WEBVTT 00:00:00.000 --> 00:00:02.180 So everyone seems to be shouting about an AI 00:00:02.180 --> 00:00:04.000 bubble these days. You hear it everywhere. People 00:00:04.000 --> 00:00:06.500 are dusting off the history books, right? Pointing 00:00:06.500 --> 00:00:10.900 straight at the 2001 .com crash. But looking 00:00:10.900 --> 00:00:12.740 through our sources today, it feels like we might 00:00:12.740 --> 00:00:14.380 be focused on the wrong kind of risk entirely 00:00:14.380 --> 00:00:16.420 this time. Yeah, it's interesting. That 2001 00:00:16.420 --> 00:00:18.359 bust, it was fundamentally about infrastructure, 00:00:18.579 --> 00:00:20.600 wasn't it? Too much supply. People call it dark 00:00:20.600 --> 00:00:23.820 fiber. All that empty, unused network capacity 00:00:23.820 --> 00:00:27.059 built on speculation. The paradox now, and it's 00:00:27.059 --> 00:00:30.719 fascinating, is what if we're seeing like maximized 00:00:30.719 --> 00:00:34.719 dark GPUs from NVIDIA? where every single chip, 00:00:34.759 --> 00:00:37.240 all the compute power is just firing on all cylinders 00:00:37.240 --> 00:00:39.549 the moment it's out the door. Welcome to the 00:00:39.549 --> 00:00:41.549 deep dive. That paradox, that's really our mission 00:00:41.549 --> 00:00:43.409 today. We're going to unpack these market dynamics 00:00:43.409 --> 00:00:46.149 that make the current capital environment feel 00:00:46.149 --> 00:00:48.530 so different from past tech booms. We'll look 00:00:48.530 --> 00:00:50.289 at the incredible speed of consumer adoption 00:00:50.289 --> 00:00:53.390 and also, importantly, the rise of these really 00:00:53.390 --> 00:00:55.869 powerful small AI models. They could really shake 00:00:55.869 --> 00:00:58.030 things up. Absolutely. And just to set the stage, 00:00:58.130 --> 00:01:00.390 our goal here is, you know, impartial analysis 00:01:00.390 --> 00:01:02.469 for you listening. We're not trying to validate 00:01:02.469 --> 00:01:05.370 the hype necessarily, but really understand the 00:01:05.370 --> 00:01:08.420 actual. mechanisms, the money, the usage, the 00:01:08.420 --> 00:01:11.840 tech that are driving this, well, unprecedented 00:01:11.840 --> 00:01:14.319 environment. Okay. Let's dive into the market 00:01:14.319 --> 00:01:16.920 dynamics first then. We're seeing just massive 00:01:16.920 --> 00:01:19.620 AI funding rounds. NVIDIA stock is hitting record 00:01:19.620 --> 00:01:22.700 highs almost constantly, it feels like. And the 00:01:22.700 --> 00:01:25.260 actual usage of the chips is described as maybe 00:01:25.260 --> 00:01:27.719 even bigger than the funding suggests. Right. 00:01:27.819 --> 00:01:31.180 And this is where comparing it to 2001 is... 00:01:31.519 --> 00:01:35.200 useful, but also maybe misleading. That telecom 00:01:35.200 --> 00:01:37.819 bust, it happened because companies poured billions 00:01:37.819 --> 00:01:40.840 into infrastructure. That empty dark fiber waiting 00:01:40.840 --> 00:01:42.780 for web traffic that just didn't show up fast 00:01:42.780 --> 00:01:44.920 enough is pure speculation on future capacity. 00:01:45.140 --> 00:01:47.680 And today, the sources we have are stressing 00:01:47.680 --> 00:01:50.140 the complete opposite scenario. There are, and 00:01:50.140 --> 00:01:53.560 this is a quote, no dark GPUs. The big players 00:01:53.560 --> 00:01:55.140 buying the chips, they're putting them to work 00:01:55.140 --> 00:01:56.799 immediately. It's actually a supply construct 00:01:56.799 --> 00:01:59.920 problem, not a demand one. Exactly. Which brings 00:01:59.920 --> 00:02:02.620 us to the hyperscalers. That's the term for the 00:02:02.620 --> 00:02:04.920 big cloud providers, right? Microsoft, Google, 00:02:05.219 --> 00:02:08.139 Meta, Amazon. They're spending just colossal 00:02:08.139 --> 00:02:10.560 amounts on data centers. Billions and billions. 00:02:10.780 --> 00:02:13.379 But, and this is key, they seem to be seeing 00:02:13.379 --> 00:02:16.139 a genuine, pretty immediate return on capital, 00:02:16.240 --> 00:02:20.460 or ROIC, from that huge spend. Because the capacity 00:02:20.460 --> 00:02:22.319 is getting soaked up instantly, either by their 00:02:22.319 --> 00:02:24.900 own projects or, you know, by their cloud customers. 00:02:25.370 --> 00:02:27.270 So it's not building empty stadiums, hoping the 00:02:27.270 --> 00:02:29.330 crowds will come later. If the infrastructure 00:02:29.330 --> 00:02:31.689 is getting used right away and generating a return, 00:02:31.909 --> 00:02:33.789 what does that actually mean for this whole bubble 00:02:33.789 --> 00:02:36.550 conversation? Why is this maybe financially safer 00:02:36.550 --> 00:02:39.789 than past manias? Well, it shifts the risk profile, 00:02:39.949 --> 00:02:41.849 doesn't it? Instead of risking oversupply and 00:02:41.849 --> 00:02:44.030 empty capacity, maybe the risk is more about 00:02:44.030 --> 00:02:47.009 depending on a really constrained supply chain. 00:02:47.129 --> 00:02:49.229 This is what one source called a fiber -rich 00:02:49.229 --> 00:02:51.629 pumpkin season. It implies it's harvest time 00:02:51.629 --> 00:02:53.750 now, not just planting seeds for some distant 00:02:53.750 --> 00:02:56.069 future. That's a powerful way to put it. So if 00:02:56.069 --> 00:02:58.409 every piece of this expensive compute hardware 00:02:58.409 --> 00:03:00.909 is being used to generate actual revenue now, 00:03:01.009 --> 00:03:03.469 the foundation of this room feels like it's built 00:03:03.469 --> 00:03:07.270 on present income streams, not just future hopes. 00:03:07.449 --> 00:03:09.550 That seems to be the core argument for why it's 00:03:09.550 --> 00:03:12.349 different. The growth is fueled by tangible, 00:03:12.449 --> 00:03:15.669 immediate use of infrastructure, not just betting 00:03:15.669 --> 00:03:18.939 on future capacity needs. OK. Speaking of you, 00:03:18.979 --> 00:03:21.379 Sid, let's pivot a bit to the cultural side because 00:03:21.379 --> 00:03:25.120 adoption is moving incredibly fast and sometimes 00:03:25.120 --> 00:03:27.199 with some pretty hilarious results along the 00:03:27.199 --> 00:03:28.560 way, which always gives us something to think 00:03:28.560 --> 00:03:30.539 about. Oh, absolutely. The speed is just incredible. 00:03:30.680 --> 00:03:33.960 Fun fact number one, OpenAI's Sora, their video 00:03:33.960 --> 00:03:35.979 generation model, it's exploded. Apparently it 00:03:35.979 --> 00:03:38.919 hit hashtag one on the App Store and they just 00:03:38.919 --> 00:03:40.819 launched character cameos. For a limited time, 00:03:40.939 --> 00:03:43.300 no invite code needed. You could literally put 00:03:43.300 --> 00:03:45.159 your cat. or I don't know, your kid's favorite 00:03:45.159 --> 00:03:48.539 toy, into a short video generated by AI. That's 00:03:48.539 --> 00:03:51.680 like immediate viral mainstream utility right 00:03:51.680 --> 00:03:54.199 there. That is a huge sign of accessibility hitting 00:03:54.199 --> 00:03:56.379 the consumer level. And then, of course, you 00:03:56.379 --> 00:03:59.520 have the famous fails, which kind of remind us 00:03:59.520 --> 00:04:01.259 these models are still pretty young, right? We 00:04:01.259 --> 00:04:03.939 saw that funny story about ChatGPT folding under 00:04:03.939 --> 00:04:06.159 almost zero pressure. That was amazing, yeah. 00:04:06.340 --> 00:04:09.639 A parent used some really complex obscure password 00:04:09.639 --> 00:04:13.520 as an anti -cheat. on their kid's computer. The 00:04:13.520 --> 00:04:15.680 kid just described the password setup vaguely 00:04:15.680 --> 00:04:18.379 to ChatGPT and asked it to figure it out. And 00:04:18.379 --> 00:04:20.240 boom, the model just gave it up immediately. 00:04:20.639 --> 00:04:24.600 Soft chuckle. Zero real pressure, total exposure. 00:04:24.800 --> 00:04:26.740 It really shows how complexity can sometimes 00:04:26.740 --> 00:04:30.100 just crumble with the right kind of simple, maybe 00:04:30.100 --> 00:04:32.579 unexpected prompt. It's a great little example 00:04:32.579 --> 00:04:36.199 of the unintended power of these accessible language 00:04:36.199 --> 00:04:38.439 models. And you see the platforms reacting to 00:04:38.439 --> 00:04:41.160 this kind of immediate access to open AI apparently 00:04:41.160 --> 00:04:43.420 had to stop chat GPT from giving out specific 00:04:43.420 --> 00:04:45.660 medical and legal advice, personalized advice. 00:04:45.720 --> 00:04:48.120 Anyway, they're clearly trying to manage the 00:04:48.120 --> 00:04:50.180 most obvious risks as people start using these 00:04:50.180 --> 00:04:52.850 tools for really serious. But the education side 00:04:52.850 --> 00:04:55.149 is trying to catch up fast as well. University 00:04:55.149 --> 00:04:57.410 of San Francisco, for instance, launched a free 00:04:57.410 --> 00:05:00.550 online AI prompt course aimed right at beginners. 00:05:00.670 --> 00:05:03.490 No coding required, just focusing on practical 00:05:03.490 --> 00:05:07.120 everyday uses. Which is needed. I mean, I still 00:05:07.120 --> 00:05:09.740 wrestle with prompt drift myself sometimes. Getting 00:05:09.740 --> 00:05:12.199 the exact output you want, it feels more like 00:05:12.199 --> 00:05:14.120 an art than a science occasionally. It really 00:05:14.120 --> 00:05:17.399 does. In this mix of real utility and sometimes 00:05:17.399 --> 00:05:19.560 just pure entertainment, it's definitely driving 00:05:19.560 --> 00:05:22.540 the money side. The culture and the capital feel 00:05:22.540 --> 00:05:25.639 completely linked. Like this AI singer, Zanya 00:05:25.639 --> 00:05:28.360 Monet, became the first AI artist to actually 00:05:28.360 --> 00:05:31.000 get on the Billboard Airplay charts. That's mass 00:05:31.000 --> 00:05:33.199 market consumption happening. And then the valuations 00:05:33.199 --> 00:05:36.399 just keep climbing. founders, all 22 years old, 00:05:36.500 --> 00:05:39.060 apparently beat Mark Zuckerberg's record. Youngest 00:05:39.060 --> 00:05:41.819 self -made billionaires. Their AI startup hit 00:05:41.819 --> 00:05:45.540 a $10 billion valuation after raising $350 million. 00:05:45.899 --> 00:05:48.779 That scale is just hard to grasp sometimes. And 00:05:48.779 --> 00:05:50.560 it's not just late stage money flooding in either. 00:05:50.959 --> 00:05:53.540 Sequoia Capital, the big VC firm, they're doubling 00:05:53.540 --> 00:05:55.939 down right at the beginning, launching two new 00:05:55.939 --> 00:06:00.240 funds. $950 million total just for early stage 00:06:00.240 --> 00:06:02.519 AI startups around the world. They are betting 00:06:02.519 --> 00:06:05.399 big right from the ground floor. And then, of 00:06:05.399 --> 00:06:07.600 course, the rumor mill keeps churning. The biggest 00:06:07.600 --> 00:06:10.699 one lately, OpenAI potentially exploring an IPO 00:06:10.699 --> 00:06:14.180 maybe in 2026 with a valuation that could hit 00:06:14.180 --> 00:06:17.459 up to $1 trillion. So here's the question then. 00:06:17.540 --> 00:06:20.649 Why do we keep seeing these like... astronomical 00:06:20.649 --> 00:06:23.269 funding rounds and valuations right alongside 00:06:23.269 --> 00:06:25.529 these immediate, sometimes kind of silly consumer 00:06:25.529 --> 00:06:27.949 fails. I think the massive investment reflects 00:06:27.949 --> 00:06:30.110 this huge belief in the long term potential, 00:06:30.350 --> 00:06:33.230 even with the very obvious, sometimes amusing 00:06:33.230 --> 00:06:36.069 short term limits we're seeing right now. Sponsor. 00:06:36.829 --> 00:06:39.189 Let's shift gears now to efficiency. Our sources 00:06:39.189 --> 00:06:41.189 suggest this might be the next really big frontier. 00:06:41.350 --> 00:06:43.290 The race isn't only about making models bigger 00:06:43.290 --> 00:06:45.670 and bigger anymore. It's also about packing powerful 00:06:45.670 --> 00:06:48.209 features into smaller, more efficient packages. 00:06:48.750 --> 00:06:51.829 Which brings us to IBM's Granite 4 .0 Nano, their 00:06:51.829 --> 00:06:54.189 smallest AI models released so far. Yeah, the 00:06:54.189 --> 00:06:57.089 Nano part is the real kicker here. And it speaks 00:06:57.089 --> 00:07:00.949 volumes about decentralization. First off, these 00:07:00.949 --> 00:07:03.750 models are open source. Apache 2 .0 license, 00:07:03.949 --> 00:07:06.310 which is pretty permissive. Developers can look 00:07:06.310 --> 00:07:09.389 inside, build on top of them freely. And crucially, 00:07:09.529 --> 00:07:11.610 they're optimized for efficient runtimes like 00:07:11.610 --> 00:07:14.689 this popular library called Llama .cpp. And for 00:07:14.689 --> 00:07:16.170 listeners, why should we care about something 00:07:16.170 --> 00:07:19.000 called Llama .cpp? Well, because it essentially 00:07:19.000 --> 00:07:21.740 lets these complex AI models run really fast, 00:07:21.899 --> 00:07:24.459 even on hardware that isn't specialized AI chips. 00:07:24.600 --> 00:07:26.879 Think your laptop, maybe even a cheap device 00:07:26.879 --> 00:07:28.769 at the edge. you know, without needing access 00:07:28.769 --> 00:07:31.550 to a massive server farm, it really democratizes 00:07:31.550 --> 00:07:34.089 who can run powerful AI. Okay. And importantly, 00:07:34.310 --> 00:07:36.089 they didn't seem to sacrifice training quality 00:07:36.089 --> 00:07:38.750 just to shrink the size. These nano models were 00:07:38.750 --> 00:07:41.990 trained on, what, over 15 trillion tokens? That's 00:07:41.990 --> 00:07:44.089 just a mind -boggling amount of data, like stacking 00:07:44.089 --> 00:07:45.970 Lego blocks of really high -quality information. 00:07:46.329 --> 00:07:48.949 And they use the same quality pipeline as IBM's 00:07:48.949 --> 00:07:51.529 much bigger models. Right. And maybe we should 00:07:51.529 --> 00:07:54.420 quickly define parameters here for clarity. When 00:07:54.420 --> 00:07:57.139 we talk about parameters in an AI model, we basically 00:07:57.139 --> 00:07:59.139 mean the size of its internal knowledge structure. 00:07:59.500 --> 00:08:01.680 Think of it like the number of connections or 00:08:01.680 --> 00:08:05.019 variables it uses to process information. Usually, 00:08:05.019 --> 00:08:07.319 more parameters mean better performance, but 00:08:07.319 --> 00:08:10.579 also way higher compute costs and energy use. 00:08:11.040 --> 00:08:13.560 But Granite Nano seems to be pushing back against 00:08:13.560 --> 00:08:16.100 that core assumption. It's showing you don't 00:08:16.100 --> 00:08:18.860 necessarily need 70 billion plus parameters to 00:08:18.860 --> 00:08:22.199 get really useful, impactful results. In benchmark 00:08:22.199 --> 00:08:24.420 tests, it actually came out on top against other 00:08:24.420 --> 00:08:26.800 models in its specific size class, including 00:08:26.800 --> 00:08:29.680 ones from Alibaba and Google like Quinn and Gemma. 00:08:30.170 --> 00:08:31.610 And what's really interesting is how well it 00:08:31.610 --> 00:08:35.090 did on what the sources call agenty stuff. So 00:08:35.090 --> 00:08:37.330 that means things like complex reasoning, generating 00:08:37.330 --> 00:08:40.269 code, tasks that need multiple steps to complete. 00:08:40.350 --> 00:08:42.190 It's not just spitting back facts it memorized. 00:08:42.470 --> 00:08:45.789 Whoa. OK, imagine staling that kind of capability, 00:08:46.149 --> 00:08:49.190 running it potentially billions of times on small 00:08:49.190 --> 00:08:52.700 local devices, even offline. That could totally 00:08:52.700 --> 00:08:54.759 change the economics for businesses wanting to 00:08:54.759 --> 00:08:57.860 deploy AI everywhere. IBM even certified it for 00:08:57.860 --> 00:09:00.740 responsible AI use, which is a huge deal for 00:09:00.740 --> 00:09:03.419 companies needing audit trails and, you know, 00:09:03.419 --> 00:09:05.759 some level of certainty. That efficiency is just 00:09:05.759 --> 00:09:08.519 critical. If you can do complex tasks, tasks 00:09:08.519 --> 00:09:11.240 that used to need these giant models on a small 00:09:11.240 --> 00:09:13.980 local chip, you slash the cost per query, the 00:09:13.980 --> 00:09:16.159 energy use, everything. So thinking about the 00:09:16.159 --> 00:09:18.460 market again, what is the success of these smaller 00:09:18.460 --> 00:09:21.700 open source models like Granite Nano mean? for 00:09:21.700 --> 00:09:23.820 the dominance of the big hyperscale players, 00:09:24.059 --> 00:09:26.299 the ones who currently own all that expensive, 00:09:26.539 --> 00:09:28.840 specialized compute power. It means competition 00:09:28.840 --> 00:09:31.340 is heating up fast, especially at the smaller, 00:09:31.440 --> 00:09:33.600 more efficient edge of the AI model spectrum. 00:09:33.960 --> 00:09:36.480 It forces the big guys to innovate beyond just 00:09:36.480 --> 00:09:38.860 sheer scale. Okay, let's try to bring this all 00:09:38.860 --> 00:09:40.889 together. Looking back at our sources today, 00:09:40.970 --> 00:09:43.909 we've really hit on three major dynamics shaping 00:09:43.909 --> 00:09:46.870 AI right now. First, the huge investment wave 00:09:46.870 --> 00:09:50.629 seems arguably justified by this heavily utilized, 00:09:50.909 --> 00:09:53.690 maxed out hardware. We're seeing real capital 00:09:53.690 --> 00:09:56.190 return, which feels different from pure speculation. 00:09:56.629 --> 00:10:00.889 The risk profile seems shifted. Second, consumer 00:10:00.889 --> 00:10:03.889 adoption is just incredibly fast. Sometimes funny, 00:10:04.029 --> 00:10:06.629 yeah, but undeniably happening. And that's driving 00:10:06.629 --> 00:10:08.169 the need for more investment, more hardware. 00:10:08.309 --> 00:10:11.019 And third, powerful AI is actually shrinking. 00:10:11.539 --> 00:10:13.580 Models like Granite Nano are proving that serious 00:10:13.580 --> 00:10:15.899 AI capability can run pretty much anywhere now 00:10:15.899 --> 00:10:18.259 and can compete effectively within specific size 00:10:18.259 --> 00:10:20.340 categories. And these three themes, they seem 00:10:20.340 --> 00:10:21.960 tied together by what people are calling this 00:10:21.960 --> 00:10:24.519 virtuous cycle that's driving the whole AI frenzy, 00:10:24.519 --> 00:10:27.240 right? More usage demands more chips, which justifies 00:10:27.240 --> 00:10:28.860 more spending on infrastructure, which leads 00:10:28.860 --> 00:10:30.940 to better models, which then drives more usage. 00:10:31.259 --> 00:10:34.330 And the cycle starts again. Yeah, exactly. And 00:10:34.330 --> 00:10:36.750 we also saw some quick hits showing the real 00:10:36.750 --> 00:10:39.590 -time risks and global spread, like Google having 00:10:39.590 --> 00:10:41.250 to pull their Gemma modder pretty quickly from 00:10:41.250 --> 00:10:43.269 their AI studio after some decimation claims 00:10:43.269 --> 00:10:46.429 popped up. Shows how volatile and legally tricky 00:10:46.429 --> 00:10:49.009 this still is. And meanwhile, you see Anthropic 00:10:49.009 --> 00:10:51.750 opening its first office in Asia Pacific and 00:10:51.750 --> 00:10:53.909 Tokyo. It underlines that this whole movement 00:10:53.909 --> 00:10:56.590 is definitely global now. So maybe here's the 00:10:56.590 --> 00:10:58.210 provocative thought to leave you with today. 00:10:58.889 --> 00:11:01.549 If immediate capital return, driven by maxing 00:11:01.549 --> 00:11:04.110 out all the current hardware, is what's validating 00:11:04.110 --> 00:11:07.090 this AI boom right now, what happens to the market 00:11:07.090 --> 00:11:09.830 structure and maybe those huge valuations when 00:11:09.830 --> 00:11:12.470 the core technology driving that return, like 00:11:12.470 --> 00:11:15.330 these granite nano models, becomes cheap, open 00:11:15.330 --> 00:11:18.110 source, and capable of running almost anywhere? 00:11:18.389 --> 00:11:20.370 That definitely raises a big question about the 00:11:20.370 --> 00:11:22.730 long -term value proposition for the hyperscalers, 00:11:22.730 --> 00:11:24.830 doesn't it? What's their unique edge when the 00:11:24.830 --> 00:11:27.190 underlying tech becomes potentially commoditized? 00:11:27.610 --> 00:11:29.690 Well, thank you for providing these sources and 00:11:29.690 --> 00:11:31.570 for diving into this material with us today. 00:11:31.669 --> 00:11:33.269 Until next time, out to real music.