AI stocks are crashing.. — Transcript

Full transcript

1. 0:00So, if you remember back in 2025 when
2. 0:02SASbased companies got ailerated because
3. 0:04of the fear around AI disrupting SAS,
4. 0:07well, this time the pendulum swing came
5. 0:09right back at AI, but now targeting
6. 0:11mostly at this layer in the AI stack
7. 0:13right here, infrastructure. And in this
8. 0:16layer right here, there's a specific
9. 0:17type of companies called Neoclouds that
10. 0:19primarily make money by renting their
11. 0:22compute to serve AI models. And
12. 0:24companies like Cororeef, Iron, Cipher
13. 0:26Digital, Nebus right here are all down
14. 0:29more than 50% or more in just the last
15. 0:31month alone. Ultimately, the problem
16. 0:33comes down to who's financing the
17. 0:35infrastructure layer. Most of us are
18. 0:37used to using AI agent in the
19. 0:39application layer and new models in the
20. 0:41model layer. And here, the cost of
21. 0:43innovation is usually in the millions of
22. 0:45dollars. And things tend to move really
23. 0:47fast in execution. But as soon as we
24. 0:49move down the infrastructure and the
25. 0:51chip layer, we're talking billions and
26. 0:53billions of dollars and few years of
27. 0:55time and both of them adds an element of
28. 0:58risk. And whoever ends up financing this
29. 1:00infrastructure layer right here has to
30. 1:03take on that risk and they can largely
31. 1:05be split between capital and debt. And
32. 1:07of course, we have more big tech
33. 1:09hyperscalers that also build data
34. 1:11centers here as well. And even for these
35. 1:13guys, we're sort of asking the same
36. 1:15question, which is who's funding this,
37. 1:17but at a much bigger scale. When we look
38. 1:19back a few years, hyperscalers like
39. 1:22Google, Meta, Amazon, Microsoft, and
40. 1:24Oracle funded most of their expansions
41. 1:27through cash flow rather than using debt
42. 1:29or external financing. And if we look at
43. 1:31that 2025 figure right here, you can see
44. 1:34that the amount of money that they spent
45. 1:36on the left, which is the capex, is less
46. 1:38than the operational cash flow that they
47. 1:41had for that year. And that was the case
48. 1:43for every preceding year as you can see.
49. 1:45But because our demand for data centers
50. 1:47outgrew to a point where this chart
51. 1:49alone can't really keep up with what
52. 1:51needs to be spent. When we look at the
53. 1:53projection going into the 2030s, this is
54. 1:55a lot more than what capex can sustain.
55. 1:58Which means now we have a funding gap
56. 1:59here which somehow needs to be filled in
57. 2:02order to meet the goal. And that's where
58. 2:04we have entities like private capitals,
59. 2:06competing neoclouds, private equity, and
60. 2:08banks providing the means to fund the
61. 2:11construction. And it can be done by
62. 2:13taking the full risks on themselves and
63. 2:15then renting them out or just servicing
64. 2:17debt to let data center builders to
65. 2:19actually build more data centers and
66. 2:21leasing the capacity out. And as simple
67. 2:23as all of this might seem, it's just one
68. 2:25of those things that deals start to get
69. 2:27really hairy. Because unless you really
70. 2:29understand how deals are made off
71. 2:32balance sheet, you're going to have to
72. 2:33wrap your head around something like
73. 2:35this, which is only one example of how
74. 2:38Meta is structuring their deal with Blue
75. 2:41Capital to build their data center.
76. 2:42Here, Meta owns 20% equity through a
77. 2:45subsidiary called Iris and through
78. 2:48another subsidiary called Pelican Leap
79. 2:50LLC. Meta is then leasing the Hyperion
80. 2:53DS center to run their models and the
81. 2:5520% of their earnings go back to this
82. 2:58other subsidiary and the rest 80% goes
83. 3:00to Bla Capital subsidiary that also has
84. 3:03obligation to pay the $27 billion debt
85. 3:06that they borrowed to construct this
86. 3:08data center in the first place. And all
87. 3:10of this is off balance sheet which means
88. 3:12Meta won't really have to show the $27
89. 3:15billion debt that was taken by Blue
90. 3:17Capital to construct the data center.
91. 3:19and more importantly their exposure to
92. 3:22pay the difference in market value if
93. 3:25they ever end up selling the data center
94. 3:27because Meta might not want to extend
95. 3:29the lease after 4 years in which case it
96. 3:32might be sold at a lower price than they
97. 3:34pay for and at that time the price
98. 3:36difference could be a big exposure
99. 3:38financially for Meta and this is just
100. 3:41one example out of many other data
101. 3:43centers that are out there now you might
102. 3:44be looking at this and say okay so if
103. 3:46the main question here is who's paying
104. 3:49for this. Why would lenders, whether
105. 3:51it's private credit or bank, actually
106. 3:53loan the money for a potentially
107. 3:55precarious situation like this?
108. 3:57Typically, lenders want certainty in
109. 3:59exchange for money. While at the same
110. 4:01time, companies like Meta want more
111. 4:03flexibility in their contract in case
112. 4:05better GPUs are needed to run their
113. 4:07models or they have to put their
114. 4:08resources somewhere else. And in this
115. 4:10case, lenders are trusting a company
116. 4:13like Meta given their track record and
117. 4:15their promise to pay the difference if
118. 4:18the data center ever ends up being sold
119. 4:20on the market. And Meta has to foot the
120. 4:22bill for the difference. In other words,
121. 4:24Meta's track record makes all of this an
122. 4:26investment grade loan. But companies
123. 4:28like OpenAI and Enthropic that are in
124. 4:30the model and the application layer
125. 4:32don't really have the same benefit which
126. 4:34means they have to lose out on the
127. 4:36flexibility in terms of contracts to
128. 4:39longerterm leases so that the lenders
129. 4:41are more willing to take the risk of
130. 4:43actually funding companies like OpenAI
131. 4:45and Anthropic. And for OpenAI's case,
132. 4:47one example is of 8 gawatt data center
133. 4:50in Ohio where OpenAI is contracted to
134. 4:52lease for 20 years which is super
135. 4:55inflexible. And because OpenAI isn't
136. 4:57considered an investment grade like Neta
137. 4:59was, Nvidia here is basically co-signing
138. 5:02a portion of this and potentially
139. 5:04guaranteeing up to $105 billion in case
140. 5:07OpenAI defaults and Nvidia can't find
141. 5:09another suitor for the space. As you can
142. 5:11see, every deal and every data center
143. 5:13more or less has different setups
144. 5:15financially to keep track of, which
145. 5:17makes this entire thing not very
146. 5:19transparent in terms of how funding is
147. 5:21actually structured for data center
148. 5:23buildouts. And this brings us to the
149. 5:25critical question now for Neoclouds. Why
150. 5:27did all of these stocks drop more than
151. 5:3050% in the past month? But first, let's
152. 5:32talk about jobs. Chances are you're one
153. 5:34of highly qualified software engineers
154. 5:36looking to work at great companies. And
155. 5:38ever since AI and agents changed our
156. 5:40industry, it's hard to get calibrated to
157. 5:42the changing job market. For example, if
158. 5:44you remember Rag, well, according to
159. 5:46Dex, demand for jobs related to Rag
160. 5:48dropped from 14% to 3% in just a few
161. 5:52months. I know things are changing. Dex
162. 5:54also found that new engineering roles in
163. 5:56major cities now have AI agents as part
164. 5:59of their role. And that's over 64% of
165. 6:01jobs, which is huge. Even something as
166. 6:03simple as using coding agents like
167. 6:05cursor and cloud code is basically
168. 6:08expected for engineers to be familiar,
169. 6:10showing more than 25% of companies
170. 6:12wanting this. Good thing Dex has your
171. 6:14back here bridging the gap for
172. 6:15sponsoring this video. Dex is your
173. 6:17talent agent where you can tell them
174. 6:19what kind of work you want, your salary
175. 6:21expectation and all of your preferences
176. 6:23around your working condition and Dex
177. 6:25matches you with high quality
178. 6:27opportunities where companies like
179. 6:28Cynthia, Lovable and Helsing on the
180. 6:30other side are looking for talented
181. 6:32people like you. Instead of going
182. 6:33through the pain in discovering jobs
183. 6:35that match you, let Dex bridge the gap
184. 6:37today in the everchanging job market as
185. 6:39AI continues to grow in adoption. Try
186. 6:41Dexay for jobs. I'll have the link in
187. 6:43the description below. When we go back
188. 6:45to the question, who's funding the data
189. 6:47centers and we trace back the dollars,
190. 6:49we can actually put together a decent
191. 6:51amount of understanding in the situation
192. 6:53right here. If you remember CDS or
193. 6:55credit default swaps from the subprime
194. 6:57mortgage crisis back in 2008, well, we
195. 6:59have the same CDS also exists for AI but
196. 7:02for SoftBank who funding a large amount
197. 7:04of data center buildouts here in the US
198. 7:07as part of OpenAS Stargate project. Now,
199. 7:09the role of the CDS is essentially that
200. 7:11it provides protection in case Soft
201. 7:14Banks ever ends up defaulting on their
202. 7:16debt. And people can buy CDS even if
203. 7:19they aren't part of the deal that made
204. 7:20SoftBank's loan going into the data
205. 7:23center. Which means outside of people
206. 7:24who actually use CDS to hedge their
207. 7:27position, people can speculate on the
208. 7:29fact that SoftBank can't actually meet
209. 7:31their debt obligation. So whoever ends
210. 7:33up insuring this loan demands a premium
211. 7:36each year from the person who's saying
212. 7:38that they want protection in case that
213. 7:40it actually does fail. And recently this
214. 7:43very premium for SoftBank has been
215. 7:45rising to 466 basis point which shows
216. 7:48that people are now demanding more money
217. 7:51to actually ensure SoftBank's credit
218. 7:53risk. And there's already been rumors
219. 7:55floating around that SoftBank has
220. 7:57already defaulted on their turbine
221. 7:59orders for one of their data centers.
222. 8:01And to make the matter worse, SoftBank
223. 8:03subsidiary SB Energy also recently
224. 8:06delayed their IPO because the $50
225. 8:08billion valuation they were initially
226. 8:10hoping to get wasn't really met by the
227. 8:12investors. So collectively, when we look
228. 8:15at the rising spread of credit default
229. 8:17swap and the delayed IPO because
230. 8:20investors weren't willing to pay the
231. 8:22high premium, you're seeing more and
232. 8:24more signs of skepticism in the market
233. 8:26right here. And if a giant company like
234. 8:28SoftBank is dealing with investors that
235. 8:31are demanding more money for taking the
236. 8:33risk, it certainly makes us think about
237. 8:35how much tougher it might be for smaller
238. 8:38players, primarily neoclouds who have to
239. 8:41fight the same battle but with far less
240. 8:43to offer. Hyperscalers that largely use
241. 8:46their own fund to build data centers and
242. 8:48neoclouds that largely rely on debts.
243. 8:51You're going to see a huge difference in
244. 8:53capital drying up. looking very
245. 8:55differently here because big techs like
246. 8:57Google, Microsoft and Meta can
247. 9:00comfortably absorb some of the pullbacks
248. 9:02and delays and data centers up to the
249. 9:05point that they've already committed.
250. 9:06But for NeoClouds where a lot of them
251. 9:08are in the middle of expanding their
252. 9:10infrastructure is sort of stuck in the
253. 9:12situation. And when we look at the
254. 9:14market price, we see a huge difference
255. 9:16looking back at the one-mon chart on the
256. 9:18left versus on the right, where the
257. 9:20impact is significantly worse for
258. 9:22NeoClouds lately than it has been for
259. 9:25hyperscalers. As you can see, when we
260. 9:26also look at the model layer, there's
261. 9:28also been more development among the
262. 9:30frontier labs to pace AI's capability
263. 9:33back in September. And this also adds to
264. 9:35the uncertainty around just what the
265. 9:38long-term demand would look like in the
266. 9:40infrastructure layer. And to add to that
267. 9:41the changing political agenda here in
268. 9:44the US is also making what's already
269. 9:46uncertain situation even more so where
270. 9:49even a state like Texas that's arguably
271. 9:51been one of the most popular state for
272. 9:53data centers but now the new governor
273. 9:55elect is proposing that they'll actually
274. 9:57extend data center restrictions until
275. 10:00later on. So ultimately, we're kind of
276. 10:02in this weird space right now where the
277. 10:04cost of intelligence is dropping and AI
278. 10:07continues to advance in technology, but
279. 10:09also intentionally pacing. And AI
280. 10:12adoption is arguably still growing in
281. 10:15the application layer, while at the same
282. 10:16time, GPU rental price for newer
283. 10:19generation chips has been rising pretty
284. 10:21fast lately. And the risk in the
285. 10:23infrastructure layer is higher than it's
286. 10:25been since AI boom has started. And I
287. 10:27think that leaves something for all of
288. 10:28us to think about because AI could truly
289. 10:31continue to succeed like it has been
290. 10:33while some companies supporting them
291. 10:35might continue to struggle financially
292. 10:37due to many reasons some of which I
293. 10:39outline in this video.