
  294. 10:51review kind of how that earnings call
  295. 10:53goes because there is now a bipartisan
  296. 10:55initiative supported by Anthropic and by
  297. 10:57OpenAI to further restrict chip exports
  298. 11:01to China. This is something that I think
  299. 11:03was likely inevitable after the Jensen
  300. 11:05scandal broke, but it's going to affect
  301. 11:08the way NVIDIA's chip sales actually
  302. 11:11work, particularly if they close down
  303. 11:13sales, say, through Singapore or other
  304. 11:15places where it's possible for Chinese
  305. 11:17firms to rent compute. There is an extra
  306. 11:20piece of news here that I don't think
  307. 11:22has gotten enough attention that is
  308. 11:24going to add fuel to the fire as far as
  309. 11:26export controls. Will DePue examine six
  310. 11:29released open source models from four
  311. 11:31different Chinese labs, including Deep
  312. 11:33Seat and including Moonshot. He found in
  313. 11:36his research numerical fingerprints
  314. 11:39consistent with NVIDIA's Blackwell GB200
  315. 11:42and GB300 chips, both of which are
  316. 11:45subject to export controls and should
  317. 11:47never have appeared in open source
  318. 11:50models in China. Now, you might wonder,
  319. 11:52how did he figure that out? Well, that's
  320. 11:54the interesting part. A model can
  321. 11:56contain billions and billions of
  322. 11:57numerical values called weights. We sort
  323. 11:59of know that that's how that works at
  324. 12:00this point. Before training, those
  325. 12:03values have to be initially set or
  326. 12:05initialized. The hardware helps
  327. 12:07determine how those starting numbers are
  328. 12:10arranged and some traces of that initial
  329. 12:13state survived training. So, the
  330. 12:16researchers worked backward from
  331. 12:18released models to identify hardware
  332. 12:21that could reproduce the patterns that
  333. 12:23they saw. So, in other words, the model
  334. 12:26itself can contain evidence of the
  335. 12:28equipment used to create it. You don't
  336. 12:31necessarily need a leaked invoice or
  337. 12:33someone inside the company to tell you
  338. 12:35to see which chips were used. The
  339. 12:38fingerprints support Blackwell access,
  340. 12:40but what they can't tell you is where
  341. 12:42were those chips at the time, how was
  342. 12:44access to those chips obtained, did the
  343. 12:46full training run take place on them?
  344. 12:48There are just a bunch of investigative
  345. 12:49questions that have yet to be answered.
  346. 12:51But what this changes is that
  347. 12:54investigators now have something
  348. 12:55technical that they can examine
  349. 12:57independently of what a company may say
  350. 12:59about its own hardware and the story of
  351. 13:01how a model came to be. And I suspect
  352. 13:04from a narrative perspective, this is
  353. 13:06going to add fuel to the fire for export
  354. 13:09control. And that's what we see with
  355. 13:10OpenAI's recent policy shift here.
  356. 13:12Bloomberg reports that OpenAI has told
  357. 13:14lawmakers it supports a package of
  358. 13:17tighter chip controls. Now, Anthropic
  359. 13:19had already backed those measures. Those
  360. 13:21measures have bipartisan support. And so
  361. 13:24when both labs support it and both
  362. 13:25parties support it, support is now
  363. 13:28growing for really closing down the
  364. 13:30routes through which restricted
  365. 13:31computing power can still reach Chinese
  366. 13:33developers. One proposal, the Chip
  367. 13:35Security Act, would require location
  368. 13:37verification for advanced chips covered
  369. 13:39by export controls. The idea is pretty
  370. 13:42straightforward. A shipment going to any
  371. 13:44approved destination should not be able
  372. 13:47to disappear from view and in any
  373. 13:49situation end up somewhere else. This
  374. 13:51follows the China security story I
  375. 13:53covered. Huang admitted an affair with a
  376. 13:55Mega Speed executive whom US officials
  377. 13:57believe worked for Chinese intelligence.
  378. 14:00And part of the allegation there is the
  379. 14:02allegation that those chips were moved
  380. 14:05other places. Now, China denies those
  381. 14:07allegations and of course, we have to
  382. 14:09see how all of that plays out. But it's
  383. 14:12in the political wind right now and it's
  384. 14:15part of the story that is driving
  385. 14:17bipartisan support for this bill. Now,
  386. 14:20there's one other story I want to cover
  387. 14:22because it affects anyone being asked to
  388. 14:24trust these companies with more of their
  389. 14:26work, which is effectively what the labs
  390. 14:28are asking companies to do. I'm talking
  391. 14:30about, of course, the Microsoft 365
  392. 14:33outage.
  393. 14:34[snorts] The Microsoft 365 outage on
  394. 14:36November 12th disrupted Outlook, Teams,
  395. 14:38and other services that businesses
  396. 14:40depend on. Microsoft now says an earlier
  397. 14:42network capacity upgrade introduced a
  398. 14:44problem that surfaced after a later
  399. 14:47network change. It says the incident was
  400. 14:49not related to cyber activity or any AI
  401. 14:52agents on the system, and I believe
  402. 14:54Microsoft. That's the technical
  403. 14:56explanation. But my larger concern is
  404. 14:59what Microsoft's engineering culture
  405. 15:01does to its ability to prevent a
  406. 15:03mistake, contain it, and recover from
  407. 15:05it. In Microsoft's developer division,
  408. 15:07some managers have been told that AI use
  409. 15:09should factor into employee evaluations.
  410. 15:11And to be clear, Microsoft is not alone
  411. 15:13on that. I've heard similar stories from
  412. 15:15Meta, similar stories from Amazon and
  413. 15:17other large companies. The company also
  414. 15:19offered retirement packages to very
  415. 15:22long-serving employees in the spring,
  416. 15:24and I think that's relevant as well.
  417. 15:26When you reward people
  418. 15:28for using AI and you encourage
  419. 15:30experienced people to leave, you need to
  420. 15:32be really sure that you're keeping the
  421. 15:34knowledge required to run extremely
  422. 15:36complex systems that you are now
  423. 15:38changing at a higher rate. Former Azure
  424. 15:40engineer Axel Rietschin has described
  425. 15:43years and years of quick fixes and lost
  426. 15:46senior expertise inside Microsoft's
  427. 15:48cloud business. His account is focused
  428. 15:50on Azure, which is a separate service.
  429. 15:53But I think the point here is that there
  430. 15:56is a larger question that many companies
  431. 15:58are grappling with about companies'
  432. 16:00treatment of engineering knowledge,
  433. 16:02particularly in the age of AI.
  434. 16:05And Gergely Orosz's recent reporting on
  435. 16:07companies that include Uber and Meta has
  436. 16:09echoed similar themes of AI use and code
  437. 16:11output without equivalent attention to
  438. 16:13whether the software involved actually
  439. 16:15gets better. Ultimately, regardless of
  440. 16:17the technical trigger, that is the risk
  441. 16:19I see in asking companies to
  442. 16:22dramatically increase AI usage without
  443. 16:25care. You can measure a lot of activity
  444. 16:28while missing what makes the service
  445. 16:29dependable, and you can walk your human
  446. 16:32talent out the door precisely at the
  447. 16:33moment that you need it the most. An
  448. 16:35experienced engineer is there for a
  449. 16:37reason. They know all of the skeletons
  450. 16:39in the closet. They remember the
  451. 16:41failures, and when something breaks,
  452. 16:43they know which thing they have to look
  453. 16:45at to fix. When I was at Amazon, there
  454. 16:48was literally a guy on a boat that you
  455. 16:50called whenever AWS went down, and he
  456. 16:53was able to come and fix it. So, this is
  457. 16:56something that we desperately need more
  458. 16:58of as we lean into AI. And none of this
  459. 17:00is saying that AI doesn't add value.
  460. 17:03I'm just trying to call out that if we
  461. 17:05are going to ask companies, which the
  462. 17:07labs are doing, to increase spend, we
  463. 17:10need to do so in a way that allows
  464. 17:12companies to improve the reliability and
  465. 17:14security of their software along the
  466. 17:16way. And that requires keeping
  467. 17:18engineering cultures intact. Amazon's
  468. 17:20own explanation of a March incident
  469. 17:22around AI illustrates why this is so
  470. 17:25fraught. An engineer at the time
  471. 17:27followed incorrect AI advice drawn from
  472. 17:29an outdated wiki. Now, Amazon called the
  473. 17:31cause user error. That's true. They
  474. 17:34claimed it was unrelated to AI, which I
  475. 17:36kind of have some questions about, and
  476. 17:38they said its controls let the mistake
  477. 17:40have too much impact. I would agree with
  478. 17:41that. But even in that account, the
  479. 17:44quality of the information and the
  480. 17:46checks around the engineer mattered
  481. 17:48enormously. It's really, really
  482. 17:51important to have strong engineering
  483. 17:53cultures that allow that kind of system
  484. 17:56to grow in reliability over time rather
  485. 17:59than to degrade. We are not going to get
  486. 18:02an official cultural explanation for a
  487. 18:04root cause for the outage. We'll get a
  488. 18:06technical one, but we're not going to
  489. 18:08know whether staffing or AI incentives
  490. 18:10contributed in any way to the Microsoft
  491. 18:13outage. But I think in a lot of ways,
  492. 18:15that is the most relevant question we
  493. 18:17can ask coming out of a global outage
  494. 18:19like this.
  495. 18:20An ordinary network error doesn't answer
  496. 18:23management questions. It doesn' answer
  497. 18:25questions around how the error came to
  498. 18:27be in the first place, around how the
  499. 18:29impacts was so large, around why
  500. 18:30recovery was so difficult and took such
  501. 18:33a long time.
  502. 18:35Every company at this scale is committed
  503. 18:38to security and quality. I do not doubt
  504. 18:40that Microsoft is committed to security
  505. 18:42and quality of their software.
  506. 18:44But my larger concern is that we have
  507. 18:47seen a lot of public statements, not
  508. 18:49just from Microsoft, but from other big
  509. 18:50companies around dramatic changes in
  510. 18:54engineering culture in order to enable
  511. 18:56AI. And I think that one of the things
  512. 18:58I'm seeing at successful engineering
  513. 19:00cultures in the age of AI is that they
  514. 19:02are not taking away deep developer
  515. 19:04experience. They're actually building on
  516. 19:06it so that they can use AI effectively
  517. 19:09to grow the reliability and quality of
  518. 19:12software in their systems. That's
  519. 19:14something that we need to think about
  520. 19:15more. And I think that is one of the
  521. 19:17things that stands in the way of the 10x
  522. 19:20growth in spend that the labs are
  523. 19:21looking for. We need to build up
  524. 19:23engineering cultures. Ultimately, I
  525. 19:25don't see some of these trends changing
  526. 19:27anytime soon. Customers are going to
  527. 19:28keep learning how to buy AI more
  528. 19:30cheaply. Suppliers have a tremendous
  529. 19:32amount of capacity coming online in the
  530. 19:34next year or so. Governments are
  531. 19:36deciding who can use some of it and who
  532. 19:38can't. That's the China export controls.
  533. 19:41And the companies building and operating
  534. 19:42these systems need to get enough revenue
  535. 19:44in the door to keep that build out
  536. 19:45going. Just underline this three times.
  537. 19:47I expect people to use a lot more AI in
  538. 19:492027. I think the question that I have
  539. 19:52is whether the spend growth is going to
  540. 19:55be as sky high as the labs need. I
  541. 19:58expect spend growth in 2027 too. I just
  542. 20:02wonder if that spend growth is going to
  543. 20:03go to model makers or if it's going to
  544. 20:05go farther up the stack to places where
  545. 20:09AI is going to actually help businesses
  546. 20:10ring the cash register. AI token prices
  547. 20:13are going to continue to fall. It will
  548. 20:14be a buyer's paradise. It is going to be
  549. 20:17up to us to build strong engineering
  550. 20:19cultures so that we can actually
  551. 20:20implement AI systems that offer real and
  552. 20:23sustained benefits. That is going to be
  553. 20:25the work of 2027. I think that's the
  554. 20:27lesson of the do's in the last week or
  555. 20:29so. I'll see you next time.

About this transcript

This page contains the full transcript of AI Just Got Cheaper. Here's Why That's a Problem for NVIDIA. by AI News & Strategy Daily | Nate B Jones, generated from the public captions YouTube serves with the video. The transcript has 3,693 words across 555 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.