WEBVTT 00:00:00.000 --> 00:00:02.819 That $200 monthly AI subscription you might be 00:00:02.819 --> 00:00:05.480 using, well, it could actually be a massive problem. 00:00:05.559 --> 00:00:07.900 Oh, absolutely. It's a huge issue. Yeah, because 00:00:07.900 --> 00:00:11.460 it could be costing OpenAI up to $14 ,000 a month. 00:00:11.640 --> 00:00:15.060 Right, for a single user. It's wild. That is 00:00:15.060 --> 00:00:18.100 the brutal math behind this current AI boom. 00:00:18.869 --> 00:00:21.690 Welcome to the Deep Dive. We have a massive stack 00:00:21.690 --> 00:00:24.789 of research today. And it really paints a fascinating 00:00:24.789 --> 00:00:27.609 picture. Okay, let's unpack this. Our mission 00:00:27.609 --> 00:00:30.489 today is to explore the hidden economics of AI 00:00:30.489 --> 00:00:33.670 subscriptions. We're looking at how massive compute 00:00:33.670 --> 00:00:37.509 costs are driving this crazy explosion of cheap 00:00:37.509 --> 00:00:40.049 open source tools. Yeah. And we're also covering 00:00:40.049 --> 00:00:42.289 how geopolitical bands are radically reshaping 00:00:42.289 --> 00:00:44.670 the whole global AI landscape. It's all connected. 00:00:44.829 --> 00:00:47.310 So let's start with that brutal financial reality. 00:00:47.549 --> 00:00:50.409 It really underpins the entire AI ecosystem right 00:00:50.409 --> 00:00:52.450 now. It does. It's the foundation of everything 00:00:52.450 --> 00:00:54.850 we're seeing. We all love a flat rate subscription. 00:00:55.070 --> 00:00:58.770 You pay 20 bucks or maybe 200 for a pro tier 00:00:58.770 --> 00:01:01.250 and you just use it. Right. It feels like an 00:01:01.250 --> 00:01:04.019 all you can eat buffet. Exactly. Like an all 00:01:04.019 --> 00:01:06.079 -you -can -eat buffet. But if you actually max 00:01:06.079 --> 00:01:10.480 out your theoretical limits on a $200 ChatGPT 00:01:10.480 --> 00:01:14.700 Pro plan... It costs OpenAI roughly $14 ,000. 00:01:15.000 --> 00:01:18.400 $14 ,000 in API compute just to serve one user. 00:01:18.680 --> 00:01:21.540 Yeah. And API compute is just the server power 00:01:21.540 --> 00:01:25.239 needed to process AI requests. It's the raw electricity 00:01:25.239 --> 00:01:28.280 and hardware. And Anthropic's top tier is not 00:01:28.280 --> 00:01:30.819 much better. Their system caps out around $8 00:01:30.819 --> 00:01:34.359 ,000 in token costs. The break -even points are 00:01:34.359 --> 00:01:36.140 actually shocking. Yeah, I was looking at this. 00:01:36.500 --> 00:01:40.319 OpenAI starts losing money at just 11 .4 % utilization. 00:01:40.760 --> 00:01:44.640 Wow, 11%. Just 11 .4%. And Claude hits its break 00:01:44.640 --> 00:01:47.500 -even point at a 20 % utilization rate. So if 00:01:47.500 --> 00:01:49.519 you use it even a quarter of the amount you're 00:01:49.519 --> 00:01:51.959 allowed to, they lose money. They're just bleeding 00:01:51.959 --> 00:01:55.579 cash. Yeah. So you have to ask, what drives this 00:01:55.579 --> 00:01:58.260 massive cost? Because human typing questions 00:01:58.260 --> 00:02:00.799 can only work so fast. Right. A person can't 00:02:00.799 --> 00:02:03.859 type $14 ,000 worth of prompts. The real culprit 00:02:03.859 --> 00:02:06.459 here is agent workflows. Let's define that quickly. 00:02:06.739 --> 00:02:09.900 Sure. Agent workflows are AI systems that independently 00:02:09.900 --> 00:02:13.300 loop and execute multi -step tasks. I have a 00:02:13.300 --> 00:02:16.219 bit of a vulnerable admission here. I still wrestle 00:02:16.219 --> 00:02:19.629 with my own token usage. when trying to set up 00:02:19.629 --> 00:02:22.229 agent loops. Yeah. It gets out of hand so fast. 00:02:22.389 --> 00:02:24.729 Oh, yeah. You are definitely not alone there. 00:02:24.849 --> 00:02:26.569 I mean, I looked at my dashboard last week and 00:02:26.569 --> 00:02:28.469 was just staring at the bill. It's because an 00:02:28.469 --> 00:02:31.889 agent breaks a task down. If you ask it to research 00:02:31.889 --> 00:02:34.969 a company, it doesn't just write a summary. It 00:02:34.969 --> 00:02:38.409 breaks that into 20 subtasks. It searches the 00:02:38.409 --> 00:02:42.050 web. It reads a PDS. It scrapes a database. And 00:02:42.050 --> 00:02:45.289 it feeds every single step back into itself. 00:02:45.569 --> 00:02:47.349 So it's constantly talking to itself. Exactly. 00:02:47.349 --> 00:02:49.810 It loops infinitely until it solves the problem. 00:02:49.909 --> 00:02:52.469 And every single loop costs money. Going back 00:02:52.469 --> 00:02:54.050 to that buffet analogy, it's like an all -you 00:02:54.050 --> 00:02:56.490 -can -eat buffet where one guy just sits down 00:02:56.490 --> 00:02:58.979 and eats the entire kitchen. Where he be? The 00:02:58.979 --> 00:03:01.000 restaurant cannot survive that. No, they can't. 00:03:01.000 --> 00:03:03.180 And we saw this happen. One firm reportedly burned 00:03:03.180 --> 00:03:06.240 through $500 million in a single month on Claude. 00:03:06.319 --> 00:03:08.569 Wait, really? Half a billion dollars? Half a 00:03:08.569 --> 00:03:11.430 billion, yeah. How is that even possible? Don't 00:03:11.430 --> 00:03:13.669 they have kill switches for enterprise contracts? 00:03:13.969 --> 00:03:16.150 Well, you would think so, but they didn't cap 00:03:16.150 --> 00:03:18.949 internal employee access, so everyone was running 00:03:18.949 --> 00:03:22.129 complex multi -step queries. Wow. An employee 00:03:22.129 --> 00:03:25.150 asks for a massive data sort, the agent hits 00:03:25.150 --> 00:03:28.250 a glitch, and it just loops 10 ,000 times in 00:03:28.250 --> 00:03:30.729 the background. That is just financially terrifying. 00:03:31.050 --> 00:03:33.469 It is. And it proves you don't need a massive 00:03:33.469 --> 00:03:35.629 frontier model for every task. You don't need 00:03:35.629 --> 00:03:38.340 quantum -level AI. to summarize a Tuesday meeting. 00:03:38.599 --> 00:03:41.520 How can frontier models survive if the business 00:03:41.520 --> 00:03:44.659 model is this fundamentally broken? They'll likely 00:03:44.659 --> 00:03:47.259 transition from flat consumer rates to strict 00:03:47.259 --> 00:03:49.979 metered enterprise usage contracts soon. Unlimited 00:03:49.979 --> 00:03:52.800 subscriptions die, replaced by strict pay -as 00:03:52.800 --> 00:03:56.219 -you -go corporate meters. Two sec silence. So 00:03:56.219 --> 00:03:58.840 because these frontier models are unsustainably 00:03:58.840 --> 00:04:01.500 expensive, the market is fracturing. Oh, completely. 00:04:01.740 --> 00:04:04.180 It is urgently fracturing into chipper, highly 00:04:04.180 --> 00:04:06.280 efficient alternatives. Let's talk about that 00:04:06.280 --> 00:04:08.780 push for efficiency. Startups are ditching expensive 00:04:08.780 --> 00:04:11.319 APIs, right? Yeah, they're moving to dirt -cheap 00:04:11.319 --> 00:04:13.580 alternatives like DeepSeek. And they're saving 00:04:13.580 --> 00:04:17.560 up to 95 % on compute. 95%. That is the difference 00:04:17.560 --> 00:04:19.800 between life and death for a startup. Absolutely. 00:04:20.000 --> 00:04:22.300 And we're seeing specialized models emerge too. 00:04:22.439 --> 00:04:26.540 Look at the new Kimi K2 .7 code. Right. The coding 00:04:26.540 --> 00:04:30.040 model. Yeah. It beats Opus 4 .8 on tool use, 00:04:30.259 --> 00:04:34.139 and it uses 30 % fewer reasoning tokens. And 00:04:34.139 --> 00:04:36.120 reasoning tokens are just tokens spent thinking 00:04:36.120 --> 00:04:38.839 before typing an answer. Exactly. Less thinking 00:04:38.839 --> 00:04:42.079 time means lower cost. Plus, Kimi can be self 00:04:42.079 --> 00:04:44.480 -hosted for entirely free. So no cloud fees at 00:04:44.480 --> 00:04:46.860 all. None. And the big players are scrambling 00:04:46.860 --> 00:04:50.860 to adapt. OpenAI just launched a $150 million 00:04:50.860 --> 00:04:54.120 partner network. To turn AI plans into actual 00:04:54.120 --> 00:04:57.240 workflows. Right. And Anthropic is testing something 00:04:57.240 --> 00:05:00.699 called Conway. I saw this. Whoa. Imagine scaling 00:05:00.699 --> 00:05:03.300 to an always -on agent environment like Conway. 00:05:03.399 --> 00:05:06.060 The possibilities are staggering. It really is. 00:05:06.160 --> 00:05:08.459 Conway is a standalone cloud agent environment. 00:05:08.720 --> 00:05:11.139 It uses webhooks and active Chrome browsing. 00:05:11.339 --> 00:05:13.560 Just to clarify, webhooks are automated messages 00:05:13.560 --> 00:05:16.639 apps send when something happens. Exactly. And 00:05:16.639 --> 00:05:18.879 this isn't just for big enterprise either. We're 00:05:18.879 --> 00:05:21.379 seeing a massive explosion of consumer tools. 00:05:21.579 --> 00:05:26.569 Like Slashy. Yeah. Slashy handles AI email triage. 00:05:26.750 --> 00:05:28.870 It writes in your voice and makes sure you don't 00:05:28.870 --> 00:05:31.230 miss follow -ups. And Tastelab. I thought this 00:05:31.230 --> 00:05:33.509 was fascinating. Oh, Tastelab is incredible. 00:05:33.910 --> 00:05:36.829 It extracts a website's design DNA. It pulls 00:05:36.829 --> 00:05:39.230 the hex codes and typography directly into a 00:05:39.230 --> 00:05:43.209 template, right? And there's Permute, which is 00:05:43.209 --> 00:05:47.399 a universal media converter. And Athenic 2 .0. 00:05:47.480 --> 00:05:49.959 Athenic is the data analysis agent. Right. It 00:05:49.959 --> 00:05:52.519 basically replaces a junior data analyst. It 00:05:52.519 --> 00:05:55.259 ships automated dashboards and reports. Are we 00:05:55.259 --> 00:05:57.819 moving toward a future of a million micro models 00:05:57.819 --> 00:06:00.660 instead of one giant god model? Definitely. A 00:06:00.660 --> 00:06:03.399 swarm of hyper specialized local models is far 00:06:03.399 --> 00:06:05.800 cheaper than one expensive monolith. We trade 00:06:05.800 --> 00:06:08.660 one expensive supercomputer for an army of cheap 00:06:08.660 --> 00:06:11.660 digital interns. Mid -roll sponsor placeholder. 00:06:11.860 --> 00:06:14.180 So it's not just cost driving people away from 00:06:14.180 --> 00:06:17.050 these big... proprietary models, it is also access. 00:06:17.689 --> 00:06:19.670 Geopolitics is playing a massive role here. It 00:06:19.670 --> 00:06:22.269 really is. Geopolitics is forcing the world to 00:06:22.269 --> 00:06:24.569 build around US tech giants. I want to maintain 00:06:24.569 --> 00:06:26.730 complete neutrality here and just report the 00:06:26.730 --> 00:06:28.329 facts. Sure. Let's look at the actual events. 00:06:28.860 --> 00:06:32.139 US export restrictions recently forced Anthropic 00:06:32.139 --> 00:06:34.980 to shut off Fable 5 and Mythos 5. Yeah, that 00:06:34.980 --> 00:06:37.480 happened incredibly fast. Anthropic is actually 00:06:37.480 --> 00:06:39.879 rushing staff to Washington, D .C. right now. 00:06:40.019 --> 00:06:42.500 They're claiming the security risk was overstated. 00:06:42.639 --> 00:06:45.040 What's fascinating here is the global ripple 00:06:45.040 --> 00:06:48.040 effect this caused. Bans didn't slow the competition 00:06:48.040 --> 00:06:50.920 down. No, they didn't. India is now actively 00:06:50.920 --> 00:06:53.240 rethinking its dependence on foreign models. 00:06:53.339 --> 00:06:55.560 Right. They are having a serious debate. Do they 00:06:55.560 --> 00:06:58.459 lean into open source or do they build sovereign 00:06:58.459 --> 00:07:02.060 AI? Sovereign AI simply means an AI model built 00:07:02.060 --> 00:07:04.519 and controlled entirely by one nation. Exactly. 00:07:04.779 --> 00:07:07.379 And Europe is taking the sovereign AI route. 00:07:07.540 --> 00:07:10.139 They are heavily backing Mistral AI. Mistral 00:07:10.139 --> 00:07:13.360 is raising 3 billion euros at a 20 billion euro 00:07:13.360 --> 00:07:15.939 valuation. Which is huge. Though, to be fair, 00:07:16.120 --> 00:07:19.120 OpenAI and Anthropix still lead the global market. 00:07:19.279 --> 00:07:21.699 For now. But the massive breakthrough came just 00:07:21.699 --> 00:07:24.579 two days after Fable 5 was banned. Yeah, this 00:07:24.579 --> 00:07:27.560 was wild. China's Zippo AI stepped in. They released 00:07:27.560 --> 00:07:31.100 GLM 5 .2, and they released it as a fully open 00:07:31.100 --> 00:07:33.319 source model under a permissive MIT license. 00:07:33.639 --> 00:07:36.339 That is the key detail. An MIT license means 00:07:36.339 --> 00:07:39.459 there are zero regional restrictions. Anyone 00:07:39.459 --> 00:07:43.529 can use it. And GLM 5 .2 absolutely dominated 00:07:43.529 --> 00:07:45.769 the benchmarks. It hit number one on BridgeBench. 00:07:45.790 --> 00:07:49.829 Right. It scored 100 .0 on BS and a 42 .8 on 00:07:49.829 --> 00:07:53.050 reasoning. So it actually beat the banned Fable 00:07:53.050 --> 00:07:56.509 5 model. It did. And it runs at 300 tokens per 00:07:56.509 --> 00:07:58.850 second. That is incredibly fast. And it does 00:07:58.850 --> 00:08:01.639 that at one tenth the cost. of the big proprietary 00:08:01.639 --> 00:08:04.680 models plus it has a 1 million token context 00:08:04.680 --> 00:08:07.480 window a context window is an ai's short -term 00:08:07.480 --> 00:08:09.759 memory limit during one conversation right so 00:08:09.759 --> 00:08:12.100 you can dump an entire massive code base into 00:08:12.100 --> 00:08:14.860 it at once we saw real world tests immediately 00:08:14.860 --> 00:08:18.319 the platform z .ai rolled it out instantly yeah 00:08:18.319 --> 00:08:20.019 developers got their hands on it right away and 00:08:20.019 --> 00:08:24.579 it coded a 925 line svg clock from scratch it's 00:08:24.579 --> 00:08:27.079 pure math and visual logic it also built a functional 00:08:27.079 --> 00:08:29.920 3d penalty kick game And a mini spreadsheet. 00:08:30.279 --> 00:08:33.159 All running flawlessly. If export bans immediately 00:08:33.159 --> 00:08:35.620 result in highly capable open source clones, 00:08:35.940 --> 00:08:38.759 do these restrictions actually work? Honestly, 00:08:38.980 --> 00:08:41.799 they seem to act as a catalyst that just supercharges 00:08:41.799 --> 00:08:44.779 global open source competition instead. Export 00:08:44.779 --> 00:08:47.960 blocks fail to contain tech and only force competitors 00:08:47.960 --> 00:08:52.389 to innovate faster. Two sec silence. So while 00:08:52.389 --> 00:08:54.850 superpowers fight over these massive models, 00:08:55.129 --> 00:08:58.669 AI is quietly rewriting our daily realities. 00:08:58.929 --> 00:09:00.830 It's happening on a deeply personal level now. 00:09:00.990 --> 00:09:04.090 It really is. It's both taking jobs and managing 00:09:04.090 --> 00:09:06.669 our personal lives. The job data is pretty grim 00:09:06.669 --> 00:09:09.470 right now. AI layoffs are rising. Nearly 120 00:09:09.470 --> 00:09:12.429 ,000 tech workers have lost jobs this year. Wow, 00:09:12.690 --> 00:09:15.289 that's a massive shift. But there is a very surprising 00:09:15.289 --> 00:09:18.909 statistic here. Almost 75 % of unemployed Americans... 00:09:19.289 --> 00:09:21.289 Never apply for unemployment benefits. That is 00:09:21.289 --> 00:09:23.789 so high. People just quietly absorb the loss. 00:09:24.049 --> 00:09:26.210 Yeah. And this brings us to what we're calling 00:09:26.210 --> 00:09:28.730 the personal assistant paradox. Right. Because 00:09:28.730 --> 00:09:31.250 while AI is displacing tech jobs, people are 00:09:31.250 --> 00:09:33.830 leaning on AI to manage their daily friction. 00:09:33.870 --> 00:09:36.529 Exactly. For example, there are these 15 chat 00:09:36.529 --> 00:09:38.870 GPT problems going viral right now. Oh, I saw 00:09:38.870 --> 00:09:41.090 these. People use them to dissect their paychecks. 00:09:41.519 --> 00:09:44.220 They use the AI to find highly specific ways 00:09:44.220 --> 00:09:47.080 to squeeze savings out of their budget. It's 00:09:47.080 --> 00:09:50.080 crazy. The same tech disrupting the job market 00:09:50.080 --> 00:09:52.539 is what people use to survive the disruption. 00:09:52.759 --> 00:09:55.080 Right. So what does this all mean? We are seeing 00:09:55.080 --> 00:09:57.899 the tension of macro -level job losses mixed 00:09:57.899 --> 00:10:00.519 with micro -level convenience. Yeah. And there's 00:10:00.519 --> 00:10:02.440 a lighter side to this convenience, too. It's 00:10:02.440 --> 00:10:06.299 not all just budgeting and layoffs. True. ChatGPT 00:10:06.299 --> 00:10:10.149 now has a dedicated World Cup 2026 page. You 00:10:10.149 --> 00:10:12.769 can track live scores and matches. And you do 00:10:12.769 --> 00:10:15.309 it without ever leaving the chat interface. It's 00:10:15.309 --> 00:10:17.710 all just seamlessly integrated into your conversation. 00:10:18.029 --> 00:10:20.330 It becomes the primary lens for how you interact 00:10:20.330 --> 00:10:22.429 with the internet. It's the ultimate trade -off 00:10:22.429 --> 00:10:25.330 of the AI era, exchanging job security for extreme 00:10:25.330 --> 00:10:27.850 personal convenience. It really seems like we 00:10:27.850 --> 00:10:30.289 are trading long -term career stability for hyper 00:10:30.289 --> 00:10:33.070 -efficient... Automated daily task management. 00:10:33.370 --> 00:10:35.750 Trading lifelong career stability for incredibly 00:10:35.750 --> 00:10:38.809 smooth daily task automation. Exactly. Let's 00:10:38.809 --> 00:10:40.909 briefly synthesize the major through line of 00:10:40.909 --> 00:10:44.169 this deep dive. The era of relying on a few massive, 00:10:44.269 --> 00:10:47.289 expensive, centralized AI models is fracturing. 00:10:47.580 --> 00:10:50.220 It is absolutely breaking apart. Between the 00:10:50.220 --> 00:10:52.919 massive token burn rates we discussed, the $14 00:10:52.919 --> 00:10:57.019 ,000 underlying cost per user, and geopolitical 00:10:57.019 --> 00:10:59.500 export bans. Right. The global developer ecosystem 00:10:59.500 --> 00:11:01.980 is just actively routing around all of those 00:11:01.980 --> 00:11:04.500 roadblocks. We are entering an era of incredibly 00:11:04.500 --> 00:11:08.139 cheap, hyper -capable, open -source agents that 00:11:08.139 --> 00:11:10.220 basically live everywhere. It's a completely 00:11:10.220 --> 00:11:12.860 decentralized future. I want to leave you with 00:11:12.860 --> 00:11:17.019 a final lingering question to ponder. If government 00:11:17.019 --> 00:11:19.519 roadblocks only motivate the global community 00:11:19.519 --> 00:11:21.779 to build smarter, faster, and completely open 00:11:21.779 --> 00:11:24.220 source models with zero regional restrictions, 00:11:24.559 --> 00:11:28.100 are we rapidly approaching a point where AI regulation 00:11:28.100 --> 00:11:30.860 is no longer a legal question, but a technical 00:11:30.860 --> 00:11:33.740 impossibility? Two sex islands. Thank you for 00:11:33.740 --> 00:11:35.779 joining the Deep Dive today. Keep questioning 00:11:35.779 --> 00:11:38.360 the information around you. Out to your own music.