To shape a future. Reaching. Sam. Foreign. You're watching TVPN. Today is Tuesday, December 2, 2025. We are live from the TVPN Ultradome. The temple of technology, the fortress finance, the capital of capital. Ramp.com baby. Time is money. Save both these used Corpor credit cards, billpad accounting, whole lot more all in one place. That's right. Why is no one talking about Armin Popperger? He's the CEO of Rheinmetall. And they've been on an absolute tear. We're of course going to get to code red. We're going to talk about OpenAI, but we talk about OpenAI every day basically. And I thought it'd be interesting to meet the CEO behind the world's fastest growing defense company. It's on the COVID of the business section of the Wall Street Journal. When I think high growth defense companies, I usually think Anduril or Saranic or there's so many other companies that are growing very fast in defense tech. Rheinmetall's been on an absolute tear. They're now basically the same size as Lockheed Martin and General Dynamics and it was a small company just a few years ago. So Rheinmetall, they make, you can see the gun that they make in that picture. They make big massive cannons, they make artillery shells. They've been very important to the Ukraine war. So in the last three years they've been on an absolute tear. They've gone from roughly 5 billion in market cap three years ago to $80 billion in market cap. We gotta ring the gong. We gotta warm up the gong. 80 billion market cap. They've been on a tear, but they had, and there's been like three, there's been basically three key drivers to the growth, to the story. We'll tell the story in three acts as briefly as we can. And while we, while we do, we will say thank you to Gemini Pro 3 Act story about Rayn Matal. 3 Act Gemini, Google's most intelligent model yet. State of the art reasoning, next level vibe coding and deep multimodal understanding. So first they had a head start. This company they actually started over a century ago. 1889. Can you believe that? Very, very old. So they spend their first 25 years basically just stacking up ammo for the German empire. This obviously comes to a head in 1914 when World War I breaks out. And at the time the company was one of the largest arms manufacturers like they were, they were pretty, pretty big. After 25 years of just stockpiling ammo, growing, growing, growing as A defense company, World war, World War I breaks out. But then after the war they got a pivot. They got a pivot because the Treaty of Versailles forces them to switch to non military products as a. Hey, cars, make some trains. They get fixated on trains and also type radars. Not the first, the first dudes to. Get fixated on trains and trains. Happens to the best of it. But they, but they have a good run. They stay in business. They keep making trains, locomotives particularly. You know, they're making big stuff. And then 20 years later, it's the mid-30s, 20, it's 1935 around there they are, they're starting to get back into weapons and ammo production. They can't stay away. Who are they really rearming? The Wehrmacht and World War II. Obviously it's massive for production. They're printing, they're making lots of weapons. But by the end of the war, their facilities have basically been destroyed by areas they need to rebuild the company from scratch. So after the second war, they get banned from making weapons again until 1950. Keeps happening. And so they have to go back to making typewriters. They keep getting relegated to typewriter like you guys, no more guns. That's enough. You have to make some typewriters. And so they get back into defense tech in the 50s, 60s. The German Armed Forces gets re established in 1956. And by 1979, Rheinmetall is making 120 millimeter guns that go on Leopard tanks that you've probably seen in that image roughly. And so there's lots of M and A, lots of diversification. Over the next few decades they expand into automotive and electronics. And that kind of brings us to the second act of the story, which is the Ukraine war. So Russia invaded Ukraine on February 24, 2022, about three years ago. Rheinmetall was around 5.5 billion market cap then. And three days later, Olaf Scholz, the Chancellor of Germany, gives what's known as the Zeitenwende speech, which is literally translates to turning point. So he says this is a turning point. Europe has been invaded. We now have a foreign army on European soil. Even though Ukraine's not part of NATO, it feels like, you know, Russia is expanding. If they keep, if they just keep going in the same direction, they're eventually going to be in our hometown. So we got to do something about it. And what does he propose? He doesn't just say, hey, this is a big deal. He says, no, we're actually going to invest $100 billion, like off balance. Sheet from some fund into defense tech. We're going to spend more money. And then of course there's a whole bunch of other initiatives that happen. There's the Trump negotiations around how much Europe should pay as a portion of GDP on defense. But basically it's this major turning point where Europe goes from spending sustainment levels, okay, we're going to spend this much every year to we are going to double or triple or exponentially grow our spending and it's all going to be net new so you can go and fight for it. And that's what Rheinmetall does. And so revenue growth, Helsing is sort. Of born out of that era. Hellsing is like the newer version of Rheinmetall. Rheinmetall is like the old roll up. It's been around for over 100 years. Helsing I think started Hellsing was 2021, was most recently in the news because they raised I think $600 million from Daniel Ek. Right? Yeah. Sparked controversy of course. A lot of people in the Spotify world, the world of music, just think that defense tech is default. Oh, I didn't realize that there was actually backlash, huh? Totally. I didn't see that. You know, if you're an artist and you, you know, believe in peace at all costs, you're gonna probably be against that. But it depends. Maybe if you're, you know, pod or you're some other, you know, musician that was played during the war on terror, you could be very pro. The Helsinki investment just depends. But yes, I understand overall. So revenue's grown 50% since 2022 and they are now guiding for sales. I think they do maybe around like 10 billion euros. I was kind of going back and forth on Euros USD, but they're guiding for SAL of $58 billion and an operating margin of more than 20% by 2030. So they have like almost a growth level, numbers of everything. It feels very similar where there's a, there's a structural change in the way their business is going to work. Same thing as Eli Lilly, same story. There's a couple of these stocks where there's now sort of a megatrend and they are in position to capture a ton of value as long as they can execute. The big question is, you know, what winds up happening. But the third leg of the stool, the third important piece in this story is the current CEO, the man no one is talking about until today, Armin Papperger. He's been called a white haired Goliath. I love that cnn. Can we Pull up a picture. Just randomly threw that in, popping through. There he is. There's some other photos. And last year he was targeted in an assassination plot by the Russians. What? So the CNN reported that Russia had made a series of plans to assassinate several defense industry executives all across Europe who were supporting Ukraine's war effort. And they also were planning to set up fires in different. There was an Ikea that got lit on fire. There were a number of different attacks, but fortunately American intelligence discovered the plot and informed Germany in time to stop the attack. And now the white haired Goliath is. Ryan in the chat. Says this feels like a paid ad. I can assure you it's not. John woke up this morning, we were at the gym and he's like, why is no one talking about Rheinmetall? And decided to write about it in the newsletter today. That's fine. No, he's on the COVID of the Wall Street Journal. And so they stop the attack. And so Russia clearly sees Armin Pappinger Pauper as a critic as critical to the European defense ecosystem. But separately, there is a debate over where the business goes over the next few years because on the one hand, like the NATO inventory requirements are growing a lot. That's going to drive a lot of net new demand for military equipment purchases. And the market's been historically undersupplied. But on the flip side, Rheinmetall may or may not be able to absorb as much of the demand as they're planning to. They have lots of integration to do between all their different acquisitions. And also with a potential, potential, potential end to the Ukraine war. Yeah, it feels like the stock would just immediately trade down on news of a peace deal. And it has even on rumors of a peace deal. Yeah, exactly. Yeah. It's down 15% over the last month. Yeah. But they're scaling up. In the Wall Street Journal it says earlier this year Armin Papperger opened a new factory that will allow his company to produce more of an essential caliber of artillery shell than the entire US defense industry combined. Surrounded that day by dignitaries including the head of the North Atlantic Treaty Organization, NATO, the Rheinmetall CEO is riding a wave of post Cold War military spending that is reshaping the global arms trade. Rheinmetall is now the world's fastest growing large defense company and a key player in Europe's quest to rearm its home country. Germany is shedding its post war reticence on military spending to lead the charge to capitalize. Pappager has pushed the once obscure Gun barrel maker into almost every part of the battlefield from satellites to war. And that's what people are kind of saying about there's a lot of acquisitions, there's a lot of new projects, there's a lot of new deals. Like, you know, the gun barrels, they've been doing that for 136 years. Satellites, they're kind of newer to it. Do they have the lineage? Do they have the experience? Can they stick the landing on those contracts? The money's certainly there, but is the expertise there? That's the big question. So his goal is to create a go to defense company with the heft and breadth to rival the American giants that have dominated the industry since World War II. And if he's writing any software, he's got to get on graphite dot dev code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. Rheinmetall's stock is up 15x since Russia's full scale invasion of Ukraine in 2022, giving it a market cap of 80 billion, roughly on par with US rivals. And when he took over the job, he started this job in 2013. So he's been CEO of Ryan Matal for 12 years. The company was 1.6 billion. Overnight success. Overnight success is right. Sort of like what happened Lisa Su, you know, she's 10x that stock. I mean, the funny thing is that Jensen's also 10x to Nvidia in that time. But the Lisa Su story is a little bit more impressive because AMD was really like down in the dumps and she has turned that company around fantastically. But back to Ryan Matal. This month, Ryan Matal set out ambitions to quintuple sales by the end of the decade to the equivalent of roughly $58 billion. Papurger, reflecting on his long tenure at the company, told investors that seeing such figures was like a wonderworld. It's a wonder world. I love when an executive is speaking a different language and it just doesn't quite translate. Is that what we say? Kind of get the gist. I get the gist. He's happy. I'm happy for him. Good job. We had a buddy of ours who actually, I'm just, I'm going to name him. I was going to keep him anonymous, but it's just too funny. Somebody's proposing 2.6 is proposing the TBPNX Standard Oil x Rheinmetall collab. So we were talking for, we were texting with Sean Frank and Connor McDonnell at the Ridge yesterday about how Their Black Friday Cyber Monday went. And they shared a bit on it and Sean ends it and says, bro, the future is beautiful and I am so happy to be alive with incredible reaction to a successful Black Friday. I'm so glad it went well for them. I mean, the. The Wall Street Journal did report that on a busy Cyber Monday, outage at Shopify halts transactions. Shopify experienced an outage on Cyber Monday that interrupted transactions for some merchants, but it sounded like it was not affected. Well, so it was. The, the. The real issue is the admin panel went down. Okay. Freaks a lot of people out because you're not able to log in and like, see what's happening. Yeah, of course. And, and also like, even if, even if everything's working as standard as expected, that's the day you're just refreshing the admin panel all day because you're just like, how much money am I making? This is really critical. Right? Yeah. So, yeah, there were some reports that a handful of merchants had actually had features on their site go down, but I didn't see any. I didn't see a ton of people saying, like, I lost my Cyber Monday. But obviously the good folks at Shopify will obviously be working extra hard to resolve any of this and provide a proper post mortem over. Overall, it does seem like Cyber Monday and Black Friday. Black Friday, Cyber Monday broadly just went very well. Like, it just, it just seems like consumer confidence was up, revenue was up, spending was up. Harley had a post, he said total global Black Friday Cyber Monday sales by Shopify merchants over the last five years. 2021 was 6.3 billion, 2022 was 7 and a half billion, 2023 was 9.3 billion, 2024, 11 and a half billion, and then 2025, 14.6 billion. So combination of execution at the company level and execution at the Shopify level. And then obviously the market plays a big role as well. Yeah, it really did seem like things are just broadly going well or at least okay. I think everyone's sort of like nervous with crypto up and down and is there an AI bubble? And how big of a bubble? What will happen? Somewhat of a code red going on. It has been a code of the world. Before we jump into that story. Profound. Get your brand mentioned in ChatGPT. Reach millions of consumers who use AI to discover new products and brands. I was going to say kind of the Anduril was covered in the Wall Street Journal. The Wall Street Journal has been. Been doing quite a lot of defense tech coverage. Anduril. They had been talking about their approach of not using government funding for testing purposes, which historically company would get a contract and then they would work to actually make it. And so the government was effectively funding R and D. Anduril has a more traditional like venture style model where they raise VC dollars, they spend much money to test and develop products. And so they gave a quote that was like we do fail a lot, but we do fail a lot. The extra context that was necessary was that that's not happening on the taxpayer's dime and it's part of their, you know, approach of doing rapid iteration and sort of like going according to plan. Of course that was taken out of context and turned into a headline that was we do fail a lot. And not so dissimilar from what has happened to OpenAI in the last 24 hours where sounds like an internal, internal meeting was leaked. We are wondering like what is it we can. Yeah, let's, let's actually, let's actually read a little bit more on the, on the Wall Street Journal anduril thing because I think it's interesting and I want to go into some of the response like how they responded to this. First, let me tell you about Cognition, the team behind the AI software engineer Devin, Crush your backlog with your personal AI engineering team. So Wall Street Journal came out with this story. We do fail a lot. It's a very funny quote. I actually think it's an awesome quote. We'll get into it. I think they should put on T shirts and hats. I think it's actually a very. It's their next campaign. I think it's a very key cultural. I think it's like a don't work at Anduril type moment. It makes a ton of sense in terms of like the culture of like fail fast. This is, this is not new in Silicon Valley and yet it's still being reframed as new. I'm surprised that there's like Alpha here still. But let's see. Palmer Lucky says the valid reasons for slowness are bureaucratic BS and cowardly executives who cater to snide analysts and public market outlets like WSJ that have nothing to say about years late programs and everything to say about a fire that covered 0.0002% of our test site. I'm not even exaggerating. That's the real number. It is exactly what anyone would expect from testing a system that violently blasts lithium powered drones out of the sky. This is what weapons development should look like. Heck, Camp pendleton has over 200 fires per year on their training range, and that is with fully mature weapon systems. Going on and on about this for paragraphs is so pathetic. Almost getting close to a fire every single weekday. Yeah. They obtained satellite imagery that reveals the damage to the grass on the weapons test range. The other examples of this story are similarly obscure. John, you're telling me that the grass was damaged at the explosives testing ground. Yeah. You're telling me something. You're telling me there was an explosion at the weapons testing facility. Yeah. Wild. The other examples in this story are similarly silly, absurd. Oh, no. An engine sucked in a piece of fod. Stop the presses. Autonomous boat behaves exactly as designed and stops moving when it receives a faulty command and are all hit by a pattern of setbacks. It's just so pathetic. The type of thing that can only be written and taken seriously by people who have no idea how hardware development actually works. And of course, a few other folks in the ecosystem chimed in mainly. There's a good post here from Blake Shoal, founder of Boom Supersonic. He says if you plan to pass every development test, you'll move slowly and expensively. It's optimal to fail many dev tests. Selective quote outtake into headline suggests a hatchet job, not an honest report on an attempt to do things differently and better. Yeah, I still just think the we do fail a lot is just. It's so ripe for a billboard campaign, a T shirt, a hat or something. Because if you like, the whole thing with Silicon Valley is that you should fail 99 times and succeed once. Because if you succeed once and fail 99 times, it's a million times better. It's infinitely better than zero failures. Zero successes. You will take a ton of failure for one success. And that's the whole ethos. That's the American ethos. Yeah, it's the American ethos. It's the technology ethos. There's a lot there. Anyway, back to. Oh, actually, we can wrap up with. You can go read the Wall Street Journal report on Armin Pappager. Pappager. If you want. We gotta figure out how to pronounce his name. He did have one fun line in here, which was, what did he say? He said something like. He said, referring to. So now he's basically the same value as Lockheed Martin in General Dynamics. And he said on the US companies, he said, they come to me 10 years ago, it was a different story. And so he's just flexing the fact that he's now big enough that he requires you can go visit him because he's made it. Always a good sign. Yeah, he's taken a little victory lap. And there's some other funny things in here, but you can go and read that. Let me tell you about Linear. Meet the system for modern software development. Linear streamlines work across the entire development cycle from roadmap to release. So let's head over to red alert territory. Gavin Baker, responding to the reporting, says, October, 1.4 trillion in spending commitments. November, rough vibes in December. Code Red. Life comes at you fast. It certainly has felt fast ever since that fateful podcast. Yes, that was a crazy turning point. Although there was plenty of conversation prior to that around what the trajectory of OpenAI would actually look like. Yeah, it's hard to actually understand the full nuance here. Like somebody in the replies, a rational analysis. Insane at insane analyst. What a crazy handle. Says debt obligations come at you fast. It's like that's not really what's happening here. The Code Red leak from the information reported, it was clearly like some sort of all hands that Sam Altman was holding a town hall with the rest of the OpenAI team. And he's kind of just saying, like, lock it. That's what he should have said. Never say Code Red. You gotta say lock in, brothers. Lock in. Don't say rough vibes. Don't say rough vibes. Code Red. Say lock in. Say, we're taking that hill. We're storming their fortress. We will grind Google Gemini team into paste and we will crush our enemies. We will see them driven. Before the hospitals learned this lesson. Yes, they used to say Code Red. Yes. That meant there was a fire in the hospital and that you would probably want to figure out a way to get out. Yes. They started. Is it Code Blue? Yeah. Now they will say Code Blue. So if you hear Code Blue in a hospital, yell Code Red. You need to be worried. You need to worry. But maybe. Maybe. Okay, Steel man. Steel man here. Maybe Sam Altman was using Code Red in the hospital sense. He didn't say Code Blue. If he had said Code Blue, we should be really worried. But he said Code Red. So he's saying it's not that bad. But don't you think they just retired? I think they just retired. Yeah. So he's saying. I'm using retired phrase. I'm not saying Code Blue. If I was saying Code Blue, he. Could have used Code Brown, which is a hazardous spill. Okay. Which Gemini 3 spilled on the timeline. It's very hazardous. We got a Code Brown. Yeah, we got a Code Brown. That's A crazy. Is that real or is that some, like, meme joke? No, this is. No, no, I'm reading the hospital emergency COD. Okay. Okay. Well, anyway, let me tell you about Restream One livestream. 30 plus destinations. If you want to multi stream, go to restream.com Cha Ching. No, I think if you're a CEO who's under incredible scrutiny, like you're Sam Altman and you have beat reporters at this point who are texting your employees every single day, hey, what's going on? What's on the ground? Give me a quote. What happened? Yeah. So to give people context, a beat reporter might reach out to. They will actually adopt the strategy of just trying to wear someone down, where they will send hundreds of messages to individual people on the team. Just over and over and over, relentless. Like email, cell phone, Instagram, DM, LinkedIn. Just like constantly, constantly, constantly flooding, hoping that at some point this person just says, like, fine, like, I'll chat with you. Well, the name beat reporter comes from them trying to beat you down. That's the whole point. Is that true? That's where it comes from. No way. Are you messing with me? Yeah, I'm messing with you. Okay, okay, okay. I have no idea, but I like the idea of it. It's like they just try and beat down new employees. It's like that. They try and beat you down. They got a beat reporter on my team. I was like, on my tail. Yeah, yeah. No. I mean, it's certainly what it's become. There is a little bit of it. No, no, I think there's be reporting. There's gumshoe reporting. Gumshoe reporting is where you report and you're actually walking around the town so much that you get gum on your shoes. That's the idea. It's like you're on the ground reporting. You're walking around the city. You're talking to people. And then I think like a beat cop. And beat reporting is like you're on a beat. Like it's a drumbeat. Like every day you report on the same thing. And so it's about consistency. It's not. It's not. What are you laughing at now, Ryan. In the chat, if you work at an AI startup and you aren't drinking Mountain Dew Code Red every day, you aren't going to make it. What if Sam was talking? What if he was just saying, we got to lock in. I bought us a bunch of Code Red? Yes. I want you all drinking it every day. It's time to really focus. Yes. And of course, that snippet got pulled out. I just want to know what person on the OpenAI team thinks it's in their best interest to be in a meeting like that and then just go share inflammatory quotes on said meeting. Just leave. Just go make 10 times as much money in a different lab. If you don't like your employer, just bounce and make more money. Why are you sitting there leaking and just dragging your company down? Don't you have stock options? Yeah. I'm so confused. It's a molecule. There's a mole. There's someone inside the organization who's working against them or something. I don't know. Seems rough. Anyway, there is some praise for OpenAI on the timeline, which we should get to from none other than Blake Robbins. Blake says OpenAI is operating on a different level. Play that sound cue, Jordy. The amount they have shipped in the past few weeks and months is incredible. Feels like we are witnessing a generational run. This was on October 6th. Okay. This was on October 6th. Sora was, I think, number one in the charts at that point. Yes. It's now 21. Yes. Pulse got some excitement early on, but I think people are a little bit not feeling like as excited. Atlas launched and then it's hard to really gauge what adoption has been like. I know some people that love it. So Eric Sufert on October 6, quote, tweeted Blake Robbins and kind of summed it up. I think he said, indeed. Impressive. But the scattershot nature raises questions about the company's discipline and ability to support these disparate initiatives. Is OpenAI a frontier research lab, social network operator, a commerce engine, a hardware company? Because it's hard to do all of that well. And then Eric goes back and finds his old. And they're trying to. They're still like very much care about competing in code gen, right? Yeah. And so if you go back, if you go back to the BG2 interview or just the BG interview, Sam's answer to the question of how are you going to support the 1.4 trillion of commitments was we're automating science. And we're making and we're making and we're making like consumer electronics. And the reason that didn't. To me, that was kind of like a concerning answer because Google has been doing those things for years. Yeah. But they've earned the right because they. Have 25 years funding it with massive cash flow. Yeah. Hundreds of billions of dollars in revenue and so much cash just to go around. And it's always been this academic lab and this sort of environment where they, they do side projects, but they've just, they, I think before they started any of that, they had firmly established themselves as like the go to search engine. And they were funding that with cash. So they were funding these initiatives with cash flow. I believe so. And even though they've been doing it for this long, it's not like Sundar is going out there and saying, guys, we're actually going to do it. We're going to do an extra hundred billion next year because we're automating science and we're, we're doing this new consumer electronic device. Yeah. No, no, it is crazy. Let's continue. First, let me tell you about Privy. Privy makes it easy to build on crypto rail, securely spin up white label wallets, sign transactions, integrate on chain infrastructure all through one simple API. I have a plan. And this comes from the chat. Of course, if Sam Altman really wants to set the record straight, everyone's saying Code Red. Oh, Code Red. It's so bad. He needs to come out with a statement. We're going to Baja blast Gemini out of the App Store. If he says our plan is to Baja blast Gemini and anthropic into the minor leagues of AI research, I think he just wins completely. What do you think? I think people are underestimating the possibility that Code Red it was actually read. Was past tense of read. Oh yes. They're talking about the code that was read by the model. Oh, yes. What was the code read in this scenario? It might have been. You were talking about the next agent. Model that was doing Code Jen. Yes, I have read the code and we're ready for the next pre training run. I listened to Mark Chen on Ashley Vance's core memory podcast. It's very good. You should go listen. Also, Ashley has a new YouTube channel for core memory podcasts. So if you want to find it, head over there. And it was interesting. Mark Chen, I really like the way he runs that organization. I liked a lot of things he had to say. He had some funny, funny takes some funny anecdotes. Basically just saying, you know, he's extremely competitive, he doesn't want to lose. He's going all out right now. And that one of the ways he's dealt with the talent wars is to just go to everyone on his team and say, hey, I'm not gonna match dollar for dollar with meta. Like if you want to make 10 times as much money, yeah, you're free to leave. You can just get go. But we are on a mission here. We're a team and we think what we're building is so big that in the long term we will be better and we will be bigger. And he also clarified, interestingly, that although there was a big raid and a lot of people from OpenAI did go to meta, he was saying like, there's been, there's been. He was basically like, there's a lot of poaching that's happened from OpenAI generally, like whenever someone starts a new lab, they always go to OpenAI. They're like, we need at least one OpenAI guy to know how they do it. Right. Makes, makes a lot of sense. He also said he didn't lose a single direct report. I don't know exactly how many direct reports he has, but he was saying that he didn't lose a single direct report. So maybe that's like, maybe he's trying to say, okay, there were people that were too loyal. Lieutenant. Yeah, yeah, his, his, his lieutenant stuck around. It was sort of interesting. But he did say that he also, he sort of echoed Sholto and said that he believes that pre training, there's still low hanging fruit there, that OpenAI will be doing new pre training runs, that they have seen that scaling is holding, that there's no platform. He also said they have models internally that outperform Gemini on benchmarks. Yes. And obviously he caveated that by saying benchmarks aren't the only thing that matter. So I do think, I mean, it's worth sharing. Let's actually play this clip from. Ashley vance here says OpenAI has seen Gemini 3 and is both moved and not. We sat down with OpenAI's research chief, Mark Chen. 90, is this right? Yeah, yeah. So to speak to Gemini 3 specifically, you know, it's a pretty good model and I think one thing we do is try to build consensus. You know, the benchmarks only tell you so much. And just looking purely at the benchmarks, we actually felt quite confident. We have models internally that perform at the level of Gemini 3 and we're pretty confident that we will release them soon and we can release successor models that are even better. But yeah, again, the benchmarks only tell you so much and I think everyone probes the models in their own way. There is this math problem I like to give the models. This is funny. I think so far none of them has quite cracked it. Even the thinking models just tell us. Yeah, I'll wait for that. Is this like a secret math problem? Oh, no, no, no. Well, if I announce it here, maybe it Gets trained on it to speak to Gemini Theory specifically. It's a pretty good model and I. Think this is looping. One thing we. Having a secret math problem that you give every model to assess. It is pretty elite. I keep reflecting on like, like. So let's read what Prinz is saying here. So, new interview with Mark Chen from OpenAI. Ashley Vance, the interviewer has apparently been spending a lot of time at OpenAI, including sitting in on meetings. He seems to be writing a book and he seems to think that OpenAI has made some huge advance in pre training. Pre training seems like this area where it seems like you've figured something out, you're excited about it, you think this is going to be a major advance. Mark doesn't spill the beans though. He says, we think there's a lot of room in pre training. A lot of people say scaling dead is dead. We don't think so at all. Big question about what that means. Is that scaling rl, is that scaling dollars in? Is it? Is it? Oh yeah. If you invest $100 trillion, you can give it one more IQ point. It's like, yeah, that would be an example of like scaling holding. But like no one's going to make that trade off. No one. No one is going to be like, yeah, I'm down totally spend the 100 trillion. Okay, so what Sam said in the internal Slack memo. Oh, it was a Slack memo. Yeah. Okay. As he was directing more employees to focus on improving features of ChatGPT, such as personalizing the chatbot for more than 800 million people. And again, we've seen them like launch more functionality around this. I think the theory is that this could be a very like make the product really, really sticky. Whether or not that's true generally is still unclear. It's certainly people have, have been very loyal to 4.0. Altman also said this is in the informations piece. Did he mention Baja Blast? He hasn't specifically said Baja Blast, but I think he's kind of alluding to it. He's warming up to talking about Baja Blast. Other key priorities covered by the code red include ImageGen, the image generating AI that allows users to create a variety of photos. You had included in your newsletter last week that you've been going over to Gemini specifically for Nano Banana. Yes. So I wonder if this is a broader trend. Does this actually matter? I think it does. I think that the image generation functionality, like fundamentally what LLMs are doing, what these chatbots are doing is they're basically instantiating full web pages. They should be able to instantiate anything that you could possibly land on, whether it's a video, an image, a blog post with images embedded, an audio format. Like it should be able to, to not just understand everything and give you the answer, but it should be able to contextualize that answer in any format. And so I do think being able to generate images at the top shelf, top tier way. The big question we were talking to Tyler was should they say hey, we're just going to use nanobanana, which is like a crazy thing but there is a world where they say hey, yeah, we're not going to focus on that. We're actually going to just vend in Nano Banana. But we are going to be the front door, the aggregator and we're just going to be. The actual Runway in the background, right? Yeah, yeah, yeah. Hand it off to a different team potentially. I don't know, it seems like that's probably a little bit too close to home. But Ben Thompson has had this claim for a While that potentially OpenAI has a strong hold on the consumer market to the point where if they swapped out the underlying model they would still accrue tons of the value because people don't really know what model is which like I think the average user doesn't do it. But first Tyler has a. Yeah, I. Mean I think that especially makes sense in the context of images and video because they're just so expensive. Yeah. Like I think a Nano Banana Pro image is like, I think it's like 10 cents. No way. It's really. Or okay, that might be per like a thousand or something. But it's still, it's still. They're really expensive. Yeah. Videos are even more expensive. Videos are like really, really expensive. Oh. So. So I think it makes more sense in that scenario because you would imagine that it's just so expensive to vend it yourself. It's like you're spending so much resources on that. Yeah. We have to look at this. I believe this is nanobanana. Let me see if I can find this, this nanobanana Pro image that. Let me see if we can pull this up. It's from John Gregorchuk. It says architects are cooked. AI is coming for you. Prepare accordingly. Have you seen this, Jordy? I did see that. You did see this one? Yeah. Did you look at the image closely? No. Okay, so is it, it's one of. The funniest images I've ever seen. So basically this image, it's like a. It has a walkway with like a 40 foot drop to the Ground. I mean, it's not quite that bad, but it's close. Yeah, I just, I just, I didn't. I don't buy the theory that architects are cooked just because you can generate like a floor plan or designs for a home. Just because the actual process is you're dealing with a city, basically. Right. And you're trying to get things permitted. Yeah. It's not like the problem is just making pretty designs. Right. It's the classic. Let me see. I'm trying to put this in the chat. It's the classic. Like, you know, is the radiologist's job just to look at images and detect cancer? No, it's way more than that. Okay, so this is the image. And I was actually crying, laughing, because the tagline is architects are cooked. AI is coming for you. Prepare accordingly. And you see this and it's like this AI generated image. And it looks like, like remarkable. Like, looks like a floor plan. It looks like a floor plan. It looks amazing. Like, it looks like. Okay, yeah, that's like, all the lines are straight. We used to be in the era of like any, any text would be typoed and there would just be crazy lines everywhere. But you zoom in and it's like one of the funniest layouts ever because you realize that it's just, it's just one massive room with, with, with like three or four. Okay, so first off, okay, so you come in through the two car garage. Then there's a powder room, the mudroom. So first off, there's this mud room, mudroom and laundry with two bathtubs in it. Scroll up to the right. Okay, scroll. Yeah, right there. So why do you have two bathtubs next to your coat closet? Right in the mud room? In the mud room. And also, like, you can't go. Normally you come out of the garage, you go straight into the mudroom. But here you have to go into the main area, which is the gallery hall. And then you go from there into there. And so scroll to the left a little bit so we can see the full. What is the coat? So then you go, there's a powder room. And then, then there's the coat bath and two toilets. Why does there have two toilets next to each other? Remember we were touring that facility and it had two bathrooms right next to each other with no line next to it? Yeah, yeah, yeah. We were in the machine in the crazy office that had the machine. One of the bathrooms just had. It was like meant to be a private bathroom and it just had two toilets. There we were like, why the two toilets? Yeah, so it's like, so you come in through your main foyer, then there's a master bathroom, then there's a coat bathroom with two more turrets. And then there's a huge walk in closet which isn't even directly attached to anything else. So you have to like go through this corridor to get to the rest. And so this master suite has three toilets. But then it gets better, it gets better. So go over to the top right hand side because this is so look at bedroom number two. It's just like off the center. Then bedroom number three is there. Then there's a Jack and Jill bath. Then scroll down, wait, three sinks, three sinks, no toilets. And then there's another bathroom. And then there's a third bathroom with a toilet. This is with five sinks. You might not like it, John, but this is architecture at its best. You have five sinks next to your two bedrooms which. And then also bedroom two doesn't have a, it doesn't have a bed, anything. It opens into the gourmet kitchen. But then if you scroll down, if you scroll down you can see that there's like this huge walk guest suite. What is the huge walk guest suite? And then you have this like massive dining room that just makes no sense. And then down at the bottom to kick it off, off, there's of course like the great room that's directly tied into the kitchen where there's just the most open floor plan you can possibly imagine. And then if you scroll down you'll see that there's like just these windows that like all of a sudden Trevor. In the chat says bathroom scaling mods. Why are all of a sudden the doors like vertical instead of. This is supposed to be a top down image. And now I'm looking at these doors and they're like present. What did the comments say? Do the comments say like, hey, hey buddy, why'd you put, you know, three sinks in that one bathroom? Well everyone, everyone gets that. It's a joke. Oh, it was meant to be. It was meant to be a joke. Yeah, yeah, yeah. Okay. Yeah, yeah. This John guy like totally thinks it's so funny and is just like joking around and so everyone's just like nightmare fuel. Like this is crazy. And John's making the same jokes. It's super convenient off the open floor plan, no kitchen, toilet. Then people just joking about all the different stuff. Stuff. And I don't know, I mean, is AI going to help with architectural design? Of course. Is Nano Banana going to randomly one shot the Perfect floor plan. No, also no. But of course there's stuff that's. That's the funniest image. It's so funny. So funny. Anyway, I was actually dying laughing at this thing. Okay, back to the Code Red Dancer. Automate compliance and security AI that powers everything from evidence collection to continuous and continuous monitoring, to security reviews and vendor risk. Didi Das is adding fuel to the fire. He says this is why OpenAI is in code red. In the two weeks since the Gemini launch, ChatGPT unique daily active users, a seven day average are down 6%. He is sharing to be clear, web traffic data. These traffic sources are so rough. I just feel like people use apps. The web traffic is probably a good proxy, it's probably a decent proxy, but even then I don't know how high intent those users are. It's like, do you think you're being tracked by Similar Web that effectively? I would hope that I don't have that much spyware on my Chrome browser that knows exactly where I am. Maybe it does, but I would think that, you know, OpenAI and Gemini and Google would be like, yeah, we're not letting you put a pixel on our site. How is that? How do. We should have someone from Similar Web on the show explain it to us. Tell us your sources. Yeah, tell us everything. How do you actually calculate all this stuff? Because, I mean, you could just poll people. You could just ask a million people, hey, what are you using? Right? I don't think that's how this works. But I wonder if any of the Chrome extensions sell your data. I'm sure a number of them. Yeah. Yeah, I have this Chrome extension installed right now. TBPN Timeline Viewer. It was vibe coded by someone sitting over there. He doesn't even know what programming language it was written in. This is not true. Did I ever tell a story on the show? I don't know. Tyler doesn't want to tell it. So Tyler gives us this. We use a Chrome plugin to track the show when we're sharing posts between us and this Chrome plugin. He vibe coded it and he sends it over and I unpack it to install it. And I'm like, why are there Node modules here? That's usually for Node JS JavaScript on the back end. And he's just like, what are you talking about? I was like, you don't know that you're using Node js. It's a good extension, sir. Because I think it was just. I trust Claude. I trust Claude to make the right decision. I don't even specify what programming language it uses, which is pretty sick. It's actually extremely bullish for Claude in Claude code. It's really good. Anyway, part of the code red, of course, is that OpenAI's Sora app has fallen out of the top 20 most downloaded apps in the United States on both the App Store and Google Play. And so things are falling. I actually opened up Sora today. I looked at it and there was some cool stuff happening. This is a little bit of a hot take. Like, it was not. There was still a lot of slop, which I would define as like, the. You know, it's a POV video of a bus driver with a bunch of cats on the bus, and it's, like, cute and funny or like, you know, it's a chipmunk water skiing, like, that type of stuff. Is that why you were late for the gym today? No, but I was sneaking a peek at Sora while I was driving. If I die and crash because I'm looking at slop, this would be extremely depressing. But at a stop sign, yeah, I stopped. But there was one cool one which was more like pixel art, actually. And it was interesting because you remember the OpenAI Super bowl ad. Yes. If you prompt Sora to make that type of content, it actually is really cool, and you can remix it in a very interesting way. And so, like, Sam had taken. Somebody else, had done, like, a bunch of geometric shapes pulsing to, like, electronic music. And then Sam was able to take it and say, make it orchestral music and make them pastel colors. And he was able to, like, remix off of that. And that felt like, okay, maybe we're getting into SUNO territory. Very odd that SUNO and Sora are so close in name. I don't know how that happened. Maybe they should team up or something. Wouldn't be the first time OpenAI has named something. Well, who? Oh, yeah, similar. Yeah. IO. I don't know, but. I don't know. I was seeing, like. I don't think it's fully over. You know, I think it's like. It's in a. It might be in just a trough of disillusionment. You don't know. This could be. This could be a trough. The trough is in a trough. The trough is in the trough off. It's entirely possible, but clearly the vibes are rough and people are taking shots terminally online. Engineer says, just put the ads in the chat, little bro, in the chatbot. Because Sam Altman says OpenAI is making a very aggressive infrastructure bet with new. Partnerships okay, but to be fair, this clip was from long time ago. That was like at least a couple months ago. Yes, yes, yes. And also so you can do both. What is interesting is that it's maybe. It sounds like they're delaying. It sounds like they're delaying ads, which feels odd because I personally was. I'm maybe the only person that's really excited about ads and chatgpt. I think it's a good thing for the business. I think it makes a ton of sense and I was excited to see where that rolls out. I hope that they don't delay it. I think that that's where they should be running. But if they really are losing ground to, to Google and Gemini and like the Gemini app so quickly, I'm shocked because it feels like the Gemini 3 news, like the launch went well, people were excited about the model, the model card looked good, the benchmarks look good. But you still have to be pretty tuned in to understand the nuances of the model one way or another. Like it's just not like the big model smell and the vibes. Like your average AI user doesn't care if the model responds with. It's not just this, it's that like most people. Clearly that's why it wound up getting RL'd into the model. Most people are like, wow. Contrastive parallelism. This is epic. I love it. Thank you. Like contrastive parallelism, antithetical parallelism. Like I've never. This is like a big, big, big phrase, big word. Like, this is amazing. And so I'm shocked that there would be such a. I'm not shocked by like a vibe shift on X and in Teapot with regard to how people have been skeptical of the OpenAI financing and so they've been looking for a crack to show and Gemini coming out and leapfrogging a little bit, even if it's just on some obscure benchmark that the end user might not even care about. I was really interested in. I understand that like X would jump on that narrative, but I'm surprised to see if it's true. This idea that like there's actually some sort of consumer shift. And I mean it seems like with the red alert comments like maybe, maybe it is, maybe it is. Do you think they have to explain the funding gap gap at this point or can we all just agree that maybe, maybe everyone got a little too excited? Yeah, I don't know. I don't know. I feel like everyone's sort of repriced everything already with the Oracle round tripping and Just this idea that some of the equity investments, they are circular, but it's basically just like a discount on their purchases. And these things probably aren't as binding as we think. And so I feel like the OpenAI is going to blow up the economy narrative. I feel like that was really oversold and is much. It should be fading, in my opinion, but I don't know. Buco Capital bloke has been digging into the funding hole. Apparently ChatGPT is also down right now. I just tested it. It's not down for me, but the chat says it's down. X is saying it's down. Oh, really? Oh, wow. Time to Baja Blast those servers back online, brother. It's time to rock. We need a pump up speech that doesn't include any negative phrases that can be taken out of context. We need to be Baja blasting. We have to Baja Blast. We have to Baja Blast our way to the top of the App Store. Sora Team, I need you to Baja Blast. It's time to. It's time to Baja Blast to the top of the App Store. You have to Baja Blast at the top of the App Store. You have to. And we're going to have to Baja Blast some funding into this company because apparently there's a $270 billion funding hole here. This is from a podcast between Ranjan, who writes at Read Margins, and Alex Cantorowith at Big Technology. They did a podcast together. And here's the quote from Buco Capital bloke says, squaring the total, it leaves OpenAI in a 270 billion, $207 billion funding hole. The math doesn't work. Maybe OpenAI should release to the world. Here's how the math can work, because I haven't seen anyone state how this can actually work. And so even if you get there, OpenAI does fall $207 billion short of the money it needs to continue funding its commitments. Right, Right. So it has. In 2030, OpenAI free cash flow will be about 287 billion. That's like insane. That's if this is. This feels like silly to me because if you're in a situation where you have $287 billion of free cash flow, like, you can't raise more debt on that. I feel like math tends to work out when you go from a nonprofit to a $300 billion cash flow a year in 10 years. Like, it just. Everything just forms in front of you. Like the. Like, yes, you are building the bridge as you're driving, but like, that tends to happen when you're on that much of a tear. The bigger question is like can they actually free cash with $287 billion in 2030? Because Amazon's free cash flow for 2024 was 38 billion. And let's see what Google's. Yeah, this is like 72. So saying that. That they're going to do three times more. Yeah. Than Google and Amazon. So this is. The HSBC report is modeling 386 billion in annual enterprise AI revenue by 2030. Enterprise AI revenue, huh? These are just huge numbers. It's almost not worth analyzing. I still think the biggest thing is just understanding how significant, how tied up are these contracts. Well, let me tell you about Fall, the generative media platform for developers develop and fine tune models with serverless GPUs and on demand cluster. So what else is going on? We should read through Ben Thompson's latest piece because he's provided a lot more context on Google, Nvidia and OpenAI with a post called Google Nvidia and OpenAI. Would you look at that? And we thank Ben Thompson for always having an even keel. Highly recommend subscribing to Stratecheri. It's a fantastic publication if you're not subscribed already. And he's a former guest of the show. So let's read through his latest Monday piece. He says a common explanation as to why Star wars was such a hit and continues to resonate nearly half a century on from its release with everyone except Geordie Hayes, who hasn't seen it because he hasn't seen any movies. I've seen Star Wars, John. How many Star wars have you seen? I have to have seen all of them except some of the more recent. All of them, except some of them? Well, no, the more recent ones are the like. Hasn't there been like a new Star wars in the last. How many, how many Star wars are there? There was. Is there six? Six. There's six Star Wars. That's how many movies they made. Six. Like real Star Wars. There's six. There's six. I'm gonna. I'm gonna. Hasn't there been like some. Everyone calls it the Septilogy. Yeah, there's six. Wait, there are six, there's nine. There's three trilogies. There's the original trilogy, prequel and then the sequel trilogy. And then there's also two spin offs. Okay, so I didn't watch, I didn't watch any of like the like new ones. But you watched the prequel trilogy, Rise of Skywalker. Jordy's talking about George Lucas directed. Yeah, okay, okay, so he's a Lucas head. I'm a. Yeah, exactly. Okay, so you've seen A New Hope, you've seen real Star wars, you've seen Return of the Jedi, and then you've seen Phantom Menace and Revenge of the Sith and Return of the Something I can't actually. I actually don't know that much about Star Wars. But anyway, you should know enough to follow along with this analogy from Ben Thompson. He says you have Luke, bored on Tatooine, called to adventure by a mysterious message borne by R2D2 that he initially refuses refusal of the call. This is the classic. This is the classic hero's journey. So he refuses the call. A mentor in Obi Wan Kenobi leads him to the threshold of leaving Tatooine and faces tests while finding new enemies and allies. He enters the cave. The Death Star escapes after the ordeal of Obi Wan's death. Spoiler alert. Ben, what are you doing, brother? What if somebody hasn't seen it and they don't know that Obi Wan dies? It's crazy. Oh. And carries the battle station plans to the rebels. While preparing for the road back to the Death Star. He trusts the Force in his final test and returns transformed. And when you zoom out to the original trilogy, it's simply an expanded version of this of the story. This time, however, the ordeal is in the entire second movie, The Empire Strikes Back. The heroes of the AI story over the last three years have been two companies, OpenAI and Nvidia. The first startup is called. The first is a startup called, with the release of ChatGPT to be the next great consumer tech company. The other was best known as a gaming chip company characterized by boom and bust cycles. Driven by their visionary and endlessly optimistic founder transformed into the most essential infrastructure provider for the AI revolution. Over the last few weeks, however, both have entered the cave. They're in the cave. This is the cave of disillusionment. And are facing their greatest ordeal. The Google empire is very much striking back and I believe didn't ingen over at a 16Z coin that the empire strikes back. Formerly a 16. He went independent. Oh, he's independent. Yeah. Oh, I had no idea. So is this what Sam meant when he tweeted the picture of the Death Star? I feel like we never really figured out what he meant by that. That was before GPT5, I think. Yes. I think he wanted the 2025 vague post of the year no award. It's so vague that Even after the release. No, no, no, no. I think this is it. I think this is it. It's Google's the Empire and he's launching the thing. It still doesn't make sense at the time because like OpenAI was like clearly in the lead of the models, like 2.5. I think 2.5 was the best general. Not on cash flow. You know, who has more soldiers, who has more researchers, who has more TPUs. Right. It would be fair to characterize, it'd be fair to characterize Google as the Empire the whole time. I guess the founding of OpenAI, then Google is definitely the desk. Interesting. Isn't that the kind of origin story of OpenAI? Yeah, yeah, yeah, it was. They were worried about that. Anyway, I enjoy the Star wars based analogies almost as much as I enjoy numeral.com, compliance, handled numeral worries about sales tax and VAT compliance so you can focus on growth. So Google strikes back. The first Google blow was Gemini 3, which scored better than OpenAI's State of the art model on a host of benchmarks, even if actual real world usage was a bit more uneven. Gemini 3's biggest advantage is its sheer size and the vast amount of compute that went into creating it. This is notable because OpenAI has had difficulty creating the next generation of models beyond the GPT4 level of size and complexity. What has carried the company is a genuine breakthrough in reasoning that produces better results in many cases, but at the cost of time and money. Time and money. Throw it in a ramp.com ad right in the middle of this article. I love it. Gemini 3's success seemed like good news for Nvidia, who I listed. Ben listed as a winner from the release quote. This is maybe the most interesting one. Nvidia, who reports earnings later today, is on one hand a loser because the best model in the world was not trained on their chips, proving once and for all that it is possible to be competitive without paying Nvidia's premiums. On the other hand, there are two reasons for Nvidia's optimism. The first is that everyone needs to respond to Gemini and they need to respond now, not at some future date when their chips are good enough. Did you know that Sundar was apparent like people were claiming that he had used the phrase code red back in 2022 at the ChatGPT was flopping. Yes, yes. And so yes, I'm remembering that now, but I missed it. Sundar came out and said he, yes, didn't use that exact term, but there was reporting that he did. And I agree that he also said that he wanted to Baja blast Sam Altman out of the atmosphere, out of San Francisco. Out of the atmosphere with a Death Star laser. Yeah, no, I want to. I want to try to find more historical examples of Baja blasting. Folks. We got a Baja Blast. You got a Baja Blast sometimes. So Google started its work on TPUs a decade ago. Everyone else is better off sticking with Nvidia, at least if they want to catch up. Secondly, and relatedly, Gemini reaffirms that the most important factor in catching up or moving ahead is more compute. This analysis, however, missed one important point. What if Google sold its TPUs as an alternative to Nvidia? We're going to talk to Tay Kim, author of the Nvidia Way, about that. He's going to tell all. He's going to tell all. He's breaking his silence. So that's exactly what the search giant is doing. First with a deal with Anthropic, then a rumored deal with Meta, and third with a second wave of NEO clouds, many of which start as crypto miners and are leveraging their access to power to move into AI. So a lot of those NEO clouds, they found a bunch of power and they don't really have the right chips yet. Or maybe they're upgrading their chips. They might be in a new cycle and so TPU could be at the top of the menu for them. Suddenly it is Nvidia that is in the crosshairs with fresh questions about their long term growth, particularly at their sky high margins. If there were in fact a legitimate competitor to their chips. This does, needless to say, raise the pressure on OpenAI's next pre training run on Nvidia's Blackwell chips. The base model still matters and OpenAI needs a better one and Nvidia needs evidence that it can be created on their chips. What is interesting to consider is which company is more at risk from Google and why. On one hand, Nvidia is making tons of money and if Blackwell is good, Vera Rubin promises to be even better. Moreover, while Meta might be a natural Google partner, the other hyperscalers are not. They're not going to be selling, you know, is we're going to have the CEO of Amazon Web Services, Matt Garman on the show in just. And AWS announced a new chip and. I don't think AWS is going to be buying TPU anytime soon, but we will be asking him that question. Exactly, exactly. And I want to get to the bottom of it. So OpenAI, meanwhile, is losing more money than ever and is spread thinner than ever even as the startup agrees to buy ever more compute with revenue that doesn't exist yet. Not mincing words. And yet, despite all that, and while still being quite bullish on Nvidia, I still like OpenAI chances more. Whoa. Oh. Ben Thompson likes OpenAI's chance indeed. If anything, my biggest concern is that I seem to like OpenAI's chances better than OpenAI itself. Whoa. Nvidia's moats. Wait, he wrote this before? He wrote this before the red alert? No. Interesting. He's really got a crystal ball over there. So Nvidia's Moats if you go back a year or two, you might make the case that Nvidia had three motes relative to TPUs senior superior performance, significant more flexibility due to GPUs being more general purpose than TPUs and CUDA and the associated devices developer ecosystem surrounding it. OpenAI meanwhile, had the best model, extensive usage of their API and the massive number of consumers using ChatGPT. The questions then is what happens if the first differentiator for each company goes away? That, in a nutshell, is the question that's been raised over the last two weeks. Does Nvidia preserve its advantages if TPUs are as good as GPUs? And is OpenAI viable in the long run if they don't have the unquestioned best moment? So Nvidia's flexibility advantage is a real thing. It's not an accident that the fungibility of GPUs across workloads was focused on as a justification for increased capital expenditures by both Microsoft and Meta. TPUs are more specialized at the hardware level and more difficult to program for at the software level. To that end, to the extent that customers care about flexibility, then Nvidia remains an obvious choice. The interesting thing about the flexibility is that isn't SSI a big TPU buyer? I feel like they I feel like SSI was maybe going big on tpu and I think of SSI is very much like we're going to experiment. We need maximum flexibility by default. I would assume that they're a heavy consumer of GPU because they want as much flexibility as possible. But maybe the nature of Ilya's research is flexibility within that is still afforded within the TPU ecosystem. They're using TPUs through Google Cloud. Yeah, but there's been no oh yeah. They'Re not buying them. But still that implies more flexibility because you can just turn it on or off. Yeah, I'm talking about the actual like the TPUS is an asic. It has literally like less features than the gpu, like a gaming gpu. And like the TPU doesn't. I think it doesn't support like FP4, right. Or something like that. There's some type of math that is harder to do on a TPU because it's making trade offs. Right. And, and so even though I don't understand it fully, I understand that TPUs are more specialized at the hardware level. And so if you were to be in the era of research, maybe you would want something that's less specialized because you'd be like, I'm going back to exploring all sorts of different types of math that aren't necessarily optimal. Again, if you're buying TPUs through the cloud, you're effectively just buying cloud services. You do have more flexibility because you can say hey, we want to use more of this or we want to use less. You're not like buying a bunch of. Servers and chips on the scale thing. That makes perfect sense. Yeah. I think also historically tpus have definitely been more restrictive just because of the software was just not as good or it was closed source or whatever. And then yesterday don't Patel was talking about how Google is slowly trying to open source more and more stuff. So you would imagine that in the future it should be much easier to use TPUs generally. Sean in the chat says flexibility in terms of hey, I have this new architecture. Do I need to write kernel code from scratch or is there a nice Cuda module I can use? Just time, right? Flexibility is engineering work from Nvidia. Yeah, yeah, no, that's a really good point. So Cuda meanwhile has been a critical source of Nvidia lock in both because of the low level access it gives developers, but also because there is a developer network effect. Dylan Patel was about talking about this. You're just more likely to be able to hire low level engineers if your stack is on Nvidia. The challenge for Nvidia, however, is the big is that the big company effect could play out with CUDA in the opposite way to the flexibility argument. While big companies like the hyperscalers have the diversity of workloads to benefit from the flexibility of GPUs, they also have the wherewithal to build an alternative software stack. That they did, that they did not do so for such a long time is a function of it simply not being worth the time and trouble when capital expenditure plans reach the hundreds of billions of dollars. However, what is worth the time and trouble changes. A useful analogy here is the rise of AMD in the data center. That rise has not occurred in on premises installations or the government, which is still dominated by Intel. Rather, large hyperscalers found it worth their time and effort to rewrite extremely low level software to be truly agnostic between AMD and Intel intel allowing the former's lead in performance to win the battle and so AMD better performance, better efficiency per dollar but didn't have the best software. But now because there's so much on the line, so many, the spending amount is so high, companies will go and work around all the bugs, develop new software that allows them to take advantage of AMD's better performance. In this case, the challenge Nvidia faces is that its market is a relatively small number of highly concentrated customers with the resources mostly as yet unutilized to break down the CUDA wall as they already did. In terms of Intel's differentiation, it's clear that Nvidia has been concerned about this for a long time. This is from Nvidia Waves and Moats which he written at the absolute top of the Nvidia hype cycle after the 2024 introduction of Blackwell. This article takes full circle. This takes this article full circ circle. This is from the previous Ben Thompson article in Sir Techery says in the before times I. E. Before the release of ChatGPT, Nvidia was building quite the free software moat around its GPUs. The challenge is that it wasn't entirely clear who was going to use all of that software today. Meanwhile the use cases for those GPUs is very clear and those use cases are happening at much higher level at a much higher level than CUDA frameworks that is on top of models. That combined with the massive incentives towards finding cheaper alternatives to Nvidia means both the pressure to and the possibility of escaping CUDA is higher than it ever has been. Even if it is still distant for low level work. Particularly when it comes to training, Nvidia has already started responding. I think that one way to understand DGX cloud is that is Nvidia's attempt to capture the same market that is still buying intel server chips. In a world where AMD chips chips are better because they have already standardized on them. NIMs are another attempt to build lock in. In the meantime though, it remains noteworthy that Nvidia appears not to be taking as much margin with Blackwell as many have expected. The question as to whether they will have to give back more in future generations will depend on not just their chips performance but also on re digging a software moat increasingly threatened by the very wave that made GTC such a spectacle. So Blackwell margins are doing just fine, I should note is he's back to the original article, the modern article. As they should in a world where everyone is starved for compute. Indeed, that may make this entire debate somewhat pointless. Implicit in the assumption that GPUs might take share from TPUs, might take share from GPUs is that for one to win, the other must lose. The real decision maker, maybe tsmc, which makes both chips and is positioned to be the real real break on the AI bubble. Interesting. So ChatGPT and moats resiliency. I can read through this one. ChatGPT, in contrast to Nvidia, sells into two much larger markets. The first is developers using their API and according to OpenAI anyways this market is much stickier and reticent to change. Which makes sense. Developers using a particular model's API are seeking to make a good product. And while everyone talks about the importance of avoiding locks, most companies are going to see more gains from building on never avoid locking, expanding, always lock from what they already always Baja blast. And for a lot of companies, that is OpenAI1 I would caveat here. We were talking to a founder yesterday who off the show who was saying he immediately as soon as Gemini 3 launched, spent like 12 hours. 12 hours like just moving over to Gemini from OpenAI. So depending on the product, I don't know that API is always going to be super sticky. Say winning business one app by one will be a lot harder for Google than simply making a spreadsheet presentation to the top of a company about upfront costs and total cost of ownership. Still, API costs will matter, and here Google almost certainly has a structural advantage. The biggest market of all, however, is consumer Google's bread and butter. What makes Google so dominant in search, impervious to both competition and regulation, is that billions of consumers choose to use Google every day, multiple times a day in fact. Yes, Google helps them helps this process along with its payments to its friends. But that's downstream from its control of demand, not the driver. What is paradoxical to many about this reality is that the seeming fragility of Google's position competition really is a click away. It is in fact its source of strength. And then there's a excerpt from we. Can skip this one and continue at the bottom, the CEO of a hyperscaler can issue a decree to work around Cuda. An app developer can decide that Google's cost structure is worth the pain of changing the model undergirding their app. Changing the habits of 800 million people who use ChatGPT every week, however, is a battle that can only be fought by individual individual by individual individual. This is ChatGPT's true difference from Nvidia in their fight against Google. And so this, I think, is the most important takeaway is just. Ben Thompson creates aggregation theory. This idea of, like, it's so important to aggregate demand in the modern Internet world. It's potentially the only thing you can do. You can't really monopolize supply. It's very hard to monopolize supply. But monopolizing demand is something that happens and the strength of habits is significant. We're watching this stuff every single day, so we can take the time to, okay, yeah, we should test out this other model. We should daily drive this app. But for a lot of people, if they have an app that's installed and they've been using it for a year, they're never changing. Even if the model is slightly better over there, they're just not even going to hear about it because they're just like, this is the thing that I use to plan my vacations. And the thing that I've heard come up multiple times is people that when Gemini 3 launched, they switched to Gemini 3 on desktop, but they stayed using ChatGPT on mobile. And I mean, to be completely transparent, the Gemini mobile app is really, really struggling to stay connected. There's something in the. When you fire off a prompt, it doesn't save it locally and then cache that and then send it off, inference it and then come back it. Like, unless you keep the app open, like, it will, it will just give you like a server disconnected error. Like, I've gotten like dozens of these and that's going to be a real. I think it should be something they should be able to fix in like a weekend. But, you know, hopefully it's soon, but for a lot of people. Logan. Yeah, well, they'll get it. But back to the moat and the map. The moat map and advertising. This is, I think, a broader point. The naive approach to moats focuses on the cost of switching. In fact, however, the more important correlation to the strength of a moat is the number of unique purchasers to users. The strength of the moat is increased by the number of buyers. Okay. So that you can see where this is going with, like, Nvidia has five buyers and ChatGPT has a billion buyers. And once it has advertised, yeah, advertisers. Might be even more. Right. So this is certainly one of the simpler charts I've ever made, because it's literally just one line, but it's not the first in the moat genre. So he talks about the moat map. I argued that you could map large tech companies across two spectrums. The degree of supplier differentiation from Facebook, where the supplier is completely commoditized, just your friend on Facebook, to Microsoft and Apple, where the suppliers are somewhat more controlled. Yeah, there's the what a chart. The more unique buyers of your product you have, the stronger your moat, because it's hard to. Because you have to convince each one of them. And then second, the extent to which a company's network effects were externalized, internalized network effects, or Facebook again, and then externalizes Microsoft. And so putting this together gave the moat map. So who has the network effect versus the suppliers map? If we scroll down there, you can see them. What you see in the upper right are platforms. The lower left are aggregators. Platforms like the App Store enable differentiated suppliers, which allows them to profitably take a cut of purchases, driven by those differentiated suppliers. Aggregators, meanwhile, have totally commoditized their supplier, but have done so in the service of maximizing attention, which they can monetize through advertising. It's the bottom left that I'm describing with the simplistic graph above. The way to commoditize suppliers and internalize network effects is by having a huge number of unique users. And by extension, the best way to monetize that user base and to achieve a massive user base in the first place is through advertising. It's so obvious. The bottom Left is where ChatGPT sits. I wonder what he thinks then about. About them potentially kind of delaying ads and delaying ads. Punch in the air. Him and Eric Suefer, probably. No. And I'm right there with them. I completely agree. Boo. Launch the ads product. Launch the ads product. Get it out. Come on. Don't delay that. That's the most important thing. So at one point, it didn't seem possible to commoditize content more than Google or Facebook did. But that's exactly what LLMs do. The answers are a statistical synthesis of all the knowledge the model makers can get their hands on and are completely unique to every individual. At the same time, every individual's user usage should, at least in theory, make the model better over time. It Follows then, that ChatGPT should obviously have an advertising model. This isn't just a function of needing to make money. Advertising would make ChatGPT a better product. It would have more users using it, more providing more feedback, capturing Purchase signals not from affiliate links, but from personalized ads. Ads would create a. Would create a richer understanding of individual users, enabling better responses. And as an added bonus, and one that is very pertinent to this article, it would dramatically deepen OpenAI's moat. Yeah, I keep going back to this idea that OpenAI needs personalized ads like Instagram. Like, that's what Sam said when he was interviewed multiple times on his ad strategy. He was like, you know, like, ads can be bad, but these Instagram ads are pretty good, you know, and people are always like, he's backtracking. And it's like, no, that's fine. Get, like, do the proper business model. Like, please implement the correct business model. I'm happy about that. But the interesting thing is that Instagram does not surface ads when you necessarily, when you search for something, like, if you're on a video for like a Ferrari, like, you don't just immediately get an ad as your next thing for, like, Ferrari of all Hollywood or Ferrari of Beverly Hills. No, you get an ad for the toaster that you were about to check out on. I actually do. Like, half the ads I get on Meta are local dealerships in la. Yes, but importantly, not when you, like. It's not tied to search. And so ChatGPT can do the same thing where they can clearly show you where they can clearly show you something that you are about to check out your shop. Shopping for Christmas for this thing you're searching for the Roman Empire. Let's show you the ad for the thing that you're chopping that you're shopping for, like right next to it. It's fine. And so I think that can work very well. Anyway, let's go to Google's advantages. It's not out of the question that Google can win the fight for consumer attention. The company has a clear lead in image and video generation, which is one of the reasons why I wrote about the YouTube tip of the Google spear. I mean, Google's advantage in data is insane. Like, you YouTube so massive. And that's gotta be just the compounding compound. I mean, we're uploading another three hours of video to YouTube today. You're welcome. This one's for you, son. This one's for you, Demis. But the flip side is, like, they also see the entire Internet because of the way the googlebot scrapes. The Google searching in Gemini is such a killer feature. It's such a killer feature if they can keep that on and they can actually surface that. And the AI search results are obviously going to get good. They're going to figure out how to surface it. I think I'm still pretty optimistic, but let's see what Ben Thompson has to say. And let's also tell you about Adeo, the AI native CRM. Adeo builds scales and grows your company to the next level. So Google is obviously capable of monetizing users even if they hadn't turned on ads in Gemini yet. It's also pointing out, as Eric Soofer did in a recent Stratecher interview, two of the best collabing we love to see, that Google started monetizing search less than two years after its public launch. It is search revenue far more than venture capital money that has undergirded all of Google's innovation over the years and it is what makes them such a behemoth today. In that light, OpenAI's refusal to launch and iterate on an ads product for ChatGPT, now three years old, is a dereliction of business duty. He's calling him out, particularly as the company signs deals for over a trillion dollars of compute. What are you doing doing? Sam, get the ads out. Put the ads in the chatbot. We love it. Just put the ads in the chat bot and go. You gotta baja blast some ads into that app. You got to. I want ads in ChatGPT, please. On the flip side, it means Google has the resources to take on ChatGPT's consumer lead with a World War I style war of attrition. Rheinmetall call back. OpenAI's lead should be unassailable, but the company's insistence on monetizing solely via subscriptions with a downgraded degraded user experience for most users and price elasticity challenges in terms of revenue maximization is very much opening the door to a company that actually cares about making money. To put it another way, the long term threat to Nvidia from TPUs margin from TPUs is margin dilution. The challenge of physical products is that you do actually have to charge people who buy them, which invites potentially unfavorable I always. So, so yeah, who, who. Both Gemini and and ChatGPT will have ads eventually, right? Yes, you can bet on that. Who will ramp ad revenue faster? It's hard not to bet on Gemini even with a smaller user base because they have the ad network, they have all the, they have all the customer relationships already. They can just say like hey, here's a pop up. Do you want to hear $10,000 of free ad credits? Try it out. Right? It's like native. It will already be in AdWords. The example that I use is like how Zuck was able to take Instagram, which had a lot of users, plug it into the meta ads platform and just scale revenue like crazy and then do it again with reels. And so, yeah, very, very clear that they both will have ads. And again, if Gemini can really ramp that quickly, they could. Again, I do feel like we're moving towards a world where every consumer will be able to get the best LLM for free. Free. Right. I don't, I don't, I don't necessarily believe that every American will be paying for an LLM in five years. And so if Google can get there first and then keep Gemini on the frontier and deliver the best free, fastest text and image model, that's going to be very, very difficult to compete with. And so again, like, getting, getting to ads fastest feels like it makes more sense. I like that point. Ivory rebuttal. First, I'm going to tell you about figma. Think bigger, build faster. Figma helps design and development teams build great products together. So getting to ads first is an advantage. That's your take. I like it. But there is a little bit of a risk with launching ads first because you could forever be branded with you're the ads. And we saw this when Arvind from Perplexity came on the show and he mentioned this idea of ads and LLM queries and we all agreed on the, on the discussion, we all agreed that ads were going to come to AI tools because that is the way to get the most people using them and make intelligence free. And you got in a debate with Mark Cuban over this, for example, and, and it seemed very logical. But there is the fact that the first major chat app to put ads in their app is going to be a massive news cycle. It's just going to be like national news. Sam Adman. Exactly. It'll be OpenAI has ads now or it'll be Gemini has ads now. And so you don't necessarily, if you like, it's much easier to be the second mover there because it's going to be less of a news cycle. And so you kind of do want to. There is a little bit of advantage to being the second mover there. Right. Because you're going to get sort of branded as like, oh, that's the ad supported one and the other one can add ads and people will be like, oh, yeah, I guess, but that's like standard. Yeah, there's going to be like a backlash and people will be like, oh, No, I don't like this company. Blah, blah, blah, blah. Yeah, I mean, the harder thing is just how do you do it? Right, right. I do feel like it's different. You're going to an LLM, people are going to an LLM for advice and recommendations. That's different than going to Google and searching and seeing ads at the top. There's plenty of surface area. No, no, I'm not saying there's no surface area, but I'm just saying like the right way to do ads and LLMs is not clear yet. Yeah, yeah, I mean they will need to do some experimentation, but I mean, just starting with Google already has retargeting information. I actually went to Gemini with a question and it clearly knew everything about me and it said you're the host of TVPN and you founded these companies and it already knew that. Probably just because I authenticated with something else. I don't know, but it should know. Okay, hey, we could retarget you with this. Let's put this in. And the same thing with OpenAI, with ChatGPT, there's plenty of spaces where it's like you're waiting for it to happen to give you the answer. Okay, hey, we're generating you the image. Why don't you show me other images of ads Right there. There's tons of surface area. I agree that there will be a whole bunch of iterations on like what the ideal ad looks like, but yeah, you can clearly get out. So let's go back to Ben Thompson. Close this out. Says the reason to be more optimistic about OpenAI is that an advertising model flips this on its head because users don't pay. There is no ceiling on how much you can make from them, which by extension means that the bigger you get, the better your margins have the potential to be and thus the total size of your investments. Again, however, the problem is that the advertising model doesn't exist yet. So he started this article recounting the hero's journey in part to make the easy leap to the Empire Strikes Back. However, there is a personal angle as well. The hero of this site has been aggregation theory and the belief that controlling demand trumps everything else. There's there. Google was my ultimate protagonist. Moreover, I do believe in the innovation and velocity that comes from a founder led company like Nvidia. And I do still worry about Google's bureaucracy and disruption potential, making the company less nimble and aggressive than OpenAI. More than anything though, I believe in the market power and defensibility of 800 million users. Which is why I think ChatGPT still has a meaningful moat. At the same time, I understand why the market is freaking out about Google. Their structural advantage, their structural advantages in everything from monetization to data to infrastructure to R and D is so substantial that you understand why. OpenAI's founding was motivated by the fear of Google winning AI. It's very easy to imagine an outcome where Google's inputs simply matter more than anything else. Which is to say one of my most important theories. Theories is being put to the ultimate test, which perhaps is why I'm so frustrated. OpenAI's avoidance of advertising. Google is now my antagonist. Google has already done this once. Search was the ultimate example of a company winning an open market with nothing more than a better product. Aggregators win new markets by being better. The open question now is whether one that has already reached scale can be dethroned by the overwhelming application of resources, especially when its inherent advantages are diminished by refusing to adopt an aggregator's optimal business model. I'm nervous and excited to see how far aggregation theory really goes. Fascinating. That's his baby. Yeah, I agree. It is the correct framing. It'll just be very interesting to see. I really wonder who's going to take the leap first. Who's going to jump and put ads in the app first? It feels like Google should do it. It feels like Google will be able to do it. Yeah. Nobody's gonna be like what Google is putting ads in a product. It won't be that surprising. So they should probably move fast. We have some breaking news. What's the breaking news? Jason Freed is joining the show at 2pm Surprise guest. Surprise guest. He's launching Fizzy today. Kanban. As it should be, not as it has. Has been. I will wait. We'll wait to talk about this till he joins in an hour and 20 minutes. So little surprise guest appearance from a. Legend calling out his competitors directly. I love when the founders do that. Founder mode. Founder Mode. Yeah, very. Should we talk about John Giandrea leaving the company? He's out. I quit. I quit. You like that one? I think that's probably been used before. Some headlines but Mar Subra. So of course Mark Gurman has the scoop. I believe the Germanator. Germanator's at it again. He says Apple AI chief John Gianandrea is leaving the company. Amar Subramania from Microsoft is joined at lead AI under Craig Federighi. And so we should dig in a little bit to this history. So SWIX has a little bit of a deep Dive here. He says Aman brings a wealth of experience to Apple. He's quoting here. Having most recently served as CVP of AI at Microsoft and previously spent 16 years at Google where he was head of engineering for Google. Gemini, wait. Oh, I guess at the end of that, because Gemini's not 16 years old, this is bearing the lead. He joined Microsoft AI four months ago. Wow, what a crazy turnover. LinkedIn says six months ago, but who's counting? That's pretty fast. But this makes sense considering Apple is partnering with Gemini and not a lot of people are going to be in a better position to help integrate that into Siri than Amar. Yeah, I mean, I don't know, maybe there's something to just having a taste of all the different big tech companies. Oh, yeah, I've been at Microsoft. I know how they work. I've been at Google. I know how that works. I'm ready to. Ready to rock over here. Gurman does need to come on back on asap. I agree, regav. He was fantastic. We'll get him in person. Hopefully before so German continues. He says strange hire for a number of reasons, but it's hard to argue that the Apple job is a bad one. Anything is an important improvement at this point. So the bar is as low as it comes. Easy to lay up on the resume. So that'll be. That'll be fun to see. I'm personally just excited to actually test drive what Gemini, how it works in Siri, how seamless that is. Because if it really is, just press the button, get Gemini and it's linked up properly and it doesn't have timeouts and it gets back to you pretty quickly. That's going to be a pretty powerful experience. That's definitely going to cut down on ChatGPT app usage for iPhone users, I would imagine. Underrated threat. I would think so. Like the re. Like there are so many moments where. People are counting out Apple. It's not. Yeah, I don't even know that. I don't even know that Apple will benefit massively from this. It's not like they're gonna sell twice as many iPhones. They're already so big. It's not like they're gonna charge for it. I don't think it's necessarily like especially bullish for Apple. It's an underrated threat for OpenAI. I would think so. There's a lot of queries that all hit open ChatGPT on mobile that are not even like super economic, but just a lot of my usage around like, hey, just trying to learn about something or research a product, et cetera. And if that's just like, again, one tap and you're in there. Yeah. I mean, yeah. The original promise of Siri was not just, hey, what's the weather today? But really asking anything. And Gemini clearly solves that. For 99% of knowledge retrieval queries, I would be. I think I'm going to be using that a lot unless they really botch it. And I don't know how they're going to botch it, but. Yeah, Yeah. I mean, I think the. Just anything's possible. You know, Apple's like, hold my beer. They could be like, every. For privacy reasons, every time you press the button, you have to E sign and it's like, why are we doing that? Yeah. Like, the original, like, Siri kind of vision was like this very conversational AI, Right. Yeah. But I don't think Gemini has a real time voice model yet. Like, I'm pretty sure OpenAI is the only one that has that. Really? Huh. I don't think it matters at all. Really. Yeah, I really think that form factor. Feels like that would be the best thing to be on. You press the button on your iPhone. And then it's just talking and then it's just on. Yeah. Yeah. I mean, I wouldn't be surprised if they can. If they can, like, get that model, that version of the model out, because it's really just. Just like distilled a little bit faster. It's not some uncanny breakthrough that the Gemini team will never be able to crack. Right. So they just have to build that. But honestly, I don't know that. That's certainly not how I would use it for most things. Most things I would say. Okay, I have one question. Get me an answer within a reasonable amount of time and maybe read it off to me or produce an article that's a pretty readable article summarizing the answer to my question. And then, yeah, maybe there is like a back and forth, but I don't know. Oh, you're. You're getting. You're getting truth zoned in the chat. Gemini does have a real time voice feature. Gemini Live. Yeah, I think it's on the app. I've tried that. I don't have an app. So. Tyler Anthropic is acquiring bun. What do you have to say? Say? Yeah, I mean, this is definitely in line with their, like, you know, focus on dev stuff. Okay. What is Bun? Bun is a. It's like a. I don't know, it's like a bundler for JavaScript, it's like a very dev. It has dramatically improved the JavaScript and TypeScript developer experience. They're going to make Claude code even better. I like the Gabriel from OpenAI. Here is OMG in the chat, which is like a pretty crazy thing to say, but I appreciate it. So Claude is one of the world's smartest and most capable AI models. For developers, startups and enterprise cloud code represents a new era of agentic coding, fundamentally changing how teams build software. In November, Claude code achieved a significant milestone just six months after becoming available to the public. Public. That's crazy. It's only been six months. Wow. It reached a $1 billion revenue run rate. We were always struggling to understand what that meant. Right, yeah. Well, so there, there are two ways to pay for Claude code. Either with your Claude subscription where you get like Claude Pro or Claude Max and there's a certain amount of tokens you can use and then there's also you can just directly wire up APIs calls essentially to cloud code and then you're being charged like directly based on you usage. So that's probably what that revenue is from. Yeah. And then, yeah, I also have thought about this thing where like, oh, you can break down the number of tokens from the subscription. So it's like your $20 subscription, 3/4 of your tokens are on Claude code. That means 3/4 of your $20 is counts as Claude code revenue. Okay, yeah, yeah, I'm not sure exactly. Yeah, I mean I'm sure that they can account for it. So. It was founded by Jared Sumner in 2021. Bun is dramatically faster than leading company. They say It's a breakthrough JavaScript runtime. Does it compete with V8? I'm very interested in like it's Node JS. It competes with Node. The thing that Tyler needs to learn. What was that company that OpenAI acquired earlier this year for like a billion? I know the one you're talking about. Isn't that analytics or something? Yeah, but again, I remember at the time people were like, oh, like OpenAI's competitors are not going to be happy about this acquisition. So. So that comment from Gabriel, congratulations to the Bunn team. Congratulations to Anthropic and everyone on the Claude code team. Very excited that you're getting to work together for your massive deal. Speaking of other massive deals, speaking of size, Alfred Lynn. Get that gong ready, John. What did he do? Alfred Lynn comes in on the board of DoorDash, buys 100 million. DoorDash Linsanity. He's not Done stewarding doordash. He's continuing to steward the company with. $100 million buy and of course sends the stock up almost 6% on that. Pretty, pretty raw. Are excited about it. He ripped. He ripped. What Is this? Another 4.6 million donated to shrimp wealth. Okay, so yeah, basically the story is in theropic. They were doing some research about smart contracts, and so they had Claude Code try to figure out issues in smart contracts. And then I'm not sure exactly where the money came from. Maybe it was for bounties, but there's some way in which Claude Code basically, basically generated like $4.6 million in cash from finding these exploits. So then they just. Did it actually generate real money or is this like the hypothetical. This is simulated testing. Okay. Huh. Yeah. So it probably means that they could have like basically stolen $4 million from people, but they don't want to do that. Maybe they should have if they really want to get up the. In other anthropic news, David Sacks says he's still waiting on Dario's support after the New York Times piece was published. Sam Altman, of course, came in and said David Sachs really understands AI and cares about the US leading in innovation. I'm grateful we have him. Of course, Dario and Sachs not the biggest fans of each other, so I don't expect that one coming through anytime soon. While we wait for our first guest, Matt Garner, let's pull up this clip from Huberman Lab, if we can play this. Rob Moore is highlighting Dr. Jeffrey tells Huberman that LED lighting in buildings is a public health crisis that could be on par with the use of asbestos. Many building contractors, designers are coming to him worried they're going to be sued and asking how to start fixing the issue. So let's pull this up when we have a second for light because I. Am very concerned about the amount of short wavelength light that people are exposed. To nowadays, especially kids. The group of us that are shuffling around, some of them are saying this is an issue on the same level as asbestos. This is a public health issue. And it's big LEDs came in and people won the Nobel Prize for this, very rightly at the time, because they save a lot of, lot of energy. The LED has got a big blue spike in it, although we tend not to see that. And that is even true of warm LEDs. And there is no red the light found in LEDs when we use them. Certainly when we use them on the retina looking at mice, we can watch the mitochondria gently go downhill. They're far less responsive. They. Their membrane potentials are coming down. The mitochondria are not breathing very well. Watch that in real time under LED lighting. Under LED lighting at the same energy levels that we would find in a domestic or commercial environment. This is why I want to rig the studio with candlescent incandescent light. Incandescent. We're going back to candles. Candle max. Do candlelight. How about a hearth? This is the way. If we put a hearth so we have lights above our heads that I'm sure are leds killing us slowly Soft. If we put somewhat a bonfire right above us, that'd be the way. And then we just. When the wood kind of burns out, the show's over. We just go until the. That would be good. I like that. Let me tell you about turbopuffer. Serverless vector and full text search built from first principles and object storage. Fast 10x cheaper and extremely scalable. Turbopuffer is at re invent. Amazing. Fantastic. Well, we are joined by the CEO of Amazon Web Services.

NEWSMAX