WEBVTT 00:00:00.000 --> 00:00:04.240 We are living in the future. Beat. But it is 00:00:04.240 --> 00:00:06.759 a remarkably messy one. Yeah, that's kind of 00:00:06.759 --> 00:00:08.880 the best way to put it. We build machines to 00:00:08.880 --> 00:00:11.779 find truth. Sometimes they just dream instead. 00:00:12.019 --> 00:00:14.640 They absolutely do. Consider a hallucinating 00:00:14.640 --> 00:00:18.300 AI trying to verify a war zone video. Instead 00:00:18.300 --> 00:00:20.640 of finding facts, it generates a fake image. 00:00:20.940 --> 00:00:23.140 Right. It literally fabricates the scene. And 00:00:23.140 --> 00:00:26.530 then it offers that hallucination as... actual 00:00:26.530 --> 00:00:29.289 proof. Which completely shatters our baseline 00:00:29.289 --> 00:00:32.229 for reality. The technology is sprinting right 00:00:32.229 --> 00:00:36.380 now. Our guardrails are Barely jogging. It's 00:00:36.380 --> 00:00:38.799 a profound tension. Welcome to the Deep Dive. 00:00:38.840 --> 00:00:40.460 We're really glad you're here with us. We are 00:00:40.460 --> 00:00:42.840 pulling from a fascinating stack of sources today. 00:00:42.960 --> 00:00:45.119 It's a heavy stack, too. We've got major tectonic 00:00:45.119 --> 00:00:47.060 shifts happening across the entire industry. 00:00:47.299 --> 00:00:49.280 Okay, let's unpack this. We're looking at the 00:00:49.280 --> 00:00:52.100 new consumer AI rankings today. We're also exploring 00:00:52.100 --> 00:00:55.740 a massive graveyard of 5 ,700 failed startups. 00:00:56.039 --> 00:00:59.020 It's a literal graveyard of ideas. Exactly. After 00:00:59.020 --> 00:01:01.200 that, we'll examine a billion -dollar push for 00:01:01.200 --> 00:01:04.069 reasoning AI. And finally, Jensen Huang's trillion 00:01:04.069 --> 00:01:06.489 -dollar theory of our physical future. That last 00:01:06.489 --> 00:01:09.590 one is a massive five -layer cake. It explains 00:01:09.590 --> 00:01:11.709 literally everything happening in the hardware 00:01:11.709 --> 00:01:14.510 space right now. It really does. Beep. Let's 00:01:14.510 --> 00:01:17.189 begin with the new global order. Andresen Horowitz 00:01:17.189 --> 00:01:19.689 just released their latest consumer AI report. 00:01:19.950 --> 00:01:22.849 Right. The A16's top 100. This is their sixth 00:01:22.849 --> 00:01:25.349 edition. And this is the list everyone in the 00:01:25.349 --> 00:01:27.700 industry actually watches. Because it filters 00:01:27.700 --> 00:01:30.579 out the empty hype. It measures raw, verifiable 00:01:30.579 --> 00:01:32.879 user traffic. It shows us what you are actually 00:01:32.879 --> 00:01:35.640 using daily. And the landscape is heavily consolidated 00:01:35.640 --> 00:01:39.180 at the top. ChatGPT still completely dominates 00:01:39.180 --> 00:01:41.599 the space. The scale is just staggering. They 00:01:41.599 --> 00:01:44.780 crossed 900 million weekly active users. Whoa, 00:01:45.019 --> 00:01:48.019 imagine scaling to a billion queries. The server 00:01:48.019 --> 00:01:50.700 load alone is mind -bending. It's a massive portion 00:01:50.700 --> 00:01:52.799 of humanity. Nearly a billion people talking 00:01:52.799 --> 00:01:55.810 to a machine every week. Two sec silence. But 00:01:55.810 --> 00:01:58.030 the willingness to pay is also shifting. It is. 00:01:58.129 --> 00:02:00.269 Claude and Gemini are seeing massive financial 00:02:00.269 --> 00:02:03.049 traction. Their paid subscriptions skyrocketed 00:02:03.049 --> 00:02:06.109 over 200 % last year. People only pay for subscriptions 00:02:06.109 --> 00:02:08.729 when a tool actively saves them time. Exactly. 00:02:08.889 --> 00:02:10.789 It has to solve a tangible, painful problem. 00:02:11.030 --> 00:02:14.389 It's graduating from a novelty toy to core enterprise 00:02:14.389 --> 00:02:18.250 infrastructure. But the report highlights a much 00:02:18.250 --> 00:02:21.750 bigger structural shift. Traditional legacy apps 00:02:21.750 --> 00:02:24.400 are suddenly dominating the list. This is a critical 00:02:24.400 --> 00:02:26.840 pivot. We're talking about platforms like Canva, 00:02:26.939 --> 00:02:30.199 CapCut, Notion, and Grammarly. Right. They aren't 00:02:30.199 --> 00:02:33.629 new AI startups. No, they are established. giants, 00:02:33.750 --> 00:02:36.610 but they're surviving because they're adopting 00:02:36.610 --> 00:02:38.949 embedded AI. They're weaving the intelligence 00:02:38.949 --> 00:02:41.509 directly into their existing workflows. We should 00:02:41.509 --> 00:02:44.689 clarify that term. Embedded AI simply means intelligence 00:02:44.689 --> 00:02:47.150 built directly into software you already use. 00:02:47.310 --> 00:02:49.389 It's the intelligence coming to your cursor. 00:02:49.650 --> 00:02:51.409 It's a brilliant defensive strategy. I mean, 00:02:51.409 --> 00:02:53.750 why would you open a separate tab for ChatGPT 00:02:53.750 --> 00:02:56.530 to write a memo when Notion has the exact same 00:02:56.530 --> 00:02:59.110 capability baked right into the blank page? The 00:02:59.110 --> 00:03:02.009 friction is just gone. Completely gone. It neutralizes 00:03:02.009 --> 00:03:04.770 the threat of new standalone tools. But this 00:03:04.770 --> 00:03:07.509 report also maps out a hard geographical fracture. 00:03:07.750 --> 00:03:11.129 The global AI market is splitting apart. What's 00:03:11.129 --> 00:03:13.030 fascinating here is the geopolitical reality 00:03:13.030 --> 00:03:17.490 of that split. We're seeing three highly distinct 00:03:17.490 --> 00:03:20.889 isolated ecosystems emerge globally. The West 00:03:20.889 --> 00:03:23.289 has its well -known sandbox, but the Chinese 00:03:23.289 --> 00:03:26.090 apps are aggressively capturing global market 00:03:26.090 --> 00:03:28.789 share. They really are. Apps like Manos hit number 00:03:28.789 --> 00:03:31.969 44 on the global list. Ginsbark fought its way 00:03:31.969 --> 00:03:34.319 to number 47. They aren't just serving domestic 00:03:34.319 --> 00:03:36.400 users anymore. No, they're competing on the world 00:03:36.400 --> 00:03:39.319 stage. Meanwhile, the Russian ecosystem is isolating 00:03:39.319 --> 00:03:41.879 almost entirely. They're building a walled digital 00:03:41.879 --> 00:03:44.939 garden. Yeah. OpenClaw is notably absent from 00:03:44.939 --> 00:03:47.120 the rankings, though to be fair, that's largely 00:03:47.120 --> 00:03:49.539 due to the report's specific timeframe. It's 00:03:49.539 --> 00:03:52.659 essentially a new digital Cold War. Every superpower 00:03:52.659 --> 00:03:55.620 realizes something profound. Whoever controls 00:03:55.620 --> 00:03:57.919 the foundational models controls the cultural 00:03:57.919 --> 00:04:00.500 narrative. Foundational models being the massive 00:04:00.500 --> 00:04:03.759 general purpose engines. powering most AI applications. 00:04:04.099 --> 00:04:07.199 Right. They dictate how a society processes information. 00:04:07.719 --> 00:04:11.680 Beat. Does embedded AI ultimately kill the standalone 00:04:11.680 --> 00:04:15.159 app? Standalone innovates. Embedded scales. We 00:04:15.159 --> 00:04:17.540 absolutely need both to push boundaries. That 00:04:17.540 --> 00:04:19.899 brings up a natural question of execution. Let's 00:04:19.899 --> 00:04:22.439 examine today's triumphs, fails, and strange 00:04:22.439 --> 00:04:24.560 hallucinations. The daily reality of working 00:04:24.560 --> 00:04:26.879 with these models is a total roller coaster. 00:04:27.339 --> 00:04:30.279 There's a fascinating project tracking dead AI 00:04:30.279 --> 00:04:33.600 startups. It's a literal graveyard listing over 00:04:33.600 --> 00:04:37.279 5 ,700 failed companies. It sounds depressing, 00:04:37.579 --> 00:04:40.139 but it's actually an incredible strategic playbook. 00:04:40.319 --> 00:04:43.160 How so? It's a map of early adopter demand. People 00:04:43.160 --> 00:04:45.879 clearly wanted these tools to exist. The timing 00:04:45.879 --> 00:04:48.079 was just completely wrong. Think about the early 00:04:48.079 --> 00:04:50.899 virtual assistants from around 2012 or 2015. 00:04:51.319 --> 00:04:54.040 Exactly. Back then, founders had to manually 00:04:54.040 --> 00:04:57.920 code every possible user intent. If a user asked 00:04:57.920 --> 00:05:00.480 a question slightly differently, the whole system 00:05:00.480 --> 00:05:03.120 broke. Right. It was brutally expensive and technically 00:05:03.120 --> 00:05:05.660 impossible to scale. The unit economics were 00:05:05.660 --> 00:05:08.319 completely broken. But today, a large language 00:05:08.319 --> 00:05:10.860 model handles that out of the box. It understands 00:05:10.860 --> 00:05:13.100 the intent naturally. And it does it for fractions 00:05:13.100 --> 00:05:15.660 of a penny. It's like finding an abandoned gold 00:05:15.660 --> 00:05:17.980 mine. You go back after the invention of modern 00:05:17.980 --> 00:05:20.720 drilling. Suddenly, what was impossible is now 00:05:20.720 --> 00:05:22.879 highly profitable. The raw materials haven't 00:05:22.879 --> 00:05:26.000 changed. Our extraction tools have evolved. Exactly. 00:05:26.019 --> 00:05:28.600 Speaking of execution, OpenAI just shipped a 00:05:28.600 --> 00:05:31.500 quiet revolution for education. Oh, this is huge. 00:05:31.660 --> 00:05:34.079 They added highly interactive math and science 00:05:34.079 --> 00:05:38.019 tools directly into ChatGPT. It covers over 70 00:05:38.019 --> 00:05:40.560 different complex concepts. You can tweak a physics 00:05:40.560 --> 00:05:42.959 equation and watch the corresponding graph change 00:05:42.959 --> 00:05:45.740 instantly. It feels exactly like stacking Lego 00:05:45.740 --> 00:05:48.279 blocks of data. It makes abstract theoretical 00:05:48.279 --> 00:05:51.750 math incredibly tactile. It transforms a static 00:05:51.750 --> 00:05:54.850 answer into a dynamic playground. But as our 00:05:54.850 --> 00:05:57.449 tools get sharper, our shared reality gets much 00:05:57.449 --> 00:05:59.769 blurrier. Yeah, the information ecosystem is 00:05:59.769 --> 00:06:02.350 getting dangerously murky right now. Fake AI 00:06:02.350 --> 00:06:04.709 -generated visuals are spreading rapidly on X. 00:06:04.850 --> 00:06:07.370 These are highly realistic images specifically 00:06:07.370 --> 00:06:10.110 depicting the war in Iran. And this is where 00:06:10.110 --> 00:06:13.769 the guardrails completely failed. Grok. The platform's 00:06:13.769 --> 00:06:16.689 native AI attempted to step in. A video of a 00:06:16.689 --> 00:06:19.129 missile strike circulated online. Grok analyzed 00:06:19.129 --> 00:06:21.990 the video to verify its authenticity. But instead 00:06:21.990 --> 00:06:24.449 of flagging it, Grok hallucinated. It actually 00:06:24.449 --> 00:06:27.889 generated its own fake AI image to serve as proof 00:06:27.889 --> 00:06:30.189 that the event happened. It manufactured alternative 00:06:30.189 --> 00:06:33.430 evidence to support a completely fabricated reality. 00:06:33.790 --> 00:06:37.709 It is a profound systemic vulnerability. The 00:06:37.709 --> 00:06:40.350 mechanism designed to protect truth became the 00:06:40.350 --> 00:06:44.060 primary engine of deception. Beat. How do we 00:06:44.060 --> 00:06:47.180 trust AI to verify information? We can't yet. 00:06:47.300 --> 00:06:50.560 AI checking AI still creates massive blind spots. 00:06:50.720 --> 00:06:53.180 That vulnerability requires incredible computing 00:06:53.180 --> 00:06:56.160 power to fix. Which brings us to our third segment. 00:06:56.589 --> 00:06:59.089 the trillion -dollar infrastructure race. The 00:06:59.089 --> 00:07:00.930 physical world is desperately trying to catch 00:07:00.930 --> 00:07:03.569 up to the software. We are seeing a massive geographic 00:07:03.569 --> 00:07:07.110 footprint expanding. In Mississippi, XAI just 00:07:07.110 --> 00:07:09.730 secured approval for a giant new data center. 00:07:09.949 --> 00:07:12.790 And the sheer scale of this facility is difficult 00:07:12.790 --> 00:07:14.649 to overstate. It's causing immense friction. 00:07:14.930 --> 00:07:17.089 Local residents are raising serious environmental 00:07:17.089 --> 00:07:19.529 concerns about potential pollution and water 00:07:19.529 --> 00:07:22.449 usage. Right. Meanwhile, regional officials are 00:07:22.449 --> 00:07:24.370 pointing to the massive economic investment. 00:07:26.050 --> 00:07:28.589 physical dispute over a digital footprint. You 00:07:28.589 --> 00:07:31.329 just can't run the digital future without aggressively 00:07:31.329 --> 00:07:33.430 terraforming the physical present. It requires 00:07:33.430 --> 00:07:35.970 a staggering amount of raw energy. Look at the 00:07:35.970 --> 00:07:38.550 new venture from OpenAI's co -founder. It's called 00:07:38.550 --> 00:07:40.980 the Thinking Machines Lab. They just signed a 00:07:40.980 --> 00:07:43.699 massive infrastructure deal with Nvidia. They 00:07:43.699 --> 00:07:47.459 plan to run one full gigawatt of AI compute starting 00:07:47.459 --> 00:07:50.720 in 2027. We should define that. Compute is the 00:07:50.720 --> 00:07:53.120 processing power needed to run heavy software. 00:07:53.459 --> 00:07:55.899 And a gigawatt is literally the power required 00:07:55.899 --> 00:07:58.839 to run a medium -sized city. It's the output 00:07:58.839 --> 00:08:02.420 of a nuclear reactor dedicated entirely to thinking 00:08:02.420 --> 00:08:05.759 machines. Nvidia isn't just supplying them. They 00:08:05.759 --> 00:08:08.810 are actively investing in the lab. Meanwhile, 00:08:09.089 --> 00:08:12.970 Meta is making very strange, aggressive acquisitions. 00:08:13.170 --> 00:08:15.769 They just bought a viral platform called Moldbook. 00:08:15.990 --> 00:08:18.170 It's essentially a social network built entirely 00:08:18.170 --> 00:08:20.569 for AI agents. It's a bizarre concept at first 00:08:20.569 --> 00:08:22.870 glance. I mean, why do software programs need 00:08:22.870 --> 00:08:25.170 a social network? Why do they? Because future 00:08:25.170 --> 00:08:27.709 agents need an environment to negotiate. Imagine 00:08:27.709 --> 00:08:30.009 your AI scheduling assistant needs to bargain 00:08:30.009 --> 00:08:32.789 with my AI travel agent. They need a shared digital 00:08:32.789 --> 00:08:36.009 protocol to exchange data and trade favors. Exactly. 00:08:36.480 --> 00:08:38.440 The Moldbook founders are now joining Meta's 00:08:38.440 --> 00:08:40.799 core superintelligence team. Beat, here's where 00:08:40.799 --> 00:08:43.940 it gets really interesting. Yann LeCun's new 00:08:43.940 --> 00:08:47.379 startup, AMI, just raised massive capital. They 00:08:47.379 --> 00:08:51.179 pulled in $1 .03 billion. Their valuation is 00:08:51.179 --> 00:08:54.220 hitting at $3 .5 billion. For a brand new company, 00:08:54.379 --> 00:08:58.259 that's wild. The money is huge, but the architectural 00:08:58.259 --> 00:09:01.509 philosophy is what actually matters. Lacoon wants 00:09:01.509 --> 00:09:04.970 to abandon how we currently build AI. He is actively 00:09:04.970 --> 00:09:08.049 rejecting the architecture behind ChatGPT and 00:09:08.049 --> 00:09:10.710 Claude. Today's models are fundamentally prediction 00:09:10.710 --> 00:09:13.309 engines. They've read the entire internet. Right. 00:09:13.409 --> 00:09:16.110 They are incredibly good at mimicking logic by 00:09:16.110 --> 00:09:18.549 simply guessing the most statistically probable 00:09:18.549 --> 00:09:21.669 next word. Right. And they sound brilliant. Yeah. 00:09:22.059 --> 00:09:24.299 But they don't actually possess a mental model 00:09:24.299 --> 00:09:26.600 of physics. They don't truly understand cause 00:09:26.600 --> 00:09:29.480 and effect. They just string plausible words 00:09:29.480 --> 00:09:31.960 together. Lacoon argues that predicting words 00:09:31.960 --> 00:09:34.559 is a dead end. AMI is trying to build a system 00:09:34.559 --> 00:09:37.320 that can genuinely reason. It needs to plan 10 00:09:37.320 --> 00:09:39.320 steps ahead in the physical world. It has to 00:09:39.320 --> 00:09:41.679 hold a persistent worldview. If it drops a virtual 00:09:41.679 --> 00:09:43.700 glass, it needs to know it shatters. Without 00:09:43.700 --> 00:09:45.860 just quoting a Wikipedia article about gravity. 00:09:46.039 --> 00:09:48.460 Two sec silence. Are prediction -based models 00:09:48.460 --> 00:09:51.059 finally hitting a wall? Yes, predicting the next 00:09:51.059 --> 00:09:53.240 word isn't the same as actual real -world planning. 00:09:53.440 --> 00:09:54.740 We're going to take a quick second, but when 00:09:54.740 --> 00:09:56.779 we come back, we're cutting into a five -layer 00:09:56.779 --> 00:09:59.860 cake. Sponsor. Welcome back to the Deep Dive. 00:09:59.860 --> 00:10:02.539 Let's move into our fourth segment. We are looking 00:10:02.539 --> 00:10:05.460 at rapid -fire empowered tools. These aren't 00:10:05.460 --> 00:10:07.679 theoretical models. These are agentic workflows 00:10:07.679 --> 00:10:10.519 you can deploy right now. Exactly. First up is 00:10:10.519 --> 00:10:13.340 a tool called Vozo. It's designed for global 00:10:13.340 --> 00:10:15.960 video distribution. It translates the spoken 00:10:15.960 --> 00:10:18.679 text in your videos. But the killer feature is 00:10:18.679 --> 00:10:21.340 that it preserves the original visual timing. 00:10:21.460 --> 00:10:23.120 Right. If you have a massive corporate slide 00:10:23.120 --> 00:10:25.740 deck, you record it once, and Vozo seamlessly 00:10:25.740 --> 00:10:29.080 dubs it into 20 languages without breaking the 00:10:29.080 --> 00:10:31.840 visual sync. It completely removes the friction 00:10:31.840 --> 00:10:34.799 of global communication. Then we have Chronicle 00:10:34.799 --> 00:10:38.250 2 .0. This acts as an autonomous AI design partner. 00:10:38.409 --> 00:10:41.190 It takes your chaotic, raw notes and instantly 00:10:41.190 --> 00:10:44.250 turns them into polished, highly visual slides. 00:10:44.470 --> 00:10:46.669 It automatically and strictly adheres to your 00:10:46.669 --> 00:10:49.309 specific brand guidelines. Next on the stack 00:10:49.309 --> 00:10:52.590 is Sitefire .ai. This one shifts from design 00:10:52.590 --> 00:10:55.539 into aggressive distribution. It deploys specialized 00:10:55.539 --> 00:10:59.080 agents to crawl the web. They analyze exactly 00:10:59.080 --> 00:11:02.080 what type of content actually drives high -quality 00:11:02.080 --> 00:11:04.980 citations. Then those agents write brand -aware 00:11:04.980 --> 00:11:08.259 articles tailored to that data. They even generate 00:11:08.259 --> 00:11:10.840 custom outreach emails to secure the backlinks. 00:11:11.039 --> 00:11:14.440 It automates the entire SEO lifecycle. Finally, 00:11:14.440 --> 00:11:18.279 there's Fish Audio S2. This one is incredible, 00:11:18.500 --> 00:11:21.740 but admittedly a little terrifying. It generates 00:11:21.740 --> 00:11:24.779 highly realistic synthetic voices. It captures 00:11:24.779 --> 00:11:27.840 deep emotional nuance, subtle breathing, and 00:11:27.840 --> 00:11:30.220 natural rhythm. And it accomplishes all of that 00:11:30.220 --> 00:11:32.559 using only 10 seconds of reference audio. 10 00:11:32.559 --> 00:11:35.519 seconds to permanently clone a human voice. Beat. 00:11:35.720 --> 00:11:38.220 But managing these tools is inherently chaotic. 00:11:38.480 --> 00:11:40.860 I still wrestle with prompt drift myself. Oh, 00:11:40.860 --> 00:11:42.779 absolutely. Everyone does. Prompt shift is when 00:11:42.779 --> 00:11:44.940 an AI slowly forgets your original instructions. 00:11:45.019 --> 00:11:47.320 You start a session and the output is flawless. 00:11:47.659 --> 00:11:50.320 20 prompts later, the model has completely lost 00:11:50.320 --> 00:11:52.279 the plot. It starts wandering off topic. You 00:11:52.279 --> 00:11:54.039 have to constantly wrangle it back to the original 00:11:54.039 --> 00:11:56.539 constraints. It requires exhausting vigilance. 00:11:56.740 --> 00:11:59.340 That's the core problem these new tools solve. 00:11:59.519 --> 00:12:02.039 They handle the complex prompting invisibly. 00:12:02.360 --> 00:12:04.879 The user just pushes a button. The agent handles 00:12:04.879 --> 00:12:07.480 the drift, the context window, and the error 00:12:07.480 --> 00:12:09.480 correction behind the scenes. What connects all 00:12:09.480 --> 00:12:12.100 these new tools? They all turn complex, multi 00:12:12.100 --> 00:12:15.240 -step tasks into simple, one -click agent workflows. 00:12:15.639 --> 00:12:18.700 That transition from chatting to delegating is 00:12:18.700 --> 00:12:22.860 everything. Beat. Which leads us perfectly into 00:12:22.860 --> 00:12:26.139 our final segment, Jensen Huang's five -layer 00:12:26.139 --> 00:12:28.139 AI cake. If you want to understand the economics 00:12:28.139 --> 00:12:30.820 of the next decade, this is The Master Thesis. 00:12:31.200 --> 00:12:34.299 Jensen Huang is the CEO of Nvidia. He recently 00:12:34.299 --> 00:12:37.379 published a rare, deeply analytical blog post. 00:12:37.500 --> 00:12:39.139 He argued that the hundreds of billions we've 00:12:39.139 --> 00:12:41.960 spent so far, that's just the warm up. He stated 00:12:41.960 --> 00:12:44.200 plainly that the build out will require trillions 00:12:44.200 --> 00:12:47.360 more. Trillions with a T. He believes AI is no 00:12:47.360 --> 00:12:50.039 longer a software category. It is core infrastructure. 00:12:50.460 --> 00:12:52.899 It is exactly as vital as the digital economy's 00:12:52.899 --> 00:12:54.539 power grid. If we connect this to the bigger 00:12:54.539 --> 00:12:56.740 picture. This perfectly explains the gigawatt 00:12:56.740 --> 00:12:59.440 deals and those fierce data center controversies 00:12:59.440 --> 00:13:01.440 in Mississippi. It's all connected through his 00:13:01.440 --> 00:13:04.019 five -layer stack model. Let's break it down. 00:13:04.200 --> 00:13:07.220 Layer one, the foundation, is energy. This means 00:13:07.220 --> 00:13:10.600 literal power plants, nuclear reactors, solar 00:13:10.600 --> 00:13:13.460 farms, and the raw electricity grid. Layer two 00:13:13.460 --> 00:13:16.889 is chips. These are the GPUs. GPUs are specialized 00:13:16.889 --> 00:13:20.389 chips designed to process multiple tasks simultaneously. 00:13:20.990 --> 00:13:24.049 This is NVIDIA's absolute dominance. Layer three 00:13:24.049 --> 00:13:27.250 is infrastructure. The massive, sprawling data 00:13:27.250 --> 00:13:30.070 centers, the advanced cooling systems, and thousands 00:13:30.070 --> 00:13:32.809 of miles of networking cables. Layer four is 00:13:32.809 --> 00:13:35.129 models. This is the software brain. The foundational 00:13:35.129 --> 00:13:38.409 models like ChatGPT, Claude, and LeCun's new 00:13:38.409 --> 00:13:40.929 reasoning engine. And finally, layer five, the 00:13:40.929 --> 00:13:43.570 peak, is applications. The specific tools we 00:13:43.570 --> 00:13:46.919 just discussed. Vozo, Sitefire, Copilot, and 00:13:46.919 --> 00:13:49.940 the chatbots. Kwong describes a massive, inescapable 00:13:49.940 --> 00:13:52.960 economic gravity. Demand sits purely at Layer 00:13:52.960 --> 00:13:55.720 5. You and I use an application to write an email. 00:13:55.899 --> 00:13:58.259 That application pulls computing power from the 00:13:58.259 --> 00:14:00.740 Layer 4 model. To run that model, it aggressively 00:14:00.740 --> 00:14:03.120 pulls on the Layer 3 data center infrastructure. 00:14:03.460 --> 00:14:05.720 Which forces those data centers to buy millions 00:14:05.720 --> 00:14:08.700 of Layer 2 computer chips. And those millions 00:14:08.700 --> 00:14:12.399 of chips put a catastrophic strain on the layer 00:14:12.399 --> 00:14:15.779 one raw electricity grid. Demand at the top strains 00:14:15.779 --> 00:14:18.299 every single physical layer below it. It's a 00:14:18.299 --> 00:14:20.980 violent chain reaction. It drives relentless 00:14:20.980 --> 00:14:23.519 physical expansion and intense hiring across 00:14:23.519 --> 00:14:26.539 every single industry in the stack. Is the energy 00:14:26.539 --> 00:14:29.320 layer the ultimate bottleneck? Absolutely. Without 00:14:29.320 --> 00:14:31.879 power plants, there are no chatbots. Energy is 00:14:31.879 --> 00:14:35.019 the ceiling. That is a sobering physical limitation 00:14:35.019 --> 00:14:38.899 to sex silence. So what does this all mean? It 00:14:38.899 --> 00:14:42.220 means we have to stop viewing AI as just an app 00:14:42.220 --> 00:14:44.539 on our phones. We started today discussing 900 00:14:44.539 --> 00:14:48.580 million weekly chat GPT users. That massive number 00:14:48.580 --> 00:14:50.740 sits at the very top of the layer cake. And from 00:14:50.740 --> 00:14:52.679 the user's perspective, the application just 00:14:52.679 --> 00:14:54.580 feels like weightless magic. But that digital 00:14:54.580 --> 00:14:57.220 magic is anchored to a brutal, heavy physical 00:14:57.220 --> 00:14:59.879 reality. It connects directly to gigawatt data 00:14:59.879 --> 00:15:02.570 centers. It demands nuclear level energy. It 00:15:02.570 --> 00:15:04.769 triggers complex community disputes over local 00:15:04.769 --> 00:15:07.549 resources and land use. The big idea here is 00:15:07.549 --> 00:15:10.649 a total paradigm shift. AI is no longer a software 00:15:10.649 --> 00:15:13.029 boom. It is a physical transformation of our 00:15:13.029 --> 00:15:16.539 planet. It is arguably the most aggressive physical 00:15:16.539 --> 00:15:19.200 infrastructure build out of the 21st century. 00:15:19.419 --> 00:15:22.000 We are pouring millions of tons of concrete. 00:15:22.139 --> 00:15:25.259 We are laying massive high voltage power lines 00:15:25.259 --> 00:15:28.379 across continents. We are literally restructuring 00:15:28.379 --> 00:15:30.960 the physical world to support a digital mind. 00:15:31.159 --> 00:15:33.620 And we're doing it at breakneck speed. It's a 00:15:33.620 --> 00:15:36.299 profound moment in our history. Beat, we want 00:15:36.299 --> 00:15:38.960 to leave you with one final slightly unsettling 00:15:38.960 --> 00:15:41.529 thought. Yeah, this is a heavy one. Consider 00:15:41.529 --> 00:15:44.669 this vast, highly interdependent, five -layer 00:15:44.669 --> 00:15:48.309 physical stack. If AI truly becomes our core 00:15:48.309 --> 00:15:51.029 societal infrastructure, exactly like electricity 00:15:51.029 --> 00:15:54.230 or the internet, what happens to humanity the 00:15:54.230 --> 00:15:57.350 day we experience our first global AI blackout? 00:15:57.840 --> 00:16:00.539 That is a genuinely terrifying scenario to picture. 00:16:00.740 --> 00:16:03.220 It's something to seriously mull over as you 00:16:03.220 --> 00:16:05.379 watch these data centers go up. Thank you for 00:16:05.379 --> 00:16:07.580 joining us on this deep dive. We always appreciate 00:16:07.580 --> 00:16:10.059 your time and your curiosity. Thanks for exploring 00:16:10.059 --> 00:16:11.659 it with us. Take care of yourselves out there.