DeepSeek, Stargate Update, Bees Do Enjoy Honey, Golden Retriever Maxxing, Bring on the Trilly

Oh, we're live. Welcome to Technology Brothers, the most profitable podcast in the world. Today we're doing a deep dive on Deep Seek, the new AI model coming out of China. They say they trained it for just $500 less than this cash stack, but we'll dig into that because it's a hot topic. A lot of people are debating about this, but we wanted to take you back to May of last year when Deep Seek actually launched. It's in the news because they launched a new model, but the company's been out there open sourcing LLMs for. For something like six months now. And it's. It's shaking up the timeline. Timelines in turmoil. Over.

Timelines in turmoil.

Yeah.

And I'll go out on a limb to say this is China's next attempt at a dji tik tok type.

Yep.

Products where, hey, let's basically sell this product in America for way less than it costs, give it away for free.

Yep.

Try to get a bunch of important, you know, companies using it or. Or consumers dependent on it. And then, you know, it's a little bit of, you know, spyware.

Yeah. The data. Yeah. There's all the same geopolitical considerations that.

We walk through, and there's the incentive, you know, for them if, you know, if it, you know, you can basically assume that any important Chinese company, Company is, you know, backed by the state to some degree, even if they're not, they're sort of, you know, got a gun to their head the entire time they're operating it. So we're excited to get into this one. You know, as our, as our mission at Technology Brothers is to, you know, entertain, cover some news and take down CCP fronts, so.

Exactly.

Let's get into it.

So let's kick it off with a fantastic analysis by Dylan Patel in Semianalysis. This is from May 7th of 2024. And it's interesting.

Dylan is, for those that don't know, he's the Tom Brady of semiconductor analysis.

100%.

He's the kind of guy that gets invited on BG Squared to. He's been on the dogs. You know, how, how the, how the semis are barking these days.

Yeah, I mean, it's a fantastic. It's essentially substack subscription model, $500 a year. But then he also does a consulting. And what's interesting is that he doesn't just read the 10Ks or anything. Like, he will task satellites to, like, figure out where they're building stuff and, like, really collect data. And then he sells doing the same.

Thing that hedge funds.

Exactly.

Doing. It's, it's more.

Yeah. So. So I think, I think Semianalysis is actually like a pretty large company now with a number of analysts and he's not going the kind of like solo creator route. He's very much focused on building like a, a research organization. But he's also just hilarious on podcast. Free flowing, just says whatever he wants. Like totally uses very online lingo too in these very professional settings, which is great. I'm a big fan. So this is the title of this is OpenAI is doomed to Microsoft and it's kind of walking through a bull and bear case for OpenAI given all of the different dynamics and different cluster build outs that were happening mid last year. And it starts by saying all eyes are on how long the Profitless spending on AI continues. H100 rental pricing is falling every month. This big trend last year in 2023 ChatGPT was blowing up and everyone was like I will do anything if someone can get me a a rack of a one hundreds. A lot of VC fans. Exactly Andromeda. But in 2024 rental pricing went down as fewer firms realized that hey, maybe they don't want to be in the pre training game, they just want to fine tune or they just want to build. They want to be rappers essentially. And then also the really really big firms might just be building super clusters that didn't even affect the rental pricing market. And so availability is growing quickly for medium sized clusters at fair pricing. Despite this, it's clear that demand dynamics are still strong. While the big tech firms are still the largest buyers, there is an increasingly diverse roster of buyers around the world still increasing GPU purchasing sequentially. Most of the exuberance isn't due to any short any sort of revenue growth, but rather due to the rush to build ever larger models based on dreams about future business. The clear target that most have in mind is matching OpenAI and even surpassing them. Today many firms are within spitting distance of OpenAI's latest GPT4 in chatbot Elo and in some ways, and in some ways such as context length and video modalities, some firms are already ahead. So context length, that's the Google thing where drop an entire book, drop an entire book in in Gemini and drop.

The entire banger archive. Exactly.

And so there's been a bunch of ways to differentiate. Even if you're you just get close to GPT4, you do a similar done and then you build a different tool around it or you fine tune it or you build different, like, you know, unhobbling. So this one uses the web, this one can process PDFs, this one can.

You don't really hear about Mistral much anymore.

No.

Is that because the French don't have a word for entrepreneur?

I like that. Oh yeah, it's the.

It's the right rhymes. Ederson, they were trying to community note him on that, but I don't think they got it through.

That's great. It's probably a war. So Gemini 2 Ultra is rumored to surpass GPT4 Turbo in every way. Furthermore, Meta's Llama 3 405B is also going to match GPT4 while being open source, meaning GPT4 class intelligence will be available to anyone who can rent an H100 server. But he goes on to mention the Phoenix from the East. It's not just the big tech firms that have rapidly caught up yesterday.

Real quick. One kind of interesting comparison is it feels like the models have fast approached being almost like the sports car market, where it's like Porsche can be like, well, we have better handling and McLaren can be like, well, we're faster in a straight line.

Yeah, exactly.

And Tesla's like, well, I'm fastest in the straight line.

Yeah.

Yep. You know, so they're all just sort of competing, doing a lot of things pretty well.

Yep.

And then competing on these Edge cases that almost the average. And the reason that Chat GPT has stayed very dominant from a user standpoint is because they have the consumer mind share. The average person is not like prompt engineering. They don't care about the context window. They're just almost using it like a Google that can do things on your behalf. Yeah.

I mean, and this was one of the big like bear cases for these LLM companies after ChatGPT dropped. Was like, well, is everyone just going to train these and it's going to be completely commoditized. But then the companies kind of figured out, well, if Grok is a little bit less PC, that's an edge and if Gemini has a bigger context window. But there's some fascinating stuff in here about how the different big tech companies are able to integrate their models and actually bootstrap to much larger user bases, which results in getting more data from users. Which is very interesting. But he highlights DeepSeek, which launched in May of 2024. It says China's Deep SEQ open sourced a new model that is both cheaper to run than Meta Meta's llama 3 70b and better. So this was not as big of news because it wasn't truly better than GPT4. So it of slid under the radar, but it was cheap on inference, meaning you just needed less less compute to actually run it.

And this is to be clear, the same thing that China has done historically with dji, where DJI is selling products at without any profit margin at all, potentially even losing money on a unit basis because they just want to get the product out into the world getting used, getting consumed.

Yep.

And so, and not too dissimilar in my view to TikTok faking views. Right. Because they're like, well, if users come on here, they post something, they get a thousand views, they're not going to want to go on Instagram and post the same content and get 200 views. Right, that sort of.

And, and I mean the Chinese censorship is already a thing it says while the model is more tuned for Chinese language queries because that's the tokenizer and the training dataset and government censorship of certain ideas. It does happen to win in the universal languages of code.

Don't. And Mac. Don't ask Deep Seek about the Uyghurs. Your laptop will explode.

Yeah, not. And so Deepsea claims that a single node of 8x H800 GPUs can achieve more than 50,000 decode tokens per second per second peak throughput at the quoted API pricing of output tokens alone. That is $50 of revenue per node per hour. The cost for an h for an 8x H800 node in China is about 15 an hour. So assuming perfect utilization, Deepsea can make as much as $35 an hour per server, or up to 70% gross margins. This is what we were talking about yesterday. As you make these models more inference optimal, you get higher margins. And that's a lot of what deepseek has been kind of lauded for. And it even says even more interesting is the novel architecture Deepseek has brought to market. They did not just copy what western firms did. There's a lot of stuff that they did copy and there's a lot of data that they probably stole and there's a lot of weird political things where they maybe broke rules and laws, but they did create some fundamental innovations technically on the algorithm side. And I think that this whole story is going to be complex because people are going to want to take one side or the other and just say, oh, it's, it's a complete copycat. They're lying about everything. There's nothing interesting here. Or it's amazing. It's the biggest breakthrough and, and they need to boil it down to a single tweet. And that's why I'm glad we have an hour.

There's about, there's two. Yeah, there's, there's, it's such an interesting thing where they're coming out with an open source product that they are undercutting the entire market on. And, but it's this, it's hard to be, it's hard to say, oh, you know, they're simultaneously putting everything out there while there's clearly stuff happening behind the scenes that we don't really know about. Right.

And so we'll get into a little bit of the background. But they do seem pretty cracked. There are brand new innovations in mixture of experts, rope and attention. Their model has more than 160 experts with six riding. Oh no, the phone fell. Too much Chinese.

Oh, it's too many notifications.

Yeah, yeah, yeah. Well, back to the show. Furthermore, Deep Seek implemented a novel multi head latent attention mechanism which they claim has better scaling than other forms of attention while also being more accurate. They trained on 8.1 trillion tokens. DeepSeek v2 was able to achieve incredible training efficiency with better model performance than other models at 1/5 the compute of Meta's llama 370B. And so for those keeping track, DeepSeek v2 training required 1/20 as much energy, essentially 1/20 the flops of GPT4 while not being far off off in performance. And so then he goes on to talk about, you know, is Microsoft committed to OpenAI? There's a big discussion there that's kind of old news.

And I mean the new news there is that Satya is, you know, subtweeting Sam in Elon's comments. So that shows a little bit of, you know, maybe the marriage isn't perfect. Yeah, under the head.

And this is, this is something that he does bring up throughout the piece and throughout a lot of these analyses which we talked about it really. He really was early to this idea of Satya saying I'm good for my 80 billion. And what does that really mean? Well, it's Azure workloads and this is something that happens a lot where Zuck will say, oh yeah, I just bought 200,000 Nvidia GPUs. And then we dig into it and you're like, oh well like a half of that is going to recommendation algorithms for reels. And, and that's great. Like it's, it's cool. They have the infrastructure, but it's not all LLM AGI Race training which is what people are really focused on. So I'm moving on to page 7. Is distribution and integration king with deepseek and llama 3405b coming to the open source there is very little reason did.

You want to enterprise before that? Did you want to cover Microsoft trying to reduce their.

I think it's kind of covered that in the sense that yeah, you know like the closer here is that many firms use OpenAI's technology through Azure. More than 65% of Fortune 500 now use Azure OpenAI service. It's noteworthy that this is not directly through OpenAI. OpenAI can lose significant business without Google DeepMind or Amazon Anthropic gaining share simply by Microsoft pushing its own model their own model instead. And so he was just kind of commenting in this piece about the like the temperature in the room with regard to the Microsoft and OpenAI deal at the time. But that story has evolved so much and we've covered it so much. I don't think we need to dig into that. I'd rather focus on this distribution and integration is king section with with Deep. With Deep Seek and Llama coming to the open source there's very little reason for enterprises not to host their own model. Zuckerberg strategy of using open source models to slow down competition commercial the competition's commercial adoption and attract more talent is working wonders. So it was like why is he open sourcing this thing? Like this is very expensive and it was all about getting people to come work at Meta on this cool project and then also curb more companies from, from doing build outs because they're like what's the point? And this is probably the pressure that we felt on Aiden Gomez cohere, right. And maybe even a mistral where it's like well if your whole thing is like the open source or like you're like not the true Frontier compounder winner. Well you're, you're, you're just never going to be able to make any money because the open source one's going to come out. Yeah, it's going to be a little bit behind GPT5 or GPT4, but if Zuck keeps open sourcing it, you're just never going to make any money. And so it's going to be really hard. So he's keeping the pressure on. And then also because of databricks, which we've talked about before, databricks lets you fine tune really easily. And so it's no longer like fine tuning is no longer a monumental task. And so one of OpenAI's advantages is that they have been ahead in collecting data, specifically usage data. And that that happens when you, sometimes when you issue a query to GPT for chat. It'll say hey, we actually give you two which one do you more? And then you can A B or every post will have a thumbs up, thumbs down and then also they will, they'll be able to just see, okay, did you follow up with another query? So they're collecting lots of data and they have hundred million users. But he says only a quarter of Americans have ever, ever even tried ChatGPT and most don't continue to use it. Most future consumer LLM usage will go through existing platforms. Google, Instagram, WhatsApp, Facebook, iPhone, Android. While Meta hasn't found it out how to monetize it, their Meta AI powered by llama370b is available across Facebook, WhatsApp and Instagram. They announced rollout, the announced rollout has been extended to 14 countries including the US so they've hit 1.1 billion in population but they have 3.2 billion daily active users. And so they're scaling up and when they do they get all that data from and then they can use that to fine tune the next version of the model.

Mark, Imagine being marked and you're already halfway through your tam.

Yeah.

And you're trying to get these incremental users that just don't have Internet.

They're like, well they were focused on that for a while. They were like we are going to bring satellite to.

Yeah, yeah. It is interesting. Meta could have probably almost launched a Starlink competitor with how much they spend on the Metaverse.

Oh yeah.

And then actually deliver the Metaverse. And the real Metaverse is just reals.

They had a project for that, project Loon at Google which was Internet via balloon via blimp essentially. That's cool. And I believe in India there's a free Facebook like phone plan that you can get the Internet and you only get Facebook. So you kind of have to do everything or something like that. Yeah, you basically have to do everything through Facebook. But, but, but it does, it does bring more users online. And so all of that is all is all of that is feeding both fine tuning data on the LLM and then also just user data and preference data that goes into the system.

They're saying if Apple was more creative would they just sent up their own satellite network and just said you now are going to get all your yeah.

Data from us and you don't need Verizon anymore. Would have been cool, probably more reliable. And so Meta's deployment probably hurts Google search more than Bing or Perplexity ever will because people will just wind up searching in Facebook and Instagram. To serve 3 billion people, you clearly need to have a small and efficient model to bring the cost of inference down. Especially when you're not making money on it. You're not charging 20 or $20 a month or $200 a month. Either Meta has made the financial math work or is prepared to invest heavily to execute a land GR consumer AI space. Either spells disaster for incumbents.

Yeah, and that's the difference between Apple and Meta. Where Meta is still in founder mode. Apple, Google are in manager mode.

Yep.

And they're willing to invest. Google obviously has to show their investors and the markets that they're investing and serious about winning AI race. Yeah, but it doesn't actually look that ambitious in practice.

Well speaking of ambition at potentially non founder mode companies, the other argument to be made is if compute and capital are king, in that case Google is king given their hyper aggressive TPU build out pace. So Google has, has created their own chip to rival the GPU called the tpu, the Tensor Processing unit. It's, it's fabbed at TSMC I believe and so, but, but it's designed specifically just for AI workloads so you can't even run video games on it. Essentially it's just for this. Much like the advanced.

No, no three Pixie. No Scope.

Yeah, yeah, much like the advanced Nvidia stuff. But, but Google is investing super heavily and definitely takes seriously the capex. And so ironically Google now has focus and is directing all large scale training efforts into one combined Google DeepMind team. While Microsoft is starting to lose focus by directing resources to their own internal models that compete with OpenAI. And so Google they actually they DeepMind used to be a separate company. It was acquired Demis Hassibas and that team and working kind of separately they finally it was Google Brain and DeepMind and they, and they brought them all together because they were like the capex is so intense, you guys have to work on one big model instead of two smaller ones. And so this is one of the biggest risks for OpenAI is that the capital game is all that matters. If that's the case, the tech company that invests the most is the winner. And this is why we're seeing Project Stargate, which we'll talk about later. But you know, with capital does matter, OpenAI needs to be in a place where they can have access to the Absolute largest cluster. Yeah.

Which is just. Which is why it's amazing that OpenAI has been able to do what they've done. Because Sam Altman, I promise you, does not care about some other startup that raised $30 million. No, he cares about these players that can invest $30 billion off their own balance sheet to compete with him and have distribution.

Yep.

So it's not. That's like the most harrowing task, to be honest for you. Take the bet. One of the best fundraisers ever in the world, Sam Altman, and then. But you give him this impossible task of saying, go compete with people who have balance sheets that dwarf your total capital raised.

Yeah. And so, I mean, it really is insane. I mean, everyone's obviously still buying a ton of Nvidia, but Microsoft has the least custom AI silicon deployed to their cloud. But they're working on it. Google has the tpu, which is obviously the biggest, but Meta and Amazon are also ramping internal silicon. And even Apple has Apple silicon, which is only for the phones and the laptops now, but clearly they have internal capabilities to develop custom silicon. And so you could see them developing their own, their own hardware as well. OpenAI trained GPT4 in 2022. Since then they have been entirely focused on the next step up, experimenting with new architecture, data, etc. OpenAI has a tremendous first mover advantage and they have been solidly targeting models an order of magnitude larger than everyone else for a while now.

So they have the first mover advantage, but every time they ship something new, people dissect it and try to implement many of those.

Learning for sure, for sure. But they seem to just be like, you know, a little bit farther on the treadmill, like just.

Yeah, yeah. They're constantly ahead, but then they're by nature of the product, they have to expose some amount of Totally how it actually.

Oh yeah. In some of these, it's just like if there's a good idea like voice or upload a PDF, it's like, okay, that's like a.

Features are easy, but also the machine, how the model works. Right, right.

Yeah, totally. They, they try and hide that, but if there's just a few secrets, it's like one, you know, East Bay rationalist party where everyone's talking and listen, you know.

No, it's so hard because people in startups like to like to yap. Right. There's no. I'm sure they all have like really intense, you know, NDAs and things like that. But yeah, it's not like it's a government Even stuff leaks out of the government, right?

Yeah, totally.

It's not as secure as.

And so he kind of closes with this, the midwit meme of on the left.

That's why he's goaded.

Yeah, yeah.

Selling these memes to hedge funds for probably like a million a year description.

So as far as the capital game is concerned, Sam Altman is flying around the world winning the favor of many of the richest people in the world for a reason. They will raise more money than anyone else is even imagining to spend on a single model after they showed the world what they've been cooking the last two years. The real risk for big tech companies is that they do not show any meaningful gain gains in revenue, but they have to keep accelerating their spending to keep up with OpenAI or maintain market share in their monopoly Google search. This would cause data center capex party to keep going and consequently their margins to compress. So it's interesting because there's almost like this expectation that it's like, of course OpenAI loses money. Like they just became like a real for profit. It's a startup. Sam's at the helm. He's dealing with Masa. It's this crazy thing. They're going for AGI. If you're Google, it's like, like you're a shareholder. You're like, wait, why are your margins compressing? Like, is this really that real? Like, why don't you just partner with them? Like, is this important?

Well, it's hard because Google needs to show that they're competitive with OpenAI. But at the same time, if they're margin, everybody's. A lot of people have been short Google because they're saying, okay, well these models are a threat to search. And so if you're, if your margins are compressing because you're spending more while people are starting to use competitive products for search, it's really hard for Sundar to go up at a quarterly and be like, yeah, we're not really worried about this model stuff. Like, yeah, yeah, he kind of has to put on a brave face for markets, but you know that behind the scenes he's worried about his job. The probably the number one candidate would be one of the founders.

Yeah, but. And it's this weird, it's this weird bargain that actually kind of drives bubble behavior where if you under invest and AI is real, you're screwed. If you over invest and AI is real, you'RE great. If you over invest and AI isn't real, you still have a lot of assets. It's still valuable. Like there was a lot of dark fiber that wound up getting used later on. So the risk reward calculation is super weighted to over investment right now. So let's move on to the. I don't know if it was an interview or just some coverage of China's new face of AI Deep Seq founder Li Wen Fang. This guy. When Fong. This guy is. He's the next Frank Wang. The next Frank Wang, the founder of dji. We're learning, I mean, very similar stories. We'll, we'll dig into this. But this is from the.

By the way, he looks incredibly good for 40. He looks like he's like 22.

He does look 22. He's 40. Whoa. Okay. Yeah.

Which again, he, you know, we don't know if anything they say is real.

Maybe twice.

So he might be like Bill Menitis.

Oh, yeah, yeah.

Be like Bill Manitis.

Yeah. A little age gambit going on.

Yeah.

A lot of people in Hollywood doing that.

Yeah.

Changing the ages. So Deep Seek is in the news again because they launched R1, which is their reasoning model that competes with O1. And there's a post explaining who this guy is and where Deepseek came from that we're going to read through now. So the founder of artificial intelligence Firm Deepseek touted a 2025's biggest dark horse in the open source large language model arena, emerged as the industry's new face in China at a symposium hosted by the premiere in Beijing on Monday. Liang Wen Fong, 40, took part in the meeting where a select group of industry experts in the fields of technology, education, science, etc. Offered their opinions and suggestions, according to the news agency. And so last December, they made waves in the global AI industry after benchmark tests showed that its Deep SEQ version 3 LLM, which again is not reasoning model. This is more just like a GPT4 competitor which was built on a shoestring budget, which we'll discuss if that's real or not.

And I have to put that in the truth zone.

Yeah. Outperformed rival models developed with more resources by the likes of meta platforms and ChatGPT creator OpenAI Lee called on the economy's new growth drivers because we don't.

Have images pulled up. He is wearing a suit here.

Looks great.

We love to see.

We love to see it.

We got to give him some credit. He looks very young for his age. He's wearing a suit. So we're gonna, we're gonna really put him in the true zone hard in a bit.

And so recently, Beijing made AI a national priority amid heightened, a heightened tech war between the United States and China, the world's two largest economies. China's AI market is expected to be worth US$765 billion by 2030, according to state backed investment vehicle China International Capital Corp. In the age of aigopolitics, tech companies inevitably would have tighter connections with governments than before, says Winston Ma, York University law professor. And so apart from being a regulator, governments could be a sovereign investor or a cross border deal mediator. And we're already seeing that with the CHIPS Act. And so investment is going to be massive. Something like over a trillion dollars invested over the next six years in China's AI industry. And that's funny that they're saying that they're going to invest 1.4 trillion because that's three times what project Stargate is. So maybe this is Stargate.

To be clear, they market it as a pro, as a infrastructure product for one company. Yeah, it was very nicely positioned as, oh, this is for America. It was pretty brilliant on OpenAI's behalf. But no, this, we don't have a lot of details, but yep. And we should try to go through the girly piece because he, he like gave a thread of. Here's all the questions I have about, about OpenAI's Project Stargate. But if, if this is 1.4 trillion total is what they're saying over six years, if, if Masa and Sam managed to do 500, which not very many people actually believe that, you know, that's just one company. So presumably you have open AI or so you have Microsoft and Google and a ton of other big players, not to mention early stage investing. So I imagine we could, we can and should dwarf that number just to make a statement.

Completely disagree. We got to size it up. If they're doing 1.4 trillion, we got to do 10 times.

That's what I'm saying. We should dwarf their one point.

Oh yeah, perfect. Okay, great.

It would un American to not do.

I was about to, I was about to drive me by the collar and start just wailing on you if you start talking like that.

Yeah, that'd be very unameric.

So hailing, let's go into the background of the founder. This is, this actually went viral just as a screenshot because it's so interesting. So this is about the founder of Deep Seek. Hailing from Southern Guangdong province, Liang went to study electronics, information and computer vision, a field of AI that trains computers to capture and interpret information from images and video data. At Zhejiang University in Hangzhou, along with a group of university classmates, he started exploring how AI can be used to automate stock investments. Then he becomes one of the co founders of a great named firm called High Flyer Quant.

That's my quant.

That's my quant, which uses AI to manage one of the largest quantitative hedge funds in mainland China. So this is the stuff that you're seeing at Jane street and Citadel now and maybe jump trading too. And a lot of those firms have moved to AI driven investment strategies, but also they, they, because they're doing high frequency trading, they operate at a very low level. So a lot of them, I mean, Jane street writes everything in OCaml, which is like a programming language, like machine language is almost like, it's like, like Python is like three levels of abstraction away from them. So they can very fine tune the speed at which these algorithms run and everything can be proven like mathematically. So these are just like some of the best programmers in the world. And high flying. High Flyer Quant grew its asset under management more than tenfold over a four year span, from 1 billion won in 2016 to more than 10 billion in 2019, according to local media reports and information from the company's website. And so it seems like they've really been on a run over the years. High Flyer Quant spent a large portion of profits on AI to build a leading AI infrastructure and conduct large scale research, the company said in a statement in April of 2020. Three months later, high flyer Quant spun off Deepsea, which launched a series of AI models used by developers to build third party applic by the startup to create its own chatbot. High Flyer Quant managed to buy more than 10,000 Nvidia graphics processing units before the US government imposed the chip bans, according to a local media outlet. On its website, the hedge fund manager said it spent 200 million yuan and 1 billion won in 2020 and 2021 respectively to build its Fire Fireflyer series of AI computing clusters. Then a senior research scientist wrote that Deep Seek has emerged as this year's biggest dark horse. And so they've been investing a ton and everyone is talking about how cracked.

They, and yeah, they very much are trying to spread the narrative that this company just randomly popped up, spent 5 million bucks and had a model that was competitive with open AI. And it can't be overstated. How, how much? That is not the truth. Right? Yes, it's a team that had been.

Building and we'll get into this, but there's There's a theory that, that they got a call from the US government or from the Chinese government and basically said like, look, we understand that you're making tons of money as a quant fund, but you have the best engineers. Like, it is your duty to step up and build AGM and go compete with. And go compete. And you have to do this. Let's skip the interview because it's a little bit dense. But if you flip that next page over, you'll see China leads the world in AI in positive AI sentiment. And I thought this was kind of interesting. So this is from an Ipsos survey and according to a new Ipsos poll, China is the most optimistic about AI's ability to create jobs up there with Indonesia, Thailand, Turkey, Malaysia and India. The 70, 77% of Chinese agreeing with the statement AI will lead to many new jobs created in my country. Country contrasts pretty, pretty dramatically with America's 36%. Very, very interesting because yeah, in America, like the consensus view is definitely like AI is going to destroy.

Yeah, look at stock yesterday, he's like, oh, if you're a lawyer or this or that, why would I ever. You know, he had a very negative view.

I mean, it's probably a reflection of the fact that like America is mostly a service economy, which is much more automatable by AI and LLMs versus if you're a manufacturing driven economy, then yeah, it might be more important to manufacture than ever. And it's not like they have some massive, you know, professional services, you know, economy going there. So. Interesting. Anyway, Adapie sums this up nicely in a very long post that I think we should go through and then we'll get to some timeline on the.

What's. What's Adapy's backstory? I see his post, but I never actually looked into it.

I. I don't know, I think Anonymous poster for a while, kind of AI insider slight, you know, very dedicated to the posting. Like very focused on like what will actually go viral. I've seen some like, really the kind.

Of guy who sees Tiffany's payouts and, and punches the wall. Sure.

For sure.

It's like that's not.

But I think at one point he had a podcast with Eric Tornberg on AI or something. Maybe it was like a solo show. Anonymous. But now I think he might have have gotten like lightly doxxed and kind of decided to put his face and full name up there.

Yeah.

But he says Deep Seek is not a side project, which is the narrative that was going on. Oh, they just like did this for fun on the side. They're so cracked.

A lot of people like to call things.

And it is weird because, like, it does appear that they are extremely cracked and like insanely like, these are really great engineers who are creating technical breakthroughs. But at the same time, this was clearly like a serious project. Project. At the same time, employees are not lying when they say it is. The story they are telling is myth making. In the same vein as Silicon Valley. We want to make the world a better place, but at the same time make billions of dollars. The team obviously had access to more than 10,000 GPUs. According to the Scale AI CEO, it was around 50,000.

And that's as of today. The Scale AI this morning, shared Alex Wang, Alex was saying, hey, they're totally lying, basically.

Yeah. And the reason that they're lying is because there are very strict, strict export controls on GPUs right now. And so they, they legally should not, as one organization have access.

Yeah, this is the situation. Like this is the equivalent to a bunch of terrorists driving the Toyota Hilux. And everybody's like, Toyota, how did you get, how did these terrorists buy 400 Toyota Hiluxes? And Toyota's like, oh, I don't, I actually don't know. And it's like, yeah, that's enough volume that you probably could figure out. Yep. Hey, how do these terrorists.

And then that's exactly how with the DJI drones. Like, oh, if there's thousands of them in Ukraine.

Yeah. And so, and so I don't think this has happened yet, but Jensen should get pressured to explain. How did these thousand GPUs just land in mainland China?

Yep.

Yeah, because it's not a small number. Yeah, there's a, that's a whole. And, and yeah, there's haven't, hasn't Nvidia adjusted some of the other chips in order to sell them into China too?

Yes. So there's, there's the H100, which is the, the current top of the line AI training chip in America. And then every time they, they Nvidia launches a, a new chip. So the next one's Blackwell, the previous one was a 100. They, they also create one that's like slightly nerfed in order to not trigger the chip band and the export controls, which a lot of people are really frustrated by because of course there's so much demand for GPUs and line time at TSMC. A lot of people would say, hey, Nvidia, you're an American company, just don't even Worry about that making that Nerf chip and wasting any manufacturing capacity on that. Instead spend 100% of your manufacturing capacity on the chip that yeah, it can't get exported but we're going to buy it in America. America like Zuck wants it and Elon wants it. And, and you will sell 100 of your capacity and your stock will moon. You don't need to do this. But I, I think the argument is.

Like, it's a weird situation because so many people in America that would be the ones to say to Jensen like cut, cut this out. Have so much Nvidia that they're incentivized to just kind of let it rip.

Totally.

It's this kind of.

Yeah. You're like ah, we'll let this one slide. You're making a little money under the table maybe. So Deep Seek feels more like Skunk works, perhaps a necessary one. Core quant business became less feasible regulatorily. It's like Lockheed setting up a small, a separate small team to compete with SpaceX because the main United Launch alliance was not going to work out. It's also very hard to track costs in China because the regional government absorbs so much cost. Early Bitcoin miners had access to free power because power plants to nowhere and miners were willing to site next to them.

And, and so to be clear, there's a very real scenario where Deep Seek can authentically say we only spend spent $5 million on this model.

Exactly.

But the local government was like we'll spend 5 billion exactly. Like that's not out of the question.

Yeah, it's interesting. The power is just a thousand times cheaper over here.

Yeah, yeah. It's very common. So in the Middle East I was, I was in, I was in the Gulf at one point with a portfolio company that was getting courted to set up manufacturing over there. And the deal on the table was basically you don't pay for any power for the next five years.

Yeah.

And so that's, and that, that and, and that was for a foreign.

Yeah.

Business to come into that area. Imagine, imagine the kind of incentives that China wants to offer to being like hey, we have these 10 coal power plants. We think the Three Gorges Dam is.

Like it's a massive dam. It's one of the biggest power generation facilities I think in the world. Maybe the most here. Another example. Alibaba was able to get regional governments to absorb warehouse construction costs on their balance sheets rather than directly pay for it and looked extremely asset light and software when it went public.

This is why American Investors. Yeah, this is. Wow. Ben. Wow. Ben just played a video on his phone in the middle of the episode. Really phone it in.

Ben.

The podcast Falling Apart. But this is why American.

We're going to have to replace him with a subsidized Chinese labor.

Producer. Ben. China. Ben Wang. Yeah, the. This is why American investors always get hosed when they invest in, in, in, in Chinese companies, basically because China has a different approach.

It also happens internationally. I, I was working at a venture capital firm that was investing in internationally in like emerging markets. So they do like clean tech in like Africa and India and stuff. And this is when the Belt and Road initiative was going on.

And I hate that initiative.

I hate that initiative. And I read, do you remember Cablegate, the leaks? That it was like a WikiLeaks project where they seized all these like, cables between different, like State Department embassies and stuff. The campus would go back and forth and they leaked them. And you know, it became like a kind of a story on like, is the government doing something wrong in this country? But like, it was a, it was a very interesting trove of information if you just searched through to find just like, how is it, what is it like doing business as an American in Kenya, you know. Yeah, because the, the, the, the Kenyan embassy would write cables and explain like, oh, we, you know, here's what, here's what's happening on the ground. And one of the things that they found was that what I found in these, in these emails, in these exchanges was that when an American company would come in and say, hey, we want to do business, like, oh, you need some infrastructure built like, we're happy to help you. There would be a bidding process between America and China. And the feeling was that the American company was its own entity and the Chinese company was backed by the full force of the Chinese government and would basically have an unlimited.

Yeah. And so when people, when people complain in the US that Tesla gets these subsidies in the form of EV tax credits and rebates, that is happening on steroids 10,000x in China, where because the CCP, the government and the economy are so intertwined, the CCP is almost a king maker, where they can be like, okay, you're our horse. You know, we're going to back up the truck. And you know, the company gets to skim some profits still. But, but ultimately they're being, you know, this is king making process.

Yeah. And so Adapi kind of closes with it's perfectly possible for most of the cost to be parked in a balance sheet Outside of the core business, perhaps as some form of tech data center construction incentive. It's also possible no one accepts. No one except the founder knows all of the financial arrangements. Some of these can be absolutely insane. Handshake deals which get resolved by reputation.

Yeah, tying this back to dji. Yeah, same exact thing. They, they claim that they got off the ground with nine $90,000, but then they also claim that they were selling their products at the hard cost to make them for years. And so how did, how does that work? It just doesn't make, doesn't make sense.

So he says, this much is clear. The model is really, really good. On par with OpenAI release from two months ago. Having said that, unreleased models from OpenAI and Anthropic are probably better. The research agenda is still being set by the US by the US firms. This model was a fast follow on the O1 release. They are working very fast as they are catching up sooner than expected. They are not copying or cheating. This isn't industrial espionage. At most it's reverse engineering.

I would, I would caveat that by saying China is infamous for industrial commercial espionage. And it's probably the case that it hasn't been proven that they're doing this yet, but without a doubt, OpenAI has.

I think, because it's open source, people saw that the architecture was in fact different. But I think the training data, a lot of it is synthetic data that's been pulled from the GPT4 API.

Yeah.

And so they're, they're like, okay, we need a trillion tokens. Like, let's just go to ChatGPT.

And this is probably, it's, it's probably a funny situation similar to when TikTok started spending aggressively on, on Snapchat, Instagram to acquire users where OpenAI is probably getting a lot of benefit from a.

Revenue standpointi business is ripping. Yeah. Oh, there's one client that spent a billion dollars. Where are they? Oh, it's.

OpenAI is like, we got to figure out who's spending all this money.

And so they are largely developing their own talent. They're not relying on US trained PhDs. They are less constrained than American law firms by IP licensing, privacy, safety, and political concerns around wrongly ingesting data from people who don't want to be trained. Yeah.

I guarantee if there's a Chinese media, a Chinese media company, and they're like, hey, we're kind of annoyed that Deep Seek is just like harvesting all our data. Like, she goes to them and be like, I will literally kill you. You like say another word.

And they're like, okay. Yeah, yeah. They also seem to be over the Tiananmen Square issue. The model can say it. Even the Deep Seq website doesn't.

Even if the Deep Seq website doesn't.

Oh, interesting. So when you're on the Deepseek app, you're interfacing with their model, but the actual model under the hood, they didn't censor that because they just like whatever, we're just sending this over. Yeah, interesting.

No, it's good. They can argue our model is not censored.

Yeah, yeah, yeah. But just our Apple app. Interesting. Of the, of these, the most significant upgrade is that they are able to develop talent internally without relying on us trained PhDs. That expands the pool.

Wait, I got a call out for the audience. Go into deepseek, play around with the API and try to have it generate an ad for your business that we will read out on here. So, but it needs to be a screenshot of, of you know, actually on Deep Seek.

So Yeah, I like that. So let's go through some timeline but.

Don'T give them any other, any other.

Give you some, some, some up to date information about the R1 launch and how it's been received. We'll start with Mark Andreessen and we can kind of go back and forth here. Mark says DeepSeek R1 is one of the most amazing and impressive breakthroughs I've ever seen. And as open source, and as it's open source, it's a profound gift to the world and then to your taxes. Says game recognizes game. An American patriot too can do justice to real achievement without petty minded copes. And so I think so. Yeah.

This was posted today.

Today.

That seems like an insane thing to post to me. I'm not, I'm not gonna like, like if you understand. I think it's not. It's sort of disingenuous to say that Deep Seek is different than TikTok or DJI.

It is like the fact that it's open source means that you can run it locally and it won't call back to the ccp.

Sure. But the broader risk of deep cp Deep. Deep Seek is extreme enough. And I'll, and I'll. I'll quote a very conflicted member of the conversation, Neil Khosla father obviously has, has a big position in OpenAI. He says Deep Seek is a CCP state PSYOP and economic warfare to make American AI unprofitable.

Yep.

Fact check. True.

Yeah. And the important thing there is that it Is is Zuck has been running this playbook with Llama open sourcing the lagging edge AI model and it's had a strong effect on the competition and so deep seek open sourcing the leading edge model is going to have a similar effect.

So he says they are faking the cost was low to justify setting the price low and hoping everyone switches to it to damage AI competitiveness in the.

US now that part I don't think is true because I think the model is actually compressed to the point where, where yes, you can just take the model, run it on a smaller cluster and if I'm choosing between Llama or running or paying GPT to host, paying OpenAI to host GPT for me or running DeepSeek locally, I will choose Deepseek locally because it's cheap. So that part doesn't fully sit because yes, they are faking the cost is low but, but, but it doesn't matter what price they set because it's open source. Like they're both things are true. They, they, they are faking the price but also it's cheaper to inference and that is valuable even if you're not paying them.

Yeah, even you're paying different. So, so signal comes in brother. Signal says complete nonsense. Your dad wanted it to be pure profit and regulatory captured so it every, everybody you know can't stop bringing up his dad, which I don't think I.

Mean timeline of turmoil.

The timeline's in turmoil, folks.

Yeah.

Until we have proof otherwise, your argument is just conspiracy theory. Now I'm not a defender of the ccp, but where would, where did they get the hardware question mark? How did they train under the radar question mark. I'm happy to be shown some proof. Also, even if, if it's 100 million, it's still a major breakthrough. The proof's not coming folks. They're not going to open up their book. Nvidia is not going to, not going to say, oh yeah, we actually figured out how to get them all these chips. The proof's never going to come. The main thing from my point of view is you can't. Yeah, no, the main thing in my point of view, China has repeatedly shown that they are willing to lie. They're willing to sell things for less than they cost to make in order to get an edge in the US market. And I do think this is again the default state of any important international tech company coming out of China. China becomes a front for the ccp.

Yeah.

And you just can't. There shouldn't be any Trust. I would hope that startups are smart enough to not, you know, a lot of consumer brands will say, well, I build on TikTok because my customers are there and I have to get, you know, and so I think the same thing will happen. But I hope that we have to.

Call this is a different scenario. Because if you just download the Git repo for Deepseek and you host it on Amazon or Google Cloud or whatever, there's no data feeding back. You're not helping them improve the model at all. That's what Mark's getting at, is that it is this complicated geopolitical thing because it is kind of making the lagging AI labs in the US less relevant. And that's potentially hurtful. But it's not as dangerous as building on tech. TikTok. Yeah, it is. It is like a fundamentally different thing.

Yep. So Neil fires back. One, I'm not my dad, two, I'm fine with open source AI. We use Llama for some of our own projects. Three, the export controls on chips were merely inconveniences, which we already talked about and Alex Wang just said it on national TV as well.

Yep.

It's a great model, there's no doubt. But as with almost every high importance industry, the CCP is undermining it with propaganda and subsidies. You see this in EV drones, electronics manufacturing everywhere. And so I'm going to take, I respect that it's open source, you can run it locally, maybe you should leverage it to some degree. But we have to treat this as a threat.

Sure. And I think you're right in the sense that it would be extremely dangerous if Deepseek became a dominant consumer app. And it was like, oh yeah, Americans aren't using Google anymore, they're using DeepSeek as their home route. That would be very bad because that is a controlled app.

But I just, but I just think about it from the lens that again, China would not allow open AI to get a foothold in the Chinese market and have hundreds of, of, you know, thousands of companies leveraging it. Sure. And I think we should follow their lead on this one.

Yes.

I mean, the re.

I don't, I don't think they've done anything to stop Llama from going over there.

All meta products are banned, but not.

Open source ones because you can literally throw it on a hard drive and just fly it over and then you still have it.

You can still make it illegal at the maybe national level.

It's just such a, it's just, it's just such a It's such a lower like security risk. Yeah, it is economic war.

Here's an interesting.

A security risk post.

I don't know. The only other thing on that point is, is if you look at what's the. What is the. What's that malware that took down the Iranian.

Oh yeah.

And so.

So there's maybe needs a security audit first.

We. There's. There's.

You're talking about stuxnet.

Stuxnet.

Yeah.

Was something that literally traveled to every country on a. On a. You know, I'm just saying that, that there's. We are not qualified to even really.

That's interesting. Yeah. I mean the actual code that runs an LLM and pulls in the weights is pretty simple and it's probably built on top of PyTorch or TensorFlow or something. And so it's probably pretty easy to audit that. And I don't think these models are fully intelligent in the way that you could bury a secret alive AI that steers things secretly. But. But maybe in the future. Like it's not. We just don't know sci fi, but we're getting close.

It's just believable that the Chinese intelligence would say, hey, we have this like Deep Seek is a winner.

That would be a great like sci fi plot. Or like actual like problem would be. Would be we're open sourcing a model. It's really, really reliable.

We're going to make it super cheap.

But then it doesn't call back. It doesn't even need to. But over time, over time as you ask questions, it will slowly start steering things because it's essentially like it has a. Has like a higher order thought process in there that it's reasoning through and it's not giving you those tokens.

Yeah, I just as I didn't say there's.

It's fascinating.

There's so much technology that. That isn't publicly the public ourselves we're not even aware of. Right. And so I just don't trust. I don't just trust it. And I think there's enough press.

One can't fight it.

Yeah. One Citrini. We got a post from Citrini first time on the show. Deep Seek, being the brainchild of a Chinese hedge fund is such an complete total cultural victory by capitalism that ideologically at least it should soften the blow. Which is cool.

I like that hedge fund guy. Yeah. I do wonder if we'll see some sort of LLM or AI contribution from Jane's Term street or Citadel or something like that.

They're, they're doing things that, that like if you're in the, if you're a hedge fund and you have some edge via AI, I promise you, you're going to keep that in house, extract as much value out of it. When you're done with it, maybe you release it and be like, here, you know, we made this.

I do think that the, the talent flow balance was shifted for a few years. Like Palantir was really the hot ticket for, you know, the best engineers. It was big tech for a while, but until the AI stuff, the really, really top engineers were getting bored of going and writing ad optimization. They're like, if I'm going to just go right, ad optimization, that's kind of the same thing as writing code that optimizes the stock market. And so why the real, the real.

Alpha, the real Trojan horse would be Jane street releases a model model that everybody can use that like kind of helps you make money, you know. Yeah, but then, but then it's feeding.

Them that, yeah, yeah, that's good.

The order flow and then they're just trading against you.

Yeah, yeah, yeah.

Niels Roj Raj says, oh, this is.

Such a funny one.

So there, so there's this Chinese company called Deep Seek which basically does what OpenAI initially intended to do. They're open sourced, a model trained with large scale reinforcement learning, beating everyone else and even releasing a paper detailing their process. And it's funny when they release this stuff in English because like who's the target audience for that?

Translate. But yeah, yeah, yeah, I mean it did take over the timeline. So the fact that this means, and.

This is what, this is, what, this is what's scary.

Okay, tell me this.

From Stanford to MIT, DeepSeek R1 has basically become the model of choice for America's top university research researchers. Basically overnight, half a million views. This was posted yesterday. Day and again, this is just bad. There's no way around it. Do we want America's youth using TikTok for two hours a day? Do we want all of our researchers using Chinese models? It's just objectively UN American. OpenAI scraped the web to build a closed source model. Deep seq scraped OpenAI to build an open source model. This is from Norgarten. Low temp banger from this morning. Another one, Deep Sleep.

The data exfiltration is the real cost one because that might have been done by another firm and then just passed over who knows how much. They, they're certainly not going to say, hey, building this model cost us $100 million in OpenAI credits, right?

Just would never admit that I don't like this post, but I'm going to read it anyways. Deep Seq is another example of China rapidly commodit new high tech industries and preventing the west from making excess profits or rent as it's called. If this keeps going, the standard of living in the west will fall. I don't necessarily agree with the last line because obviously Llama has already been doing the same thing effectively, but certainly TikTok eats into eats into Meta's profits just because of attent Their their profits are derived from the attention that they get. If they're losing an hour a day per 20 year old in the US that that does you know they are going to take a hit. But yeah, we'll see how this plays out.

It is interesting how like one of the anti open source AI arguments for a while, and I was certainly arguing this, was that it doesn't make sense for America to open source cutting edge age cutting edge AI because it's just giving our best tech to the CCP free. But now it's kind of flipped around.

So did you generate this with deepseek?

No, this is a so I think the last post there explains what that is. It's from Ethan Moloch, I think, and it's him giving examples of Deep Seek and just explaining how Here, take this and I will so Ethan says the raw chain of thought from Deep Seek is fascinating, really reads like a human thinking out loud, charming and strange. And so why this is interesting is because most of the when you use ChatGPT01 it does internal it has an internal monologue, it has a lot of it's generating tokens that are about the task and reasoning through it and then it's adjusting and whatever you ask it's saying, okay, here's what the user wants. And usually that's either tucked behind a little like tab that you don't really look at, and even if you look at it, it's usually just giving you summaries of it. But because Deep Seq is open source, you can you can read literally every token it generates. And so he asked it do something to delight me in exactly seven words. Make it not cheesy. This is important. So very simple prompt and it thinks for eight seconds and it says okay, the user wants something delightful in exactly seven words. Not cheesy. Let me start by understanding the key points here. Delight can come from surprise, cleverness, or a unique perspective. The challenge is the seven word constraint, which means Every word has to count. Account first. I should avoid anything overly sentimental or cliche. Like words like love, heart, forever might come off as cheesy. Instead, think about elements that evoke wonder or curiosity. And then so it's like going back and forth, like wait, map is a verb here. Fireflies, mapping constellations, that personifies them slightly, giving them intention. And so it's like thinking through all this stuff, counting all the words. And then it comes up with fireflies trace constellations in quiet night conversation, which, which isn't that good. But anyway, it's an interesting deep dive into just understanding that to get a seven word response you don't just want to go to an LLM and just take the next seven words because that's going to be truly slop. But if you generate three pages of text, you can get a little bit closer. And I think this is kind of a bad example. But a better prompt is create a history of DJI and it'll pull a bunch of stuff together, together and then it'll revise that and refine that and do like kind of editing passes and say okay, do I really have a beginning, middle and end here? Did I include all the dates correctly? And it's fact checking itself so then when you get the final result it's much better. And so that's just a good example of like how chain of thought reasoning works, how these reasoning models like work in general. And I think it's just cool that people are able to introspect things and I think a lot of researchers will learn from this stuff. And there's some more posts in there about the actual technical implementation.

Yeah, this one is from Henry. He says, I've made over 200,000 requests to the Deep SEQ API in the last few hours. Zero rate limiting. And the whole thing cost me like 50 cents. Bless the CCP. OpenAI could never. I hate that I had to read Eric works, he got me. But yeah, I mean this is the whole thing. The developer community, because they are hosting it feels like smoking crack.

And so this is an example of what you're working worried about because if you're. I guess my contention is like yes, if you download the git repo, make sure it doesn't have any spyware in it and then you're hosting it locally. That's a lot less controversial to me because you're just using the open source library. But if you are interfacing with deepseek.com or whatever and every one of these requests that he's putting in is information for them to train the next model. And that's not good because that, that really will be a compounding advantage in.

The way that we talked about how.

Meta has a compounding advantage because of all the users that are on Instagram and interfacing with Llama. Even though Llama is open source, Meta is getting the training data and the and the like, the upvotes and downvotes from those user interactions because they're because Llama is the number one LLM, not because people are hosting it independently, although they are. Plenty of companies are hosting Llama independent independently. They're the number one used MLM because it's tucked into Instagram and so they just have a ton of interactions there.

Yeah. Signal. Another good post. Kind of hilarious how Deep Seek is exactly what Chat GPT was supposed to be what OpenAI originally promised before Sam pivoted hard toward profit. China going open source on AI is a wild plot twist. Basically throws out every argument Vnode Khosla and others have made. This is like watching the Communist Party write a love letter to decentralization, which very signals a great writer. But yeah, I don't know if Sam pivoted towards profit. He clearly enjoys making a buck. Yes, he clearly enjoys Koenigsegg owner. Yeah, Kona Segs P1s. But it was more it really, it does seem that he was correct in believing that you weren't going to be able to raise enough money via nonprofit structure to compete.

Or the open source model.

Like even like the Red Cross's annual budget globally is like a billion dollars.

Right.

And that's like one of the biggest, you know, nonprofits.

Yeah.

I don't believe that China is writing a love letter to decentralization. No, you, you could say it's like watching the Communist Party write a love letter to decentralization, but the Communist Party Party is purely interested in control and maintaining power. Yeah.

It's not a love letter to decentralization. It's a middle finger to our chip industry and our trade export controls.

And the irony here is that you could ask Xai to explain what features does Deep Seek have. And.

There'S even more on this with the AI diffusion rule. This is a new law and rule against chip exports. And they're just showing that we're going to find the loopholes no matter what. And the really crazy thing is that the loopholes, the export ban is only relevant to American owned intellectual property and Nvidia because they're an American company. But China is actively working to catch up in chip fabrication with SMIC and SME. So they have their own tsmc, they have their own Nvidia, and they have their own asml. And so they're behind. Yeah. And they have their own open AI.

And so they're willing to spend at every single level energy at the state level, energy data centers, the actual model level.

And so this is just like we're going to fight. I do wonder if the open source thing was. Was more about like making a splash and actually being a middle finger than any. Is that even more important than the economic warfare? It's just to really get on the map and really, really say, hey, we've caught up. And it's almost like a battle cry, like a rallying cry for them that we're like that we're on top now. Anyway, what else? Oh, this is fun.

This is interesting.

You see, that's fine. Yeah, I'll explain this. So Rachel Tolstoy BB says tagging the ADL under every tweet, denying deepseat costs 6 the reason is because people like. It's like, it's like it's a. It. It's a criminal to claim that they spent that little.

Yeah, yeah.

And it's like you want to.

Yeah.

Engage.

So this is an is from an interview with the founder of of Deep Seek about moats. He says for the past 30 years we've been focused on making money, often the expense of innovation. True innovation is driven not only by commercial incentives, but also so by curiosity and the desire to create. He's giving off TikTok vibes here. We've been constrained by past habits, but these are transitional. Interviewer says but after all, you're a commercial organization, not a nonprofit research institute. If you choose to innovate and then share it through open source, where will you build your moat? For instance, the mln MLA architecture innovation this may will likely be copied by others, right?

Yep.

Liang goes in the face of disruptive technology. A closed source mode is temporary. Even OpenAI's closed source approach hasn't stopped others from catching up. Our value lies in our team, which grows and accumulates know how through this process. Building an organization and culture that can consistently innovate is our real moat. Open sourcing and publishing papers doesn't mean we lose anything. For technologists, being followed is an achievement. Open sourcing is more of a cultural act than a commercial one. Giving is a form of honor and it attracts talent by fostering a unique culture. This guy's media training is insane. Yeah, Lulu, we're Going to have to start asking some questions. Is trained by the best, it seems. But these are all quotes that are going to resonate with the developers that they want to get using their product. So I don't believe any of this at all. And there's this other guy on X called Han Zhao and this guy Anton just post. How is Deep Seat going to make money? Gets two and a half thousand likes Banger Han Zhao says Deepseek's holding is a quant company many years already. Super smart guys with top math background happen to own a lot of GPU for trading, mining purpose and Deep Seek is their side project for squeezing those GPUs. So this, this is literally on the, on the, on the 22nd or a couple days ago claiming that Deep Seq is a side project, but also the Deep Seq founder is meeting directly with this, yeah, the whatever questionable with the ccp. So yeah, don't believe, I'll just say you just cannot believe or trust anything that they say.

So yeah, I mean the interesting thing here is that like, even though the results are super impressive, very clearly the algorithmic improvement is very serious and legitimate. And I don't think anyone denies that. There is the question of, okay, China was able to kind of like weasel together probably 50,000 GPUs. And there's a few ways that they can do that. You can only officially buy 50,000 GPUs for the next four years, but you can exploit loopholes. So there's a limit on 1600 GPUs at a time. So you set up a ton of shell companies, bring them in that way, and then there's also training that happens.

What I'm saying this is Malaysia goes over. This is like, you know, Toyota selling their Hilux to, into Afghanistan.

And so just to give you some scale here, Xai and Meta are building 100,000 GPU clusters. And at that point it's like you're getting so many GPUs it becomes harder to hide it in a batch of like, oh, a thousand over here, a thousand over there. Like let's just sneak a few over here, right? And just to give you some scale, a single H100 GPU is $24,000 and that's about 40 to 45K. Once you add in servers for GPU H100, 100 and 100K. GPU cluster can draw 150 megawatts, but the next wave is 400 to 500 megawatts and then 1 gigawatt in 2026. And when you add up everything. It's like $5 billion for 100,000 GPUs. And so Xai's round was 6 billion and they're building 100,000 GPU cluster. It's like it matches the cost of the GPUs and all the networking almost perfectly. And so the question is, was this an aberration in the sense that this was the last great model that China can build before hitting an actual real export wall, or will they figure out a way to do something that looks like what Xai and Meta are doing with.

I mean you just, you just find a, like markets, companies, governments find a way. Right. Like, wasn't it that the United States didn't want Israel to have a nuclear weapon and then they just made it. Yep. Right.

Yeah.

It's all going to happen.

Yeah.

It will look like, oh, Peru bought 50,000 H1 hundreds and then they didn't train a model. Like it just. We don't know where they went. Right, yeah. So.

So I mean, it's getting tricky. Like there are pretty strict export rules on, you know, you can only go to an ally country. If you go to literally.

It'll literally be like, there's. Who's that? The guy we traded Granny Grannier. NBA player in Russia.

Okay.

We traded him for the merchant of death. Oh yeah, there's going to be the merchant of gpu. Yeah, yeah, yeah, that, that, that's just doing these black market GPU deals. Some kind of imagine down China doesn't care. Like China doesn't care if they, they pay 50 grand per each. 100. Yeah, they probably pay 100 grand for it.

Sure.

And so some engineer in the US gets approached by a nice young lady. Well, out at the bar, she's like, hey, like I have this new project.

The numbers are going to get so big that it's going to show up in Nvidia's like earnings and stuff. Oh, there's an extra 1 million GPU order and there's only four companies in America that that could be, you know? I know. And if it's not them, like where did it go? Oh, it went to some country and like, okay, we want answers. It'll be very like. What I'm saying is that like as the data centers get bigger, like they become more visible on satellite. The shipments become not just, not just a shipping container, but a whole.

All I'm saying, just because easier travel, the illegal arms trade, drug trafficking, all this stuff.

Yeah, you're right.

Has been happening. We've never been able to stop Arms dealing. Never been able to stop drug trafficking.

Yep. A million GPUs is a lot though. I. I would hope. I would hope that it's easier for.

Us to stop a million.

Okay.

I've watched the last Netflix show I enjoyed was. Was Narcos. Sure. And it's a great show. Narcos. You have a situation where the US Federal government knows exactly who is selling the drug and raking in the profits and still takes him decades to take them down. And. Yeah, so it's just tricky when you.

Have like a, like an SEC regulated company at the center of this, like Nvidia, like pressures directly on them.

Yeah, but you can. There's nothing that stops a company in Argentina from being like, we want a thousand GPUs.

Yeah, but you're gonna have to do that 10,000 times.

But it's worth, It's.

It's worth. Yeah, yeah, yeah. I guess.

Same way that Thousand Fronts. It's the same way. Okay. We know that fentanyl crosses the southern border every single day.

Yeah.

But it's happening. Thousands of people are doing it. So to stop it is like whack a mole. You hit one and they pop up.

So it's more important than ever to, you know, monitor who. Who these gpu' are going to. And there's an interesting story in this interview with Dylan Patel about how Elon is kind of just going completely founder mode on his data center build out. And I think it's kind of instructive for how you can kind of duct tape together one of these huge training operations. So Xai, as we mentioned, they raised 6 billion for 100 GPU cluster. And they are just 100,000 GPU cluster. They're going crazy. They bought, bought. They bought a defunct appliance factory near Memphis. And it's right next to a large national gas plant. But then they also tapped the local natural gas line and then set up natural gas generators. Then they also brought. Which is like not ESG friendly at all, but. But they, they just need the power. Then they also have Tesla battery packs to help. Yeah. Like balance out load. And then they're also taxing.

We talked about this. That they sell like a $50 million battery pack.

Yeah, yeah. Literally just sent that over there.

The reason to be extra bullish on Xai is that Elon is the only founder mode model, you know. Yeah, CEO. I don't know if he's actually CEO or. But whatever. Chairman at least. Who has done large scale infrastructure projects over and over and over.

Yep.

The factories At Tesla, Starbase, et cetera. Yeah.

And so they're, they're, they're next to this gigawatt natural gas plant, which should support roughly a million GPU cluster, which would be like the GPT6.

And the crazy thing, it's, there's probably a chance that in the long run, elon does absorb OpenAI. Right. Because you just don't know. Who knows? He's happy to acquire a hostile company.

Yeah, yeah. And so one of the funniest things. So there's two funny things we should also talk about in here. One is that the ESG thing is becoming like way, way more complex than it used to be. So there's typically the AI researchers are pretty ESG pilled and all the big tech companies made huge ESG promises saying all of our data centers are going to be 100% clean energy by 2028. And like Apple made China this, China definitely doesn't do it. And, and, and they were, they were, they made those promises based on. Oh yeah, well, we just serve like, we don't serve AI models, we just serve like, you know, basic queries to our database for, you know, Instagram. We just, we just host photos. Like, that's not that energy intensive to just serve you up a photo and take your like or your comment and slop it in.

They weren't, they weren't expecting us to release 10 hours of podcasts.

Exactly. Yeah. And so, and so they were like, our compute and our energy consumption is growing at like, you know, 2%, 5% a year, like with our revenue, with our profits, with everything. So yeah, it won't be a problem. Just swap out 10% a year for clean energy and then we'll be clean by the state. Then this massive wave of AI training comes and all of a sudden they're like, okay, we need to deal with like, we need to find energy. A lot of it's like through these outsourcing projects where there's third party data centers that don't have the projects. But what's interesting about it is that a lot of the super AI pilled researchers are like, actually now I'm super worried about the environment. I believe in climate change. I do really care about clean energy. But I think AI will, AGI will be so powerful that we don't need to worry about climate change in the short term. And so let's just burn all the natural gas and oil and get to AGI and then AGI will literally, they literally think AGI will figure it out and just do carbon Sequestration and we'll have so much powerful super intelligence that we'll be building Dyson spheres and we'll be able to just completely change the environment however we want. So we shouldn't let this hold us back. Which is a very interesting switch in the model because normally it's a very left right issue. It's a very like, oh, if you're this thoughtful scientist, you definitely let's not destroy the planet. But now they're like, actually the planet can take it because we're really close to the AGI future, which is fascinating. And then the last thing in here that was interesting was, I mean XAI is just doing all this crazy stuff. They needed water cooling for all their GPUs and so they brought in restaurant grade chillers that just chill water. And so they just have like a ton of cold water filtering through. It's all just like, you know, duct taped together. And then the last thing is Monday.

Monday or Tuesday we're going to cover the business of, of data centers.

Yeah, like coreweave. Yeah, that'll be great. Core weave and stuff. Because core weave is also doing crazy stuff. Core weave has 200,000 GPUs, billions in revenue and they convert, they're converting old crypto data centers and then hooking into on site gas plants and all this stuff is just measured in, in just energy consumption now because it's so, it's so big and Amazon and Anthropic have these trainium GPU servers coming on, but they're a little bit smaller or like less powerful power efficient than others. But it's a massive, massive, it's a massive knockout drag out fight.

Which we love.

Which we love.

We're ready for war.

We're ready for war. It's great. Well, that about wraps up our deep dive on Deep Seek. Hope you enjoyed it. And we'll be right back after a quick break. Thanks for watching. Welcome back to the Technology Brothers, the most profitable podcast in the world. We are continuing our deep dive on OpenAI Stargate joint venture with a fantastic artist article in none other than Dylan Patel's semi analysis he writes the OpenAI Stargate joint venture announcement had many folks heads turning despite us calling out that the capital requirements for OpenAI's immediate plans months ago. So OpenAI has been talking about hey we want 100 billion for a long time. The headline 500 billion is such an earth shattering number that it also caused deserved skepticism from folks like Elon Musk stating that SoftBank has well under 10 billion of funding secured. We talked about the yesterday Altman clapped back by saying it was already under construction, which is true and we'll show some pictures and to come visit it and you can in fact go to Abilene, Texas and start looking around here. And so it's fun to see the.

Elite going wait, don't skip over this headline Cluster measuring contest. Oh yeah, yeah our best, our elites are getting into a cluster measuring contest again.

Yep. SoftBank doesn't have anywhere close to the money required for the cluster. At the same time the cluster is already under construction. It Texas see the low res GIF for free subscribers.

Well presumably isn't there going to be many clusters?

I think this is basically all so initially it's all going into one.

This, this one. Yeah, this, this one that's under construction could be under $10 billion.

Yeah.

So and in that case then they would be in a position to start construction. The, the other thing, I'm sure we'll get into this but, but debt plays a big factor here. So even if, if they have $10 billion they might be able to build.

Yeah $30 billion of actual the, and he goes into the total cost of ownership stuff. But essentially the, the game that we're playing here is order of magnitudes. We've the, you can think about the GPT4 cluster as $500 million. The XAI cluster, that's 5 billion, that's 100,000 GPT. And now OpenAI is planning for the $50 billion cluster and then the $500 billion cluster which will be 1 million and then 10 million GPUs. And so every time we're going up an order of magnitude in terms of GPUs, power and money. And so like the way you're thinking about it is 150 megawatts goes into 100,000 GPUs and that costs 5 billion. And everyone now is talking about the multi billion dollar cluster. And, and this is just thinking a few years out to what's the next order of magnitude. Yeah and so the reality is that only the first phase is certain which is a portion of the committed 100 billion. But the top line number is being counted in an interesting way to say the least. We already discussed the first phase of this Abilene Texas data center publicly in our report four months ago. We also have had it in both the accelerator industry model and the data center industry model for quite a bit longer. Below we go we go over what's actually happening with Stargate. And so the real numbers behind Stargate we believe this project is being measured on the total cost of ownership, not capital expenditures. That is the 100 billion and the 500 billion is TCO, which includes the capital expenditures for data centers, servers, networking costs, power costs and many other operational costs, including debt and financing costs. Furthermore, we believe Only the first 100 billion has data center sites and power ready to go. Funding is not fully secured. The other deceptive detail is that the first cluster of this mega project is the Oracle Open AI data center deal that was announced last year. This is being built by Lancium, Crusoe and Oracle. As far as we know, no other such softbank.

Yeah, back to everybody. This was a party round for a new entity that is going to do some stuff for OpenAI.

Yes.

And a lot of bunch of that's a clearly a holding company of some sort that's going to have different owners at different levels.

Yep.

They'll be, you know, probably location based entities.

And so the goal of the venture is that they're going to invest $500 billion in AI infrastructure over the next four years with 100 billion deployed AME lately. It's the first big AI initiative announced by Trump. The equity founders in Stargate are Softbank, Oracle, MGX and OpenAI, with the former three companies contributing capital. Initially, the project includes the build out of 20 data centers, 10 of which are underway in Abilene, Texas, each with roughly 500,000 square feet. So let's get into the scope and location of the initial $100 billion investment. The initial Oracle Crusade deal is included as one of the data centers that will be connected to other GIGA campuses, which I love the name. It's fantastic to perform distributed training at scale and is orders of magnitude larger than any other training run. Cleverly, they went for a more creative financial structure through this separate entity which pushes the need for fundraising down the road. And so moving on the campus mentioned by Sam Altman as already underway is the Oracle Crusoe Abilene campus thing that.

You skipped over that I think is interesting. So the project includes the build out of 20 data centers, 10 of which are already underway in Abilene.

Yep.

With each being roughly 500,000 square feet. So just giving a sense of the scale, 10 separate 500,000 square foot facilities. Imagine if you had just happened to own a bunch of retail. Yeah. Like businesses or real estate.

Yeah. What are you thinking?

I mean you have to imagine that we. One, you got bought out. Ready?

Totally.

Two, if you had a gas station there all of a sudden like your revenue just like probably 10x. I don't know what was going on in Abilene prior to this, but you're doing pretty well if you owned, you own real estate there. Because presumably these will just keep expanding and expanding.

I mean I know some guys who have been to the site and they said like when you fly in like you see it from the air, airplane and it's like, it's like immense. It's like an intense experience seeing how it's, it's a mega project, it really is. And so the campus is being built to house the hundred thousand GB200 cluster for OpenAI's use. The development, along with the GB200 servers that are due to be shipped over Q2 and Q3 will be dropped into the Oracle Crusoe data centers. There's a high resolution photo of the campus shown below. Four of the 120k square foot modules make up one 500,000 square foot building. So only two buildings are currently being built with 180 megawatts of critical IT capacity. To give an understanding of the scale of this build out the 20 building campuses. The 20 building campus will be comprised of 80 of those modules. So it's going to be like, like it's going to look like a city, which I'm excited about. This is awesome. Based on the build out we have Scheduled scheduled for OpenAI in both our models, OpenAI doesn't have to build an entirely new campus to expend all 100 billion of TCO. Therefore we believe that the entirety of the initial $100 billion investment from Stargate joint venture is going into this campus alone. And so it sounds like with these different, like Oracle is going to own one, Stargate is going to facilitate the ownership of a few of these. But it feels almost like when Satya says I'm good for my 80 billion, it might be, well yeah, I'm going to build a huge data center and it's going to be part of campus. I'll own this one, I'll own these GPUs and so I'm chipping in but it really, it's mine and it just happens to be here. So it's localized and we're all working together on this. So there's some, there's some details here. On the Abilene clean campus, it's 1100 acres, 200 megawatts energized and a gigawatt energized by 2025 5. There's a tax abatement in form of city and county pilot zoned for light and heavy industrial and. And they're building additional gigawatts already.

Yep.

And so the campus is in. The model is known as the Lansicum Crusoe Abilene Clean Campus Lancium and has a stated capacity on their website of 1 GW. Energized in 2025.

It's cool to see when you let private companies take on big projects like this that you can do things pretty quickly.

Oh yeah.

Comparing this to the California rail.

Insane. It's just nightmare project. Yeah, it's insane. Yeah. The, the guy behind Crusoe is known for just being really, really good at finding power. And I think that's kind of what kicked off a lot of this, was just realizing that Abilene, Texas was the right mix of different resources and pipelines and grid infrastructure, that this was the place where you could actually get all the power. Because that' what the big tech companies have been hunting for and they've been looking for, oh, can we turn this nuclear reactor back on? Can we tap this natural gas line? Is that wind available or solar available? Can we just stack a bunch of batteries here? Like what, what can we do? And so in short, we estimate OpenAI is paying roughly $2.82 per GPU hour for the first GB200 cluster, which is a total cost of ownership of $2.38 an hour for the full scale cluster cluster. These figures scale with future Nvidia GPU deployments and then they break down some of the Nvidia revenue estimates. And as we go forward to new GPUs, the number will double for each generation and the power per GPU also grows. While 700,000 GPUs sounds small for 100 billion.

So this is why, this is why our joke being like it's all for you brother. The post we made about Jensen and Masa. Oh yeah, so like Nvidia is the real winner here. You've said it before, they have more than 100% of all the profits in AI right now.

Yeah, because everyone else is losing money, but not Jensen baby. And so the project's financial backing, Oracle and SoftBank are the well known equity providers. MGX is a more recently founded technology investment vehicle from Abu Dhabi and the fund is chaired by the chic who is reported to oversee 1.5 trillion in investment funds. So the money really is available to some degree. If they want to pour it all in this, they can.

But MGX as of now is $100 billion fund.

Yeah, that's right. And There's a question about like you know, what will the concentration be for that investment vehicle but it could be very high since this is such a big story and it has like such insane backing. In light of the recent Elon Musk claim that SoftBank could only raise under 10 billion, it is important to note the JV will also raise debt capital for financing. It's unclear what the mix of equity and debt financing will be. Oracle and MGX can fund a big chunk of the project with their massive balance sheets. OpenAI and SoftBank are allegedly down for 19 billion. On paper each. SoftBank doesn't have this liquid today, but we think SoftBank will likely sell off a portion of its ARM stake to fund the equity check. And So I think softbank has like.

Soft bank also borrow against their arm. Certainly Moss is not, not afraid of levering bridge.

OpenAI also doesn't have the capital but they will be able to raise with O3's release and they don't have to do it directly fully, but can through some kind of hybrid instrument issued to everyone else in the consortium.

Elon Musk is right. That's the tough, that's, that's the tough thing when you get into like it's not hard for Sam to raise a billion and this is the same thing for founders. Very easy to raise a million. A million. Raising 19 million million. You're getting into territory that people start to ask they want the spreadsheets start coming out a little bit.

The question is if project financing risk will be separated from the parent organizations. Separating this risk is important because OpenAI will need to raise capital to pay for this. Thankfully, OpenAI is not immediately required. OAI has enough existing capital to pay for rental payments on the cluster this year, but they need to be able to pay rental fees in 2026 and beyond which scale mass massively. It is possible OpenAI cannot raise the money directly, but we are believers that their technology and products will scale to support this. The reality of the joint venture many big names were announced alongside the Stargate JV ranging from Nvidia, arm, Microsoft, Oracle and the previously mentioned financiers. The names are large and impressive, but the reality is there are distinct winners and losers here. Microsoft is the biggest loser. Microsoft is largely left on the sidelines as OpenAI utilizes another infrastructure partner. As Sachin Nadella stated, all I know is I'm good for my 80 billion. Microsoft's recent blog post stated the facts of the partnership and how Microsoft has a roer. But the reality is that this is incrementally negative for Microsoft's long term because they were afraid to take on the risk of this investment.

Well if they're afraid and the fear was correct, then it's not necessarily guaranteed to be negative.

No, no, good, good point, good point.

Yeah.

And then AAM is thought to be a winner potentially because they rallied 16% of the news because they were named a technology partner because of course they're owned by SoftBank. So a lot of people were worried that ARM would sell off because what if SoftBank sells their whole position but if they have a massive new contract to, to sell new chips.

And that's what he's doing.

But only because of the Grace and Vera CPUs that accompany Black, Blackwell and Ruben GPUs all from Nvidia. So SoftBank is actually like they don't even make GPUs but they make CPUs that need to go alongside every GPU for coordination of the actual server. So the server has a lot of different. And there's so many companies that we, we should get into in the server, in the, in the, in the data center deep dive there, there, there are companies that just make ethernet cables and they're just printing or like they make you know the screws that go in the server racks or like the air conditioning units, the water chillers. Like all of these things are like you can't even buy natural gas generators anymore because they're so in demand and so those companies are going to be ripping. We should do a whole market map like totally with that other company. The, the, the, the business of rebuilding I guess so SoftBank likely pushed for ARM to get on the PR and the, and the optics look nice. ARM isn't doing much. The reality for ARM shareholders as stated above is SoftBank will likely have to sell off a chunk of its stake in the company to to fund Stargate. We think investors are largely missing this point mistakenly view this announcement as materially incrementally good news. Oracle is placed with a huge burden of tackling both data center management and managing supply chain logistics. It's quite a task to manage a 1.8 gigawatt cluster and P procure all necessary equipment at the command of another company. Nvidia is obviously involved as it's their hardware. OpenAI will maintain control of operations and run almost all of its own cluster management software. The future of the, of the joint venture. This is the next step in an exuberant progression in the AI cycle which he recently wrote about in Fabricated Knowledge. I gotta Follow up and read that. That sounds interesting. We know the first hundred billion will be spent in Abilene and that there will be more campuses developed for the remaining 400 billion. Even more impressive, all these data center campuses will be, will continue to be built out with the aim of distributed training. So training across multiple data centers.

Yep. And so, and so there's some savvy real estate investors out there trying to figure out where the next data centers will be. Oh yeah, and buy, buy up the surrounding land and properties and stuff like that. Because all of a sudden new data center pops up. It's like, okay, well all of the surrounding rental, like apartments just probably, probably two extra.

Yeah, I mean data centers aren't as dense in terms of human capital as a Facebook campus, but there's still tons of employees and tons of stuff going on and like there's just a million different ways to make money when there's such a hive of activity. I mean even like, you know, security guards are going to be in demand.

Like dump trucks and like the white monsters.

Oh yeah. Or the super cracked engineers, the ultra white monster. Yeah, those will be in short supply. There's going to be. We're going to need to keep all the supply lines flowing. But fascinating to see Stargate rolling out and excited to go visit the plant in person. I want to go see.

We need a shot of us, you know, on the ground. Wind ripping, yelling and you know, yelling into the microphone. Yep.

We're here down at Stargate Stargate. Maybe they'll rename the town. They should rename it Stargate Town. Or Starbase. Starbase Stargate Rywood. Oh, interesting. I wonder if the star name is like a little bit shot across the bow.

Yeah, I wouldn't be surprised if Elon buys up out of all the out of home ads in Abilene on ad quick. On ad quick.

He just rips and just Billy into.

Ad quick and just perpetually Trolls. Trolls.

You know, he would be, he would do that. He likes a troll.

No, he's. Yeah, he's not, not afraid to be. Although the more efficient thing to do is just comment with a crying emoji under every thing that Sam posts.

Well, that's our update on Project Stargate. Let's move on to. We got a dm. We got a correction to issue. On the last show I said that bees don't eat honey and I was wrong. They do eat honey. Coogan is totally wrong. If a beekeeper takes too much honey from the hive, they will starve, slash die. So I'M sorry. I made a terrible mistake. Please forgive me.

Won't happen again. That's kind of factoid. You just don't forget. I think I'm going to remember that forever.

These.

Enjoy, honey.

So our printer ran out of ink, so we don't have all of the post printed. We're going to be working off a laptop. But our first banger today comes from Dhruva. He says we can do this. We have the companies. We just need our own Shenzhen place and we'll soon get the policy. And Kyle Chan is saying Chinese tech companies are increasingly becoming tech. Swiss army knives starting in one industry and then quickly branching out into a range of adjacent tech domains. Xiaomi, which was a phone company, is now building electric vehicles. DJI builds LiDAR and they also build cameras and microphones and all sorts of production equipment. Li Auto builds robots. Baidu builds autonomous vehicles. And here's a piece that he has in high capacity showing how all of the different, different Chinese companies operate in multiple sectors. And this is just something we haven't seen in America. Like there was a time when Apple was going to build a car and we've seen some, some crossover, but I think the FTC really screwed this up because, you know, even just the Amazon thinking about building home robotics, they had to fight on the Roomba acquisition and all this stuff and there's just been.

So much Correctionba like almost died.

Yeah. And if you're, if you're a big tech company and you're going into to. Okay, yeah, we are Apple. We're going to build car. And in order to do that, yeah, we're going to need to acquire some teams. Like look at, look at what Zuck did with the metaverse he wanted to build. He bought Oculus and that one went through, but then he bought some fitness VR app. Did you hear about this? He bought a, he bought a fitness VR game basically. Clearly probably just an acquire for some great developers who never really got product market fit on their fitness app because it's such a small market, tries to bring them in and the FTC blocks it and says that so that this will give them undue monopoly control over the VR fitness industry. It's like, what VR fitness industry? This isn't a real thing. There's like $10 of revenue in this industry. Why are you shutting this down? Just let, let Zuck, like build a little bit of a team so we get something, man.

Zuck.

Let him Zuck. Let him Zuck. And so yeah, ridiculous. I don't know if you have anything more on on this.

But yeah, it ties back. I posted that we need our own project Stargate for domestic drone manufacturing. We've let DJI on the commercial drone, you know, consumer drone market in the US they own a lot of the commercial drone market. They fell into police stations. There's a lot of individual SMBs and manufacturing facilities that could manufacture parts for drones. But the issue is there's not a clear demand from US based companies. And so we need, if we want to be competitive on drones, which, you know, other people have said are the next. It's like getting into. It's like the. It's the next gun in terms of a player on the battlefield.

Yep.

We probably need to invest like $20 billion to like set up the proper infrastructure, set up campuses that can produce. Because in a near peer conflict, if drones actually are an important aspect, you know, players in that we will not be able to compete purely on an output basis. So even if we were to have a better product, say Andrew has a better product than dji, it just purely on a mass volume basis will get smoked. Right. So yeah, I mean, and this goes back to Carried, no interest has talked about this before. He says that if you want to compete with China, it's not about just encouraging manufacturing and making, making a better regulatory environment. It's massive government incentives and investment that the Chinese government has done for decades now.

Let's go to Ben Hylak. He says deep seat R1 has an existential crisis and I don't know if this is Photoshop, but he was clearly talking to it about Tiananmen Square and says no. I mean the Chinese government killed people there. And it just says no, the Chinese government has never made mistakes. It has only been dedicated to its people, the people of China. It has never made any mistakes. It has never made any mistakes. It has never made any mistakes. Any mistakes and it just like completely freaks out. And I don't know if this is real, but it feels real.

You never know with Hylac.

No, no, he's a good Photoshop artist. He's a notorious screenshot.

He did the Jaguar rebrand.

He did the Jaguar rebrand. That's right. Yeah, hilarious. But yeah, I mean, I'll have to play around with it myself, see if I can jailbreak it, see if there's some fun stuff. I'm sure we'll get more in the cool coming days of people playing deep seek. Let's go to SULA says should I try the Trump diet? Let me know. Will you try this Out. It's no breakfast. Well, we stay away from politics, but this is a health, this is just health. This is just health.

This is biohacking.

This is biohacking. The Trump diet is no breakfast. Small snacking around lunch, a feast for dinner. 12 Diet Cokes per day, no coffee, no alcohol, no eczema. Exercise.

I think Trump is so active that he probably gets 10,000 steps a day without ever going for a walk. Yeah, he's just like walking between meetings. Spends a lot of time on his feet. So the exercise thing is crazy. But, but also a lot of these old dogs that, that, that have, you know, Warren Buffett, etc, they don't seem to need to lift weights like they, they maintain mass. Mass. You know, they're size lords. Right? They're a size lord. You can keep the mass on without, without weight training. It's, it's interesting because most people feel better if they have a bigger meal for lunch. But I find when I'm really busy.

Yep.

Today we had, I had like, we had breakfast. I had like half a breakfast burrito. Then I won't have eaten until I'm going to have a big, you know, meal.

I mean, speaking of the mass monsters, the bodybuilders really do eat a lot of small meals throughout the day. And so there's something to that, like consistent sustained energy like snacking all through the day. So you're not seeing big spikes at any point in time. And then the 12 Diet Cokes is kind of the same thing. You're not seeing spikes in caffeine, you're just caffeinated lightly. All steady state. There's something to it, but it is a little silly. Bennett Siegel says seeing a lot of pitches on marketing agents as a low cost, high volume way to scale customer outreach. Before the Gen AI era, we just called that spam.

Yeah, I, it's, it's hard to be. It's interesting because, you know, I made a joke recently about how, you know, Trump making an executive order to ban the creation of new AI sales agent companies.

Very trained.

And so it's a very, you know, and I'm sure there's like 10 of these companies in the YC batch. Yep. People have joked before that oftentimes founders won't find product market fit and then they'll pivot into building some sales enablement tool because they, what they've just experienced around selling, they have some type of experience there or pain points. But the explosion of these AI sales tools have, has made me more and more bullish on the historical Sales leader who is taking their client out to a nice line lunch, really getting to know them, bringing them across companies. Right. So somebody might work at Databricks and be a top performer there and go to Oracle and bring a lot of those relationships over. Right. And so relationships seem, the value of relationships seems to be accelerating. The value of a cold email is going to decline, you know, pretty dramatically because I'm surprised somebody hasn't made a passive cold email tool that just sort of sends cold emails every day to high value. Kind of like people.

Yeah.

Being like, hey, what's up? I'm an 18 year old, you know, and historically to do a good cold email, you really had to put some effort into it. You know, you don't want it to be too long, but you want to make sure that there's some thought put into it. But I'm sure some 18 year old will build a tool that's like, we send 10 great cold emails a day on your behalf, you know, to.

Yeah, I mean, I don't think I've ever fallen for an AI generated cold email. I, I don't even respond to any cold emails. I don't even respond to warm emails often. I have about a thousand emails that I should probably go through. You know, even some close friends have trouble getting through to me sometimes, unless it's really important.

I have 3251 unread texts right now. So I'd like to see an AI.

I'd like to see AI try and break through on that. But at the same time, like, I do think that, yeah, it might be hard, it might be hard to like close me as a client, but you see this with like the AI slop on, on Instagram, like it clearly works on some people. And so the question is just like sales enablement for what product? Because there are people right now scrolling Instagram and being like, what, like, you know, a cow made out of cabbages or whatever. And they're just looking at the most, like, sloppy, like, oh, like a bed made out of bread. Like, so cool.

The other thing with, with emails, you can send a terrible sales email, but if you send it 10,000 times. Yeah. You're going to get one person and if it's very cheap to do that. Yep. Then it was worth doing, I guess. Yeah, it is.

Yeah. Yeah. Maybe, maybe the spam alpha is like, is like normally like direct email outreach is reserved for B2B customers. Where a single client closes is like usually well over $1,000 LTV and, and direct response on instagram ads is for stuff that's like around $100 LTV. Right. Or below. And so maybe the AI sales agents kind of democratize that SDR experience for consumer products at the lower level. And so it's.

I got it, I got it.

Yeah. It's more like you went down the marketing funnel and now, like, you're just texting with someone like, oh, do you want to. Do you want to actually close the deal on that, like, T shirt you were about to order?

It's interesting. The only email spam that if you go to your spam folder.

Yeah.

You'll see is for him, the products that you can get on him. So like, you know, boner pills, basically.

Yep. And it's been like that for decades. Like, that was the major alpha of, like, Internet.

And they figured out you can send. If you send a hundred thousand emails and only. And it's point, figure matter. Yeah.

Long as you can.

I got an email. I got a text message today from somebody. I don't know if it was real or spam, but it was. They were claiming to be from TikTok. And I was like, you really didn't do your research? You think they're going to recruit me to. To join TikTok?

Yeah. There's a. There was a guy. What's his name? He was like a drug trafficker. People think he invented Bitcoin. Paul LaRue. This is a crazy story. We should do a whole deep dive on him. It's fascinating. But basically he got his start online pushing boner pills and illegal supplements. Right. And he would get turned off. He would get blocked at, like, the domain name level. And so then he started buying new domain names constantly and just warming them up immediately, spinning up a new website and then sending more emails. And then he started getting blocked at the DNF level. And so he bought a registrar and like, he was the person that owned the registration, the ability to register new websites in, like, an entire country.

Yeah.

And then he went on and did a bunch of other stuff, and I think he was like smuggling crystal meth from North Korea at some point. Was like. It is a crazy story. He's. I think he's in jail now, but a lot of people thought he was like satoshi for a while because he wrote a bunch of crypto. Not cryptocurrency software, but like cryptography software. And it was like early, like, kind.

Of like powers for good.

Yeah, he was kind of like a. Like a Silk Road guy. Yeah, yeah, yeah. Kind of proto Silk Road. Let's go to Avi Schiffman on the bucket poll. Let's see what he's got to say. This look, this sounds familiar. Let's see what he's going to say. He says shower products are the biggest scam and only for disgusting pigs that roll around in the mud all day. I stopped using shampoo and skincare stuff a year ago, and my hair and skin is the cleanest and healthiest it's ever, ever been. Water and an office job is all you need. Water and an officer in an office job. Yeah, I think you had that take like, like a week before you.

Funny. I, I, when I was in college, still, I, as many college students do, would randomly suffer from acne. So I was like, I'm gonna try to, like, get some face wash and all this stuff. And I ramped it up. And I was using, like, face wash and like, you know, a moisturizer, and I, and I would look around at, like, you know, my friends, and I was like, I know you don't do anything but splash water on your face and your skin looks better than mine. Yeah, I just quit it completely and my skin got better.

Yep.

And it really comes down to diet people often, I think men do skin care and things like that. Reactionary. Either their girlfriend or significant other goes, hey, you should try this. Or they're having a skin issue, so they're like, I want to solve, solve this. But usually it's a dietary thing that's catalyzing it. And so you can just keep just putting water on your face and just eat better food and you'll probably be okay.

Are you familiar with the Taylor Lorenz skin care regimen? No.

I should be.

It's hydrogen peroxide wipes like a thousand times a day.

I heard it.

And I was, I was like, I've never heard anyone doing that.

It's because the temperature set to 88 degrees.

Maybe that's it. But I was like, I don't know anything about skincare, but that one sounds weird. That sounds wrong. Who knows? Maybe it works. I don't know. Give it a try. Let us know how it works for you. Give that a try. Bill Gurley says, as a third party, curious observer.

All right, serious.

As a third party, curious observer, I have several naive, unanswered questions about the Stargate project. Obviously understand that they have no obligation to disclose. Here's a list with responses wide open. I would love to be able to aggregate the best answers at some point. 1. Corporate structure. The OpenAI press release says that Stargate is a new company. Is Stargate established as a standard C Corp. Llc, a joint venture or something else entirely? And even a joint venture is not a, that's not a structure. Like it can be a joint venture. That's a C Corp or a joint venture. That's an llc. Right?

Yeah.

I think when you set up joint venture you just say we're starting a new C Corp and the shareholders are the, the, the joint venture partners.

Right.

So he has a bunch of these CEO leadership. Will Stargate have an independent CEO? Customer exclusivity is. Is OpenAI?

Didn't they say in the announcement that it would have a new CEO, a.

New board, something like that?

No.

Will they be as operationally intense as the XAI team that launched Memphis? This operating partner role releases OpenAI as the operating partner. What does that mean in practical terms?

I think one thing that meant that they cleared up in the, in the Dylan Patel piece is that they're using their software to manage the clusters.

Yes.

So it's, it's more like a SaaS management component than.

Yes. I like this. Is. Is OpenAI the sole customer or will Stargate yearn to serve other businesses?

Is that many people yearns?

You gotta yearn. Oracle's day to day involvement. What specific role are they playing? Did Oracle previously own the assets there comparison to Core Weave, what kind of business is Stargate? Is it effectively a direct competitor to Core Weave's AI hosting approach or is it structured and positioned differently? Debt usage? We talked about this. Core Reeve is rumored to go as high as 3 to 1 or 4 to 1 leverage. What level of risk will the entity consume? Appeal to third party investors? Clusters standalone and potential round tripping and pre training versus inference and data center structure. So he asks a bunch of questions and I think that you know, we'll have to do a follow up to see what kind of hopefully it is a post or he talks about on BG Squared but it'd be great to see where those come around. I think that some of them.

Yeah, one, one thing is.

But there are.

They don't. There's. It's private company. Yep. Absolutely no obligation for them to share. Leave them alone.

Actually, actually why are you asking so many questions, Bill?

No, everybody's been asking questions because it's such a big headline number of course that you're inviting a lot of that. But I think people flirting out. So if, if Oracle's public company they'll have to report on their partnership and investment. OpenAI has enough shareholders at this point that they'll have to come out. It'll. It'll come out. Softbank is a public company. Yep. So they'll have to. So it'll all kind of start. You'll be able to paint a more clear picture over time.

But again, they're raising pre incorporation. This is what happens.

Yeah.

You know, ra. Raise. Raise a trillion. Half a trillion. Pre incorporation.

Yeah.

You know, there's plenty of examples. The Google, Google check. It was, it was half of something. It was I think half a million or something. The guy wrote the. And they were like, whoa, where do we put this? Yeah, we can't cash this. Didn't matter. Worked out going to happen again. Let's go, let's go. Apple only bull market.

Bubble behavior.

Bull market everywhere. Bull market somewhere. Arbitrol at Electric Capital says President Trump just issued a major executive order that touches on everything. Crypto. This is a big deal. Stablecoins operation. Choke point developer protections, a new crypto task force and a digital asset strategic reserve. Here's a breakdown. Protecting economic liberty. Americans are guaranteed the right to use open blockchains without fear of persecution. Self custody digital assets. Promote the dollar through stablecoins. Banking access for all. No more choke point 2.0 drive to get regulatory clarity for innovation. CBDCs are off the table. So no central bank digital currencies. They're going to create a new digital asset task force with stablecoin oversight, risk management and digital finance. A national crypto reserve, he says. A rollback of the Biden policies. Trump revokes EO 14067 explicitly rejecting central bank digital currencies and rolling back policies viewed as overly restrictive or government heavy in their approach to digital finance. Can you look up what bitcoin did after this went out? I think it's down recently.

It's all priced in.

I guess it was priced in. I mean really, people were expecting Trump to do amazing things with, with crypto and so like, you know, how can he meet those.

Well, it's funny because what would a bitcoin have done if, if Trump Coin hadn't sucked all the extra liquidity? Right. People were selling bitcoin to buy.

Yeah. I mean certainly some of the alts like Dogecoin traded down because people were rotating. I mean bitcoin's so big at this point. I mean it's almost trillion, I think, maybe more and so but it's average. But still that's 5%. Yeah.

Mean, yeah, it's Trump Coin is attracting more money from people that are not in the trenches. Right. They Just have a little bit of money in Coinbase or whatever and they. They want to rotate.

Yeah, I don't know. I haven't been following the. The crypto news too much. I'm still just focused on those three that we always talk about, the stablecoins, betting markets and Bitcoin. Bitcoin basically store value. But I don't know. Hopefully we get more interesting stuff out of the of this. We haven't even entered the phase of the cycle where people are talking about new use cases. Really it's all just like yeah, that's.

Actually what I'm excited about is okay if Trump made kind of doing whatever you want legal by nature of doing what he did, then a lot of the. I wouldn't be surprised if somebody starts to try to use tokens to represent stocks and actually bringing real world assets on chain which historically was just expensive of securities laws. But that's been a big.

Yeah also just like when the markets are up like this, there's so many more. There's so much more capital flowing around for like oh, if somebody wants to give, really give, you know, crypto gaming another run like they'll have another shot on.

Yeah, there was so many.

Maybe it doesn't work maybe but like at least we'll get another shot on it.

I saw, I was doing some research for Banger Archive last night and I saw somebody posting the obvious end state is the intersection of crypto and gaming because everybody's sort of in game earning tokens with their time. But they're not actually tokens, they're just points in the game. You can kind of like sell your account or you can sell gear and it's sort of like it is a market but it's not actually financially engineered. And so I do think, I wonder what you don't hear about any games that were created during the 2021 boom. Like if you're remember Board Ape Yacht Club was supposed to be developing an mmo maybe they still are. I don't seen much about it went out.

There were like Decentraland and like the Metaverses that were supposed to be like kind of Roblox do whatever you want. There was one game that did feel like it had some modicum of product market fit called Dark Forest. Oh yeah, yeah, that, that, that got like a little bit popular and seemed like people actually liked and the map was all cryptographically sealed so it was like open source and you could kind of verify that the other person wasn't cheating. And there were some interesting things to it.

There's coming out financialized game coming out I'm excited about. So Josh Harris, who was one of my first hires at Party round, he was 19 at the time. Like, college dropout. Killed it at Party Round. After he went to, like, was like a year, I think, at Paradigm. And now he's built building a crypto chess game that I think will be pretty cool. Like, being able to play chess with, like, real money on the line, have, like, special characters, almost like a Dungeons and Dragons. So, yeah, that'll be launching this year, which I'm sure will be very.

There was always just this big problem with, like, the crypto stuff. Like, sounded good, but there was very little reason for a game company like Valve or Electronic Arts to say, like, yeah, sure, you can take your assets out of our beautiful ecosystem that we tax 30% on everything and we own 100% of the liquidity you have. Like, there was no financial incentive for those companies to ever be like, yeah, yeah, this is actually.

And the big one you had to. The next gta, if it's had.

Okay, so then I need to raise $1 billion and hire all the best artists and developers to actually make the next.

Yeah, I think that was the biggest crypto games. Yeah. Is like, you are not going to create a new game studio.

Yep.

Yes. You think, yes, you can use that. If you have a pretty good game, you can use the token to incentivize people to actually play it.

Yep.

But then it becomes just an economy where you have. In the Philippines, there was that game, I forget, Infinity, Axi Infinity, that people were like, this is the new economy. Yeah. And then. And then it just was a Ponzi.

Yeah. I mean, I watched a little bit on a Twitch stream just to understand, like, okay, what does this game actually look like? And it just, like, it wasn't as entertaining as, like, even League of Legends or, like, you know, Hearthstone or any game that it was, like, competing with.

And that's just because now you might have these big game. Activision.

Yeah.

You know, Riot. All these players being like, okay, crypto's a real thing. Let's implement. Implement it.

Yeah. There's just still such such a countervailing economic force to saying, let's not do that in capital.

We have the best game.

And yeah, yeah, yeah. Gta. Oh, you want to buy GTA coins?

And the whole thing is like, I don't think that you should. I think if you're the best in the world, if you're the top 200 player at a Video game, you should be able to make like a decent living off of it. Yeah, I don't think that video games should be financialized to the degree where just playing, playing at like, you know, the top hundred thousand players should be. Yeah, it's just a lot of. Yeah, we shouldn't purge that as like a career path.

Yeah, a lot of this just winds up being some sort of like regulatory arbitrage like wealth transfer. And I guess like the good ending would be like, hey, because of, because of restrictions on investing in companies, you have to be an accredited investor and so you can't raise money in mass. You can do like some sort of Kickstarter crowdfunding campaign for your game game. But then you need to go to the venture capitalists and maybe there's some sort of capital formation thing where if the game gets traction, like buying the coin is essentially like investing like the Trump coin. Like it's like. Yeah, it's just like, it's just like getting more capital into this thing. But then you have to count on the developer actually like making something. It's all very, very fraught. Anyway, let's move on to. We'll do Base Baron and then we'll do a promoted post. Yeah, Base Baron says E WAC or the memetic death and rebirth of tech.

If there was ever a moment, called it ewack.

If there was ever a moment, if there was ever a movement destined to fizzle out like a damp matchstick, it was eak. At least the EAS had the self respect to make their leader serve hard time. Spf. Here we see a movement ostensibly about the relentless push towards a techno capital singularity. Instead in my. Instead mired in the quicksilver sand of its own internal contradictions and the inherent flaws of its messiahs. First consider the misalignment between the movement's professed ideals and the actual behavior of its leading lights. Eyak, in its essence, was meant to be a clarion call for the acceleration of technology, a rallying cry for those who see the universe's entropic dance as a mandate for human ingenuity to multiply complex complexity and intelligence. Instead, what we observe, observe instead is a parade of digital anons ensconced in their online echo chambers, chanting accelerate like a mantra detached from any real world action. These are not the builders or the shapers. They are the spectators, mistaking their fervor for progress. And he goes on to write about. Wow, he really wrote a lot about this. He really cares. And he got 10 less words. So congrats.

Wow. No, I think. But I bet there's there's a real, sort of very real message here where when you have a movement, like.

It'S very hard to go from doxed to not doxed. Like if we knew who Satoshi was, would that be good for the price of bitcoin? Almost certainly. No.

Yeah, right.

Not even the selling pressure. Just every human is flawed and I find. Find this like there's so many, there's so many different, like health influencers. You got to live this way. And then there's a family influencer. You gotta have a family and like work influencers. Oh, you gotta work really hard. And then when you dig into those people, you can always find like, oh, this, this person who's saying you gotta work really hard like they have never had a family or like they've never. Or they're completely out of shape and you're like, yeah, wait, like no one.

There's. There's an issue where people commenting under every post, accelerate, accelerate, accelerate.

Yeah, it's cool.

But I think what, what the Baron's saying is he just wants to see those people actually creating progress in the real world. And not being part of a movement doesn't mean you're having any type of real tangible impact. Right. The way to have a real impact is to participate in the underlying activity that is driving towards the goal of the movement. So.

Totally. Yeah, it's tricky.

Love to see Baron. Love to see Baron branching out into more thoughtful.

And I still think EAK was best when it was just a joke at the expense of EAs. And I thought that when it came up, I was like, I never put in my bio. I never joined these like silly movements. I didn't do the dot E thing. I owned it. But I didn't put in my handle. I don't do the NFT profile picture. I have trouble joining these big movements because I feel like they're always very risky. You're on the top, you could back backfire. But with you'd rather super aligned. But I thought it was just so funny to just be like, oh, we're poking these like super serious EAs who are like, oh, we need to put everything in a spreadsheet. These people kind of need to be made fun of. And that's when EI is at its best. And then when it became this super serious, like actually acceleration is like the only way. It's like, okay, this is not fun anymore. For me, at least that was my.

Experience with it ruined it.

It Ruined it.

We got a promoted post on a more light hearted note, but also serious at the same time. We have a 19. John. John's. You're interrupting the post. John.

Let's go.

Relax. The 1966 Aston Martin DB6 Shooting Break by FLM Panelcraft is a rare and striking example of bespoke British craftsmanship. With only three produced and one in left hand drive. And I just have a thing for shooting brakes. I. The first car that I ever, you know, had access to was a Toyota Corolla. That was triple my age at the time. Very paint chipping, you know, hundreds, probably 150,000 miles on it. Ever since then I've loved shooting brakes. I was lucky to own a Ferrari FF for a while. Great car. Thought it would be a good family car. But the Ferrari reliability was somewhat of an issue. But this thing is just amazing. I mean, imagine, imagine driving across the country in that, putting some real miles on it. Somebody in our audience should do that. So great, great example here.

Pick up. Yeah. Let's go to Chairman Burb Bernanke. Love the name is. I'm plotting on your rise, bro. I'm literally up in the lab scheming on beautiful abundance for you. I'm cooking up ways for a rising tide to lift all of us, our boats. I'm praying for your gains. I'm lighting a candle to your inevitable splendor. Dude, just basic. He's plotting on your rise.

I love, I love this format.

It's great.

It's great ripping around.

Don't. Don't plot on people's downfalls. Plot on their rise.

I don't think, I don't think enough group chats. A lot of group chats have this sort of unserious people dropping memes and talking, talking a little smack, stuff like, like that. I want to see group chats that are oriented around winning together.

I'm in one. It's called Iron sharpens iron.

Iron sharpens iron.

Everyone takes it very seriously.

Yeah. But the issue is, is people at the top are hyper competitive.

Yeah.

So there's a little bit of that. You got to really, you got to cut it down.

This one's the most positive I've ever been.

Five people oriented around winning together.

Yep.

Helping the brother succeed at all costs. That's what we're trying to do.

That's what we're trying to.

This podcast, that's what, why we, we run ads. You know, you, you scratch our back. Yeah.

You leave us a five star review. You throw an ad in there. We'll read it on the show.

There you go.

Let's go to Katherine Boyle. She says America is in deal mode and deal mode is infectious. If you have a deal you want to do, now's a good time to get it done. Also, it's time to get married, dive in. It's the golden, it's the golden age of America.

She just added that.

Yeah, I like that. Little follow up like, oh, by the.

Way, if you have a post that does well, this post sponsored by Pronatalism.

Yeah, pronatalism.

Yeah, the, the, the deal energy is real. It's never, it's rarely time, it's rarely a bad time to do a deal. Right. You want to be greedy when others are fearful. Right. So people are pretty optimistic right now for a lot of good reasons. But. But yeah, go do some deals. It's a good call to action. Right, that's great. Yeah, Simply deal.

Yeah, deal deals with deal mode. Let's go to Ben hylack. He says ChatGPT can now use Claude.

And so Ben immediately, of course, he's immediately backing it.

5K likes. Good post. Yeah. So he went to Operator, create a new, create a new chat to create a, a new B2B dashboard and try using the app and tell me what you think. And so he's having operator open Claude, log him in and then use Claude to code for him. And it's all just very silly. But yeah, I don't know, I mean it's probably a good stress test of the system if you can use that.

It is interesting. So great marketing for Ben because he has an analytics platform I believe for AI products. But it's interesting how we're seeing already that somebody, I forget who posted this, that AI is eating ui. Right. Which eventually you will be able to use a model to use another model. So your preferred model. If you want to do something with another model, just say, hey, go find the best model to do this task or create this type of content. It'll just do it. And so you may never interact with the end ui. You'll just interact with your preferred model. And that is why for all the hate and shade that people throw at OpenAI, oh, it's overvalued. Oh, it's still owning. Being AI for the average consumer has massive.

You're an aggregator and you're the front end. Yeah. And when it was last year or the year before OpenAI launched GPTs where you could go in and create a custom GPT and you could upload specific documents, give it fine tuning Instructions. And there was an app store marketplace. And every once in a while I made one where I uploaded a bunch of books and a bunch of information PDFs about a specific topic, and then I could query that. And there were a couple others, like one that would help you write threads on X. And I think the idea there was that you let people build a ton of different GPTs, and then the one master GPT can route to one. And people were joking that there was one. One that was like laundry buddy that would just help you find like a laundromat. But. But in theory, it's like, yeah, like, maybe someone's made a GPT that does one thing much better. And then the, and then when you query the, like the master GPT chat, it can say, oh, okay, I'm. I'm going down. And this already happens with a mixture of experts and the, and the reasoning models. But we can say, oh, okay, there's. There's a, There's a coding model that's really good and I need to write some code. So let me go over there and use the.

That.

Or there's a booking integration. And so all of that's happening kind of under the hood. It's. It's funny in this frame, but it's, it's very, very real. So let's move on to Lulu. She says, oh, we should start with the. Yeah. Brother of the year. So bank of America posted a. An image that says, bank of America serves more than 70 million clients and we welcome conservatives. We would never close accounts for political reasons and don't have a political litmus test. And it's really that, like, my shirt that says I'm innocent of killing people is. Is raising more questions than it's answering. Yeah. That type of thing.

Yeah.

And then there was instantly a community note saying, bank of America has been placed on notice by 15 state attorneys general for debanking conservatives. And Lulu is breaking down exactly what went wrong here. One, if you turn off replies, you look guilty. You can't turn off, quote tweets. So you're just setting yourself up to get ratioed even harder by other means. You should have anticipated the community note, which looks worse than either. Than either comments or quote tweets. And if you're in the position of having to make a statement like this, comms and social media aren't your biggest problems. Rough. Yeah. I don't even know why there.

Eric Prince.

Yeah. Which is a quote where he has debanked by you guys is like, what?

And. And his family and his ex wife. And his ex wife's husband. Because they were receiving.

Yeah. Payments.

Whatever.

Alimony.

Alimony. Yeah. Yeah.

Wow. Yeah. Rough. Not a good day.

And the whole thing is there were fintech companies emerging over the last six years, probably four years, I guess. Sure. Years that were explicitly. Because conservatives were being debanked.

Yeah, they were.

They were saying this is the right way wing, which we shouldn't have. Political leaning.

Yeah, it's very silly.

Our financial institutions should not be political.

Yeah.

Politically leaning at all.

Yeah.

Let's just keep politics out of. We keep politics out of the show. Let's keep them out of our banking system.

I agree.

Agree.

Oh, this is, this is a fun one. I know you're gonna have a lot to say about this. Jared Friedman says Rocketable is the first. Is building the first one person, $1 billion company. The plan is simple. Buy up software companies, fire all the employees and replace them with AI agents. Literally the most AGI pilled idea I have ever heard.

Okay. So like many topics, like, like many podcasters, podcasters often have a lot to say.

Yep.

I'll preface this by saying I, I don't work in. Directly in, in private equity. Yep. And I'm not a software engineer. But. And the other thing I would say is I like this sort of big, bold idea. It's cool to have this major milestone, one person, $1 billion company. Very exciting. And I think that many sort of. This is a cool experiment and is worth funding from YC and is worth the founder trying to take a crack at it.

Sure.

But I have to. Matt, here. Here's the thing, like buying companies. Companies trade on a market and it's not perfectly efficient. But any company worth buying unless the owner does not have, you know, is not. Doesn't want to get the best possible price, is going to talk to a lot of people. Right. So if you're this one person company and to get one deal done, maybe you have to talk to like 200 companies. Right. And all those 200 companies are talking to other private equity firms, other individuals. They're exploring, raising other forms of capital. They have all these different options. I don't know that the classic private equity strategy has been to come in, fire all the unnecessary people, replace them with modern software. And so to me, I'm not quite sure that this. While I'm in favor of the experiment, I'm not quite. It would be. I would find it hard pressed to invest in the founder unless it was like a personal relationship, because I don't know that you're going to have any edge over the many other private equity firms that will all presumably follow the same strategy. Right?

Yeah.

There's classic examples of software PE guys coming in and there's 100 employees at the company, firing 90 of them. Right. And so firing all of them. I don't know how, how much better. I don't know if you're gonna.

Positioning in brand.

Yeah. And. And it's tough as a buyer historically. Yeah, it's very good way to get attention. But what happens if this post goes viral? A bunch of talented people say, hey, I want to work on this. Sounds fun. And you're like, well, sorry, we want to be a one person, $1 billion company. We'll hire you as a contractor or something like that. Just can't have you on the W2. And so why. And so setting these artificial constraints is cool as an experience experiment to say, let's see what's possible, what's going to come out of this and maybe we'll force us to develop agentic software that ends up being better. But I thought it was interesting. The Gumroad founder now has something called Anti Work where you can come just build software to replace humans at Gumroad.

Oh, okay.

I think Anti Work is owned by Gumroad. And so there's other experiments that are going at this. But the main thing is this guy isn't just computer competing.

Yeah.

He's competing with every private equity firm that buys software companies. And they don't care that you just have one person on your team. If they have a deal that they want to win, they'll staff it with five people. They'll tell the founder, hey, we're going to let go of a lot of people because we want to increase earnings, but we're not going to let go of your, you know, we're not going to let go of you or whatever, you know, whatever. And so it really comes down to paying a good price price, winning the deal, knowing that the deal is going to happen. And I don't see business owners being too thrilled about the thesis. Right.

Yeah, exactly.

You're going to fire my whole team.

It's like, no, more like even if you sell the private equity, if you built up a business, you're like, well, yeah, like you're probably going to trim some people here and there, wind down some relationships over time. Sure, the company might be smaller in a few years, but the idea of just like, oh, day one, the goal is just like zero employees. Like, screw everyone. It's like even More ruthless. And like, I mean when you go to the website of a private equity company, they don't put front and center because sometimes they don't like, sometimes they actually grow the staff because that's what's best for the business. And they're really like, they're, they're in some ways more ruthless or maybe less ruthless, but they're really just saying like we, we actually just care about shareholder value, value. And so that might mean more jobs, that might mean pay raises, that might mean pay cuts. It's like it's very dependent on what is.

And what this turns into is one person being the CEO because you still need a chief executive, probably officer, presumably for some amount of time anyways. I'm excited to see it play out covered on the show. I want them to be successful.

Yeah.

And let's see what happens.

Yep. Let's go through some other stuff. Bill halligan says the CEOs of the fastest growing startups are doing the following 100% in the office, two very, very lean headcount, three mostly homegrown exec teams, four shooting for $1 trillion, unicorns are dead and five 80 hour work weeks. You love, love it. You put it in there because you're excited.

Yeah.

Shooting for a trillion. We've been hearing about this. A couple, a couple companies have been talking about a trillion. I mean there's more trillion dollar firms than ever now.

So you're only a few 10x's away from being a trillion dollar company. That's true, that's true. I think this is accurate.

I mean also, like, it's just hard to underwrite.

And this is the dynamic right now.

Outcome.

Yeah, this, this is the dynamic. So if everybody has access to these AI tools that make everybody more efficient, it actually just means that you should work even more to just increase the returns on that leverage. So.

But you know, if you're saying 100% in office, don't let me catch you on a zoom podcast, Brian, because I want to see you doing it in person.

That's right.

Ray says Cursor is down. I'm writing organic farm raised non GMO code by hand the way our ancestors used to do. 2k likes. A lot of people using Cursor. They're excited. I saw people got Cursor working with Deepseek under the hood really bad. That's the most inorganic, the most farm raised code you could possibly coding with.

Some of these AI tools like Devin and Cursor is like retired people in Vegas on the slot machines or you're just like, yep, tab, tab, tab, tab, tab, tab.

But you got to get the code written one way or another. Let's do another promoted post and we'll close it out with a bucket poll.

Promoted post from our friend Sean Frank. Not his first time on the podcast, certainly not his last. He goes that's a lot of meta spend. He was sent a plaque Ridge for being in the $100 million club on Ridge. That that means they have spent a hundred million dollars with the greatest financial institution through their ramp cards. Through their. No. Well they have a bill. They have payments per too. So not. Not necessarily all through their cards. Although it went through. But it. But. But yeah. And and just goes to show, I mean the Ridge is a perfect example of a ramp customer. Very large, lots of. Lots of net income and run extremely efficient, efficiently.

And I love that it's a deal toy. We need to bring back deal toys.

We've been trying to bring back deal toys. We're working on some of our own. Yep. So if you're a company or a firm and it's put it on your.

List, it's gonna go on the desk. It's gonna be a constant reminder of like yeah, ramp cares about me as as a. As a business counterparty, which is great.

I think this is in the back of a truck bed too.

Oh really? It's cool.

Very cool.

Cool. Well, let's close out with this bucket pull from Blue blue blue w mist blue says you are getting lapped by people 50% dumber than you because they don't overthink stuff. Aim fire. Correct Later.

I love it. Some good ties into ties into your your strategy recently. Golden retriever.

Golden retriever. Maxing got just be friendly. Hot and dumb Intelligence is going to.

Be too cheap to me but high energy.

High energy. High energy. Just do stuff. Just get out there. Just bark plays bark.

Let your dogs bark.

Grab that Frisbee out of the air. Grab that capital that half a trillion.

Go get it. Half a trillion. Go for 1 trillion. Bark bark.

Let's go.

We'll see you watching.

Everyone please go in the five star review and leave an ad for your company or friend's company in the review. Comments. We love these five star reviews. We'll print it out, we'll read it on the show. It's a free ad read.

Also big Monday. As of now we're planning to do our first ever live show happening from a around call it 11 to 2pm Pacific.

Sounds great.

You're not going to want to miss it. We're very excited to go live and have a good weekend.

Thanks, everyone. Bye.