WEBVTT 00:00:00.000 --> 00:00:01.820 I want you to visualize two very different things. 00:00:01.899 --> 00:00:06.620 On one side, you have this plan for a gigawatt 00:00:06.620 --> 00:00:09.240 scale supercomputer in India. They're calling 00:00:09.240 --> 00:00:12.779 it Stargate. It's massive. Or city level power 00:00:12.779 --> 00:00:14.980 consumption. Exactly. And then on the other side, 00:00:15.099 --> 00:00:17.399 you just have a tweet, a warning from Andrew 00:00:17.399 --> 00:00:21.399 Yang about the, quote, disemboweling of white 00:00:21.399 --> 00:00:24.239 collar jobs. It feels like two completely separate 00:00:24.239 --> 00:00:27.260 worlds. It does. But they're not. They're actually 00:00:27.260 --> 00:00:29.940 the same story. That huge piece of infrastructure 00:00:29.940 --> 00:00:32.320 is being built to automate the exact kind of 00:00:32.320 --> 00:00:35.020 cognitive work that keeps our offices running. 00:00:35.899 --> 00:00:39.520 Welcome to the Deep Dive. It is Wednesday, February 00:00:39.520 --> 00:00:42.719 18th, 2026. If you're listening, you're probably 00:00:42.719 --> 00:00:44.700 what we call the learner, someone who wants to 00:00:44.700 --> 00:00:46.500 see the mechanics underneath the machine, not 00:00:46.500 --> 00:00:48.439 just the headlines. And the headlines are just... 00:00:48.700 --> 00:00:50.840 They're deafening right now. I was looking at 00:00:50.840 --> 00:00:52.479 our sources for today, and it really feels like 00:00:52.479 --> 00:00:54.719 the tectonic plates are shifting. We're talking 00:00:54.719 --> 00:00:57.100 about physical AI infrastructure in India, a 00:00:57.100 --> 00:00:59.539 total change in how software thinks, and then, 00:00:59.539 --> 00:01:02.420 well, the economic fallout when all that hits 00:01:02.420 --> 00:01:04.140 the job market. Yeah, it's a chain reaction. 00:01:04.319 --> 00:01:06.459 So here's our plan. We're going to start in Delhi 00:01:06.459 --> 00:01:09.459 with this whole Stargate thing and what sovereign 00:01:09.459 --> 00:01:11.439 AI even means. Yeah, then we'll hit the agentic 00:01:11.439 --> 00:01:13.980 shift when tools start doing things on their 00:01:13.980 --> 00:01:15.939 own. And the cost of that intelligence, which 00:01:15.939 --> 00:01:18.140 is... Honestly, it's pretty staggering. Yeah. 00:01:18.200 --> 00:01:19.980 Then we'll land on that warning from Adrienne 00:01:19.980 --> 00:01:23.200 about the end of the middle manager. It's a lot, 00:01:23.200 --> 00:01:25.260 but you can draw a straight line through all 00:01:25.260 --> 00:01:26.519 of it. Okay, so let's start with the hardware. 00:01:27.000 --> 00:01:29.799 OpenAI for India, we hear these big user numbers 00:01:29.799 --> 00:01:32.739 all the time, but what's the actual scale we're 00:01:32.739 --> 00:01:34.500 talking about here? The scale is kind of hard 00:01:34.500 --> 00:01:37.239 to wrap your head around. It's 100 million weekly 00:01:37.239 --> 00:01:41.659 chat GPT users. Just in India? 100 million. Yeah. 00:01:41.719 --> 00:01:43.939 I mean, that's like a third of the entire U .S. 00:01:43.959 --> 00:01:46.640 population. But that's not even the big news. 00:01:46.799 --> 00:01:49.719 The real story is the partnership between OpenAI 00:01:49.719 --> 00:01:52.359 and the Tata Group. Which is basically the industrial 00:01:52.359 --> 00:01:55.379 backbone of India. Right. Specifically, Tata 00:01:55.379 --> 00:01:58.599 Consultancy Services, TCS, they've announced 00:01:58.599 --> 00:02:01.079 the Stargate Initiative. And this isn't just 00:02:01.079 --> 00:02:03.780 a cool name. They are literally building the 00:02:03.780 --> 00:02:06.700 infrastructure. OpenAI is the first customer 00:02:06.700 --> 00:02:09.439 for their new HyperVault data center. And they 00:02:09.439 --> 00:02:12.460 are starting at 100 megawatts of capacity. For 00:02:12.460 --> 00:02:15.939 anyone listening who's not an engineer, 100 megawatts. 00:02:15.939 --> 00:02:18.819 Is that a lot? It's huge for a starting point. 00:02:18.919 --> 00:02:21.000 Your typical big data center might be, I don't 00:02:21.000 --> 00:02:25.259 know, 20 or 30 megs. But the plan here is to 00:02:25.259 --> 00:02:29.400 scale to one gigawatt. A gigawatt. That's a nuclear 00:02:29.400 --> 00:02:31.699 reactor. It is. It's the kind of power that lights 00:02:31.699 --> 00:02:34.539 up a major city. And all of that energy is going 00:02:34.539 --> 00:02:38.530 to be dedicated to one thing. Thinking. Processing 00:02:38.530 --> 00:02:40.930 Tokens Which brings us to the key concept here, 00:02:41.110 --> 00:02:44.030 sovereign AI. I keep seeing that term, sovereign 00:02:44.030 --> 00:02:46.069 AI. It sounds very official, very diplomatic. 00:02:46.430 --> 00:02:48.550 What does it actually mean, though? Think of 00:02:48.550 --> 00:02:50.449 it like a national border for digital thought. 00:02:50.590 --> 00:02:53.370 For years, if you used a big AI model, your data 00:02:53.370 --> 00:02:56.069 was probably being sent to a server in Virginia 00:02:56.069 --> 00:02:58.509 or Dublin. Right. But if you're the Indian government 00:02:58.509 --> 00:03:01.270 or a huge bank there, you just can't have your 00:03:01.270 --> 00:03:03.150 sensitive data leaving the country. That's a 00:03:03.150 --> 00:03:05.530 non -starter. So sovereign AI means the whole 00:03:05.530 --> 00:03:08.349 stack, the computers, the storage, the model. 00:03:08.560 --> 00:03:11.199 It all lives inside India's borders. Exactly. 00:03:11.300 --> 00:03:13.800 And that suddenly unlocks all these use cases 00:03:13.800 --> 00:03:16.180 that were impossible before. Mission critical 00:03:16.180 --> 00:03:18.080 government work, for example. You couldn't do 00:03:18.080 --> 00:03:20.340 that before because of security. Right. Now you 00:03:20.340 --> 00:03:23.960 can. So you see Tata rolling out chat GPT enterprise 00:03:23.960 --> 00:03:27.060 to hundreds of thousands of their own people. 00:03:27.460 --> 00:03:29.780 A's New Delhi is using it for medical education. 00:03:29.979 --> 00:03:33.159 The data stays local. The intelligence is world 00:03:33.159 --> 00:03:35.560 class. So it's about nations controlling their 00:03:35.560 --> 00:03:38.889 own what? cognitive infrastructure? That's the 00:03:38.889 --> 00:03:40.849 perfect phrase for it. Yeah. Just like you need 00:03:40.849 --> 00:03:43.110 rows and bridges, nations now feel they need 00:03:43.110 --> 00:03:45.729 their own sovereign AI capacity. Otherwise, you're 00:03:45.729 --> 00:03:47.509 just renting intelligence from somebody else. 00:03:47.669 --> 00:03:49.669 Okay. So you have this massive engine being built 00:03:49.669 --> 00:03:52.550 in the East. Let's pivot to the software that 00:03:52.550 --> 00:03:54.909 runs on these things. The big news right now 00:03:54.909 --> 00:03:57.509 is what's being called the agentic shift. We 00:03:57.509 --> 00:04:03.030 saw XAI drop Grok 4 .20 into public beta. And, 00:04:03.050 --> 00:04:05.949 you know, Aside from the classic Elon Musk version 00:04:05.949 --> 00:04:07.889 number, what's actually different under the hood? 00:04:08.069 --> 00:04:10.110 The whole workflow is different. I mean, until 00:04:10.110 --> 00:04:13.129 now, we've treated AI like a very smart encyclopedia. 00:04:13.409 --> 00:04:15.610 You ask a question, it gives you an answer. It's 00:04:15.610 --> 00:04:19.610 one to one. Grok 4 .2 changes that. When you 00:04:19.610 --> 00:04:22.290 give it a complex task, it doesn't just answer. 00:04:22.410 --> 00:04:25.370 It actually spins up four separate AI agents 00:04:25.370 --> 00:04:28.209 that work in parallel. So like one agent is doing 00:04:28.209 --> 00:04:30.189 research while another one is reasoning through 00:04:30.189 --> 00:04:31.769 the logic. Exactly. It's like a little committee. 00:04:32.329 --> 00:04:35.209 One researches, one reasons, one might draft 00:04:35.209 --> 00:04:37.569 some code and another one reviews it. It's a 00:04:37.569 --> 00:04:40.170 shift from just chatting to actually collaborating. 00:04:40.529 --> 00:04:42.649 That makes sense. But my first thought is cost. 00:04:42.949 --> 00:04:45.769 If I ask one question and it kicks off four different 00:04:45.769 --> 00:04:48.470 processes, haven't I just like quadrupled my 00:04:48.470 --> 00:04:51.050 compute cost? You absolutely have. And that is 00:04:51.050 --> 00:04:53.430 the elephant in the room. You know, we saw Apple 00:04:53.430 --> 00:04:56.230 integrating ChatGPT and Claude into CarPlay, 00:04:56.350 --> 00:04:59.209 which is great. Tap to ask is way better than 00:04:59.209 --> 00:05:01.939 yelling at your phone. Sure. But the ubiquity 00:05:01.939 --> 00:05:04.600 of it all hides the incredible expense. I was 00:05:04.600 --> 00:05:06.519 just reading the projections for Anthropic. The 00:05:06.519 --> 00:05:08.620 numbers, they're just eye -watering. They're 00:05:08.620 --> 00:05:11.660 heavy. Anthropic is projected to pay over $80 00:05:11.660 --> 00:05:14.860 billion. That's billion with a B. $80 billion. 00:05:15.000 --> 00:05:18.240 To Amazon, Google, and Microsoft by 2029. And 00:05:18.240 --> 00:05:19.839 that is just the cloud computing bill. That's 00:05:19.839 --> 00:05:22.639 the rent they pay just to exist. $80 billion 00:05:22.639 --> 00:05:25.060 just to keep the lights on. And look, their sales 00:05:25.060 --> 00:05:28.060 are exploding. Projected to hit like $6 .4 billion. 00:05:28.180 --> 00:05:31.259 That's amazing growth. But just... Do the math. 00:05:31.560 --> 00:05:34.259 The cost of generating this intelligence is astronomical. 00:05:34.639 --> 00:05:37.420 That explains moves like Mistral acquiring Coyab. 00:05:37.540 --> 00:05:40.500 Right. The French AI lab. They're trying to build 00:05:40.500 --> 00:05:42.959 their own infrastructure muscle because renting 00:05:42.959 --> 00:05:45.439 compute forever is just not sustainable. This 00:05:45.439 --> 00:05:47.540 really isn't a software bubble then. It's an 00:05:47.540 --> 00:05:50.519 infrastructure arms race. Totally. Okay. So we 00:05:50.519 --> 00:05:53.500 have these huge expensive models. Yeah. How is 00:05:53.500 --> 00:05:55.699 that actually trickling down into the tools that 00:05:55.699 --> 00:05:57.629 you and I use every day? It's trickling down 00:05:57.629 --> 00:05:59.850 fast, and it's changing what the tools do. Look 00:05:59.850 --> 00:06:02.509 at something like WordPress. For 20 years, it 00:06:02.509 --> 00:06:05.310 was a blank page. You had to do the work. Now 00:06:05.310 --> 00:06:08.209 they've launched a built -in AI assistant. It 00:06:08.209 --> 00:06:10.730 can rewrite your text, change the layout, all 00:06:10.730 --> 00:06:12.889 without you leaving the dashboard. So the tool 00:06:12.889 --> 00:06:15.910 is no longer passive. No. Or look at Flixier 00:06:15.910 --> 00:06:18.089 for video editing. It automates trimming and 00:06:18.089 --> 00:06:21.980 cutting. Or Figma. which is using Claude code 00:06:21.980 --> 00:06:24.980 to turn a UI design directly into editable code 00:06:24.980 --> 00:06:27.879 layers, the tool is becoming an active partner. 00:06:28.180 --> 00:06:30.399 I have to. I have to make a bit of a vulnerable 00:06:30.399 --> 00:06:32.899 admission here. I honestly still wrestle with 00:06:32.899 --> 00:06:35.540 prompt drift myself. Oh, yeah. As these models 00:06:35.540 --> 00:06:38.160 get smarter, I find myself kind of struggling. 00:06:38.279 --> 00:06:41.019 Like, I get used to prompting Claude 4 .5, I 00:06:41.019 --> 00:06:43.480 learn his little quirks, and then 4 .6 drops, 00:06:43.639 --> 00:06:46.060 which it just did, and all my old prompts, my 00:06:46.060 --> 00:06:48.819 spells, they don't work the same. The model just 00:06:48.819 --> 00:06:51.459 thinks differently now. It's so different from 00:06:51.459 --> 00:06:54.060 any other software, right? When Microsoft updates 00:06:54.060 --> 00:06:58.319 Excel, SUM still means SUM. But when an AI model 00:06:58.319 --> 00:07:00.779 updates, its whole reasoning path can shift. 00:07:00.959 --> 00:07:02.560 It's like your colleague got a brain transplant 00:07:02.560 --> 00:07:04.319 over the weekend. Yeah, you have to relearn how 00:07:04.319 --> 00:07:06.519 to talk to them. And that leads us right to the 00:07:06.519 --> 00:07:08.939 heaviest part of this. We've got hardware scaling 00:07:08.939 --> 00:07:11.800 to gigawatts. We have software becoming agentic. 00:07:11.819 --> 00:07:14.589 And that brings us to Andrew Yang. Yeah. That 00:07:14.589 --> 00:07:16.910 piece he wrote on X about the end of the office, 00:07:17.089 --> 00:07:19.370 it went absolutely viral. And the language he 00:07:19.370 --> 00:07:23.089 uses, it's visceral. He calls it the great disemboweling 00:07:23.089 --> 00:07:25.290 of white collar jobs. Disemboweling. He's not 00:07:25.290 --> 00:07:27.529 pulling any punches. No. He's trying to shake 00:07:27.529 --> 00:07:29.629 people up. His whole argument is that we are 00:07:29.629 --> 00:07:31.790 witnessing the extinction of a very specific 00:07:31.790 --> 00:07:34.529 kind of worker. The middle manager. The coordinator. 00:07:34.910 --> 00:07:37.639 Exactly. But why the middle? For a decade, we 00:07:37.639 --> 00:07:39.399 heard that robots were coming for the blue collar 00:07:39.399 --> 00:07:42.079 jobs, for truck drivers and factory workers. 00:07:42.579 --> 00:07:45.779 Why is the target now the person in middle management? 00:07:46.079 --> 00:07:48.120 It all comes down to what he calls cost logic. 00:07:48.660 --> 00:07:50.879 Replacing a truck driver is actually really, 00:07:50.959 --> 00:07:54.139 really hard. You need robotics, sensors, insurance. 00:07:55.399 --> 00:07:57.639 It's a mess. Right. Physical world problems. 00:07:57.779 --> 00:08:01.639 But a middle manager, what is their primary output? 00:08:01.839 --> 00:08:04.480 It's not a physical thing. In theory, it's strategy 00:08:04.480 --> 00:08:07.649 and guidance and practice. It's a lot of emails. 00:08:07.850 --> 00:08:10.310 It's summarizing reports from one team to send 00:08:10.310 --> 00:08:12.569 up to the next level. It's high cost coordination. 00:08:12.810 --> 00:08:15.290 It's routing information. And Yang's point is 00:08:15.290 --> 00:08:17.009 that the AI agents we were just talking about, 00:08:17.110 --> 00:08:20.029 they are infinitely cheaper and faster at routing 00:08:20.029 --> 00:08:22.589 information than a human who is earning 150 grand 00:08:22.589 --> 00:08:25.810 a year. The math is just, it's brutal. It's a 00:08:25.810 --> 00:08:28.050 spreadsheet decision. He says if five managers 00:08:28.050 --> 00:08:30.649 can be replaced by one operator with a fleet 00:08:30.649 --> 00:08:33.330 of AI agents, the market will force that to happen. 00:08:33.720 --> 00:08:35.679 You know, this gets to what I call the CEO's 00:08:35.679 --> 00:08:38.600 dilemma. Think about this a lot. Let's say you're 00:08:38.600 --> 00:08:41.220 a CEO and you want to do the right thing. You 00:08:41.220 --> 00:08:43.759 don't want to fire people, but your main competitor 00:08:43.759 --> 00:08:47.059 just automated their entire middle layer and 00:08:47.059 --> 00:08:50.440 their profit margins just jumped 20%. Suddenly 00:08:50.440 --> 00:08:52.600 your investors are on the phone screaming at 00:08:52.600 --> 00:08:54.700 you. They don't care about your morals. They 00:08:54.700 --> 00:08:57.019 care about the returns. You're trapped. You have 00:08:57.019 --> 00:08:59.419 to make the same cuts just to survive. That's 00:08:59.419 --> 00:09:01.919 the trap. It becomes a race to the bottom on 00:09:01.919 --> 00:09:05.620 labor costs. And the ripple effects, they're 00:09:05.620 --> 00:09:07.980 frightening. If you hollow out that whole middle 00:09:07.980 --> 00:09:10.580 layer, what happens to entry -level jobs? There's 00:09:10.580 --> 00:09:13.659 no next step. You hire a junior analyst to eventually 00:09:13.659 --> 00:09:16.700 become a manager. But if that manager role doesn't 00:09:16.700 --> 00:09:19.679 exist, why hire the junior in the first place? 00:09:19.919 --> 00:09:22.019 And what does that do to the value of a degree? 00:09:22.509 --> 00:09:25.090 Yang makes the point that if knowledge is basically 00:09:25.090 --> 00:09:27.950 free and reasoning is cheap, does a generic business 00:09:27.950 --> 00:09:30.570 degree even hold its value? We could see some 00:09:30.570 --> 00:09:33.330 colleges, the weaker ones, just shut down. It's 00:09:33.330 --> 00:09:36.090 possible. We're moving from an economy of credentialism 00:09:36.090 --> 00:09:39.769 to an economy of capability. It's not about what 00:09:39.769 --> 00:09:42.250 you know. It's about whether you can direct the 00:09:42.250 --> 00:09:45.309 machine. It's a heavy thought. It really shakes 00:09:45.309 --> 00:09:47.350 the foundation of what we've all been told is 00:09:47.350 --> 00:09:50.110 a safe career. It does. It means the safe middle 00:09:50.110 --> 00:09:52.669 is now the most dangerous place to be. And that's 00:09:52.669 --> 00:09:55.009 the thing that keeps me up at night. OK, let's 00:09:55.009 --> 00:09:57.230 try to tie all this together. We started with 00:09:57.230 --> 00:10:00.450 gigawatts of power in India and we ended with 00:10:00.450 --> 00:10:03.029 the deaths of the middle manager. How do those 00:10:03.029 --> 00:10:05.870 two things connect? They are the hardware and 00:10:05.870 --> 00:10:08.789 the software of the exact same phenomenon. That 00:10:08.789 --> 00:10:11.330 Stargate initiative, that massive amount of compute. 00:10:11.929 --> 00:10:14.570 That is the engine. Right. It's being built specifically 00:10:14.570 --> 00:10:17.850 to run the kind of heavy reasoning models that 00:10:17.850 --> 00:10:20.129 can perform that high -level coordination work. 00:10:20.429 --> 00:10:23.950 So TATA and OpenAI are literally laying the physical 00:10:23.950 --> 00:10:26.490 groundwork. For the exact economic shift Andrew 00:10:26.490 --> 00:10:28.789 Yang is warning us about, it's a move from AI 00:10:28.789 --> 00:10:31.009 as an assistant, something that helps you work 00:10:31.009 --> 00:10:33.990 to AI as an operator, something that does the 00:10:33.990 --> 00:10:36.509 work. The infrastructure is being built to support 00:10:36.509 --> 00:10:39.009 the replacement of cognitive labor. That's the 00:10:39.009 --> 00:10:42.029 big idea. That's it. And for you, the learner 00:10:42.029 --> 00:10:44.649 listening to this, it's not about panic. It's 00:10:44.649 --> 00:10:47.090 about recognizing that the middle is not a safe 00:10:47.090 --> 00:10:49.830 place to be anymore. You either need to be building 00:10:49.830 --> 00:10:52.529 the machine or directing the machine. Which leads 00:10:52.529 --> 00:10:54.110 to the final thought I want to leave you with. 00:10:54.409 --> 00:10:57.289 If that middle part of the corporate ladder just 00:10:57.289 --> 00:11:01.429 disappears, what does a career path even look 00:11:01.429 --> 00:11:05.110 like in 2030? Do you just go from intern straight 00:11:05.110 --> 00:11:08.750 to executive? or do we need a completely new 00:11:08.750 --> 00:11:11.250 structure for how people contribute value? That 00:11:11.250 --> 00:11:12.929 might be the question of the decade. Maybe it's 00:11:12.929 --> 00:11:14.490 not a ladder anymore. Maybe it's more like a 00:11:14.490 --> 00:11:17.230 network of specialists. But the days of just 00:11:17.230 --> 00:11:20.370 climbing the rungs, those are probably over. 00:11:20.690 --> 00:11:22.649 So I want you to think about your own job this 00:11:22.649 --> 00:11:25.730 week. Look at your calendar. How much of what 00:11:25.730 --> 00:11:27.789 you do is coordination, just moving information 00:11:27.789 --> 00:11:30.909 around, and how much is actual creation or direction? 00:11:31.649 --> 00:11:33.649 Because the coordination part is being solved. 00:11:33.750 --> 00:11:35.789 That's the work. Thank you for diving in with 00:11:35.789 --> 00:11:38.629 us. It's a complicated new world, but at least 00:11:38.629 --> 00:11:40.350 we can try to parse it together. See you in the 00:11:40.350 --> 00:11:41.330 deep end. See you next time.