YouTube2Text

HBM Under Threat? Are Microsoft, Intel & Marvell Changing the AI Memory Game — Transcript

by Market Signal · 3,745 words · 553 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Hey guys, welcome back to Market Signal.
  2. 0:02Hope you're doing well. It's another
  3. 0:03late night here in Phoenix, but there
  4. 0:05are a few key updates that I want to
  5. 0:07walk each and every one of you through
  6. 0:09to make sure that you stay informed and
  7. 0:11updated in this memory sector. You know,
  8. 0:13in my video last night, we talked about
  9. 0:15the Nvidia leadership shock and what it
  10. 0:18meant for the $500 billion AI buildout.
  11. 0:22But today on Saturday, November 14th, we
  12. 0:25are seeing a new batch of supply chain
  13. 0:28reporting that is dropping. And if you
  14. 0:30own Micron, SKH, and Samsung, you need
  15. 0:32to pay close attention to the bigger
  16. 0:35architectural picture that is emerging
  17. 0:37around us right now. There were three
  18. 0:39separate announcements over the last 48
  19. 0:41hours that are showing us some of the
  20. 0:43clearest signals yet that hyperscalers
  21. 0:46and chip designers are continuing to
  22. 0:49pursue a way to engineer around the HBM
  23. 0:52bottleneck that we see in the
  24. 0:53marketplace. For example, Microsoft just
  25. 0:56signed with Cerebrus. They dropped their
  26. 0:58earnings on the street at the end of
  27. 1:00this week. I'm going to show you that
  28. 1:01and what it means for the sector. And
  29. 1:03guys, I actually think Cerebrus might be
  30. 1:05at a really great entry point right now.
  31. 1:07But we're going to walk through all that
  32. 1:08data together in this video. And Intel
  33. 1:11is also pitching an HBM free inference
  34. 1:14GPU architecture. And then finally,
  35. 1:16Marll just laid out a blueprint showing
  36. 1:18that AI workloads are outgrowing the
  37. 1:21memory inside individual racks. They are
  38. 1:24now requiring optical memory fabrics and
  39. 1:27CXL expansion just to keep compute fed.
  40. 1:30So tonight, we're going to connect all
  41. 1:32of these dots and these signals that we
  42. 1:34see in the market. I'm going to separate
  43. 1:36the real architectural threats to HBM
  44. 1:39from the early stage hype and we're
  45. 1:41going to break down what it means to
  46. 1:43Micron and the other memory companies.
  47. 1:45So, make sure you hit like and make sure
  48. 1:47you subscribe to Market Signal. Guys, I
  49. 1:49hope you're having a great day. Let's
  50. 1:50dive in and let's talk about all of the
  51. 1:52signals and the architectural updates
  52. 1:55that we are that we need to know to stay
  53. 1:57informed in this sector overall. All
  54. 1:59right, guys. The first thing I want to
  55. 2:00jump to and talk about is the cerebrous
  56. 2:03earnings that they dropped on the
  57. 2:04street. So, let's walk through that
  58. 2:06together. I'm going to pull up their
  59. 2:07earnings statement here, and we'll walk
  60. 2:09through some of the key details. You can
  61. 2:11see the piece that I highlighted here.
  62. 2:13Surging demand for high-speed inference
  63. 2:15fuels growth from new and existing
  64. 2:18customers. So, they dropped their a
  65. 2:20record revenue in the third quarter of
  66. 2:222026. In addition, Cerebrus, they raised
  67. 2:25their fullear outlook overall. But a
  68. 2:29couple elements they their c as their
  69. 2:31CEO mentioned Andrew Feldman who's the
  70. 2:33co-founder and the CEO of Cerebrus he
  71. 2:35goes on record to say that Cerebrris
  72. 2:37technology remains the undisputed
  73. 2:40champion of speed. He went further on to
  74. 2:42say that they've exceeded our guidance
  75. 2:44across the the board for the third
  76. 2:46quarter for a third quarter in a row and
  77. 2:48delivered sequential improvements on all
  78. 2:51core financial metrics. And we we won
  79. 2:54new marquee customers including
  80. 2:56hyperscalers, neoclouds, and large
  81. 2:58enterprises. And we saw our pipeline for
  82. 3:00high-speed AI inference increase by more
  83. 3:02than 10x in the quarter. Take a look at
  84. 3:05a few of the metrics overall. So I'm
  85. 3:07going to show you their non-GAAP, their
  86. 3:09core financial uh results. the non-GAAP
  87. 3:11results. You can see in the first part I
  88. 3:13highlighted here the core cloud and
  89. 3:15other services revenue 140 almost 145
  90. 3:20million. It's up 272% year-over-year.
  91. 3:23But take a look at core total revenue.
  92. 3:25It's 231 million. It's up 70%
  93. 3:29year-over-year. And they have a core
  94. 3:30gross margin of 41%. And that's both
  95. 3:34sequ it's it's up both sequentially and
  96. 3:36year-over-year.
  97. 3:38But one of the other key things I want
  98. 3:40to show you is not only their remaining
  99. 3:42performance obligations, but what they
  100. 3:44said in the business highlights first.
  101. 3:47Two 26 billion in remaining performance
  102. 3:49obligations as of September 30th of this
  103. 3:52year. But really the piece that jumped
  104. 3:54out to me was this business highlights
  105. 3:56and the customer. So they went on record
  106. 3:58to say that Cerebrus is expand is
  107. 4:01rapidly expanding its customer base,
  108. 4:03broadening its reach and diversifying
  109. 4:05its revenue. So Oracle is integrating
  110. 4:07Cerebrra's fast inference into Oracle
  111. 4:09cloud infrastructure. But it's this
  112. 4:11second bullet point that really piqued
  113. 4:13my curiosity when the CEO said it in the
  114. 4:16in the analyst call. He said that
  115. 4:18Cerebrris announced an important cloud
  116. 4:20deal with a top three hyperscaler. More
  117. 4:23will be shared next year. It made me
  118. 4:25think who they're going to be working
  119. 4:27with because we already see Cerebrris
  120. 4:28working with Open AI, Amazon, Oracle.
  121. 4:32who would be the next top hyperscaler
  122. 4:34that they they would be working with as
  123. 4:36we move into 2027. And I saw that just
  124. 4:39the other day as it popped up. Excuse
  125. 4:41me, just saw it in another article as it
  126. 4:43popped up, but you can see it is
  127. 4:44Microsoft. So Cerebrus raises its 2026
  128. 4:48forecast and signs Microsoft as a cloud
  129. 4:51customer. We see cerebrus they expect up
  130. 4:53to 960 million in core revenue this year
  131. 4:57and its CEO says the Microsoft deal was
  132. 4:59signed actually in October. So if we go
  133. 5:02down into some of the details, the
  134. 5:04company now expects core revenue of 940
  135. 5:06to 960 960 million this year. And the
  136. 5:10results the results statement named a
  137. 5:13new customer only as a top three
  138. 5:15hyperscaler as I showed you which piqu
  139. 5:17my curiosity. But on an interview with
  140. 5:20CNBC, the chief executive Andrew
  141. 5:22Feldman, he mentioned that it is
  142. 5:24actually Microsoft which signed in
  143. 5:26October and he expects the revenue from
  144. 5:28the deal to start arriving in the first
  145. 5:30quarter of 2027. As I mentioned,
  146. 5:32Cerebras already works with Amazon with
  147. 5:35or Oracle and we know they work with
  148. 5:37Open AI which is a big part of that uh
  149. 5:40remaining performance obligation for the
  150. 5:42company. But we are seeing in the fourth
  151. 5:44quarter they forecast their revenue now
  152. 5:46to be 310 to 330 million. The analysts
  153. 5:51were expecting 271. So this is
  154. 5:54significant guidance above what the
  155. 5:56analysts expected for the upcoming
  156. 5:58quarter. Much of the growth rests with
  157. 6:00open AAI. So it's that compute deal that
  158. 6:03they have that underpins Cerebrus's 26
  159. 6:06billion in remaining performance
  160. 6:07obligations. the contract of revenue
  161. 6:10that is not yet booked according to
  162. 6:12Reuters.
  163. 6:13But there's a few other things I want to
  164. 6:15walk through, guys, with the deal. I
  165. 6:16actually think this is a very
  166. 6:18encouraging and positive earnings that
  167. 6:20Cerebras dropped on the street. And why
  168. 6:22does it matter so much that they won a
  169. 6:24new hyperscaler contract with Microsoft
  170. 6:28this cloud contract? Why does it matter?
  171. 6:30Why does it matter to Micron and SKH
  172. 6:32Heinix, the shareholders or investors in
  173. 6:34those companies? because we know how
  174. 6:37different the physics and the chip
  175. 6:38architectures are between these two
  176. 6:40companies. And just to remind you, I can
  177. 6:42pull it up because we know when you take
  178. 6:44a look at the difference and this
  179. 6:46graphic shows it perfectly. It's the
  180. 6:48memory bottleneck solve wafer scale
  181. 6:50versus the GPU mode. You're looking on
  182. 6:52the left side, you see the GPU plus the
  183. 6:54HBM cluster. So this is the Nvidia the
  184. 6:57AMD model where we have stacks of high
  185. 7:00bandwidth memory that are clustered
  186. 7:02right next to the compute. It is me
  187. 7:04according to the article here by uh by
  188. 7:07cerebrus it is memory traffic. It's
  189. 7:09memory bound and you could see the
  190. 7:11traffic and that has to data has to move
  191. 7:13back and forth from those stacks to the
  192. 7:15compute. Overall the cost of the bill of
  193. 7:17materials in this model is about 60 to
  194. 7:2070% for memory and we know that an HBM3
  195. 7:23stack cost about $300 each. If you look
  196. 7:25on the right side you can see Cerebrus'
  197. 7:28wafer scale engine that chip on a disk
  198. 7:30approach that they have. It's a very
  199. 7:32different architecture overall. They use
  200. 7:34this massive wafer scale engine with
  201. 7:37huge amounts of SRAMM integrated
  202. 7:39directly on chip. And by keeping memory
  203. 7:42directly on the silicon wafer, Cerebrus
  204. 7:45avoids the external HBM supply
  205. 7:47bottleneck and eliminates the energy and
  206. 7:49the latency penalty of moving data
  207. 7:52across off-chip interconnects for ultra
  208. 7:54fast AI inference.
  209. 7:56But we have to keep everything in
  210. 7:59perspective because as I just showed
  211. 8:00you, Cerebrus is at under 1 billion in
  212. 8:03annual revenue. It is still just a
  213. 8:05fraction of what Nvidia or Micron's
  214. 8:08footprint is. However, the microscope,
  215. 8:11excuse me, and this Microsoft deal, it
  216. 8:14doesn't mean that the hyperscalers are
  217. 8:16abandoning GPU clusters. However, what
  218. 8:19it does show us is it proves that when
  219. 8:21HBM supply is as tight as it is in the
  220. 8:24market and as expensive it is with the
  221. 8:26memory manufacturers having all the
  222. 8:28pricing power right now, the largest
  223. 8:31cloud buyers in the world are willing to
  224. 8:33fund alternative architectures to get
  225. 8:36inference capacity online. And I think
  226. 8:38this is what this Microsoft deal is
  227. 8:40showing us right now with with uh
  228. 8:42Cerebrus is they are working to try to
  229. 8:45get capacity online in many different
  230. 8:47ways. They're looking at alternative
  231. 8:49architectures. But as I said, Cerebrus's
  232. 8:52uh uh revenue is still under a billion.
  233. 8:54It it pales in consideration to the
  234. 8:56footprint that we see with Micron,
  235. 8:58Samsung, SKHEX, even Nvidia and AMD
  236. 9:01overall. But so I wanted to share that
  237. 9:03update with you because there's a couple
  238. 9:05things that I want to jump to next is
  239. 9:06just to take a look at uh Cerebras in
  240. 9:09general and is it a deal right now for
  241. 9:11us as investors? If I jump over to take
  242. 9:13a look at how the stock is trading, we
  243. 9:15know it had a couple good days to end
  244. 9:17the week. It was actually up by over 11%
  245. 9:19on Friday. It closed at a at 10872. But
  246. 9:22if you look to the right, you can see we
  247. 9:24are far off from where we are after the
  248. 9:27launch this year. When they had the IPO
  249. 9:29earlier this year, the stock got up to
  250. 9:31about $386.
  251. 9:33It reached 120 on Friday as the intraday
  252. 9:36high. But I want to call your attention
  253. 9:38to just what the fair price and
  254. 9:39ultimately what the analyst targets are.
  255. 9:41Now, fair price. Cerebrus is right in
  256. 9:44line with what the fair price is.
  257. 9:46According to the model on investing.com,
  258. 9:48you can see it's $106 a share. The stock
  259. 9:51is trading at about $109 a share right
  260. 9:53now. But analyst estimates have this
  261. 9:55stock at about $219 a share. $101%
  262. 10:00upside today from where we are trading.
  263. 10:02And just take a look at how Cerebras
  264. 10:04actually has been penalized compared to
  265. 10:06some of to the broader chip market from
  266. 10:08the from July to where we stand today.
  267. 10:11Cerebrus is in the orange line right
  268. 10:13here and it's comparing against the S&P
  269. 10:15500 as well as the chip index and you
  270. 10:17can see it's significantly down. This
  271. 10:19stock has been put under pressure way
  272. 10:23way further and harder than we see with
  273. 10:25the broader indexes overall. Was that
  274. 10:27pressure justified? We know it jumped
  275. 10:29out of the gate in its IPO. I stayed
  276. 10:31away from it in its IPO and and in full
  277. 10:33transparency, I'm not a shareholder in
  278. 10:35Cerebras at this moment. I've been
  279. 10:37waiting for the right price point to
  280. 10:40move in on the stock and I'm going to be
  281. 10:41honest with you, I think right now is a
  282. 10:43right price point for me. Make sure you
  283. 10:45do your own due diligence. Remember, I'm
  284. 10:47not a financial advisor. Make sure you
  285. 10:49do your own research and it fits your
  286. 10:51own individual risk tolerance, but where
  287. 10:53the stock is trading right now, the
  288. 10:55earnings that they dropped on the
  289. 10:56street, the deals that they continue to
  290. 10:58sign, the most recent deal they just
  291. 10:59signed with Microsoft, I do think we are
  292. 11:01in a position to consider this stock as
  293. 11:04an entry point at 108. And I just wanted
  294. 11:06to show you where the stock is sitting
  295. 11:08from the analyst perspective. So we have
  296. 11:1111 analysts that are following the stock
  297. 11:13all listed as a buy. As I mentioned, the
  298. 11:15average 12-month target is about 219.
  299. 11:18But you can see the range. We have
  300. 11:20everything from 170 a share for Rosenlat
  301. 11:23all the way up to 205 for Craig Harlem,
  302. 11:26which is representing anywhere between a
  303. 11:2856 to 130% premium for where the stock
  304. 11:31trades in the market today. So guys, I
  305. 11:33wanted to cover that first and foremost
  306. 11:35because we are seeing great earnings
  307. 11:36with Cerebras. We are seeing
  308. 11:38hyperscalers sign with them as well.
  309. 11:40It's just another strong sign for the
  310. 11:41demand, the memory demand that we have
  311. 11:43in the market. And the fact that we're
  312. 11:45seeing hyperscalers and other customers
  313. 11:47not only work with Micron, but trying to
  314. 11:49solve the the bottleneck by pursuing
  315. 11:51other architectures that we see in the
  316. 11:54market. But it wasn't just Cerebrus
  317. 11:56today too. We had other signals that I
  318. 11:58want toh show you around this
  319. 12:00architecture focus around what customers
  320. 12:02are doing to try to solve the high
  321. 12:04bandwidth bottleneck that we see in the
  322. 12:06market. There was one from Intel that I
  323. 12:08want to cover with you and then the last
  324. 12:10one from from Marll and then I want to
  325. 12:12show you a little bit of the supply side
  326. 12:14that we're going to see as we move into
  327. 12:162027. The fact that we are going to be
  328. 12:18continue to be constrained as we move
  329. 12:21into 27. But first, let's talk about
  330. 12:23what Digi Time showed with regards to
  331. 12:25Intel and what their approach is or what
  332. 12:27they're putting together right now with
  333. 12:29a partnership over in Korea. So, you can
  334. 12:31see Intel formed a 23 company Korean AI
  335. 12:35alliance as it pitches HBM free
  336. 12:38inference GPUs. So, Intel is courting
  337. 12:41South Korea's AI ecosystem on two
  338. 12:43fronts. The first is a 23 company
  339. 12:46alliance built around Korean-made
  340. 12:48servers and an inference GPU that swaps
  341. 12:51costly HBM for low power DRAM.
  342. 12:54That's the one kind of the front that I
  343. 12:57want to focus on from this article
  344. 12:59overall because what we are seeing is
  345. 13:01they announced the launch of a KAI full
  346. 13:04stack alliance describing it as a
  347. 13:06commitment to creating innovation
  348. 13:08together with Korean companies on a open
  349. 13:11infrastructure and ecosystem built
  350. 13:13around domestically manufactured
  351. 13:14servers. Inference an inference GPU
  352. 13:17without HBM. The centerpiece project was
  353. 13:21was not a CPU but Crescent Island. the
  354. 13:24inference GPU that Intel renamed the GPU
  355. 13:27D series in October. It uses LP DDR5X
  356. 13:32instead of HBM and draws 350 watts,
  357. 13:35slightly over half the 600 watts of
  358. 13:38Nvidia's RTX Pro 6000. Intel said memory
  359. 13:42capacity could reach 480 gigabytes, up
  360. 13:46to five times that an Nvidia's card. So
  361. 13:49that is a key piece that we are seeing
  362. 13:50with Intel, this alliance that they're
  363. 13:53they're launching right there in Korea.
  364. 13:54And I think for for us Micron investors
  365. 13:57or advanced memory investors, this is a
  366. 14:00particularly for Micron investors, this
  367. 14:01is a classic two-sided coin. On one
  368. 14:04hand, if HBM free inference chips gain
  369. 14:07traction for lighter enterprise
  370. 14:09workloads, it will reduce HBM content
  371. 14:12per accelerator. But on the other hand,
  372. 14:15deploying lower cost inference hardware
  373. 14:18at scale still consumes massive volumes
  374. 14:21of standard DDR and LP DDR bit capacity.
  375. 14:25So right now, this is just a proposed
  376. 14:27roadmap that we're seeing Intel try to
  377. 14:30put together with the Korean companies.
  378. 14:32It is not a proven commercial threat
  379. 14:35matching the HBM accelerator
  380. 14:37performance. So, I would treat it as a
  381. 14:39long-term watch list rather than an
  382. 14:41immediate order cut or an immediate
  383. 14:43threat that we see in the marketplace.
  384. 14:46And in addition to this, we're also
  385. 14:47seeing Marll come out. We know that
  386. 14:49Marll is going to be at at the at an
  387. 14:52upcoming conference. They're actually
  388. 14:53going to be at the super, it's called a
  389. 14:56supercomputing conference, the
  390. 14:58international conference for high
  391. 14:59performance computing. It's actually
  392. 15:01starting tomorrow. It runs from November
  393. 15:0315th through the 20th in Chicago. And
  394. 15:06what Marll came out today is as you can
  395. 15:09see I've have it highlighted here. They
  396. 15:10are mentioning that as AI workloads
  397. 15:12outgrow single rack memory and
  398. 15:14bandwidth. Marll demonstrates
  399. 15:16interconnect and memory technologies
  400. 15:18engineered to scale together. So as a
  401. 15:21leader in the connectivity and the
  402. 15:23memory innovation, Marll is addressing
  403. 15:25the connectivity challenges with
  404. 15:27advances in switch and interconnect
  405. 15:29technologies the and memory technologies
  406. 15:32with innovations in CXL memory expansion
  407. 15:35and acceleration. Additionally, Marll is
  408. 15:37pioneering technologies to unify these
  409. 15:40domains and scale memory and
  410. 15:42interconnect as an as a single system.
  411. 15:45So Marvel will showcase the technologies
  412. 15:48enabling the next generation of AI
  413. 15:50infrastructure across scale up, scale
  414. 15:53out and scale across architectures.
  415. 15:56Optical interconnect technologies may
  416. 15:58allow memory and compute to scale
  417. 16:00together across multiple racks, growing
  418. 16:03the size of the scale up domain. And the
  419. 16:05demonstrations that Marll will have at
  420. 16:08their booth, include the Marll photonic
  421. 16:10fabric memory module, the CXL memory
  422. 16:13solutions, including the Maravell
  423. 16:15structer, the addin card, the near
  424. 16:18memory compute and expansion and
  425. 16:19compression solutions, and partner
  426. 16:21memory demonstrations with Structura X
  427. 16:24memory expansion controllers featuring
  428. 16:26both DDR4 and DDR5. So what Marll is
  429. 16:31telling the market is really simple
  430. 16:33here. What they are saying is that the
  431. 16:34next phase of the AI buildout isn't just
  432. 16:37about stacking more HBM directly onto a
  433. 16:40single GPU. It's about disagregating
  434. 16:44memory and pooling it across racks using
  435. 16:46high-speed optical interconnects and CXL
  436. 16:49controllers. It expands the total
  437. 16:52addressable market for optical and
  438. 16:55networking players like Marll, Broadcom,
  439. 16:58Arista, Coherent, Lummentum, Credo. And
  440. 17:02it also proves that the HBM the excuse
  441. 17:04me the memory bandwidth remains the
  442. 17:06central bottleneck of AI scaling. So
  443. 17:09while all this is happening right now we
  444. 17:11are also seeing signs on the supply
  445. 17:14side. So we're the supply side is
  446. 17:16showing us that the HBM super cycle is
  447. 17:19heating up as we head into 2027 and
  448. 17:222028. Just take a look at a few of these
  449. 17:24pieces that I put together to just give
  450. 17:26you a demonstration of how this is
  451. 17:28heating up and what we are seeing from
  452. 17:29some key providers right now. You can
  453. 17:31see on your screen HBM supply heats up.
  454. 17:34The HBM super cycle's next phase for
  455. 17:372027 and 2028. On the left side, we know
  456. 17:40that SKHX is reportedly just pushed its
  457. 17:44HBM4 production yields past 70%. Giving
  458. 17:47it giving its stronger unit cost uh
  459. 17:50economics and volume output as HBM4
  460. 17:54ramps. In the middle, we're talking
  461. 17:56about Samsung and how it's accelerating
  462. 17:58its new HBM packaging facility in Asia
  463. 18:00by a full year. They are targeting June
  464. 18:032028, opening with an initial 4.48
  465. 18:07billion facility investment and roughly
  466. 18:1030 billion in subsequent advanced
  467. 18:12equipment spending discussed. And then
  468. 18:14finally on the right side we see CXMT in
  469. 18:17China is pushing into the higher value
  470. 18:20DDR5 server modules and the LP DDR DDR6
  471. 18:25products. And while CXMT still faces
  472. 18:27steep hurdles in commercial volume and
  473. 18:30US export constraints, it reinforces
  474. 18:33that conventional DRAM pricing won't
  475. 18:35stay immune from com from competition
  476. 18:38forever. And when you combine these
  477. 18:40three, you can see at the bottom the big
  478. 18:42picture. HBM supply is expanding with
  479. 18:44more capacity, higher yields, and new
  480. 18:48competition, setting up a more
  481. 18:49competitive memory market in 2027 and
  482. 18:522028. We, as I mentioned, we see
  483. 18:55Samsung's accelerated 30 billion plus
  484. 18:57packaging buildout for 2028. SKHik 70
  485. 19:02plus HBM4 yields. And the core debate
  486. 19:05overall for Micron, for us investors, it
  487. 19:09needs to shift in our mind. It is no
  488. 19:11longer about can we sell every wafer
  489. 19:13that the company has today. It is
  490. 19:15shifting. And what that shift is is
  491. 19:17continuing to show us is just how tight
  492. 19:20HBM pricing can stay once competing
  493. 19:23capacity hits its full stride in 2028.
  494. 19:26That's the central question. When we hit
  495. 19:28our full stride in 2028, can Micron and
  496. 19:31the other companies keep that pricing
  497. 19:33leverage? Will they be able to keep that
  498. 19:35HBM pricing, that premium pricing in the
  499. 19:37market? because we're going to see more
  500. 19:40capacity hit in the marketplace. But as
  501. 19:42of right now, we are still constrained
  502. 19:45as ever. We are going to be constrained
  503. 19:46in this market as we move into 2027 and
  504. 19:502028 despite what we are seeing from not
  505. 19:53only SKH, from CXMT, from Samsung, from
  506. 19:56Micron as well. Even what Cerebrus is
  507. 19:58doing right now. You think about how the
  508. 20:00other players are trying to work around
  509. 20:02the HBM bottleneck. I gave you two
  510. 20:04examples tonight. Intel as well as what
  511. 20:07Marll is talking about as well. So to
  512. 20:09put it all together in a clear kind of
  513. 20:11framework for us as we head into
  514. 20:13Monday's opening bell, first and
  515. 20:15foremost, we have to pay attention to
  516. 20:18Tuesday's catalyst moment with Nvidia
  517. 20:21when they step to the podium with their
  518. 20:22new leadership team. They share their
  519. 20:24earnings and they discuss what their
  520. 20:27future is going to be without Jensen.
  521. 20:28It's going to be the biggest moment next
  522. 20:30week. We have Enthropics IPO that will
  523. 20:32mean a ton for this sector as well. But
  524. 20:35we have to realize that HBM demand is
  525. 20:37not collapsing. It is just diversifying.
  526. 20:40Near-term HBM3 and HBM4 supply remains
  527. 20:43tightly sold out, which is why Micron's
  528. 20:46fundamentals remain strong despite the
  529. 20:48pullback we have seen in the sector
  530. 20:50recently. And secondly, watch the
  531. 20:53architectural hedges that we're seeing.
  532. 20:55Microsoft's deal with Cerebrus and
  533. 20:57Marll's CXL and optical memory push.
  534. 21:00They prove that hyperscalers are
  535. 21:02actively building complimentary ways to
  536. 21:04scale memory bandwidth outside of
  537. 21:07standard GPU plus HBM architectures. But
  538. 21:10guys, let me know in the comments below
  539. 21:12what you think, what your thoughts are.
  540. 21:14What did you think of Cerebrus's
  541. 21:16earnings? Do you think the wafer scale
  542. 21:18chip that Cerebras has in the market or
  543. 21:20maybe even the CXL memory pooling that
  544. 21:23Mel Marll is talking about, do you
  545. 21:25actually think that could dent the HBM
  546. 21:28growth rate that we may see in 2028 or
  547. 21:31will HBM continue to be the undisputed
  548. 21:33king in the AI compute? Guys, that's the
  549. 21:36update I have for tonight. I hope you
  550. 21:38enjoyed the video. Make sure you hit
  551. 21:39like. Make sure you join Market Signal
  552. 21:41if you haven't done it already. and I'll
  553. 21:43see you on the next

About this transcript

This page contains the full transcript of HBM Under Threat? Are Microsoft, Intel & Marvell Changing the AI Memory Game by Market Signal, generated from the public captions YouTube serves with the video. The transcript has 3,745 words across 553 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.