Google I/O Reaction, Microsoft Build Recap, OpenAI's Stargate Datacenter | Austen Allred, Jeff Morris Jr., Cliff Weitzman, Logan Kilpatrick & Tulsee Doshi

You're watching TVPN. Today is Tuesday, May 20, 2025. We are live from the temple of.

Technology, the capital of capital.

And we are recapping Microsoft Build, which was yesterday, and Google I O which is today. We have a bunch of folks coming on from Google but we're going through what happened at Microsoft. Obviously they're going even deeper into artificial intelligence. Talking about the agentic web, the future where the robots do your work for you and work while you sleep. Google Ads.

This is why we do three hours a day, folks.

Yep. So we can cover it.

You're gonna have time to watch, listen to the full thing every single day.

There's a crazy amount of news and deep dives this week specifically into AI. Obviously Microsoft and Google are doing big keynotes right now, but Bloomberg has a fantastic deep dive into OpenAI's Stargate. And then there's also a profile by the same journalist in Bloomberg about Dario Amade over at Anthropic telling his soul story. And so basically everyone's getting coverage and AI is the only thing people care about. And so that's what we're talking about today on the show. But let's give you a little bit of a recap of the news. So Microsoft Build 2025 open Satya Nadella calling for an open agentic web. There's new copilot features, open protocols. They're going all in on MCP model context protocol, that standard, not a particular tech product obviously defined by Anthropic but then open sourced and then adopted by both Microsoft and Google. So the narrative is MCP1 basically and MCP became a very hot button issue. Lots of people posting threads on it.

It's goaded yet, but it's definitely in.

The conversation it seems, it seems like it could be goaded. It's doing quite well. It's getting massive adoption and most important importantly I think is like it's not, it's more of a standard than a particular product. And so you're not seeing a wave of startups, but you're seeing startups use mcp. And then Elon Musk joined Satya Nadella on stage via Zoom presumably or probably Microsoft Teams, I guess they called in from and confirmed that azure will host Xai's Grok 3 and Grok 3 mini. And he kind of talked about his vision for GROK being a truth seeking LLM and focused on foundation, a foundation model built on physics and first principles which has been his, his goal with that. But if you look at Satya Nadella's pitch now for the host. The models that are hosted on Azure, it's really just everything you can. They even have. So they have. You can swap quickly between even deepseek is on there and there's a version of Deep SEQ that's been fine tuned by Microsoft AI internal team. There's all the Llama models, There's all the OpenAI models. Now the Grok models.

Maybe they found him, they asked him to leave.

Now they're hosting him. We're good. Yeah. And then there's even. There's so many models now that Azure has a model router for OpenAI models. So you can import the Azure OpenAI model router and then it will decide what the cost benefit is for the individual model that you're using. Oh, do you have a. Do you have a plug or. I guess we're. Oh, no, I got this one. Cool. Okay. And so yeah, you should go watch the full conversation between Microsoft CEO Satya Nadella and Elon Musk. He says with Grok 3.5, which is about to be released, it's trying to reason from first principles. Also, I saw a ton of Tesla battery packs showing bad news for product.

Managers who have been spending their entire life figuring out how to think from first principles. That Grok 5 will just do it out of the box.

Yeah, it's a golden age for people that never learned first principles. You can just say, think from first principles. Do it for me.

Yeah, just do it.

Golden retriever mode all over again. Bold.

We've been calling it bold prediction at the time.

Dan Shipper summed it up. Yesterday, Microsoft rolls out MCP to support to Windows. Today, Google officially supports MCP and the Gemini SDK. It's over. MCP wins and that makes a lot of sen. It's a standard that everyone's building against now and we imagine that that will continue because once you have Google and Microsoft and Anthropic, everyone has to conform to it.

Coordinating with Dan but hoping to get him on the show Thursday. He's got some news to share.

Another big announcement from Google IE. They have real time Google Meet Translation is live now. I feel like we've seen this demo 10 years in a row, but it feels like it's finally here. And so there's a video where you can be talking to somebody in a different language and it will live translate what you're saying into the local language so you can just have a conversation. Obviously this is a. It's like one of the most obvious AI use cases. It's something that Google should do. It's something that seems like extremely beneficial for humanity across the board, just like pure alpha. So excited to see that they're also releasing a diffusion language model that I want to dig in with the Google team more on because we've seen image models move away from diffusion into more token based structures and transformer based architectures. Now language models are going into diffusion. So I don't know enough about this, but I'm going to be digging in this week. They also launched DeepThink 2.5 Pro, new enhanced reasoning mode for using Google's research in parallel thinking techniques, meaning it explores multiple hypothes before responding, basically branching all the different trees. There's already a little bit of this going on in OpenAI's deep research product, but obviously it's a very competitive space. This enables it to handle incredibly complex math and coding problems more effectively. And I was excited about Google Gemini. I've become more bullish on the product since I sent you those incredible AI generated videos that are generated from VO2. And so I went in to, to Gemini because remember I generated three or four of those videos and then it gave me a timeout and it said look, I know that you've spent millions of dollars with Google over your career, but $20 is the most money I could accept from you. I couldn't possibly charge you an extra dollar to make one more of these videos. 200 just doesn't feel like the GPUs are on fire.

Even though you would pay 2,000 and.

We have, even though we have $100 billion of cash, we have to rate limit you. Yeah, we couldn't possibly light another GPU on fire for you right now. We should ask no price.

We should ask Logan later around rate limiting and the plans there.

Hopefully it was very frustrating experience for me because I would have paid more in that moment to generate one more photo of a Lamborghini URUS with a TVPM livery driving around a track. But I couldn't do it and it was, it was not possible. I mean I guess I probably could have if I got a separate Google account with a separate Gemini Pro subscription, maybe on a separate phone and then how to do that.

Yeah, huge opportunity right now to create a, a basically a bundled Google subscription where you subscribe to a service that then aggregates like a thousand, you know.

Yeah, yeah, yeah, yeah. It's kind of like how there's how the banks have FDIC insurance and you're insured for up to like $100,000, then $250,000. So some people would take, you know, oh, spread. Spread a million dollars across 10 accounts or 5, 4 accounts. You need to do that with Google so you don't hit the rate limits, clearly. So they put me on a timeout. They said, john, you have generated too many of these. Of these GT3 RSS liveries. You need to take a break. We'll see you one day from now. Okay? One day.

Come back, come back.

It's been three days now since I did that.

And you should be okay with that because that's kind of a typical workflow with a human, right? You maybe get a few iterations and they say, hey, I got to go to sleep.

I got to take a break.

See you tomorrow.

Yeah, you're overworking me, boss.

Yeah.

And so it told me to take a break. It said, one day.

That is the agentic web.

Monday night. I remember what it said. It said, Monday night you can create more. So I wanted to go and create one today because it's Tuesday. And I said, generate a vertical video 16 by 9 of Cinder Pichai arriving at Google IO in a Waymo and waving to a crowd of adoring fans who are taking pictures with Android phones and says, too many requests in a short time period. Try again later. So again, a weird thing where like, the product is amazing, but I can't access it enough. And I keep being frustrated with the actual rollout of these products, but hopefully it's getting better. I do wonder if they're getting slammed right now because of Google I O. Everyone's testing out the new model and so they really are under a ton of demand. What was interesting is that it didn't generate any videos for me today. And immediately it just said. It just said too many requests in a short period of time. It didn't say if that. If it's too many requests from me because this is the first request I've sent in today. And so. So I was very unsure of what's going on. There is just their systems under load.

Can you prompt it to address you as boss man? Because it might be more. It would feel a little bit better if it said, hey, boss man, that's enough for today. Instead of just hitting I'm working hard, feeling like you're hitting those rate limits.

I'm working hard.

Going back a little bit. One thing that stood out to me, seeing Elon and Satya having a conversation working on this product launch felt significant to me because it was not too long ago that Microsoft was named as a defendant in the OpenAI lawsuit. Right. Elon's lawsuit against OpenAI Microsoft was a defendant. And at a certain point, I believe it was in 2023 as well, Elon was threatening to sue Microsoft for training on X or Twitter.

Oh, interesting. Interesting.

I didn't know that. And so these two haven't exactly had the best relationship. I mean, clearly, like, they're fine on an individual level, but the two companies, the partnership now is cool to see, just given that there had been quite a bit of friction over the last couple of years. But the funny thing is, Musk was actually way back in the day, an intern for Microsoft windows in the 90s. So very.

Isn't there some story about him trying to get a job at Netscape with Marc Andreessen and he didn't make it past the first round and so he went to Microsoft. That might be wrong, but there's something about Elon. Before he started Zip2, was kind of moving around the valley at different tech companies. It's funny to imagine intern Elon.

Let me look it up.

Writing some code for Microsoft Office at the time or something. But now they're partnering and I mean, it's all part of Satya's strategy to be model agnostic. It's always felt like Microsoft Research has been behind the curve on LLM training like Zuck. Although Llama has kind of hit a rough patch with llama 4, apparently.

So Elon was rejected from a job at Netscape.

Yeah, that's right.

He sent in his resume, showed up at the office, even lingered in the lobby, but no one acknowledged him.

Wow. Just Marc Andreessen walking by, mogged. And now. And now Mark's in xai.

Yeah, now they're boys.

Ripped a big check and a bunch of other companies too, but it's great. What a full circle story. But yeah. So, I mean, Satya obviously is positioning Microsoft to be model agnostic and wants to be a vendor of this in the cloud platform. And I think that's. It seems very good for Azure in the idea that if you're a business and you're building on. On Google, you might feel like, oh, they're going to really push the Gemini models on me. Maybe I get locked in at some point, maybe that gets expensive. But if I'm on. If I'm on Azure, I can dance around between different OpenAI models and different Llama models and different Grok models and. And have a lot more flexibility because Satya has kind of said like, hey, we're not, you know, gonna Push you towards a particular model or particular.

They also release an open source VS code, you know, sort of like Copilot product, which is big. So anyways, reaction to that was. Project Padawan was positive as well.

GitHub Copilot graduates to a full coding agent. Project Padawan GA for, for Copilot Enterprise and ProPlus capable for autonomous feature work, bug fixes and refactors. And so everyone's kind of moving into these autonomous coding agents. We're having Scott Wu from Cognition on the show on Wednesday tomorrow to talk about the landscape and we're also having Lee Marie from Kleiner Perkins on Friday to talk about the same thing. This idea that how is the AI coding landscape shaping up? Is it 1, Market, 2, 3, 4. It feels like there's. As we were talking about it a little bit earlier, it felt like there are, there's AI code that's written. I mean you just gave an example where you asked it to count. You asked ChatGPT to count the number of.

Yes. This is funny or something.

What happened there?

So basically Xai got received 168 Tesla megapacks, which is going to power Colossus 2, which is their second data center.

And.

And there's this sort of above ground image. And I initially saw the image and somebody else had sort of counted them up. So I asked ChatGPT to count and it spent 13 minutes attacking the problem by writing code. In that time, of course, I just looked at the X axis and the Y axis and multiplied and got the right number. But ChatGPT was running in the background, you know, attacking it in the most sophisticated way possible. Ultimately what the number I wanted to figure out that was interesting is that the retail value of the megapacks, which of course you can buy on the Tesla website, $5 million comes out to around $850 million.

Wow.

Just on megapacks, which is about half of their 2024 revenue. Right.

Wait, half of who's not X?

AIs, X's revenue, which is now combined.

Yeah, yeah, yeah, yeah. XAI doesn't have any.

X has graduated.

Yeah. But they raise a lot of money so they're spending on Tesla megapacks. That's wild.

Yeah. And I'm interested to see how ultimately exactly how this new round shakes out. There's been rumors around the new fundraise for a while.

The final details of that round haven't been announced yet.

Right.

Yeah.

It's still just rumors. That's interesting. I wonder. Thirteen minutes is fast, but not that fast. I wonder if there's a world where OpenAI has a function it can call to like Mechanical Turk or scale AI to just have put a human on the case.

Well, it failed.

It failed. So it didn't even do it in 30 minutes.

It just, it just basically timed out because I could see it working and just writing code, taking all these different cracks.

Yeah. So.

So imagine to its credit, it would.

Have taken a human not 13 minutes to count those.

No, just manually counting them. Like what I did.

Yeah.

No, I didn't actually go line by line, but like it took me whatever, 20 seconds.

20 seconds.

But if, if a human had to run, take that many attacks at solving it.

Yeah, they wouldn't. Like. What I'm saying is that, is that you kick off this. How many whites or how many white squares are in this image and it just displays that to a mechanical Turk who's just sitting there doing random tasks all day long and they just one shot it in, you know, two minutes.

Yeah. This is what I was talking about over the weekend. I was joking around. I'm sure some people took it seriously. But I was saying I met somebody in the Philippines whose job is to make investment decisions for a firm that claims to use AI. That's the final, the final boss of, of, of using AI. Yeah, but, but it's really, you know, Mechanical Turk is just, you know, using.

Yeah.

Using outsourced labor to just make all investments.

Yeah. The overall announcement at any of these keynotes is always we're 90% of the way there. In the last 10% will be another 90% of the way there. But I mean there are some, there are some impressive, you know, announcements and, and steps forward, but obviously there's a lot, there's a lot left to do. So Copilot Studio just running through. What else Microsoft announced Copilot Studio now supports multi agent so developers can build agents that delegate tasks to one another. With Microsoft 365 agent builder NL Web debuts as HTML for the agentic web letting any size, any site expose a conversational endpoint discoverable by AI agents. So an AI agent shows up and it can just chat with your website. We talked about project Padawan GitHub Copilot graduating to compete in that space with Codex from OpenAI. And now just in the OpenAI world, you have three products that effectively write code. You can just go into O3 or even 4.0 and it can write a little bit of Python like you experienced, which can have mixed results but is sometimes extremely useful. You have Codex that can actually plug into a GitHub repo and go and write some code and fix bugs and whatnot. And then you have Windsurf which can sit in your IDE alongside and be kind of that ground up adoption. And so there's one mental model where there's like the consumers are going to ask questions and they don't even know that they should be thinking about writing code. Code will be written. They might never see that code if they don't unfold the reasoning tokens. On the flip side, you might have someone who's using codecs to change bugs and make small changes in a GitHub repo without really opening up an IDE constantly. But they're just interacting with code, but they're aware of what should be done, what should be built. Then you have a product like Windsurf where for basically a full time programmer, they're working in code constantly and they're solving problems at a much higher level than any of the agents can actually do. But they are enhanced and sped up by these AI IDEs that are speeding up their workflows. And then potentially you have this top down enterprise AI which is what Scott's building with cognition, where there isn't as much of a groundswell around Devin as a consumer tool. But as I understand it, Devon is something that's been pushed top down on big corporations as something that is, that needs to be rolled out and it has much more of a like an enterprise sales like go to market and sales cycle. So interesting to see if that, if that model holds of like four distinct markets like consumer prosumer, ground up enterprise, top down enterprise. I don't know if that's the right way to think about it, but I think that's something we'll be digging into over the, over the next couple days.

Yeah, I mean ultimately some combination of product LED and enterprise is probably the winning combo long term.

They all seem to be getting adoption and printing money. The question is just yeah, churn and how the market looks in the really long term. It is interesting because like AI coding tools, it does feel like it's a massive market. You looked at the numbers for how much money spent on software engineering. It's huge. And then it's also just a completely new market in the sense that there isn't anything that's like really established, like they're not really displacing spend on anything else. It's all just incremental and additive. So it's kind of hard to, hard to handicap exactly what you're going after.

I thought this post from Caleb Harris was an interesting call out. He says. Last night I dropped 12 backlog bugs from Linear into Codex and launched Greptile. Found some nits in the reviews but they were almost entirely correct. And then Spur QA tested almost entirely human out of the loop at this point.

That's awesome.

And yeah, well good time to tell you about Linear.

Linear is a purpose built tool for planning and building products. Meet the system for modern software development, streamline issues, projects and product roadmaps.

Anyway, thank you John.

They also launched Copilot Tuning which allows companies to fine tune their models with their own data to create domain specific agents. Windows AI Foundry launches as a local platform on Windows and Mac for training and fine tuning and running LLMs with foundry local Azure AI foundry models adds Grok 3 which we talked about. MCP gets first party support, Windows subsystem for Linux, a bunch of other stuff SQL Server 2025 hits Public Preview let's go SQL Server. Huge for the database engineers in the crow.

Let's give it up, let's give it up.

Anyway, I think the big definitive analysis of the post game on Microsoft build came from Ben Thompson. But Derek Thompson. No relation I believe, but maybe Thompson boys brothers.

That would be wild.

He wrote Abundance with Ezra Klein, he says Fascinating. Ben Thompson's vision of the future of the Internet the I. E. The agentic web scenario is Before I go to sleep I tell ChatGPT plan my 5 year old's birthday next Saturday. Budget $500. When you've made the reservation, email these two these 20 people a printable invitation to attend. Also my wife wants to go to England in mid July, find five plausible flights for the family and make several distinct itineraries. Finally, please edit this work memo When I go to sleep, the AI agent negotiates slots with two bowling alleys, buys a cake, emails, printable invites, plans the trip copy edits, etc. This presents an interesting economic challenge. What happens ad revenue when more traffic is just AI? Ads make little sense when the reader is a robot. Well, maybe instead sites ask agents to make tiny payments that are fractions of a cent every time they call up an article. Ben suggests the use of no fee stablecoins blockchain dollars, but I think microtransactions could work with old fashioned dollars. I have no idea if this cashes out, but very interesting vision of how AI would necessarily transform Internet economies as traffic shifts to from humans to agents. Uh, interesting. So I want to I want to dig through this strategic article, but I.

Also, and this is significant because so every business, every consumer tech company eventually becomes an advertising business, from Instacart to Uber to Apple.

Yep.

And if you remove these ad businesses, it typically will just rip out a lot of the actual earnings. Right. Because advertising tends to be very high margin. And so what I think we're going to see is ultimately this kind of drawn out war between agentic software and the old Internet, where the old Internet doesn't want to let go of advertising. And so I think that microtrack transactions are potentially somewhere that you could land.

But it's interesting to dig through. I'm not entirely sure that language models and AI agents will be entirely resistant to advertising, because I imagine.

No, but it's more so if you're Uber.

Yep.

And you have a great ad ads business.

Yep.

Are you thrilled to have an agent trawling and booking things and buying groceries and all this stuff? You're not. Right. You're like, I would prefer unless.

Unless the ads that you run on your platform influence the future training of those AI models. So, for example, if an LLM is training on transcript of this show right now. And then we start talking about how Ramp can save time and money.

Well, it does, John.

Yeah, yeah, exactly. We're stating the facts here.

Yeah.

And we're explaining that Ramp has easy to use corporate cards, bill payments, accounting, and a whole lot more all in one place.

Not even to mention ramp travel.

We didn't even get to mansion ramp travel, which makes travel so seamless and streamline.

We use it every time we hit the road.

We do. And so an LLM, as it's trolling through the Internet, it comes across a host red ad like that that influences the future recommendations it makes. Right. It gets baked in the training data. So perhaps ads could be just as influential in the sense that LLMs are simulating human behavior. Everyone likes to say that advertising doesn't work on them, but time and time again, data shows that it does. And so if you have an agent that's actually perfectly simulating a human, it should also be simulating receptivity to advertising. You would think that it would.

That's right.

But. But let's dig into Ben's article because I think I agree and I think you agree more than we're letting on right now. But let's dig into it. So he starts with this, with this 2014 article in the Atlantic by Ethan Zuckerman about, like, the death of the web. I remember this, this iconic article he says, from Ethan Zuckerman I have come to believe that advertising is the original sin of the web, the fallen state of our Internet as a direct.

Nobody's taking this seriously.

Yeah, I forget what Jordy has. Jordy has a laugh track now. It's called the Internet's original sin. And keep going, John. You're gonna make this very hard. The fallen state of our Internet is a direct, if unintentional consequence of choosing advertising as the default model to support online content and services. Through successive rounds of innovation and investor story time, we've trained Internet users to expect that everything they say and do online will be aggregated into profiles which they cannot review, challenge, or change that shape, both what. Both what ads and what content they see. Outrage over experimental manipulation of these profiles by social networks and dating companies has led to heated debates amongst the technology savvy, the technologically savvy, but hasn't shrunk the user bases of these services as users now accept that this sort of manipulation is an integral part of the online experience. Mark Andreessen hit the bomb. Bomb. Marc Andreessen. Marc Andreessen, who was there when the web was born, explained in a 2019 podcast why this sin was committed. The quote is edited lightly for clarity. One would think the most obvious thing to do would be building in the browser the ability to actually spend money. Right. You'll notice that didn't happen. And in a lot of ways we didn't even think it's unusual that it didn't happen because maybe that shouldn't have happened. I think the original sin was we couldn't actually build economics, which is to say money, into the core of the Internet and there. And so therefore advertising became the primary business model. We tried very hard to build payments into the browser. It was not possible. We made a huge mistake. We tried to work with the banks and we tried to work with the credit card companies. It was sort of this classic single point of failure bottleneck, or at least in this case, two points of failure. Visa and MasterCard essentially had a duopoly at the time. And so they were just literally, if they didn't want you to be in the switch, they didn't want you to be able to do transactions. You simply weren't going to do it. And what's interesting is like, I feel like the modern entrepreneur, like the Sam Altman style deal maker, if you transported the back in time, you put the you can just do things mindset in there, you'd just be like, oh yeah, like the whoever's building the browser did some crazy deal with Visa and is bundling these transactions. And so yeah, they're totally waiving the 3% fee, but they're aggregating this way and they're also investing, it's interesting watches the other.

It's interesting that the credit card companies at the time too, you can imagine I would pay an incremental fee to have my credit card data perfectly stored in my browser. Like you could capture bips on that pretty easily.

Totally, totally.

Like so you think there would have been a deal to be done? Yeah, but remember this was at a time when people were scared of the Internet. People generally weren't even fully convicted at the time that it was. It still had non believers.

I wonder if it was more of a failure of imagination from Visa and MasterCard executives or a failure of aggression in the deal making of the Web 1.0 entrepreneurs. Because the web, whatever generation we're in.

Right now, you think you're basically saying entrepreneurs today are built different.

It's not just that they're built different, is that they're very willing to go and do crazy deals with huge companies.

Yeah.

Like I mean think about one thing that's real.

One thing, that one thing that's real. Back then there was more, there was more almost technical risk to something, to something like this.

No, it's simple now, didn't happen. But yeah, yeah, I mean a lot of the maturity today there's people take startups seriously. But I'm thinking of that. What is that Anchor Jane company where you can pay your rent with credit card and Wells Fargo like totally took a bath on the deal, at least in the short term. You remember this, it was some sort.

Of credit card built.

Built. Yeah, built rewards. And so like doing a deal with Wells Fargo at that scale as like a startup is crazy but like Ankur is clearly just a really, really great like dealmaker and so he was able to get it done. And if, if points on housing actually does plant like play out to be a real thing in a big business. Like it will be on the back of that crazy deal that needed to get done to make this happen. Like it wasn't just going to be some ground up small company deal. Like you had to go in and do this, this big thing.

Yeah. Interesting. In some ways, Apple Pay enables what would have been the goal at this time, but it does it effectively at the hardware layer.

Yeah, yeah, yeah, yeah. It's not even that good.

I think it's pretty. It's getting to the point where it's pretty good.

Yeah, yeah, it's pretty good. But I mean still like you, you hit web checkout, I basically time and have to hit the auto fill, which is like such a kludge because it could just not even be there.

Yeah.

It could just be as native as you click a link or you know, even just like the integration of email or the integration of the share sheet. Like there's so many different aspects that are built into the web browser natively. Like the ability to play video on the web. Like that was something that all the browsers needed to figure out. And like it just works now. Like every browser supports video. Mark Andreessen was talking about all the different things like, like why the hyperlinks are blue. Like he's the one that decided that because he was like, we need hyperlinks and we should make him blue. And it's like there could have been a version of money. So anyway, Ben Thompson goes on to say, I think Andreessen is too hard on himself and I think Zuckerman is too harsh. On the model Andreessen created. On the model Andreessen created. The conditions for the original web was the human web. And advertising was one of the best possible ways to monetize the scarce resource in digital human attention. The incentives all align. Users get access to vastly. To a vastly larger amount of content and services because they're free, which I really like. It's great that anyone can access the Internet for free, basically free. And they can go all over. And most of these services are free. The information spreads very quickly. Content makers get to reach the largest.

Information doesn't spread quickly. There's a paywall.

Yeah.

But there's the black market of PDF as well that I think it's Paula has talked about. Yeah, yeah, yeah, definitely in sv.

Yeah, yeah. Get to reach the largest possible audience because the access is free and advertisers have the opportunity to find customers they would never have been able to reach otherwise. Yes, there are downsides to advertising Zuckerman fretted about, but everything is a trade off. And the particular set of trade offs that led to advertising, the advertising centric web were on balance a win, win, win that generated an astronomical amount of economic value. Yeah. It's hard to go back and play, play out the Google story and say like, oh, they left a lot on the table.

Yeah. In so many ways. Content. Content wants to be free. Right. The second that you put web pages or anything behind a paywall.

Yeah.

The attention they get drops by 99% usually more. Right.

And yeah, it's just. It's just way better to be advertising led. It's way better to just, you know, as you're talking about one story, just start talking about figma. You know, think bigger, design build faster. Figma helps design and development teams build great products together. You can get started for free anyway.

It is the backbone. It's the best of our show and we can't thank them enough.

Okay? Moreover, I disagree with Andreessen that we could have ended up with a better system if the banks and credit card companies had been willing to play ball. In fact, over the last 30 years, the credit card companies have in particular have in part thanks to companies like Stripe, gotten their digital acts together and are integral to a huge amount of web based commerce, which itself is driven through digital advertising. The largest category of advertising for both Google and Meta. That too is human in the biggest in that the biggest outcome of digital advertising is physical products and real world experiences like travel. Digital products like apps and games, meanwhile, are themselves pursuing human attention. What was not viable in the 1990s, nor at any time since then, was something like microtransactions for content. One obvious problem is that the fee structure of credit cards don't allow for very small transactions. Another problem is that costs to to produce content are front loaded and the potential payoff is both backloaded and unpredictable, making it impossible to make a living. The biggest problem of all, however, is that microtransactions are anti human. Forcing a potential content consumer to continually decide on whether or not to pay for a piece of content is alienating, particularly when plenty of alternatives for their scarce attention exist.

I really believe that a large amount of the substack economy is effectively crowdfunding somebody's ability to just nerd out on a specific set of topics. I agree, and I think it's beautiful.

I agree. Yeah, it does seem like if there's.

Yeah, sorry, yeah, I was just going to say, but a lot of the content creators that are paywalled on substack would be bigger and more influential if it was just completely free, right?

Yeah. Yeah. There's a big question in my mind of like, where does investigative journalism go in the era of going direct in modern tech media? Like, where is tech? Seymour Hirsch, like, you need someone who can just be like, Seymour Hersh was just on the payroll of big newspapers for a long time and then could go and like hang out at a bar and talk to former soldiers about what they did in Vietnam and then figure out that the My Lai massacre happened and like uncover this like massive scandal and we See this in Silicon Valley where there are companies that are like, you're not going to be able to put a bunch of sponsored content on a takedown of some company that's defrauding their investors. The, the win case is the John Carrey Roo story with Theranos where he was able to be on the Wall Street Journal payroll for a long time. Hunt that story down for a long time, do a lot of sourcing and a lot of confirmation because it's extremely risky to write something negative about a $10 billion company that has incredible resources. I mean they hybrid, you kind of.

Need the heat shield.

Totally.

A big organization.

Totally. Yeah. And also just you need, you need to be able to pay your mortgage while you're doing the research and not publishing and just doing nothing basically. And, and being 100% focused on, on chasing down a really big story. And then when you do it, there's a huge outcome because he of course sold a book and then the rights to the documentary and the movie and the TV show and all this stuff. And I don't even know if he's, I don't know what he's up to now, but it seems like should reach out really well. It'd be great to get him on the show here if he's getting back in the game. Digging into some of the. Maybe Theranos 2.0 is coming right now.

He's like, I'm back.

He's like, I'm ready to start talking to the employees of this new blood testing company.

It would have been huge missed opportunity to not just name the company. Theranos 2.0 Inc. Yeah, that'd be great.

I mean, do we know what the name is yet? I don't know.

I missed it.

But we'll have to look it up anyway. Subscriptions do work at smaller scale, says Ben Thompson, because they are ultimately not about paying for content, but giving money to another human, which you mentioned or human institution. From the local news business model which Ben wrote, it is very important to clearly define what a subscription means. First, it's not a donation. It is asking a customer to pay for money for a product. What then is the product? It is not, in fact any one article a point that is missed by the misguided focus on microtransactions. Rather, a subscriber is paying for the regular, regular delivery of well defined value. Each of those words is meaning is meaningful. Paying a subscription is an ongoing commitment to the production of content. Not a one off payment for a one off one piece of content that catches the eye. Regular delivery. A subscriber does not need to depend on the random discovery of content. Said content can be delivered to the subscriber directly, whether it be email, a bookmark or an app. Well defined value A subscriber needs to know what they are paying for the last point is the crux of why many ad based newspapers will find it all but impossible to switch to a real subscription business model when asking people to pay. Quality matters far more than quality and the ratio matters. A publication with one valuable article a day about a well defined topic will easily earn More subscriptions than one with 3 valuable articles, but 20 worthless ones covering a variety of subjects. Yet all too many local newspapers built for ad based business model that calls for daily content to wrap ads around spend their limited resources churning out daily feedback filler even if those ads no longer exist. Ben says, I expect that this model will endure in the age of AI. Obviously I'm biased at this point, but in a world of infinite content on demand, common content becomes community. And if I'm successful, this essay will generate a lot of discussion amongst a lot of people precisely because it is both original and widely accessible, funded by an audience that wants me to keep writing articles exactly like this. So of course Ben writes free articles and then also private ad supported articles and goes back and forth. So he gets a little bit of the reach.

So he goes in here the death of the ad supported web the ad supported web, particularly text based sites, is going to fare considerably worse. In fact, the most substantive pushback to my defensive advertising was in my own excerpt. Most ad supported content is already terrible thanks to the bad incentives both Zuckerman and injury and bemoaned and the impossible economics enabled by zero marginal cost content generation and consumption. Consumption Google in its most idealized form aggregated content consumers Aggregated content consumers by mastering discovery in this world of abundance, directing users to exactly the site they were looking for, which was monetized through ads that were sold and served by Google. Indeed, this is the great irony in the ads antitrust case in which Google is currently embroiled Suert asked on Mobile.

Dev Memo I've heard arguments that because Google suppressed competition in open web advertising markets, those markets should flourish when Google's monopoly is broken. But my sense is that this ignores two realities. First, that consumer engagement has shifted into apps and walled gardens. Irreversibly, of course, like most people get their news from X or Instagram or Facebook.

Yeah, and this was the logic for X to say yeah, we don't Want you posting links, we're fine for you to post an excerpt or a story.

Just build those walls a little bit.

Higher because we want you to click into the content and see another ad. And then click out, go see something else, see another ad, and just do that forever.

And second that Google was keeping the open web on life support. And the open web's demise will be hastened when Google no longer has an incentive to support it. What happens to the open web when its biggest, albeit imperfect, benefactor loses the motivation to sustain it? Well, everyone's going in the walled garden. Walled gardens, like social networks, are both more attractive to most users and also better for advertisers. Google might soon lose what little motivation they had left to support the open web. However, that's not Google's and the web's only problem. Why go through the hassle of typing a search term and choosing the best link, particularly as search results are polluted by an increasingly overwhelming amount of SEO spam now augmented by generative AI, when ChatGPT or Google itself will simply just give you the answer that you're looking for.

In short, every leg of the stool that supported the open web is at best wobbly. Users are less likely to go to ad support supported content based websites, even as the long tail of advertisers might soon lose their conduit to place ads on those websites, leaving said websites even less viable than they are today. And they're barely hanging on as it is.

I mean, can you think of a site?

The classic example of this is like historically trying to find an ingredient, like a recipe, right? You're like, I want to make oatmeal.

Yeah.

And the site like buries the only information you actually need, which is like how much, how many cups of oatmeal to water the ratio or whatever.

Ridiculous.

And you have to see, you have to scroll by, you know, two pop ups and six ads to get to that little nugget of information. And now it can just be immediately surfaced by an LLM.

I'm trying to think about the last time I visited like a truly ad supported website. Like I'm scrolling through here and it's just, it's just nothing in my tabs. Like everything is either a, like a, a marketplace that monetizes by transaction fees or a paywalled website for news or just Google search generally finding stuff. But like there are very, very few websites that I visit today on a regular basis that are purely ad supported. Like I do go to the Wall Street Journal or the Economist or SHRUTAKERI or any of these, but they are, none of them are purely ad supported. The Wall Street Journal does have ads, but they also have a big paywall in an actual subscription fee. So it's interesting.

Yeah.

The open web, I mean, is it, is it dying or is it dead? Like it seems like, like. No, like, like you know, the, the, the Vice News and the buzzfeed, like that boom and bust has already happened.

Yeah.

So I don't know. Anyway, he moves on.

I mean, what do you think? It'd be interesting to try to honestly run a deep research report on trying to understand traffic to these websites that are effectively dying but not dead.

Right.

I wonder what's happening if you have.

A website that's, I would imagine like how many websites have recipes for, you know, cooking various things that still get people trickling in and they're technically still monetizing.

Maybe web apps.

Yeah. The question is like at what point do you, does, does it cost more to even host the website than it's generating traffic? Right.

Yeah. And there's been a couple of those.

Because if you're making whatever, if you're making $1,000 a month in ads and it costs you a couple hundred bucks to maintain the website, in theory somebody would just keep that running indefinitely.

Yeah. I mean TechCrunch was obviously famously ad supported. They were bought by Regent, which is a private equity roll up of brands. Let's look at their portfolio. It's all very hyper specific stuff. They own TechCrunch now, but they also own Computer World, Network World, PC World, Mac World, Tech Hive, they own Cheddar News, they own Defense News, Air Force Times, Navy Times, Army Times. So I imagine that Navy Times is probably ad supported. I mean I'm just clicking around on the website and there's not a lot of like paywalls I'm hitting. But they are running and I'm sure that they have other ways to monetize. Whether it's through premium, you know, reports or maybe conferences or something like that. But in general it seems like there aren't that many companies I would struggle to name companies that are generating over a billion dollars in revenue off of purely ad based website. That's not an ad exchange, for example. Yeah, right.

I doubt it exists.

It doesn't really exist. Anyway, so Ben takes all of this, all this like preliminary thesis and weaves into Microsoft and the open agentic web, says this reality is the fly in the ointment of an intriguing set of proposals that Microsoft put forward yesterday at the BUILD 2025 developer conference about the open Agentic Web. And so the quote from the CTO is the thing that is super important if you think about what an open agentic Web could be, is you need agents to be able to take actions on your behalf. And one of the really important things about agents being able to take actions on your behalf is that they have to be plumbed up to the greater world. So you need protocols, things like MCP and A2A, and things that are likely going to be emerging over the coming year that will help connect in an open, reliable, interoperable way the way agents that you are writing and agents that are being used so actively now by hundreds of millions of people to be able to go access content, to access services, to take action on behalf of users in fulfilling the tasks that have been delegated to them. One aspect of this vision of the agentic Web, says Ben Thompson, was Microsoft's commitment to mcp, which we talked about a little bit created by Anthropic. Scott told Neelay Patel in an excellent interview on the Verge, that while MCP wasn't exactly what he would have designed from scratch, ubiquity is more important than semantic differences, particularly when you're trying to create HTTP for AI agents. The second part of Scott's vision was something Microsoft created called NL Web Natural Language Web. And these are natural language interfaces for websites that make them more directly accessible for agents. He says. If you think back to the web, we have HTTP and then we had things that sit on top of HTTP, like HTML mainly, that are opinionated about the payload. So today we're announcing an NL Web. The idea behind NL Web is that it is a way for anyone who has a website or an API already to very easily make their website or API. An agentic application lets you implement and leverage the full power of language, large language models to enrich the services and products that you're already offering. And because NL Web endpoint, the every NL Web endpoint is by default an MCP server, it means that those, those things that people are offering up via NL Web will be accessible to any agent that speaks mcp. So you really can think about it a little bit like HTML for the agentic Web. I always thought this was odd because I thought that the, the layer to go between APIs and HTML for an agent to just transform HTML into something that an agent could understand would be very, very simple. I didn't think that we'd need a new protocol, but it seems like it's getting tons of adoption. So I guess this makes sense. We've done a bunch of work already with our partners who have been really excited to be able to very quickly get to implementation and prototypes using NL Web. They've worked with TripAdvisor, O'Reilly Media, a ton of really great companies that offer important products. So Scott concluded by re emphasizing how important it was that layers. That the layers of the agentic web be open and use the evolution of the Internet as his reason of why.

So he says. So the last thing that I want to say before handing things back over to Satya is just sort of press on these two points about why Open is so important here. You know, it is unbelievable what can happen in the world when simple components and simple protocols that are composable with one another are out there exposed to the full scrutiny and creativity of every developer in the world who wants to participate or who has an idea. This thought game that I play with myself all the time is trying to imagine what the web would have looked like if one of the actors in the early development of the web, say the browser manufacturers. Browser manufacturers is great terminology.

Yeah. Mark Andreessen, a browser manufacturer.

Yeah. He made his money in browser manufacturing, domestic browser manufacturing, Domestic browser manufacturing, American dynamism browsing. So had decided that they wanted to vertically integrate and own the entire web, a hundred percent of the web would have been dictated by the limits of their imagination. And it's just obvious with 30 years of history that that wouldn't have been a very interesting web. The web is interesting because, yeah, think of the iPhone app layer. If Apple wasn't controlling, it would be extremely chaotic, but there would undoubtedly be sort of cool things that would emerge.

Yeah, I'm kind of split on it because, like, there's so much creativity that you can do within a jpeg or an MP4. And so if you get on YouTube, I mean, it requires more production value. But, you know, we, like, we are not constrained by anything other than just pixels here. And we can put tickers and do all sorts of things within the confines of the MP4 box. I guess it is a little bit more narrow than a web UI that anyone can use and click around. So in some ways we are being constrained. But like you just scroll on any social network and you see the amount of creativity that occurs even within the constraints of like, no links. Right?

Yeah. Despite Apple's control of the App Store, you still get the benefits of the web, which he goes on to describe here. The web is interesting because millions, tens of millions, hundreds of millions of people are participating to make it this rich, dynamic thing. That's what we think we need with the gentic web and that's what we're hoping you all can get get inspired to go work on a little bit to riff on and to use the full extent of your imagination to help make this thing interesting.

Yeah, so he says. So back to Ben Thompson and he says, I think the widespread adoption of MCP as a protocol layer and NL web as markup layer sounds excellent. The big hole in Scott's proposal however was pointed out by Nilay Patel in that interview. That's the piece that on the web right now seems the most under threat. The underlying business dynamics of if I start a website, put a bunch of schema on it that allows search engines to read my website and surface content across different distributions. I might add an RSS feed which is a standardized distribution everyone agrees on. There's lots of ways to do this. But if I wait, if I make a website and I open myself up to different surfaces, what will I get in return for that is not necessarily money in almost every case not money. What I'll get is visitors to my website and I'll monetize them however I choose. Selling a subscription, display ads, whatever. That's broken, right? As more and more of the answers appear directly, particularly AI based search products, traffic to website has websites has generally dropped. We see this over and over again. What's going to replace that in the agentic era where we've created a new schema for agents to come and talk to my website and receive some answers. What's going to make it worth it? And this is a good question. Scott in his answer noted that websites would be able to communicate to agents that that they wanted to make available and on what terms? Terms along with some vague hand waving about new advertising models and transactions. That last point is valid. TripAdvisor sells hotel rooms. O'Reilly sells training courses. And you can see a world where websites based on transactions can not only benefit from exposing themselves to agents, but in fact transact more and potentially pay an affiliate fee. Patel so yeah, I mean if you're, if you're searching for a car and you're searching across ebay and auto Tempest and cars and bids and bring a trailer, all of a sudden it doesn't really matter if someone's landing on your website. All that matters is that their transaction transacting and that they found the car that they're looking for. And so for all those transaction based websites they're going to do fine. Yes. Someone who's just making content and putting ads on it is kind of screwed.

I don't know about these transaction based websites, right. I mean, I'm thinking of cars as an example, right. You have car gurus, you have probably what, cars.com bring a trailer.

Cars and business models.

Yeah, the auction ones. But imagine a world where I just have. Cargurus is like an ad based business model. They want to drive leads to dealerships and get that in a world where I just say, hey, find me every single Ford GT between this year and this year that's publicly available online. Bring me the contact information and actually reach out to the people.

Yes, yes, yes, yes. But I don't think you can dis. Intermediate entirely. I think that, I think that the agent would. Would say, hey, there's a Ford GT that's available on bring a trailer right now. Here's the current price. Do you want me to sit here and bid on it for you?

Yeah, but here's the issue though, because I can have the. You could have the agent. Like, in theory, the agent would go to Cargurus and fill out the form and tell the dealer, process a lead to the dealer. And the dealer's like, thank you for this lead. Car gurus, if I sell the car, I will pay you a fee. Or maybe it's just paid up front. But the issue is an agent could theoretically just like find the. See the image of the car and then find. Do an image search maybe. Yeah, I'm just saying do an image search and be like, oh, I found this exact car and this exact photo that the dealer posted on their own site.

Own website.

And I'm not gonna pay Cargurus, I'm just gonna find. And I'll actually call the dealer and confirm that they have it and I'll negotiate on your behalf. And then.

But people already cooked.

Yeah, yeah, they already do that. But it's just much more scalable to.

Say, yeah, but I mean, that's not, that's not a marketplace. Like, like if you are on bring a trailer and you see a car, like, you generally can't disintermediate.

Bring a trailer. Sure.

Like, so if you are a platform that's, that's bringing liquidity to the market and bringing listings that don't exist elsewhere on the Internet. But yes, I mean, people see this with Airbnb where they see a beautiful place and they're like, I'm gonna go just look up this place and see if they have their own website. But a lot of people that are selling a one off car don't have a website where they're listing their car. Like somebody might do that. Maybe in the future you would just, you would just, you know, post on X hey, I'm selling my car, I have it listed on bring a trailer. But you can just negotiate.

Well, it's funny. So one of the, one of the companies that's probably pretty cooked in this new era is Internet brands and they own WebMD, they own but all the way through they own Ford Truck Enthusiasts, they own Renlist, they own Rough, they own the Dodge Forum. Like some of these more community based sites I think are gonna be fine.

Yeah, yeah.

Because it's a place where discussion is happening. But again, oh well, well, sites that are purely indexing.

Yeah. Ben Thompson goes on and says the original sin of the Internet lacking native payments was not, in my opinion a sin at all. Advertising supported the human web not because Andreessen failed to make a deal with the credit card companies, but because it was the only business model that made sense. No, the real neglect and missed opportunity in terms of payments is happening right now. Microsoft is on to the right idea with this adoption of MCP and in an introduction of NL Web, but its proposal, by virtue of not including native payments, isn't nearly compelling as it should be. As compelling, as compelling as it should be. The key difference from the 1990s is that on the agentic web, native digital payments both are both viable and the best possible way to not only keep the web alive, but also in the process create better and more useful AI. And so he goes on to talk about stablecoins and agentic microtransactions. Let's start with the viability. This is from Bloomberg. Stablecoin legislation overcame a procedural blockade in the U.S. senate, marking a major victory for the crypto industry after a group of Democrats dropped their opposition Monday. The industry backed regulatory bill is now set for a debate on the Senate floor with a bipartisan group hoping to pass it as soon as this week, although senators said a final vote could slip until after the Memorial Day recess. So hopefully this passes and we can talk to everyone on Crypto Day about this next week. That'll be a lot of interest, a lot of fun. Ben goes on to say, I know I've driven longtime Stratachery readers a bit batty with my long running and still enduring in the face of massive grift and seemingly unending scandals interest in crypto, but stablecoins are genuinely a big deal. I wrote a brief explainer last fall when Stripe acquired Bridge. And so I'm sure most people are familiar with stablecoins. He goes on to say, stablecoins solve several of the microtransaction problems I listed above, including dramatically lower or no fees and the fact that they are infinitely divisible and thus can scale to very small amounts. Stablecoins, by virtue of being programmable, are also well suited to agents. Agents, meanwhile, are much more suited to microtransactions because they are not in the end simply software making a decision, because they are in the end simply software making a decision, unencumbered by the very human feeling of decision paralysis. So an agent can just go around and make a bunch of microtransactions. The entire digital ad ecosystem, he writes, is an example of deterministic agents making microtransactions at scale. Every time a human loads a webpage, an awe inspiring amount of computation and communication happens in milliseconds. In milliseconds as an auction is run to fill the inventory on that page with an ad that is likely to appeal to the human. These microtransactions are only worth fractions of a penny, but the aggregate volume of them drives trillions of dollars worth of value. The problem, as I and Patel both noted, is that this ecosystem depends on humans seeing those web pages, not impersonal agents impervious to advertising. Which destroys the economics of ad supported content sites, which in the long run dries up the supply of new content for AI. OpenAI and Google in particular are clumsily addressing this supply issue by cutting deals with news providers and user generated content sites like Reddit. This however is bad for the sort of competition Microsoft wants to engender and ultimately won't scale to the amount of new content that needs to be generated. What is possible, not probable, but at least possible, is to in the long run build an entirely new marketplace for content that results in a new win win win equilibrium. First, the protocol layer should have a mechanism for payments via digital currency stablecoins. Second, providers like OpenAI should build an auction mechanism that pays out content sources based on the frequency with which they are cited. AI answers and this is kind of already happening with Spotify, right? Like there's a whole bunch of music on Spotify, they have a whole pool of capital from all their subscribers and then the more listens that they get, the more money flows in. It's not really an ad based auction system, it's more like just dividing up the pie.

The thing there, and it's a good comp, is that that was happening from the very, very beginning in the initial deals that Spotify did with record labels to get the music on in the first place. And now we're in a situation where all the LLMs are just using the Internet.

That's not true, though. Like, OpenAI has a deal with the Wall Street Journal, and every time they.

Have some deals, but they certainly don't have deals with every book they've ever ingested.

No, no, no. But you imagine that they'll kind of go down the power law curve of like the most important information sources and broker deals more and more and more, which is what Spotify did at the beginning.

They'll do that if they're sued.

Remember, like they had and they have been sued. They've been sued a ton. And then over and over, so they get sued and then they settle and then they do a deal. And like, yeah, it's a little messy, but it's fine. Because at the end of the day, the Wall Street Journal has a deal with OpenAI where OpenAI can ingest information from this surface it in ChatGPT results and the Wall Street Journal gets paid for that work. And I don't know the structure of the deal, but I imagine that it's something like we're generating this much money from people searching and you're driving 1% of our value. And so we'll give you 1% of the cut or half a percent, we'll split it with you. Just like if you're Taylor Swift on Spotify and you're getting 1% of the listening time, they'll give you half a percent of revenue or something like that. Is that not crazy?

I love it.

Theoretically, I think that's like what's happening. And the problem is go down the stack until it's self serve. Because if you want the open web, you can't do individual deals. And so eventually anyone can go on Spotify. Like, we're on Spotify, we get, I think we get a check. We got to check from Google at least. And in theory it's like, we don't need to do a deal with Spotify. Of course, Joe Rogan did do a deal with Spotify directly. Like the Wall Street Journal did a deal with OpenAI. Yeah, but plenty of podcasts go on Spotify. Spotify automatically runs ads and does premium plays and then sends them a check.

It's just such an interesting dynamic. Right, because why would it play out different so often? LLMs are not just serving the content, they're effectively remixing it. Totally Right, but the precedent, historically, they're hitting the page historically, if the journal went into the, you know, somebody from the journal goes into the library, public library, reads a bunch of books, forms its own kind of summary of. Nobody's getting paid for that. Right? Yeah, but so there's an argument, the models can make an argument that there's enough precedent that says there's nothing illegal about ingesting.

Yeah, that's true information. But there is this win, win, win that Ben Thompson's talking about, this idea that Google has an incentive to keep the open web alive and the LLM providers might have a similar incentive because they want more journalism and more facts to hit the Internet that they can scrape and pull into results. And so what is the value of a single Wall Street Journal page loading logged out. Right.

Yeah, we know. I could see a scenario where it's like net new content or something like that where like OpenAI does a deal with substack.

Why not every time? Why not? Why not on a user basis? I asked for a summary of the Wall Street Journal mansion section.

No, I'm saying.

And it went red 10 articles.

I'm saying every time. I'm just saying it's possible that there needs to be some type of cutoff date where again, how does this. Actually it sounds really simple and finding this win, win, win. But the actual mechanics of getting a deal done between the Internet and a relatively small group of foundation model providers.

But again, are you saying that there's a tremendous amount of complexity?

I'm saying there's a tremendous amount of.

Complexity, John, because that's exactly what Ben Thompson's concluded with. He said there is to be sure a tremendous amount of complexity in what I'm proposing and the path to marketplace for data generation is quite uncommon. Unclear at the moment. Who however could have predicted exactly how the ad supported web would have evolved or centrally designed the incredible complexity that undergirds it. This is where Scott's exhortation of openness is spot on. A world of one dominant AI making business development deals with a few blessed content creators and scraping the carcass of what remains on the web. For everything else is a far less interesting one than one driven by a marketplace, auctions and aligned incentives. To get there however, means realizing that the Internet's so called original sin was in fact key to realizing the human web's potential. While the actual mistake would be in not building payments now for the coming agentic web and so bull market in stable coins. We can see table coins trading at $1.0001.

It's so funny because everybody's so bullish and convicted on Stable coins now. But it's like, how do we make money on this? How do we make money? I don't think it's launching the next Circle. Right. Or another stablecoin provider. Even though that clearly is a great business. If you look at, you know, some of the different providers I was actually wanting to get. There had been rumors that Circle is talking with Coinbase and Ripple about a sale because they put out their S1. The market hated it. And they're basically like, you are completely dependent on Coinbase in many ways and paying all such a large part of your revenue to Coinbase. So anyways, could make a lot of sense for those two to join forces. But Coinbase is also dealing with a lot right now.

Yeah. Well, that concludes Microsoft. And we've talked about Google a little bit. Should we move over to Stargate?

And I would like to do that, John.

We're taking a full tour today. We're just going around the horn to all the big labs, one story after another.

We got to get boots on the ground. And Abilene.

Yeah, we got to go there. I mean, Chase was. Was texting me earlier. We got to have him on the show. But I think visiting would be worth it. It seems like Shirin Ghaffare. I don't know if I'm pronouncing that right, but Shirin over at Bloomberg got an exclusive look inside project Stargate, the $500 billion AI infrastructure project, including it on the ground tour of its first massive data center in Abilene, Texas, with interviews of Sam Altman. Love it. She's been on absolute tear. Wrote this today. Dropped the anthropic profile of Dario yesterday. Just banger after banger. So inside the first Stargate AI data center, OpenAI, Oracle and SoftBank hope the site in Texas is the first of many across the United States. Let's dig in. Trucks carrying concrete and electrical wiring plod over red clay, weaving in between cranes and excavators. Two symmetrical buildings stand in a massive plot of land where thousands of people in brightly colored vests work day and night to construct six more near identical structures that will make up the first site for Stargate Project Stargate is collaboration between OpenAI, Oracle and SoftBank, with promotional support from Donald Trump. Love it.

Throwing that in. Still unbelievable that they got him to just go and announce it.

Even though the guy loves building in America. This is like his whole brand. Showbiz loves build big Things to build data centers and other infrastructure for artificial intelligence throughout the United States. The companies have pledged to spend as much as $500 billion, a number so large it's hard to believe it'll actually happen. But at least for one. For this one in Abilene, Texas, 180 miles west of Dallas, Chase Locke Miller says they're good for the money. Chase Lock Miller.

This is an incredible.

He chases the energy, he locks it down. Then it's Miller time. Chase Locke Miller, baby.

Determinism is wild.

Let's pull up the photo. Chase, he's looking great. He's hanging out.

An absolute dog.

Yeah, he's been out there for so long.

Locke Miller is the co founder ago.

You can go to the next slide. It's. He's more centered. I met him four or five years ago and they were doing oil and gas flaring for crypto mining. And it was.

Yeah, there was.

And it was. I mean the business was doing.

Funny because it's funny because it was very low status to do the crypto to AI pivot.

Totally.

Yet you get CoreWee and Crusoe out of it.

Yeah.

Kill absolute monsters.

Always ignore the meme. Whatever the meme is, just ignore it. Build a GPT wrapper and sell it to OpenAI for 3 billion. Like pivot from crypto to AI and IPO for 40 billion.

Do it.

It's fine. It doesn't matter. Ignore what the haters think.

Yeah. So Locke Miller is of course the co founder and chief executive officer of Crusoe, a startup that helps develop AI data centers. Crusoe's cost to build the one in Abilene are expected to reach the 12 billion lock, Miller says. And that's not counting the billions of dollars worth of Nvidia Corp chips that will be installed in the finished facility. I've never built anything of this scale, Lock Miller says. I've had to learn a lot on the job. So the three main companies behind Stargate could say the same thing. OpenAI knows AI, but has relied on Microsoft for data centers. Oracle knows databases, but only holds 3% of the cloud market. SoftBank once raised a $100 billion fund, but that didn't go so well. Let's give him some time. Let's let him cook. You know, 10 year, 10, 10 year horizon. We got a. Got a few years left on that one.

Oh yeah, we got to pivot really quickly to the, to the Economist because the Economist has an article about will OpenAI ever make real money? We should do a whole deep dive on this one. But the way they describe SoftBank is hilarious. Here it says they're talking about the latest open air fundraising. Happily for The CFO of OpenAI, money men swept up in the AI mania need little persuading. They are falling over themselves to fund OpenAI. On May 13, SoftBank, a Japanese tech piggy bank, said that its $30 billion investment in the firm was unaffected by OpenAI's recent decision not to ditch its odd governance structure. A nonprofit board will keep control of its for profit arm. It's a whole, it's a whole great article.

I mean, we were talking about this off air. I mean it's absolutely credit to Sam.

Yeah.

For not completing the conversion but not having to renegotiate at least the top level numbers. Yeah, the 30 on 330 or whatever. It was one of the best dealmakers of all time. And Masa doesn't clearly, clearly doesn't care what price he pays.

Just wants simultaneously. The SoftBank deal is like crazy and unprecedented and wild and Japanese piggy bank and all that. But the premise of like, will soft. Will OpenAI ever make real money? Is like ridiculous to me.

Like, will the next, will the next consumer tech giant make money online?

Will a company, 500 million users make money?

I would say, saying, saying this yesterday off air, it was like, you remember, it was only it was less than a year ago, people anthropic was ripping Yahoo narrative and everybody was just saying, yeah, open AI is Yahoo. Open AI is Yahoo.

I was thinking about this a lot and I was like, what, what was the problem with Yahoo? Like they got like, they got disrupted by Google, obviously. But like, more specifically, what happened? Like, it was, it was a shift from, from non algorithmic links, right. To a better algorithm in Page Rank. Like it was a development of a better algorithm that really smoked them. And I'm wondering, like, is that, is that a point of fragility for OpenAI? Like you could imagine someone comes up with a different model for AI. Like they say, like, the transformer architecture is fundamentally not it. It's V1 and we're coming out with V2 and it's completely different architecture. Fine, maybe that happens. Is OpenAI agile enough to implement that within six months? Like probably, right? If Deepsea can clone Transformers and reasoning models in just a few months, you would think that, like, I guess the question is like, the Yahoo failed not just because Google invented Page Rank, which sorted the links by the number of links to a specific website. And so if everyone is linking to strategy and you search for Tech News that increases the page rank and that shows up at the top. And that was the, that was the key insight by Larry Page, the creator of PageRank. And it. And it created much, much better results. Like when you went to Google, it was dramatically better of a product. I remember when it came out, like I was a kid and I remember using AltaVista or what, like 30 at that point? Yeah, yeah, exactly. Like little late 18th, I think I was actually like 8 or something. But anyway, so I had been using Yahoo and AltaVista and you would search and it would just be kind of like random links or they'd be like kind of curated. It would literally just be. It would basically just search based on the keyword. So if you search for like tech News, you would more than likely just see like technews.com come up because it had really strong, like SEO just from keyword searching.

Yeah.

Google brought in PageRank and said, well, if, if it doesn't matter that strathechary the name, the word doesn't make any sense. And Ben Thompson never actually writes, I am tech news, technology news, hashtag technology. It doesn't matter that he says that all that matters is that everyone is linking to tratechery and it's actually a great source for tech news. And so we're going to put that at the top of the results. Now, that was a better algorithm and it was a better product very quickly. But if yahoo had implemented PageRank, which I don't know if it was like patented or something, but if they'd been able to keep up with that, people probably would have just stuck with Yahoo for a lot longer. But something about where the company was and the culture didn't allow them to upgrade their search results and so they fell behind very quickly. And so for OpenAI to become the Yahoo of AI, there would have to be a new company that comes out with a dramatically different technology or paradigm and either is able to patent it or, or keep it locked up in some way. That OpenAI can't just port that innovation back because they've obviously been able to do that with other advances in image models that are happening elsewhere. Reasoning or tool use, like, it would be crazy. And this is something that's always interesting in tech is like, there just don't seem to be that many patents that hold. Like, how amazing would it be if you're just like, I'm the person that thought of MCP or I'm the person that thought of image diffusion and I have the patent and so I'M the only one that can build the company or you have to pay me to license this. Like, that's just not a conversation at all.

How many designers, for example, have patents from big tech companies and then can leave and effectively rebuild the same features?

And the patents don't even seem to hold. Like, like, did you know that the swipe down to refresh, pull to refresh, that was patented by a designer at Twitter, like 10 years ago or 15.

Years ago and essentially they created the slot machine for the mind?

Yeah, Endless scroll was another one.

Slot machine for the mind.

Yeah, Endless scroll now. Now you don't even really pull to refresh. I mean, I guess you do every once in a while, but mostly it's just Endless Scroll. But that was another thing that was invented at one of the big tech companies and then immediately ported across every other platform. And like Stories. Snapchat. Evan Spiegel created the Stories format. Thank you, Evan Spiegel, for creating Stories. We love innovation.

Thanks for creating a low risk way to just share what you're up to.

Really show you stories. But he wasn't able to patent it or keep Meta from rolling it out to every single thing. LinkedIn has stories now. And there's this question. It's this weird failure potentially of the IP system. Maybe it's better because stuff just goes everywhere. But if you're, if you're OpenAI and you're on top of the, of the innovations that are happening even at other labs, you would imagine that you'd be able to port back anything and stay on top. As long as you're that front door and you have the scale that Yahoo or Google eventually got to hundreds of millions of users, you should be fine. Anyway, that was just my thinking on the Yahoo thing. It feels like a stretch.

Yeah.

Especially given the recent, the recent innovations.

Totally. Anyways, back to, back to Abilene, where we don't use Manus, we use. We use chat GPT around these parts. Anyway, so the article here is basically pointing out that OpenAI, SoftBank and Oracle are funding this $500 billion project and yet none of them have hyperscalers, Hyperscaler, sort of data center.

Microsoft, but yeah, experience.

No, Microsoft. I mean, Microsoft is not in Stargate. It's OpenAI, SoftBank and Oracle.

Yes, yes. True, true, true. And so the only companies that have built cloud infrastructure on Amazon, Oracle, Microsoft.

Yeah, Oracle knows databases, but they only have 3% of the cloud market. Right.

Oh, there's Larry Ellison.

Don't count out Larry.

Oh yeah, right. Oh, Sharon you're in the truth zone. Larry's good.

Larry scouted.

Don't. Don't come for Larry. Larry's on top of it.

We will defend Larry.

Seven years old, and he looks like he's 42. Great.

Incredible. And he follows TBPN, so he's gonna be fine. We love you, Larry.

He can build some cement. He's got Chase on the case.

That's right. He's got Chase on the case. Anyway, so in interviews with Bloomberg Businessweek, the CEOs of OpenAI and SoftBank acknowledge that they've promised a lot and that to some extent, they're figuring things out as they go. And we knew this, right? They announced this. It was like, you know, Elon was pointing it out. He was saying, where's the money coming from? It's very, you know, pulling $500 billion together is not the easiest thing in the world. So they dismissed critics, the loudest being Elon Musk, who says the full scope will never be realized because they don't have the money. We don't. And Masa Says, founder of SoftBank, the infamous Japanese piggy bank, is this article. And the Economist says we don't need $500 billion in one day. We'll go step by step.

Do you think more founders should be pitching stuff like Stargates? Yeah, to realize my full vision, I need 500 billion. I don't need it right now. So if you're going to come at me for that, that's ridiculous. I didn't say I needed it right now. But over the course of this company, we will consume 500 billion capital. I mean, a lot of the nuclear.

Companies might, who knows, produce trillions. So OpenAI CEO Sam Altman knew they'd need construction crews and lots of computers.

Yep, we're going to need computers.

Hey, folks, we're going to need lots of computers.

You're going to need computers.

But data centers also have extensive electricity and water needs to keep the machines running at a safe temperature, often requiring infrastructure upgrades that can take years. For Abilene, which will have a mighty capacity of 1.2 gigawatts, the team built its own gas power plant to get the place running more quickly. Yeah, we talked a little bit about.

Augustus, but there's this whole meme about, like, every time you search it, like, pour empties out an entire lake and, like, we're running out of water and like, I mean, I'm sure it uses some water, but I have no idea exactly how that's calculated. And that feels like Something that's.

Yeah. The question is what percentage evaporates during cooling process. Yeah. And is it like a lot of.

A lot of cooling systems are pretty closed loop because you're just piping, you're piping the water through. Through tubes and then the water goes onto the chip, the chip heats up, the water goes. The water gets hot and then you pipe the water out. That's like a cool GPU rig on like a gaming PC. And like. Yeah, you need to change out the water every once in a while but like not constantly. And so it's not like. I do wonder how what the actual impact scale is here and I bet there's a lot of people with like the incentive to. I wonder if you use any water or. Yeah, I wonder if you could water.

I wonder if you could rip out the, the parts of an air cooled Porsche and potentially use those.

Yeah, air cool.

I want an air cooled Elon. Could take air cooled Porsches and put, put engine swap Tesla motors into that.

I think we wanted a supercharged AI factory. Just pipe the air back in to accelerate, take the exhaust fans and reroute them right onto the chip.

Yeah. So anyways, Sam knows he's going to need a lot of computers, but they also have extensive electricity and water needs. So Abilene is going to have 1.2 gigawatts. The team built its own gas power plant to get the place running more quickly. One of the things that really surprised me as we were starting to dig into this, Altman says, was just how many things feed into the main line. The main line is still being assembled at the Abilene site in March. Some rooms are in the process of being wired and are off limits due to the risk of electrocution. In the future, tubes will pump cooling liquid around servers which will be loaded with Nvidia graphics or sorry, GPUs. And the physical GPU actually stands for.

Graphics processing unit, as they thoughtfully noted in this piece.

It's good to note. It's good to note if you've been contextual living under a cluster.

They also contextualize it by telling you what a GPU here is. It's not only a graphics processing unit. You can think of it as the physical brain of AI.

Something there, something there. PBAs.

This is, this is, this is ripped straight from a softbank deck. Yeah. It's like they're talking the same audience retail.

Yeah.

For now. Those tubes dangle from the ceilings beside half lit hallways. We're trying to deliver on the fastest schedule that a hundred megawatt or greater data center has ever been built. Locke Miller says from the front seat of a buggy on a ride through the construction site. In this moment, speed matters a lot. The big reveal for Stargate was at the White House on day two of Trump's second term. I didn't real. I forgot it was so early. Altman Sohn and Oracle chairman Larry Ellison stood behind.

This was so funny too remember because it was Elon and Sam like both around the White House on the same day in the midst of this battle. Dude, a little bit of an awkward.

Have you seen this picture of Masa getting hype mogged so bad by Larry? It's brutal.

Billy. John, you're going to bring up hype mogging.

It's a sensitive topic on the show.

It's a sensitive topic but we, but.

I mean we should pull it, we.

Should pull it up.

We got to call it out anyway. Lavishing praise and crediting him with help making the project possible. The whole production came together in a matter of days according to a person familiar with the events planning who asked not to be identified Stargate's real origin. According to Peter, the vice president of infrastructure and strategy at an operations at OpenAI. Hoshle. Hoshole. I'm not sure how to pronounce the last name. Goes back to a research paper published at the beginning of the decade written by a group at OpenAI that included Dario Amadei, who'd go on to start one of the company's main rivals, anthropic PBC. The paper describes so called scaling laws. This is GPT2, which we've talked about a bunch which assumed that the more capable AI requires ever more data and computing resources. What's interesting is like we're talking about increasing the ooms, but like this has to be synthetic data at this point because we've kind of already hit the data wall I believe. Or are we? Because I don't know that there's ever anyone expanding it. Yeah, but I don't know that anyone's really mapped out like 10x. Yeah. Check out that photo. Rough podium situation. Trump's also pretty big too. Larry's pretty big, but yeah, Masa well.

Presentation still a size lord.

Yeah. Makes up for it with the checks he rips.

Yeah. OpenAI, as many know, already had a contract designating Microsoft, its primary financial backer, as its exclusive cloud provider. But Altman decided he'd eventually need other options. Oracle was looking for partners for the future Complex and Abilene and was talking with Musk before Altman came in, says a person familiar with the discussion. Musk obviously chose Memphis for xai, so went a different direction.

Oh, this is cool. Altman picked the name Stargate because of one of OpenAI's early data center designs resembled the ring shaped portal that could open a wormhole in the 1990s. Science fiction movie of the same name, also a TV show. You ever watch Stargate? Absolutely not. Next question. After half a year of Altman selling the idea to prospective backers, Stargate really came into focus in mid 2024. And he says he found himself in a lot of Zoom meetings.

Wasn't Stargate also a CIA project around mind reading?

Oh, I don't know. I'm not familiar with that one.

Look it up, Chad Can.

The hardest part was figuring out what shape of the, what the shape of the deal would be. Would it be part of OpenAI? Should it be a separate entity? You know, they love separate entities over at OpenAI, so we know where they landed on this one. Should we just try to get one company to build it all for us?

That took. The Stargate project was a classified United States army program, wasn't affiliated with the CIA, but it was initiated in 1977 by the DIA and SRI International to explore the potential of psychic phenomena in intelligence gathering and military applications.

How'd it go? Did they. Were they successful?

I mean, seems like it. I'm sure, I'm sure they. I'm sure they cooked.

So. OpenAI is responsible for the operation of the business and will be its main customer. SoftBank handles the financial side, including raising additional funds for Abilene. Oracle will lease the data center and fill it with servers. Crusoe declined to comment on whose else is going in. Open Air and SoftBank will each put, put in 19 billion to start. Oracle and another equity partner, Abu Dhabi based investment firm MGX, are on the hook for 7 billion a piece, according to a person familiar with the matter. So, Yeah, I mean, 19 and 19 plus 7 and 7. You're getting up into the 40s, 50s, 60s, billions. Not bad. It's not 500, but they're getting there. They're within one OOM. That's what they got.

Striking distance.

Each project will be financed individually with a mix of equity and debt with data center firm. Primary Digital Infrastructure is helping orchestrate the financing. JPMorgan Chase is doing some loans.

Yeah, so if you have money over at Chase, you're, I would say at this point entitled to say that you're a financier of Abilene of the project in Abilene.

That's great for future sites. SoftBank has held talks with dozens of lenders and alternative asset managers, although it's yet to secure a deal. And some of those conversations have slowed due to global market volatility. But now we're kind of flat. Maybe the markets open back up and things move on. That's a structured financing that SoftBank has done many times in the past. So we are very confident. And Yeah, I mean SoftBank took arm private and so they're clearly, they're clearly capable of operating at this scale. It's not. Not crazy. Not everyone is so optimistic about SoftBank's ability to execute. The last time so did anything close to this scale was the Vision Fund, which depended on a $45 billion contribution from Saudi Arabia's largest sovereign wealth fund. Imagine clipping 2 and 20 on that. SoftBank's fund racked up losses most disastrously, most disastrously on the real estate startup.

Sometimes I get crazily overexcited and I make a mistake. Like we work. But when you have the conviction and the passion. Moss's mistake was not backing. What's the new company? No, no, no. The new real estate company by Newman.

Oh, I don't know.

Andreessen did flow, flow. He should have doubled down. But when you have the conviction and the passion, you actually learn from those mistakes and the scars and that will make you stronger. I've made so many more mistakes that I think I'm a little stronger than in the past.

He's been very wrong, but he's also been very right. And he's been right more than he's wrong. And so he's still in the game. And as long as SoftBank is cooking as a company, he can keep ripping checks.

Yeah, I mean, it's. His calculus is, I believe can be done on a napkin. And this is just my guess.

Huge napkin.

He has a very large napkin. And it's basically if I buy 10%, if I buy 10% of open AI and it's a trillion dollar company, someday I'm going to do very well and I'm going to lever up to do that. So my actual return on equity is going to be fantastic.

It's amazing.

And he's going to be laughing. And this is, I mean this is.

Like, this is like he's a gambling.

One of the best books on Masa is called the Gambling Man. That's the context you need.

Yeah, we did a whole deep dive on it. Complicating matters are trump's tariffs. Cost calculations swing every time the President changes his mind. But the trend line is expected to go up on is expected to go up on aluminum, steel and other construction supplies, not to mention Nvidia GPUs which are made mainly by Taiwan Semiconductor Manufacturing Co TSMC. Sohn says he expects no significant impact from the trade policies and I think he's right. I think this is what Trump wants and I think all of these will get, will get carve outs and they've already been negotiated down so I'm not too worried about that. Trump. I think the bigger question here is just like, like what is the return to scale because we've seen the scaling laws kind of maybe potentially drop in terms of pre training wall and whatnot. And so the question is just like what does a 10x larger facility actually get you? It better be worth it because we know the, we know the economics of GPT4 training run were great. Like they, they printed tons and tons of $20 a month subscriptions off of that and I think the total cost was like $100 million. Now you're up at 10 billion. Can you get everyone on a $2,000 a month subscription if you can, no big deal.

Well yeah and a lot of, I mean won't a huge amount of Stargate's resources just go towards servicing models?

Potentially. But I mean I feel like a lot of the narrative is like training but you know, obviously, obviously it will be both.

Alibaba chairman Joe Tsai has raised the possibility that we're in the middle of a data center construction bubble and obviously Amazon, Microsoft have pulled back on some data center plant.

Altman is undeterred, still seeing compute capacity as precious resource. We need more compute, more capital, he says. We want to have, we want to have access to a lot more of the machinery to make AI and run AI. And I mean yeah like the GPUs do go on fire and there are limited resources still even with the models. I mean obviously there's a ton that you can do on computer.

The fact that you're getting rate limited by Google as a paid customer, it's insane.

Yeah, it's insane. Like it is the biggest bull case for build more data centers but I honestly put that less on their cloud offering than just some PM was like oh no one would want to generate more than three videos in one day or whatever.

Totally.

Or like let's just roll this out slowly. Altman is undeterred, still seeing computing capacity. The Abilene skyline is made up largely of buildings constructed in the 1920s during the Texas oil boom. Among the city's major employers is a nearby Air Force base and cheese production plant that opened in 2022. Well, meanwhile, activity at the Stargate site, spanning 900 acres larger than New York City's Central park, moves at the pace of a boom town. And I was, I was talking to some real estate folks and I was saying, like, you should probably just go to Abilene and start building, like, Walmarts and gas stations and everything because, like, it really will be a boom town. They're going to be building this for a long time. Yes, these are going to be somewhat human resource light, but there still will be a ton of people working there. Ton of people getting paid a lot. They'll want bars and restaurants and like the local economy should, like, it would be impossible for it not to flourish like the economy booms. When you put in, like a soccer stadium, like you're talking about.

Yeah, I wonder if there's billionaire. I wonder if there's been any. I wonder if there's been any people going. I mean, sure, the Stargate project in general has just been buying up surrounding land to just sort of mitigate potential risks or future issues, but if you own like an acre, even just a single acre that's not connected to roads at all within this area, you could probably end up selling it for, you know, millions.

You got to just hold on and make them build the data center around you. So you're just.

Do you need a helicopter?

And you just look every. You have this quaint little house on.

The prairie and then it's its own little microclimate because, like, the heat coming off.

Have you seen that photo in Virginia where basically there was like this nice little house and then AWS and Amazon just built up these massive. And so you just see this nice house and you pan up and it's just as far as the eye can see. Data centers. It's like, beautiful, rough. The team behind the Abilene data center, conceived under the code name Project Ludacris. Nice, nice code name. First put shovels on the ground. In June of 2024, Crusoe decided to fly in workers from all over the country because of the cities, because the city of 129,000 people didn't have a labor force large enough to support it. Locke Miller says it's Project 129,000 is.

Way more than I originally kind of imagined for this.

Probably got some good high school football.

That feels more like a city than people. People like to Position Abilene as like, oh, this queen don't hate on Abilene.

It's projected to take slightly more than two years for the entire project to go from start to finish. Although parts of the data center are slated to go online sooner. Towns like Abilene are usually on the losing end of technological change, which tends to create most jobs in the largest cities. And data centers traditionally don't employ that many people. Crusoe has promised officials that the Abilene site will deliver 357 full time jobs once construction has finished. That's not that much actually. That's pretty small. I thought it'd be higher.

A $500 billion project?

Well, I guess it's 10 billion in this version or what are they a little bit more 10 or 50? But yeah, that is crazy low numbers.

But the AI agents hate to be contributing to this. I backed a company that makes robots for data centers.

Reducing it from three. We're going to get that number down to five. Yeah, that's the goal.

Yeah. I mean it's basically the idea that like data centers are flat and there's like routine maintenance that needs to be done.

No, it's reasonable and it's like very.

Much a controlled environment and just driving a little wall E car robot around makes a lot of sense.

We're talking about like how much of a boom town is like 375 jobs.

For now on a base of 129,000. That's like nothing existing. Nothing people.

I feel like there will be more knock on effects though. Anyway, the Abilene.

What about the chip? There's Chipotle opportunities. Yeah, those 357 people are going to be eating a lot of double steak walk burrito bowls.

Yeah. So the Abilene government is hopeful that it is the start of something bigger. The city and county agreed to extend significant tax breaks to get the deal done. And local officials saw an influx of requests in the days after Stargate was announced from housing developers, Infrastructure, energy infrastructure vendors would be investors and other municipal governments desperate to bring jobs to their town. It will impact the rest of the economy. Our restaurants, our home builders, our Chipotle. With many new people coming in and taking these jobs, the city was able to accommodate Stargate's waters water needs which are small compared to a typical server farm due to a novel cooling method that recycles most of the liquid. Let's hear it for the fake news that ChatGPT is destroying the ocean. It's not true. Like they figured out how to recycle the water, of course, because water's expensive and why would you waste it? The trade off is that the data center.

Yeah, the water is not being massively contaminated, like in oil production, fracking, and.

So they need a ton of electricity. Data centers at this scale are unprecedented. This was cause for concern where the memory of 2020, 21 Texas blackouts, which led to the death of six people in the county, still fresh in people's minds. Much of the blame for those outages belongs to Texas's deregulated energy grid. But the flexibility that comes with the lack of regulation was part of the allure for Stargate. Crusoe went out and bought gas turbines and oversaw the construction of a natural gas plant on site, which was supplemented the what the data center can get from local utilities. Yeah, I talked to an investor who was. Who knows Chase and was like, he's just. He has an uncanny ability to find energy, which is like, I don't know, something I wouldn't even think about in terms of, like, what the skills of a founder are necessary. But he's like, extremely good at finding, like, opportunities and pockets of energy that are being underutilized.

So I'm technically millennial, but I was fairly close to being Gen Z few days. There's that, like, rising energy. Like, I'm just trying to put it. He's an energy. He's an energy.

No, it's like.

Is that where you're.

It's like. It's like there. I think there are a lot of. There are a lot of companies out there that think kind of along the lines of like a single energy source. So there is a lot of. Of natural gas here. Chase is very good at putting together different opportunities. There's a existing plant. We can build a new plant. There's wind, solar, like very energy agnostic. But how do you piece all those together to create the. To meet the actual demand on an ongoing basis? Stargate's. To Stargate's critics, the idea that the project's backers could maintain this base and scale across the country is reason to think they're living in a fantasy. Amadei, co creator of the scaling laws, publicly questioned Stargate's seriousness, calling it chaotic musk breaking with Trump called it fake. Altman's retort. He says all sorts of crazy stuff. Salesforce CEO Marc Benioff said in a post on X in April that Stargate signaled an end to the honeymoon between OpenAI and Microsoft. The project does reflect a shift in the relationship between the two companies. They revised their Cloud contract in January to allow OpenAI to use other vendors such as Oracle, as long as Microsoft has still has right of first refusal. Microsoft doesn't intend to invest in Stargate, but was listed by OpenAI as a technology partner in the project. And this was interesting because they.

Can I just put your logo on the deck, please? Can I please just put your logo on the deck?

Yeah.

There's this amazing.

Microsoft is providing exchange.

There's this amazing. There's this amazing line from this guy Aaron, buddy of mine.

We'll be using Microsoft Excel on site. So they're a technology partner.

Yeah, that's it. He wants a quote.

I'm good for my 80 billion. He's just like, I'm good for my 80 billion. Generically. Capex. It'll happen somewhere. It's not necessarily an investment starting.

I'm going to butcher the exact clip, but Aaron Frank has an iconic post. At one point, or maybe he just said this to me in person. He just said, some people just want to see my name and some people just want my name in their deck. Because there was like this fintech boom and people would just come to him and be like, please, can I give you 50 bips to just like, put your name because like that, like, at some point, please do. He, he created the card that became the Apple. Okay, what, what was it called again? You don't even see that. The Apple Push. No, Apple Card.

Remember Apple Card?

I think it was called the Apple Card, but his, his company, like, became that. Yeah, Apple Card. You don't hear much about the Apple Card.

They had a deal with Goldman and then they kind of like fell off and maybe they're not doing it anymore. Did people get those? I never had one. Was it good to have, like, cash back on stuff?

It was pretty good. Okay, anyway, unlimited daily cash back.

I'm not really a credit card points maxer or into.

Never been big into that world.

Never to management consulting coded. It's like every, every guy I know who works like McKinsey or Bain or BCG is like, points. I'm just like, I don't care. Just tell me which card I'll get and I'll use that for a decade. And they'll, they'll squeeze me for annual fee, probably. And I'm probably, like, way underwater on it because I'm not. Points maxing.

Yep.

Anyway, if things go as planned, Abilene could serve as a model for the next of Stargate's developments. The day after Altman first visited the Abilene site, In early May, he testified before Congress asking for help to fast track data center permitting on future locations. The fat the first site was incredible. We need a lot more of that. Crusoe is evaluating another potential spot in Amarillo, Texas, which it hopes could be part of Stargate, according to a person with knowledge of deliberations who asked not to be named. OpenAI has said is also looking in several states, including Oregon, Pennsylvania and Wisconsin. And while Stargate might rent data centers already built being built elsewhere. Meanwhile, OpenAI and Oracle are taking the spirit of Stargate International with an AI data center under development, Abu Dhabi, which will not be part of Stargate LLC but will count OpenAI as a customer. That's simple. If all that wasn't enough, Stargate could even expand to include semiconductor production, according to Altman, perhaps as soon as next year. I think of Stargate as the AI factory. If we don't have to be involved here, maybe we wouldn't. But if we could just magically spin up all the compute that we needed in the sky. But at this kind of scale, he can't do that. And so there's another interesting angle here, which is the idea of building an AI data center in space. We're having a couple folks to come on the show maybe next week to talk about that. There's a company that's thinking about doing it. There's some investors that are looking into it. It's sounds silly, sounds extremely buzzwordy and sci fi. It's like hard tech space and AI and data centers. But think about it, you get energy from the sun, from solar panels, basically for free. And then you also get cooling because you're in the vacuum of space and it's cold for free. And so you don't need probably water cooling and you don't need as much energy. And so the economics might work out potentially in gta.

Don't need land.

Pretty. Yeah, you don't need land. And so you've cut out a bunch of costs and then added Starship or Falcon 9 launch costs which are significant, maybe small.

I cannot wait for the day where a satellite is ghibliing for me, just knowing that I can create the most silly picture in my mind and that send it up to space to get made and then have it sent down. Just send it back whenever you get around to it. Could be tomorrow.

And then it says you've run out of requests. We don't too much Ghibli anymore.

When's the last time you saw Ghibli?

It's been a while I have been using. I've been using images in ChatGPT. Chat GPT a lot, but on a lot more like practical sense. Just trying to visualize different things. I like the live.

I told you I've been using. I've been using ChatGPT image images to get my 3 year old to eat dinner.

Yeah, you said that was working really well.

By bite. I will generate a new. If you take another bite of chicken and chew it well, I'll generate another picture of me as a dinosaur.

My new, My new. Go to Ghibli. Instead of asking, turn this into a studio Ghibli, I just say turn these, turn the image into bodybuilders. And I've been having a lot of luck with that. I sent one to the chat. We'll try and pull it up, but.

Get this pulled up. It's.

It's great.

Oh, but that's, that's the original picture.

Yeah, this is the original picture.

So we'll pull up the original picture first of the OpenAI team. There they are.

Yeah.

Okay. And then can you. Can now can you make him look like bodybuilders?

Yeah, for sure. Yeah, yeah, I enjoy that. Also just making. Yeah, just a lot of Cadillac Escalades with GT3 liveries and a lot of TBPN merch ideas. A lot of stringers, a lot of suits with ads on it. Speaking of which, we should do some ads. Go to vanta.com, automate compliance, manage risk, Prove trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance.

If you're doing anything important, you probably should be on Vanta for sure. You heard it here first. Go to vanta.com should also get on.

Numeral sales tax on autopilot. Spend less than five minutes per month on sales tax compliance.

Benchmark Series A.

Anyway, do we have time to do. Well, we only have 15 minutes until our next guest.

We do a little timeline.

We should do a little timeline. We got. We got Austin Allred coming in. We're going to ask him about AI coding in the era of agents. We can cover Dario. Maybe tomorrow. I also want to go through. Oh, we didn't get a chance to dig through Craig Federighi and the Apple AI piece in Bloomberg. But there are some really hilarious points in here. One is that apparently when Steve Jobs was trying to acquire Siri, he called the CEO and said, I'd like to buy the company. I think this is the future of computing. I think voice interfaces, artificial Intelligence. This is what Apple needs. We want to buy your company. Bring them in house. The CEO of Siri said, no, I don't want to be part of Apple. I'm good. And Steve Jobs proceeded to call him every single day for 24 days straight before he got the deal.

Done.

Done.

Insane.

Absolutely insane. You don't hear about that a lot.

Yeah, you. You have. You get this idea in your head of somebody that is a demigod figure and you think that they don't have to be cringe to get things done.

Yep.

And turns out sometimes you got to cringe. Max.

Yeah.

And. And do whatever it takes. And sometimes that's calling somebody every single day.

It must just be so funny to pick up the phone.

Imagine you're like 24th. Imagine you're like, yeah, walking your dog. Or, you know, it's like Steve Jobs calling again. Oh, he's leaving a voicemail.

Oh, hey, bud. Still not interested. Still building my company, still grinding.

Steve Jobs would get sort of total privileges to call people, bud. Right.

Totally one of the few when you're at the top.

One of the few that earned it.

When you're the greatest entrepreneur of all time.

So if you're not Steve Jobs, don't use.

Remove that from your vocabulary. Let's go to some timeline Zach Kukoff's talking about. Oh, my God, we're really going to speedrun 2008 with collateralized.

Both losses have widened as more consumers fail to repay loans. Unfortunate learning, you know, basically, probably speed running risk management that credit card companies have done for a really long time.

I wonder what's actually going on here, because the whole like, yes, you can buy now, pay later on your burrito. That's silly. But that's probably not the vast majority of their underwriting.

Right.

Like, that's the meme. But that's not actually what's driving buy now, pay later activity. And I wonder if in the lead up to ipo, they. They loosened underwriting restrictions to try and ramp volume and because it seems like Klarna. Have they gotten out? No. Right. They're not public yet. Right. They have not successfully IPO'd. And so I imagine that in the lead up to the ipo, they're trying to do a bunch of things. There's that whole news cycle around, oh, we're not going to hire anyone. We're going to use AI for everything. Obviously, that's sending a signal to the market that, hey, we're going to be an ultra efficient business. The margins are going to dramatically improve. We are taking AI super seriously at this company. And so, you know, our opex should be really low over the long term. You can underwrite the the IPO against a really low SGNA cost, right?

Yeah.

And then also on the revenue side, they're probably saying they're probably trying to do everything they can. Maybe they went too far. And now they're generating losses, which is obviously not.

Yeah. So the issue is they doubled losses. So their net loss for the first three months of 2025 was around 100 million a year ago. For the same period it was 47 million.

Is this from like the S1 or something? Like, why is this data public?

I think they're trying to get out and so I imagine they need to do some type of reporting. Revenue increased 13% year over year to 700 million for the quarter. And again, this is a company that in 2021 was valued at $45 billion. Most recent private mark was around 15, so. So unclear what they'll do. They put their IPO plans on hold last month along with StubHub. And again, they had been on this marketing tear. Right. Remember the CEO came out and he was like, we're using AI for everything. We're letting everyone go. And then he says, actually, no, we're not doing that. It's not good enough yet.

I don't like when the IPO window is closed. I like an open window.

Yeah.

I like companies going out every day.

You punched a hole in the wall at the gym when you heard.

I remember when the IPO window closed. When it closed, Yeah.

I was like, john, why is your hand two feet into or sorry, why? Why is your fist two feet into the wall, John? You were trying to get it out too.

Frustrated?

Pissed off.

Well, if you're trying to track all these IPOs, trying to get it on the action, head over to public document investing. For those who take it seriously, they got Multi Asset Investing, industry leading yields, and they're trusted by millions. Also. Jordy, how'd you sleep last night? I put up some good numbers. I'm climbing up the ranks.

I'm getting back in the game.

Beating you back to back days. Let's see. How'd you do? What's your number? What's mine?

No, you go first.

Okay, well. One second. Get a pod 5 ultra. They got a five year warranty, a 30 night, risk free.

I got an 80, John. I'm sure you beat me.

Oh, I got a 94. Smoked.

Smoked.

Get out of here.

Get out of here.

Try 96. Actually, 96 there you go. Boom.

Look at you, John. I'm proud of you.

Yeah. Still not good on the consistency. Going to bed at all random times. Waking up pretty consistently. 545. Anyway.

All right, coming on. We kind of covered this already. This from the lion does not concern himself with SOC 2 compliance.

That's because the lion uses Vanta.

Right, exactly.

So I think he'd be concerned.

Maybe we need a promoter.

Yeah. This is an ad.

Hey, this is an ad for sure, Andy. You got to be disclosing.

Yeah. You have to do hashtag ad.

Or maybe he just loves Vanta so much.

Possible it could just be UGC.

Yeah.

But anyway, obviously, don't concern yourself with SOC 2. Be a Lion and get on Vanta.

Art over at Brex announced a partnership with Zip, a big procurement platform. Apparently, Zip is taking a step back from the corporate card market, and so this partnership allows them to.

Oh, interesting. Okay. That makes a lot more sense. So Zip had a corporate card, and now they're just partnering. Rolled it back with Brex for this.

And Brex had a procurement product.

Yeah.

Which they roll.

And that kind of happened, too, with. With Stripe. Had a corporate card for a little bit and then stopped that. Right.

Yeah. Everybody was like, wait, corporate cards, Good business. We should do corporate cards. Turns out, hard to. Hard to do.

Yeah. Well, there's always a question about, like, where does the corporate card sit? Because Stripe does not sit in the same place as, like, Ramp in terms of, like, the CFO suite. Like, it's a. It's more of, like, almost a developer tool.

It's less their initial wedge.

Yeah, yeah, but it's. It's less about, like, an accounting product. Product. It's more about, like, the. Like the. The payments and engineering product, although obviously it's very tied. But I don't think Stripe's ever gone into travel or, you know, bill pay. Like, they. They haven't focused on that as much. Anyway, Art says this is two YC companies with cultures of shipping fast and collaboration creating something truly special.

We love to see YC companies, obviously, the most fame. The recent collaboration between Deal and Rippling. Also YC companies.

Yes.

Hasn't that one more tumultuous. Yeah, a little tumultuous, necessarily.

Call it collaboration.

Well, they were partnered in some ways because they were both.

There was an employee. Yeah, they were the same employee.

Yeah.

Anyway, back in 2021, Art approached Zip. CEO. CEO. Because I saw Zip doing something remarkable in the enterprise segment. A market Brex was determined to win. At the time, the CEO of Zip. Zip pushed back saying Brex was too focused on startups. What would he get from this? Fast forward to March 2024. Brex has built their base of enterprises and many were demanding our products to work seamlessly together. Both companies had also retrenched to focus on core strengths. Zip stepped back from card, we stepped back from enterprise procurement. This is a classic competitors turned co creator story. Zip brings world class procurement orchestration. Brex brings global card capabilities and spend management across 50 countries. So so congratulations to Art and Brex.

Art Fellow, LA Guy and I have.

Some advice for Brax. Why don't you run a billboard to promote this on Ad Quick? Yeah, get some. Brex actually has a ton of billboards, right?

I mean they ran, they invented the.

Billboard in many ways. They did invent the billboard. They should get on it.

They should get the billboard economy on their backs. They did in 2020 19. Right? It was like it was hard to get a billboard because they were on all of them.

So head over to ad quick.com art out of home advertising made easy and measurable say goodbye to the headaches of out of home advertising only Ad Quick combines technology, out of home expertise and data to enable efficient seamless ad buying across the globe.

Great, great transition. John. We got a post here from TJ Parker who will be on the show Thursday again. He's coming back on. He says one thing I'll never understand about healthcare startups is the instinct to take a strategic route rather than an administrative tactical approach. You don't need a top to top meeting to get in network with a payer, just fill out the paperwork. You don't need a strategic relationship with Quest LabCorp to integrate scheduling or get competitive pricing. Just work with an existing vendor. If things go well, you can elevate to strategic relationship over time. But you'll never get off the starting line with the strategic approach. So Will has a, has a funny reply here. It's very funny. When people tell me they're struggling with getting in touch with payers. It's like have you tried going in the front door? Timeless advice, Timeless wisdom.

I wonder what's going on here. Is this just because there's a lot of like tech people who think I'm going to build tech for health care and then they just completely unfamiliar.

I think it's all about partnerships versus just.

Or they're just unfamiliar with this market structure because TJ has been, I mean he was a pharmacist before, right? Like he really understands the industry and has continued to understand the industry or even post exit. Whereas I think a lot of people are like I build software, I know that healthcare is a problem so I'm going to build healthcare software. But they don't understand like the structure of the market at all or how the deals get done and so they got to learn stuff from from TJ's posts.

Anyway, we got a post here from Andrew Reed in the news. Authentic Brands buys Dockers from Levi Strauss. A little mix up in the pants market. Little switch up and Andrew says Docker the containers overlapping with Dockers, the pants both making hundreds of millions selling to system admins.

Yeah, Andrew, it's really beautiful. He's a great investor but he puts his pants on one leg at a time like anyone else. Yep, let's get him some Dockers. Let's see.

Get him some Dockers.

I somehow think he's.

We should get him too stylish for that. Yeah, I can't see him wearing Dockers but we should get him a TVPN tie.

Yeah, tie would be working on with.

Ads on last time.

Yeah, last time he came on the show he was wearing a actually really nice Figma sweatshirt but didn't have a collar. So you know, docked some points from that.

Yep, docked, docked.

Well you know we should also tell.

You about Wander Find your happy place Find your happy place Book a wander.

With inspiring views Hotel graded minis, dreamy beds, top tier cleaning and 247 concierge services. Vacation home but better folks.

John Andrew founder Wander was teasing how excited he is about some big news and I couldn't possibly guess what the news is but we're going to have him on the show.

Toddler show says OpenAI tried to partner with Google to use its search API to par to power real time results in ChatGPT's search offering. Google said no citing too many complexities. And this is from Michelle Fradden or Frayden. Interestingly she was at Sequoia, she was on the FTX deal and then went over to OpenAI right before the Sam Altman firing. And so it was just like chaos for a while. But it seems like she's settled in over at OpenAI and is doing well but getting some pushback from the Google folks. She says we hope you're doing well. We wanted to follow up on our discussions over the past few weeks regarding using Google Search API to help power Chat GPT Search GPT prototype and search functionality in Chat GPT. As mentioned, we are currently testing a prototype search PT and would and to learn more about how Users and publishers use AI chat experiences coupled with real time information and search like features. On July 25, the day of our public announcement, we reached out to Google about using its search API. We believe having multiple partners in particular Google's API would enable us to provide a better product to users, not just destroy you. We also understand Google publicly offers its custom search JSON API to developers looking to incorporate Google search results into their products, providing additional precedent for us to work together. So basically asking like, hey, you have this service that you offer to a lot of people like, can you offer it to us? And they're like, absolutely not.

Hey, do you mind letting the fox into the hen house? Gemini, summarize this email. Fox sending the email to a chicken in the henhouse. Hey, I'd love to come in.

I'd love to come in, check out the house.

I thought I might ask. It seems like a great house. I think my, I think the other foxes in the Fox house would like it.

Yeah.

Google's like, the chickens over in the hen house are like, I think it'd be kind of cool of complicated. Google said, we're going to decline this.

Because we believe that you are a wolf in sheep's clothing. Well, cheers to Sundar for not letting the fox in the henhouse. And he lives to fight another day. Built his own Gemini, built his own apps. And we're excited to talk to some folks, frankly.

Built different.

Google built different. We're excited to talk to some folks over at Google about with the henhouse protection strategy and the Fox defense defense.

System, the Fox defense network, the Fox defense engine.

Should talk about Coinbase. This is a rough one. Markle Arrington is very upset with the Coinbase news. He says, I'm a longtime investor and champion of Coinbase. Something that has to be said though, this hack, which includes home addresses and account balances, will lead to people dying. It probably already has. The human cost denominated in misery is much larger than the $400 million or so they think it will actually cost the company to reimburse people. The consequences to companies who do not adequately protect their customer information should include, without limitation, prison time for executives. Very disappointed in Coinbase right now, using the cheapest option for customer service has its price and Coinbase's customers will bear that cost.

And so, yeah, so for context, if anybody was living under a data center, Coinbase effectively was offshoring their CX function. Certain CX reps had access to consumer data, including home addresses and account balances, which they then effectively sold or were bribed into passing to nefarious groups who then were threatening Coinbase, saying they were going to leak it. Coinbase came out and said, kick rocks. We're going to pay basically a bounty that leads to the arrest of these people. The issue, obviously, with home addresses and account balances being online is that bad actors could just find a home address, somebody with a high bitcoin balance, go there and say, with a, you know, the threat of physical violence, and say, send me your bitcoin right now. Or even worse, abducted. And so Michael is basically saying that there's going to be a big, you know, real, real consequences here. Coinbase came out and said, anybody that sort of falls to various social engineering hacks due to the, you know, they're basically going to reimburse them. Apparently, they expect to spend quite a lot of money, hundreds of millions of dollars to reimburse those people. But Michael is saying that the. The true cost is much higher. And apparently there had been just like a bunch of other issues here. I mean, and who knows what's actually going on? I'm sure that I do wonder how.

Much of a cost center is customer service for Coinbase? It seems like it's a big company, very profitable, has done very well. If they doubled the cost of the expense, would it get better? Is customer service really something that we're spending more money on? It is a lever, or is this something that's a little bit more intractable? The other interesting thing is, like, I do wonder, like, the Internet is pretty locked down. I feel like. Like, it is. Like, yes, there are addresses that can be sold in the dark web, but, like, it's pretty hard to just put up something that is illegal on the Internet. Like, like all the major platforms take it down. Google Shadow bans it. The links get banned. The browsers can be like, it's not.

But it's fairly crazy for this information to just be like a file.

Yes.

An Excel file.

Yes.

But at the same time, criminals can transact and sell multiple times at the same time.

Like, let's say that I'm like a, you know, the type of criminal that does. That does just show up to a house and put a gun in some.

The real, the immediate risk is social engineering. It's somebody calling you from a number that's spoofing Coinbase and saying, hey, I noticed some activity. I need you to verify this. Give me this code. So that's the immediate risk. Second order effect is that a criminal could just show up at somebody's house.

And demand yeah, yeah, yeah, yeah. So the type of criminal that just shows up at people's houses, like how are they actually going to get this information? Because they're not just going to Google it and it's not just going to show up there. They're going to have to find it somewhere on like through some sort of hacker network. Are they paying for this? Like, it feels like this information can be somewhat controlled. Like credit card numbers leak all the time and, and like the, the global cybersecurity industry is pretty good at making these not just proliferate to like script kiddies and like random criminals. Like yes, there are organizations that try and like programmatically like run up a ton of charges and just like spam the network. But in terms of like these one on one dangerous interactions, like I'm not so sure that, that, that it's, it's going to happen like in mass. Obviously it'd be terrible if it did and it's a very big deal and it should be taken very seriously. But, but, but, but I do, I do wonder if there's a way to, to kind of like, like not allow this information to proliferate ironically. Like what, what if it goes on the blockchain? It's just available. It's like, it's a very naughty issue with like, you know, censorship and, and things. But biology. Had a good response here to Arrington. I don't have, have it clipped, but we'll have to talk to him about like the future of kyc. His argument was that like part of the problem is that Coinbase is required to keep all this information when, when they, they might not actually need to if the government didn't require it. And so it's this double edged sword of the more information that you keep on someone, the more information that can be leaked. Anyway, we'll dive into that more next week. For now we have Austin Allred coming in to the studio. How you doing Austin? Good to hear.

Hey, good.

How are you guys doing?

What's going on? How are you?

I'm good, yeah. Hanging out.

Yeah. I wanted to have you on the show for a few reasons. Obviously there's a lot of stuff happening with AI coding agents.

What's this whole AI thing?

Yeah, what's this whole AI thing?

I think.

Haven't heard about it. Yeah.

I mean maybe start with like introduction on yourself and what you're up to now, what you've done in the past and then we can kind of go into the structure of AI in the workplace, how people are learning to use AI and then kind of some of the disruption that's potentially happening with AI tools actually replacing jobs.

Yeah, for sure. So a little bit of background. Co founder of Bloom Tech, which used to be known as Lambda School, before a trademark lawsuit. It's technically still the same company, but about 18 months ago, we started talking to people about this newfangled AI thing and if it had any impact within engineering teams. And frankly, I was pretty skeptical in the early days, but we sent out a research team, they looked at what everybody was doing, and at the time, it's not quite that way anymore. But there were a handful of people who had sat in the basement playing with LLMs, seeing if they could make them write real code for a couple of years. And when we saw that, we were like, like, oh, wow, this is actually working. It's not mainstream at all yet, and at the time even less so. And then working with a couple of those companies that we were. So we started training engineers to, you know, use AI really well. And it's one of those things where there are no best practices. Everybody's learning more every day. The game completely changes every day, let alone every week. And then one of the companies we were working with brought us in to do a, you know, fully intensive 100 hour a week, find all the people with 98th percentile IQ and above, fly them into Austin, train them over the course of 10 weeks to be expert at using AI and kind of staying on that cutting edge, riding that wave up, and then they hire them on the other side and pay us to do that. Now we're expanding that to more companies, but I basically spent the past several months hold up in a, a little office with a bunch of sweaty people trying to figure out how to max out AI and figure out how to stay on the cutting edge of using AI to build software.

Is the shape of that more like PM who's now able to write code or instantiate their ideas in code? Or is it software engineer who understands some basic programming that can then write more functional code or just be more performant? Like, like which angle? Is it both or is one narrative driving more of, like AI adoption?

Yeah, I'd say it's more the latter. So the more experience you have as an engineer, the more it sucks to start using AI because, you know, you're going from here to here as far as code writing ability goes. So in the early days of Gauntlet, we force everybody, you know, and the more senior you are, the more you hate it. We say you can't Write code manually. We put software on everybody's computer that, that watches them and AI has to do all of the code writing for you. And everybody thinks that that's the dumbest thing they've ever heard. And they fight against you for a week, week and a half. And then the more senior you are, the longer it takes for you to actually get to parody. But you'll never go back once you get there. It has fundamentally changed the software industry forever in ways that people don't fully appreciate yet. Entire companies are going, companies are going to have to rethink the way they do everything. And we're only seeing the very, very, very early stages of that.

Do you guys have your own kind of messaging internally with the students around the framing of like vibe coding? Because it maybe is, it's like this cool viral phrase people.

Yeah, but it's kind of different.

But it's also different if you're doing it in this hyper intentional way, trying to effectively get to the same quality and consistency as regular software engineering, but just doing it in a super AI native way.

Yeah, I feel like that's muddied the waters quite a bit. I mean when I say Vibe coding, most people imagine somebody who doesn't understand software kind of of blindly trying to one shot applications which it's incredible that that works at all. But when we think of, you know, whether you call it vibe coding or some people started calling it super building internally, I don't know if that ever made it out. But using AI to write performant enterprise grade bulletproof software, there's not a great distinction for that out in the market. But we do view it very, very differently. There's a time and a place for both. You know, if you're writing your own personal software or little internal tools like yeah, just blindly telling an LLM to do something might work. But most of what our engineers are doing would work in, you know, very enterprise, very legacy, very big, large scale applications. You know, security is taken care of. The, the formatting's done correctly, it's, it's scalable and secure and all the things that software needs to be.

What are you seeing the most success with in terms of tools? There's so many different options. Every week there's a new tool. Right. You know, maybe they were using Devin at some point and then they're using Claude and then they're using now Codex. What, what, what's your kind of how, how are people prioritizing? Are they using things like replit, you know, what does that actually, what does the stack look like week.

Yeah, it's not an exaggeration to say that it changes every week in a very fundamental way. So I remember when Claude 37 came out, before Claude 37 came out, pretty much everybody was using Claude 35 as their daily driver. And I feel like Claude 37 was the first kind of breaking point where that became no longer true anymore. So before CLAUDE had basically a monopoly. If you were to walk around the Gauntlet offices and ask everybody what they're using, 95% was Claude 3.5. With 3.7, it worked better for some people's workflows and worse for other people's workflows. GROK is gaining ground. Most people don't have API access to it yet. So that's a. You have to use repo prompt or something to get it into your ide. So that's a separate thing. I would say. On average, for people who are working on really big, really enterprise applications, Claude 3.5 still seems to be winning out, even over Claude 3.7 and Gemini, I think kind of changed the game. When was that? Was that two weeks ago or a year ago? I kind of lose track at this point. But it has a much broader context window, so if you have a small code base, you can get most of the code base inside the context window. As opposed to Claude 3.5. Cloud 3.5 has proven to be better for more surgical stuff. Cloud 3.7 is like cloud 3.5, but it kind of runs away and tries to do a lot. So some people like that, some people don't. Every engineer ends up with their very ordinate, well orchestrated workflow at Gauntlet, and they do things their own way, whether it's, hey, I build all the front end stuff really, really well and make sure that's all working. And then I'd try to build a backend behind that. Or I diagram everything and plan everything out with a model like O3. And then I'm creating a checklist and pulling pieces off the checklist. Everybody has slightly different workflows, but it's.

What are you seeing on the other side, on the hiring side? Are companies expressing preferences already to you to some degree around, hey, this is the stack we're using internally. We want people to kind of be ramped already on this set of tools. Or is it more kind of bring your own toolkit?

Yeah. So there's a spectrum, of course, on one side, I would say the companies that, I call them, the companies that don't really get it yet, they're like, okay, give me somebody who has five years of experience with Typescript and they understand a little bit of AI and we'll drop them into our traditional engineering Org. The companies that are on the other side of that spectrum are more actually what was this entire team can now be one person. And what language we use matters a whole lot less than how good they are at utilizing AI because anything that you don't know AI is better at writing code academically than any engineer you're going to talk to. Knowing how to get it to write the right code at the right time in the right place is where all the magic happens. So on the other end of that spectrum, there are companies who are. I mean, the average company that we bring into Gauntlet and we show them what students are doing, they walk away saying, oh my gosh, I need to rethink our hiring. I need to rethink our entire strategy. Our roadmap needs to change. It is a fundamentally different world when you're utilizing AI in the correct way. And then I guess even further on the other end of the spectrum, there are a lot of companies who tried it once and wholesale rejected it and aren't going to really use it. Or if an engineer does it at all, it's on their own time and against the will of the manager. There's more of that than people would anticipate if you're spending all day on X like I am. But. But yeah, there's a pretty broad spectrum and I go back and forth with frankly how much time I'm going to spend trying to convince people to do it all the right way or to experiment with new ways of building companies versus just okay, you know what? I'm going to go create a holding company and spin up 10 companies myself and show everybody what can happen. I go back and forth on that. But yeah, it's wildly, wildly different.

What's your read on the hiring market right now? Obviously, there was big headlines, maybe it was last week, I believe midweek, Microsoft again. Our read on that situation was Microsoft has always taken taken the approach of kind of cutting the bottom 3%. There still was a bunch of stories that came out from people being like, I've been here however many years, working 12 hours a day.

Florida says they're not going to hire anyone, but kind of went back on that. Like, yeah, overall hiring market temperature, what's it looking like?

I mean, I think so. First of all, I agree with you that when there's a 3% riff, that's rarely actually a riff that's performance management being done in as kind of way as possible so you can give people severance which if you're a company like Microsoft with as much cash as they have, you probably should do it that way. That said, generally say, generally speaking, what we're seeing is a. If you're really, really good at being an AI engineer. I mean we had, there were times when we had multiple billionaires wandering around our office with printed out offer letters for people begging them to come work for them them. That's what it looks like if you're really good at AI. Most companies are not really understanding that. So that's not everybody. It is interesting. I, I think it's fair to say the junior engineering market is completely decimated. It's pretty well gone even. I mean if you're at, at Stanford, maybe you can get an okay internship somewhere. But it is, it's Armageddon to be.

What does that mean? Does that mean that like junior engineers should be vibe coding companies and starting more apps and trying to just like build lifestyle businesses? Almost like what, what is your advice for people entering the workforce in software engineering capacities?

Yeah, that's a really good question. It's, it's a tough one because the way that I describe it is someone who's not very good at AI and AI engineering can still get roughly the output of, you know, starting junior engineer out of AI. And that's what a lot of companies are doing. Instead of hiring a junior engineer, I'll just get, you know, the output that I need. There's still, and there always is crazy demand for senior and you know, high end and everything else. I think it'll take a while to figure out how that plays out because the flip side of that is you don't get super senior engineers if you don't have junior engineers. And then you talk to the universities and the average university that I talk to is closer to explicitly and permanently forbidding any type of AI usage in the classroom than they are to interweaving it or adopting it in any material way.

Why is that?

Because it's cheating. It's so much better at writing code than you are that the professor has no idea if you know how to write code because the AI will write the code for you.

I talked to his latest post. I was at the park a couple days ago and was at the swings pushing the little guy and there was another dad there who was a teacher and he said that a high school teacher. He said they basically now wait homework at effectively zero like it's like kind of check the box. It maybe contributes to like a few percentage points of your grade. And then tests are now just 100% because it's the only environment in which you can actually prove.

Would you advise people to learn to code now or has the idea of learning to code changed in some fundamental way?

So I still think, I feel like I'm almost on an island on this. I still think you should learn to use AI and if you want to use AI in software environments, you should learn to code. Now to me, that looks very different than it did five years ago. Our learn to code business, we've basically shut down. It no longer operates. I'm not interested in teaching people to code in the traditional way. Now the flip side of that is normally when you would train people to code, you would start close to the metal and at the bottom of the stack. And so you would start with, okay, let's write binary and then we'll figure out how to write some low level Java. And eventually a year and a half later you start actually building programs that can do things that are interesting to you. The right way to learn to code is actually in my mind the inverse where you start by building applications and then you figure out what's broken and what's missing and what you don't understand. And the thing that is different about that than just creating a different curriculum is AI understands that AI understands. So we had software that would basically watch what you were doing on your computer and from that we could derive what CS principles you understood, what CS principles you didn't understand. We could have AI generate a custom curriculum that meets those guidelines specifically and we can fill any gaps in a way that I would have killed to a decade ago training people how to code. So I think you can actually just start by building stuff and then figure out how it works down and start at the top of the stack and slowly work your way down. And to what extent you get down the stack depends on what you want to do. Nothing. The average engineer today isn't playing with a Linux kernel or writing c. They're writing JavaScript and they're building applications to be at that level, you can probably get there relatively quickly. But I think I'll credit Martin Casado from A16Z with this, who said basically to be a good engineer, you've always wanted to understand one layer of the stack beneath what you were actually working on. And I think that remains true.

Yeah, give me your view on the AI software engineer market. OpenAI now has like three offerings between, you know, O3 will just write code randomly for you. Codex is a new product. They now own Windsurf. There's top down enterprise companies like Cognition and Devon. There's bottoms up enterprise like Windsurf and Cursor. How does this all play out? What do you look. What are you seeing in this entirely new market of AI tooling?

Yeah, my first point here is we're in an incredibly lucky position because as a consumer, you don't have to predict who will win. You can just figure out what's gaining momentum and latch onto the best. And so I have a lot of empathy for the companies trying to build those products and spend a billion dollars building a model. And if something is 1% better, I'm switching to it tomorrow. So that's a great place to be in as a consumer. How is that playing out? I think everybody's trying to fight similar battles. Cursor and Windsurf, I think, are uniquely positioned because they're focused primarily on the UX of the experience as opposed to the data. It's kind of in AI company's DNA to think that every problem is a data problem and they're generally speaking. Right. But where Krishna and Windsurf are getting better and better is okay. It's really cool that there's this model. I'd say on average, that is the problem that AI companies have. People have no idea what they can do with the models or how they should be using them. I think that gets simplified over time. The same way every product becomes. I shouldn't have to know whether I want to use O3 mini or 4.0. No normal human's ever going to care about the difference between those. So I think they'll just get better and better. Who wins?

High eyes?

I don't have a clue.

Trillion dollar question.

Yeah, it is if you'd probably be a VC if you're in that game. But thanks so much for stopping by. We'd love to have you back. This is great.

Yeah, keep us posted.

We'll talk to you soon.

Cheers, Austin.

Bye.

Have a good one.

Next up, we have Jeff Morris Jr. Himself coming in for Chapter One Ventures.

It's got some big news today. We'll let him.

Are you ready?

I'm gonna press this button.

Okay.

Welcome, Jeff.

Jeff, how you doing?

Welcome to the studio.

Welcome. Come on, guys. There he is, our technology brother.

This is a life, life goal achieved.

I feel like I'm like, let's go. No, I've seen, I've seen this progress and I'm just such a big fan of you both and I won't flatter you for 20 minutes, but really, you guys are killing it. It's awesome.

Well, I remember I think we got breakfast burritos in Santa Monica right around the time that we were starting the show. And you probably thought I was a little bit crazy for going full time on a podcast, but you always had faith. I think you saw the potential from the beginning. So it's great to have you on and maybe give a quick intro and then I'll let you talk about the, the, the news today and then we'll get into a bunch of other stuff.

Yeah, so I have been an investor, I guess for the past five years full time, started a firm called Chapter One. We call ourselves the Product Fund, mainly because all of our backgrounds are in product engineering, design, data science, and kind of take a product centric view towards what we invest in. But part of that grew up in the Bay Area, probably most famously as an operator was a VP of product at Tinder for a bunch of years. Ran revenue over there kind of during like the hyper growth period of 2015 to 2019 when the company was growing. And then now we're on the venture firm full time. And yeah, we've evolved quite a bit and today we're sharing some of that news.

Amazing. What's the news? And we're not quite set up to have multiple guests on, but we'll have to have your new partner. Well, new gp. So why don't you break down the news and kind of talk about the evolution of the firm.

Yeah.

So the big news today is Jamison Seidel is being promoted to a general partner. We started the firm in 2019 when. There we go. I think it kind of marks a moment in time for us as a venture fund, but also maybe more broadly in venture. We started Chapter One as a solo GP firm when Venus Solar gp, I think was pretty new, but also I'd call it trendy in 2019. And the it's been just really interesting to see the peer group go in different directions. So people who were silly GPS at the time, some have continued down that path, other people have gone back to operating. And then I think there's another cohort of people who want to build partnerships and firms. And so I think, you know, there's all these right ways to do venture, whether it's building a firm, equal partnerships, so the GPs and there's this constant debate, but I think we've really been clear on what we want to become and Today, building a partnership is obviously what we're announcing today.

That's amazing. Had this been your plan from the beginning? Did you realize quickly that you wanted investing at the firm to be a team sport versus something that that was more of this, again, solo endeavor, at least like it was in the beginning.

I think a lot of us in 2019 really, like, we had a plan, but it wasn't concepts of a plan, what we become. And so I think I realized in like 2020, probably a year and a half, two years in, that I didn't want to just do it on my own. And I think there's a lot of people who are. Who are really happy with that lifestyle. And frankly, it's probably a better financial decision just to do, raise funds, do it on your own. But I think to be competitive and to build a great product for founders, you likely can do a better job of that if you bring on partners and build a firm. And so that was the goal in 2021. It took a while to figure out who that person would be and also the right way to do it. So a lot of people will tell you to kind of like promote people from within. Build. You know, you kind of like, people buy into the culture. They know what you're doing, and you kind of elevate internally. Then other people say, go find a superstar externally, which at times we also thought about doing. Do you want to go bring like a fancy GP spinning out or maybe former CEO type? And those things are just really hard to get right. And so I think we were really lucky where Jameson was doing an outstanding job internally bought into what we were doing. And it's kind of like a head coaching search, right? You guys kind of in the TV pan model. Like, it's really, I think, alluring to go hire, like the big fancy head coaching name. But often times, like, promoting the person from within is the best move because they know. They know your players, your players trust them, and they know the system. And I think, I don't know, maybe like, the Boston Celtics did that really well.

Right?

And so I think it's really. I think that's kind of the decision most. Most firms have to make as they're growing their teams.

She's focused on. I mean, in her bio, she says, AI ML infra How you guys making money off of Stargate? There's 500 billion floating around. What's the angle? We were talking about just go set up a Chipotle there. Become. Become a. You know, don't just set up a.

Chipotle owner set up like a. Basically recreate the product. The 2007 Chipotle in Abilene.

In Abilene.

Massive opportunity.

There's 357 workers there. But I mean, seriously, like, obviously there's this AI boom. A lot of the, A lot of the trains have left the station. I don't imagine you're participating in the next open AI round. But, but, but how do you make money in the, in the current AI boom? Or, or is it. Or are we into like the next, Next trend?

Yeah, I think there's a couple angles. Like one is location. So there's. Jameson lives in London. And I think there's a popular narrative right now that all the, like, all the value is happening, value creation is happening in the Bay Area. And so obviously I grew up in the Bay Area and I live in Los Angeles. And we've made a decision to focus kind of on the other most important areas in tech. And so her view, and it's my view being spending more time in London, is that there's actually a lot of talent in the universities of Cambridge, Oxford, etc, And also we had a conference in London maybe two months ago and we had the founder of Granola. We had Anthropic, we had a lot of people from Jeep. Mind, is there a plan to help.

Those founders escape the backwater that is the UK and the Europe? Because I can't imagine that they're going to try and build businesses there. Right. That's crazy. If you go there, you give them money and you tell them here, go to San Francisco, because then you have an actual shot. Is that the plan?

I think, I think it's. It's a bit like there's the, there's like the Project Europe movement, which obviously like love Harry and think that's important, but there's another view. It's like meeting founders where they are today and often that, that is, hey, we want to raise the seed round and go, you know, build.

I mean that's the story of anthropic and granola. Right. Like international founders who eventually came to the U.S. there's always this. I have about.

There's also capital. Right, right.

Sweden's pretty goated. Pretty good.

The Nordics. The Nordics. There's something.

There's special. Yeah, there's always the capital flow. Like there are a lot of great LPs there that want to invest in American companies, a lot of founders that want to come invest in America. Obviously there are some great companies in Europe. We're joking around. But, but but getting that capital flow right, so you don't just expatriate US dollars to underperforming countries is probably, probably a risk you want to avoid. But at the same time it's probably really underpriced assets that are going to go find niche markets all over the world. And you know, who knows, maybe the next power law company comes from some bizarre country.

Like do you have. Do you have strong beliefs around the intersection of AI and crypto? There's been a lot of different attempts and people attacking it from different angles. You obviously have backed a bunch of crypto companies historically.

Yeah. How do we make money off of stable coins? Ben Thompson was writing about this today how stable coins are going to pop up. We were joking web and I want to get in on the action. I'm super long stable coins but they haven't moved at all. My portfolio is flat.

No, it is. It is a tough thing where everyone's just so bullish on stables and yet, and yet it's kind of unclear.

Can we do like a meme Coinbase stablecoin USD on pump.fi just USD coins.

Potential potentially just 1 trillion. There might be something there.

USD should be a dollar eventually.

No, but more seriously, zero FTF backing. It's just. It completely unstable.

It's just complete me.

Yeah. Unstable coins.

Yeah.

I think people have tried.

I think people have tried. Everything's been tried. But seriously, like stablecoins what is exciting? Or again like with the bridge acquisition, is that the end of the story or the beginning?

I think you're seeing a lot of app application layer payments applications are using stablecoins to enable their payments. But there's not. There hasn't been a lot of value accrual to net new circle competitors or trying to compete with a tether or usdc. And so stablecoin is like the easiest thing to talk about at a AGM or LP conference because it's like okay, I can kind of see that use case and so I think there's a shift to emphasizing stablecoins today. I think the combination of stablecoins and agents starts to get like. It does make sense but it gets pretty blurry in the current use cases of agents and what's possible today. I don't. For what's worth, we don't have a strong belief on the intersection between crypto and AI today being there aren't a lot of the primary use case we like is giving open source developers a way to actually make money. I think that's a pretty clear thing that's been missing if you look at any open source ecosystem. So you look at like now's research or different like Bittensor is trying to do this a little bit too. But the, you know, I think the worlds are very separate today. If you go spend time in the barrier and actually talk to the best AI researchers, like there's nobody cares about crypto in the Bay Area for the most part. And that's totally okay. Like I kind of like that. Crypto's still very weird and unlike by most people.

Let's get Linus Torvald's a mega yacht. Let's get Guido van Rossum a private jet. This guy created Linux Liquido, created Python open source software. He didn't get comped.

Let's get him some stables.

Immense, immense value.

Let's get him some greenbacks on chain.

Exactly. I want to talk about the venture dynamics for these mega platform funds and smaller funds. There was this, there was this trend during the ZIRP era where it seemed like the mega funds were just preempting everything you had the crossover investors. And that made it pretty hard for early stage investors to kind of make the decision on whether or not to write pro rata checks into the, into their earlier portfolio. Because they just wrote the seed check. Maybe they did find the great company early. But then there was this pressure to say, hey, these companies are graduating to Series A immediately.

Yeah, you want to defend your ownership.

Exactly, exactly. And so has that dynamic changed or has the early stage market adjusted to spinning up growth vehicles so that they can ride that wave as the overall market gets hotter? I mean some of these rounds for these AI companies, they get up Series A at 100 million. Like happening happens pretty frequently now. And how do you set yourself up for that as a manager?

I think that's probably the, the hardest part of 2019-2021 vintage, right?

Yeah.

And you look back at your follow on decisions and at least speaking for ourselves, I think that was probably the, the part of our investment we wish we could get back the most.

Yeah.

And so like today the markups are, are just as crazy, except there's obviously more revenue that you can kind of lean on to underwrite the companies. So it might be harder today because you have before it's like, okay, Excel or Index is marking up the company. We did the seed. We should maybe do the Series A out of principle now it's like, okay, one of those firms is doing it, plus they have 10 or 20 million dollars in revenue. And so it's equally hard And I think for seed managers, the growth fund thing for seed managers is pretty much done from what I've seen. People aren't going and raising growth funds and if they do, they have to have a very special reason why they're the seed fund who can do both. So I think there's a lot more. We've seen SPV volume pick up again, that's become a bigger part of the ecosystem. And then I think a lot of LPs are looking for co invest. That's always been the case, but especially so today. I think you see a lot of LPs coming to funds so they can co invest. And that's, I think, a bigger trend today than when I started in 2019.

What's the sophistication level when they're looking at co invest opportunities? Are they saying, okay, I'm going to back an emerging manager, let's say a $30 million fund, and then I'm planning to do one or two, take one or two co invest opportunities per fund and really just try to get into those winners. Or is there more of a. We just want broad direct exposure. And I'm curious what you've seen or what you're seeing broadly.

Yeah, I think it's both our preferences that you do broad exposure mainly because the odds of doing one co invest and having that be the right company are just extremely low. And so we, you know, we don't, we don't do a ton of SPVs, but the messaging is always like, please don't overshoot on any single deal. Because if you're doing a series bspv, we all know the amount of risk even at the series B that especially today, like that's even more pronounced. And so, but yeah, I think series.

B risk is more pronounced now.

I think if you look at the, the amount of competition, the amount of change happening on a daily basis in, in just technology advancements and then the pricing and size of these rounds, I would think if we look back two or three years from now, we'll will say the series b risk in 2025 was probably on par with the 2021 vintage would be my guess.

Call it a bubble.

It's crazy. I mean, I just see this having probably 60 plus different angel investments now. I'll get an update and see that a company is getting a markup. And I'm like, I don't really think they're there. They've made progress.

Yeah.

Are they any closer to being a business that truly has the intrinsic.

Used to be like pre ipo. It used to be like this is an extremely solid business within sight lines of true profitability.

And I think we need a new stock exchange probably in Texas, just called Dogs, which is like venture backed companies that have a little bit of revenue, still got a lot of, of work to do. Let's just get them out in the public market. Let's let them trade at like 2,3 million bucks, clear the prep stack and then maybe they 100x right. Maybe they figure it out. Right. So anyways.

Well, what was that, what was that SPAC that, that holds.

There's that one, there's that one Destiny. Right. I don't know, I don't know if you've seen this. Yeah, yeah, yeah.

The inverse destiny of the dogs, the companies that are underperforming private markets. You wrap all those into a spac.

Yeah, investors can kind of sell to Dogs, take a little bit of a loss. But it has such a penny stockflow, like maybe one of the. There's one banger in here.

It's actually, there's something there.

There's already businesses that buy, that buy underperforming. But yeah, this index, Destiny, I don't know if you've seen it, they hold a bunch of basically SP.

Like they have like some exposure to SpaceX, but they traded like 10 times the underlying asset value.

Yeah, that's crazy.

I mean people want access to this stock so they're going for it.

Destiny Tech 100. Anyway, last question I have. What's your updated thinking on incubations? I know there's some stuff in the works that you probably can't announce, but is that is part of bringing on a GP to free you up, to kind of be able to spend a bit more time on internal stuff?

That's been a part of our stretch we want to expand on. I think we're doing one incubation, which I'd love to come on the show in like three months to announce, partnering with a really big piece of ip, like a globally known piece of IP to build an application for that property. And so I'm really excited about this.

Because obviously Mickey Mouse payroll software. It's Mickey Mouse payroll software for sure.

Honestly, good execution and payroll. You're at least $100 million company and so you slapped a Mickey Mouse IP on that.

That's actually kind of true genius.

It's Mickey Mouse, just hyper commodity. Like you go after these markets that are just hyper commoditized. Yeah, we're partnering with, with Disney. It's just going to be Disney themed.

Oh, you don't want the Batman vpn. Why don't you want the Batman vpn? You don't want the Dark Knight protecting you while you're browsing online. You're going to go with what? Nord? What's nord?

Honestly, why not Batman? This is your new. This is your new playbook. You live in la. Let's go on a road show.

Let's build.

Is genius.

Enterprise SaaS meets legacy. IP. Buying it for pennies on the dollar. You give them like, 5% of the company. Yeah, good to go. Anyways, I do know. I do know what, Jeff?

What if Tony Stark built your erp. Iron man erp.

Stark.

Stark erp. Stark Enterprise erp. This is the.

Jeff could actually run this playbook.

Be great.

I kind of love this.

Yeah, definitely, definitely come back on. I think it's gonna really break the Internet when you launch it. So excited to see it.

All right, thanks, guys.

Congrats again to the whole team. Cheers.

Next up, we got Cliff Weitzman coming in from Speechify. Great entrepreneur, great founder, good friend of mine.

Are we talking about business or lifting?

We're talking about lifting mostly. So he's been on an absolute tear in the gym. He's super jacked. No, it's not a joke. It's not a joke. Hit the Ashton Hall. Hit the ashl. Bring. Bring Cliff in. Welcome to the Stream, Cliff. Welcome to the Stream. Oh, we were expecting Shirtless John said.

You'Re going to be shirtless as well.

Yeah, no, no, we're making.

We're making TVPN stringers.

Yeah. I worked out with him in. In this building where we record this. He put up some fantastic numbers on the bench press. Showed me a spreadsheet where he's. What's it called? Thor 2.0.

That's right.

He's trying to become Thor 2.0. And so I've been telling him I'm trying to do Thor 3.0 and it's been getting under your skin.

Has a list in his gym. How's it been all the PRs of anyone who's ever worked out in that gym?

Yeah, yeah, yeah. We're getting there. Are you in LA now? Where are you at these days?

I'm in la.

Nice.

I'm in LA and Studio City, but I'm about to. We move every four months or so. So I'm about to do sf, New York, Prague, back to New York, London, Rome, back to London, back to New York.

Those are all four month. Those are all four states.

So you're playing like 10 years out.

That'S the next 40 days.

Okay.

Back based in Florida for a set period of time.

Okay. Yeah. So I mean, take us through the structure of the business. Obviously you can travel a lot, but it's a fantastic story and I just want to hear what you built, how you wound up there and kind of the state of the union with Speechify.

Yeah, I guess to give people a sense for the scale, which I think is important because typically people just think companies are big if enough VCs post about it. But you haven't raised very much money. Speechify has over 500,000 5 star reviews, over 50 million users. Chrome extension of the year, app of the day. I'm just gonna keep. I'm hit this a couple times, but anyway, anyways, continue.

Yeah, take us through it.

Yeah, happy to share more. So I'm super dyslexic and I have adhd. So first, second, third, fourth grade I had a really tough time learning how to read. And when I was right about to start college, I built this text to speech tool that would read out everything to me. And in high school my mom used to read my summer reading books to me and just we didn't have time to finish the summer reading book for college. So I cobbled together this thing that would read the stuff into my iPhone and then I'd listen to that on the plane and it worked. And when I was in school, I studied renewable energy engineering at Brown. I ended up building about 36 different products. Everything from 3D printed skateboard breaks to iPhone, apps and websites and payment systems.

Wait, skateboard breaks?

Skateboard breaks. I'll show you an example.

Yeah, tell me about that.

You want to go slower.

So if you ride a longboard or a skateboard and you go down concrete hill, you don't want to stop by putting your face against the pavement. So what I did, 3D printed the series of brakes that would attach to the back axle.

Oh, wow.

That actually works really well. That's great. That's a good product. That's probably huge on Shark Tank. That's like the perfect product for that.

Could have been on the tank.

Oh, you did.

That's the most truck. I was considering going on a show, but it's like you end up sitting for like 48 hours in a trailer waiting to get called.

Oh wow.

Another big opportunity that I had. At the same time I was like, shark Tank isn't worth it enough. So at a certain point I realized that software is much better. So I built this kind of meme Maker on a flight from Logan Airport to SFO, published it, checked in on it 30, 40 days later, and it had 90,000 users. And I was like, oh, software is way easier than injecting molding stuff and way easier than like boron doping silicon wafers. My thesis was on a more efficient solar cell that I was building. I was like, all right. And around this time I read this paper about narrow applications of deep learning. A bunch of academic papers. One of them was called WaveNet about autoregressive speech. It came out of DeepMind and I was like, I can make a 100x better text to speech experience and a 100x better audiobooks experience. And I wasn't sure what to work on. I knew I didn't want to go get a job at Google or Palantir or Meta. And I was like, all right, well, I don't know what I want to do, so let me just write. So I wrote a 30 page paper about my worldviews. And the conclusion was I wanted to be the person that I needed the most when I was a kid. And the thing I really needed was someone to do my readings for me. Okay, I'm going to fully send it on this. So I convinced two of my professors to sponsor me to stay in school as a visiting scholar. Basically, I guess, taught classes. I got to be a meal plan, live on campus, use the gym, but not pay tuition or do homework.

Gym's important.

Amazing.

I was going to do that work teaching computer science over the summer, unlimited, indefinitely, until something took off. And so six months in, it took off. And so now there's like 50 million people who use it. But the goal is to make sure that reading is never a barrier for learning for anyone, no matter what your background is. So if you download the Speechify iPhone app or Chrome extension or Mac app or Android app, it lets you take a picture of a physical book. It reads, it gives you play buttons throughout the Internet. It's like the voice of the Internet. You click play, it reads and it coaches you to listen fast. So I listen to two audiobooks a week. I've done that since I was 14. So I've read more than 1, 800 books. By the way.

You listen on like 3X or something?

Yeah, I listen on 3X.

So intense.

And so you can listen to the.

Whole show in just one hour. Because it's a three hour.

Yeah.

Why?

Streaming doesn't work for me. It's gotta be something that I can react.

We're in RSS, we're in. We're in. We're on YouTube so you can pull it in there. Listen.

That's exactly what I did do. I show up to live streams, you know, 80% yeah, yeah, yeah, yeah, yeah.

Catch up at the end.

But look, I, I learned English when I was 13 so in the beginning I would listen at 0.75x speed built up to 1.52x, 2.53x and I'm obsessed with biographies, theology, philosophy, fantasy. And so in the beginning, nobody wanted to back a dyslexia education startup AI was not yet hot. The toughest point was around 2017 we didn't have the level of engagement and retention that I wanted. And there were a bunch of other sexy projects I could have done but from first principles my conclusion was I think that the trend that I most back is narrow applications of deep learning. Today we call it generative AI from a supply side and from a demand side, it's audio as a user interface. And I was like the intersection of these two. If I wanted to build something in that intersection, it's Speechify, so I might as well stick with it. It I ended up putting a gigantic help button on the app. It made like 20% of the screen real estate, bright red help Slack message us and if you clicked it, it just put you into an imessage conversation with me. So like I didn't need you to enable notifications. There was no intercom, there was no email. And if I messaged you back and you didn't respond, I would FaceTime audio call you and spreadsheet of all of our users. And if for whatever reason you didn't use the product, I would call you every day until you used it. Didn't use it. And then the product became really good.

That's amazing.

When we were 25, 18 of the folks who worked at Speechify were previously either CEO, CTO of VP of Engineering at their last company. A lot of them were XYC founders or folks who exited their company after their series. A nice tidbit on fitness. 12 of the original teammates got six packs within 10 weeks of joining and we had at least 17 people who gained 10 pounds of muscle in the first 10 weeks.

Everyone.

And we get a big Airbnb in a different city every four months or so. And so at the time we were in LA and I convinced one person to move from India, one person to move from Bulgaria, one person to move from Mexico, one person moved to San Francisco and I went and bought a bench press and I rented a truck Installed it in the apartment in my room. And so we had. Me and Valentin were sleeping in.

And that's like the only kind of weights you need, right? It's just to Ben.

And actually it gets even more deep because I went up to Marin to make sure that I was buying food for my parents because I didn't want them to go out during COVID And so two of the guys were in the apartment working out just body weight. And they did a DEXA scan before and after and they found that they had like not gained any muscle even though they were working out like animals. And I was like, no, you guys need more weight. And so I brought the press and we had two guys sleeping in a bedroom in the living room that we partitioned with a window, two guys sleeping in the bed in the, in the room, and then one guy on an air mattress. And that was like the best time ever. Like incredible time. And then we found just very strong product market fit ended up scaling really quickly. We closed partnerships with all the top publishers to resell their audiobooks and ebooks. My little brother Tyler started coding when he was 7, building Dragon Ball Z websites. When he was 8, he taught himself assembly to hack video games. He went to Exeter for high school, skipped four and a half years of math, skip five years computer science, did Stanford as an undergrad, dropped out to run a cyber security company, went back to Stanford to do his master's in AI. And I was like, hey, if you can help me build a model that will compile in 3x real time, has better quality than any API that's currently out there and meets the following requirements. I will do anything. I gave him a very nice compensation offer and he did. It took him about 10 months. He joined Speechify five years in as head of AI. So now we have a 40 person AI engineering team and we make the highest quality digital voices in the world. We are the largest supplier of speech AI in the world for consumers. And then so that's like level one, level two, level three. The vibe now is we make everything multimodal. So if you imagine that 5% of the population would read books for fun on their own, and let's imagine 15% of the population would listen if you provided it as audio. Now we're using text to video models to turn everything into like a full audio and visual experience. That's like super high level for Speechify but happy to riff on anything.

How big is the company now?

We're 176 people now.

Wow, that's huge.

32 countries. So 105 software engineers, 40 engineering team, and the rest is growth.

Yeah, so talk about, talk about the things that Venture has tried to make you, that the venture industrial complex has maybe tried to make you do that you've rejected because in an alternate reality, you would have raised $500 million by now and, you know, and, you know, decided not to do that. And that's influenced a lot of decisions, I'm sure.

So part of it is the fact that between the ages of 14 and 18, I read 461 books. A lot of them were about economics, philosophy, biographies, and I also listened to a lot of fantasy. And when you listen to fantasy, you think, would I make the decision this character would have made in this point in time? So my favorite book is the Way of Kings by Brandon Sanderson. I'm obsessed with this character named Kaladin. I based my entire leadership style after Kaladin. And so when I was writing that 30 page paper to figure out what I wanted to do in my life, I became one of those lucky kids who, when they were 21, figured out what they were put on earth to do. And for me, it was to solve dyslexia. And so I love what I do. Like, I wake up every morning early and I go to sleep very late. I sleep very little because I just, I can't get enough of working on this. And so I don't want to sell it. I want to own as much of it as possible. And I consider equity. Holy. Mind you, I was a solo founder for five years working on Speechify. And again, we had an amazing leadership team that blood, sweat and tears. We figured out how to hire. Last year, 50,000 people applied to work in the technical positions at Speechify. 10,000 people to the asynchronous engineering challenges. But we had incredible offers for our Series A from like, literally the top firms. We said no. They doubled the offers. We still said no because I didn't want to sell 20% of the company. And even with Seed, we just picked folks who we thought were really good. So founders of Instagram, of Twitter, of Robinhood, of all these companies that I could learn from, who did consumer subscription really well. And that was key. So, for example, Dylan Field from Figma taught me a ton very early on. And my nuance is if you take money from a fund, by and large, they have a fiduciary duty to their LPs. If you take money from an individual, they have just a duty to themselves. And so you get a Rolodex of someone who's really great, but you're not beholden. That being said, there are some businesses where it makes a lot of sense to raise money. In our case, we have a very simple business and we understand the user extremely well. And it's a very niche audience initially for people who have dyslexia, adhd, low vision, autism, anxiety, concussion, second language learners, and then the entire productivity suite. And at this point, we've worked on speechify for eight and a half years, so it's really hard to catch up to the product innovations as well as the AI side. And so we are always talking to investors. And don't get me wrong, we have raised money from funds as well, but we've just done it where we typically look for three things. Number one, it's folks who either have an evergreen model so they can ride with us because I'm going to be CEO of Speechify in 80 years, they have a history of philanthropy and education, or an experience with dyslexia or ADHD and their family. And we think really highly of them. They can see around corners that we can't. Either they've done an amazing job taking multiple companies public, or they've backed founders in ways that.

So you would go public?

We would go public. But into the future, not in the short term.

Yeah.

What are you thinking about competition from the hyperscalers, the big tech companies. Google I O is today. Microsoft had their build keynote yesterday. It seems like an obvious place that competition could come from. And yet on the product side, a lot of the big tech companies just don't seem to be iterating on the application layer side as fast as most people expected. So what's that been like? Has anything slowed you down over the past two years with competition from the big guys?

Oh, they've helped us a ton. So I'll give you two heuristics that are relevant to any founders, then I'll give you the specifics on Speechi5. The first one is, in my personal opinion, value accrues at the application layer more than it accrues at the model layer and models become commoditized.

Did you ever doubt yourself? Did you ever think, oh, maybe value is going to accrue to the model layer? Because there was like an 18 month period where everyone was like, models will do everything.

Well, I'm giving you what my opinion is today. And the only reason I have a strong opinion is I've gone deeper into it than almost anyone has. Right. We bought millions of dollars of H100 GPUs we set them up in our own data center. We have a 40 person team who does research and development on models. So the second part I was going to say is, but if you really want to be indestructible, you got to own both the model and the data and the application layer. And that's what made ChatGPT and OpenAI so successful, is they had Brockman work on the user facing side and they had Ilia work on the research side. And if they just had the research, they'd just be another laptop app. It's application layer that made it incredible. And that's the first thing because you want to be able to control the quality and the cost as well as the user experience. And then why did WhatsApp get acquired by Meta for $19 billion? Is because they had all the users. And so you own the end relationship. You want to have the phone number, the credit card, the email of that user, which we talk in a minute about kind of the changes that came with Apple. So that's the first thing. The second thing is, and this is very unique. Unique, sorry, not unique, ubiquitous. If you are Apple, Meta, Google, Netflix, there is no point in doing any project unless you think that project is going to make you one to a hundred billion dollars per year in the long term. We had a point where we introduced translation into Speechify. This is in 2020, 2021. And Mike Krieger, who is the founder of Instagram, CTO of Instagram, was like, you should consider removing this feature. And I'm like, are you kidding me? Do you know how long I spent building this feature? But I have respect for Mike. So I took it out install to trial. Increased activation, increased users use the product more. I was like, shit, I gotta remove this.

Why?

Why? Because it was distracting. Speech Advisor. One thing to give you the best experience listening to something, the second you add an extra button, users go down that path. So you know from Words you guys.

Have such insane scale that you could immediately see something that again, just removing some complexity.

Back then we didn't have mad scale, but you could see it in three days of usage, like there's nothing.

No.

So Clippy is horrible because Word is horrible because the design is terrible. But you know who has amazing design? Google one one text field. Who else has amazing design? ChatGPT one text field, one button. And that's why they're successful. If they had multiple buttons, they would be far less successful. Now here's the key. Text to speech is buried five levels deep in the menu. When I was 19, I made a video on YouTube under an anonymous account because at the time I was embarrassed for people to learn I was dyslexic. And it was called how to Text to Speech Mac Free. And if you search that phrase on Google, even today, my video from when I was 19 is number one. Why? Because it's so difficult to activate Text to speech that you need 19 year old Cliff to show you how to do it. And you can't pause, you can't play. And by the way, if you want to listen like me, what I ended up doing in college is I would use the terminal to change the speed because the GUI didn't go fast enough and it backed down. I had to restart my computer so I was restarting my MacBook multiple times a day when I was in college. And then I was like, screw this, I'm going to build a Mac app with keyboard shortcuts. I'm going to build a thing that ocrs the screen. And so my dream is for Google in mobile Chrome to add a gigantic play button that appears in every website that lets you listen to it. Why? Because it will make everyone addicted to listening, listening. And then when they want to listen to a long PDF and they want to do it offline or they want to do it in my voice or Snoop Dogg voice or their own voice, or they want to translate it to another language or they want to scan a physical document, Google doesn't cover that, Apple doesn't cover that. And people use multiple platforms, so you need to use Speechify. And so Speechify is the premium experience for text to speech. What I need is for people to become educated about the idea that you can listen to stuff. And so a lot of our work over the last eight years has been to educate the market. And so there's really two, two goals. Number one, build the most exceptional product. Like Extreme product. We have four core principles. Extreme product quality, leading with love, frugality and speed. So we talk to users a lot, we take really good care of each other, we don't waste money on anything. We build our own SaaS tools and we move really fast. We ship a lot to production. And the second goal is to educate people that hey, you can practice being a fast listener and if you practice, anyone could do it. My dad is 65 years old and he English is not his first language and the joke we have in my family is he wishes he had a mute button for me and I wish I had a speed up button for him. He now listens at 2 1/2 x speed to everything. And so my life is better because my dad listens and talks faster now. But anyone can learn how to do this. It's just a matter of practice. When you come into first grade, no one expects you to be a good reader. First, second, third, fourth, fifth, twelfth grade. We expect 12 years for you to become a good reader. But you can become a good listener. After listening to 15 audiobooks. You're not going to be good on the first one, you're not going to be good on THE tenth one. And being a good listener means three things. Number one, you can do something else and listen at the same time. Drive, cook, walk, whatever. Number two, you can listen to more than 2x speed. Number three, five weeks later, I ask you about the book. You have great retention of everything that you read in the book. And it will not happen when you start. You have to practice. But here's the thing. Reading is a hack we invented, right? 24 characters and some dead trees. Listening. We evolved to be good at listening, right? Telling stories over the fire. So if you were a bad listener, you were removed from the genetic pool. If you're a bad reader, you were not removed from the genetic pool, because otherwise I wouldn't be around. And so when you read, 30% of your brain is dedicated towards decoding. 70% is comprehending. When you listen, like 3% is dedicated toward decoding. The rest is comprehending. And so you can understand a lot better. And once you practice listening fast, you have that skill for life. And especially if you have adhd and you get distracted easily. The speed of listening matches the speed at which their mind is working. And that helps people focus, especially in school, but also in work. And so that's the goal of the company, is to just make that accessible to everybody. And Right now, there's 450,000 audiobooks, but there's 100 million books. We make all those accessible, plus your emails, plus PDFs and everything else.

It's amazing.

Incredible. You are cracked. Wish we had more time.

Yeah, we got to get a workout in soon. Soon.

We will.

Yeah. Come by before you. Before you take off.

We got to make sure you break that pr.

Yeah, yeah, I'm working on it. I think I'm pretty close. We'll talk to you soon. Cliff, thanks.

Thanks for coming on. You're the man.

You got it.

Bye.

A bunch of VCs are going to reach out to you.

He really is a fantastic business. Let me tell you about Bezel. Go to getbezel.com your Bezel Concierge is available now to source you any watch on the planet. It seriously, any watch. They got Rolexes, they got the GMT Master ii. The Batman. The precursor to the Batman payroll suite or the Batman.

You're really onto something there.

I think it's good. I think that's the future.

The Disney payroll.

I mean, Logan Paul. Logan Paul has a energy drink.

Why not Jake, Paul has deodorant.

Jake Paul has deodorant. Why doesn't Batman have a ERP system? Why not?

Or shack payroll?

Shack payroll would be pretty good. Before we bring in the Google folks, we should talk about Joe Wiesenthal. He says your neighbor is online. He thinks the BRICs are creating a new currency that will soon replace the dollar. He thinks Blackstone owns 60% of single family homes. He's long cardano and this is a response to Buco Capital. Bloke says your neighbor isn't online. He doesn't think about tariffs. His Honda CRV only has 80k miles on it. Good for another 100k. He doesn't have an opinion on the rating of our sovereign debt, only the only inflation he cares about is shrinkflation. Yet his 401k goes up just the same. I think the. The Blackstone owning 60% of family homes is the funniest meme. This whole idea of like, you know, like Blackstone's the most powerful financial institution in the world just because they index.

Everything, they're turning single family homes into pods.

They are.

All right, let's bring in, you know what time it is. There they are. Huge day.

How you doing?

What's going on?

Congratulations on today. What's new? Would you mind introducing yourselves and then just giving a quick breakdown of what's going on today over at Google?

Yeah, I can go first. I'm Tulsi. I lead the product team for Gemini.

Cool.

I work on some of the Gemini developer tools.

What are the top 3 announcements today? What's the most exciting thing? What is the. The one takeaway for the audience?

Ooh, I'll start with one takeaway. I think the one takeaway is like Gemini, at least from my perspective. Gemini 2.5 is just getting better. And it's not just getting better in terms of the quality of the models themselves, but we're now bringing that capability everywhere. So AI mode is coming to everyone, which is kind of our advanced version of search. Gemini is coming to glasses, which is going to be amazing. Gemini is coming in even better ways to Gemini app with more rate limits and just More access and better features.

I was going to have to spend glasses.

Is this like a return of Google Glass? Is this a new.

Yeah, so, so part of the issue for, for us today is that we got to watch like 20 minutes of I.O. and then we started our show. So we still have to catch up. But. But yeah, and then, and then it's. It's good. You mentioned rate limiting because I was worried I was going to have to spend this entire conversation moderating John, asking for higher rate limits from you guys.

Whatever you need.

Thank you.

Anyways, sorry. Tulsi, continue.

No, no, I think it's like an exciting time for us because I think the models continue to get better. We released a new version of Gemini 2.5 Flash today. We introduced like improved versions of 2.5 Pro with the deep think kind of deeper reasoning version of the model. And so I think the models are getting better, but it's also amazing to now see them actually really landing in Google products and in a very real way. I don't know what are your takeaways or top three are.

Yeah, I think my, my TLDR is Google's firing on all cylinders, across every vertical, across every dimension. And like I.O. is this like weird manifestation of this actually happening where you just like, I'm like, holy. Like even as someone who, who spends all day working with Tulsi, working with our teams and seeing this happen, like IO is this crazy reminder of just how much breadth Google has in the product areas and just how many people the model innovation touches. And it's. Yeah, I think today's a celebration of all that hard work and making all that happen.

Yeah. How do you think about the messaging from the continuum of Google's so big that consumers will watch a developer focused key. Some of the products that are launched are very consumer facing. But then there's obviously developers that want to build on top of these tools and then there's also stock analysts that are probably watching for how things are going on benchmarks or MAUs or any new data points. And so what are you focused on? Kind of messaging to the different audiences. Is there a concept of speaking to multiple audiences or do you just try and laser focused on dev specifically?

No, I think we really do want to speak to multiple audiences. I think you hit the nail on the head in saying there are definitely experiences that we talk about here at IO that are very consumer focused. We also know though that developers are trying to build for consumers. So when you showcase amazing consumer experiences, you are also able to speak to devs because you can speak to them about how you actually were able to build that experience and what they can do to build experiences like the ones that you're trying to build. Right. And so I think there's an art to try to figure out what are the right demos, what are the right stories that really can land cross audience. And then we try to create sections that are maybe more targeted. So for example, I spoke earlier today, Logan spoke in the developer keynote, really targeted at developers and what developers can do. And then there are obviously sections around Search or Glasses or Gemini app that are much more targeted at consumers.

Walk straight off the dev demo stage right here.

Yeah, it's true.

You said you could do two and you also said your keynote was starting at 1:30. I was like, are you worried somebody might, might ask an extra question in.

Multiple places at once? It's a skill.

This is amazing. How important do you think those demos are? Because you know, we saw ChatGPT break through with the Studio Ghibli moment. Everyone was memeing the Ghibli's. I was recently using VO2 to generate video of cars with specific liveries and decals on them driving around and the result was fantastic. But it hasn't broken through in the same way. Even though I feel like the model and just more importantly the accessibility. Because I was able to generate a video from a Gemini chat box and I didn't have to go to a different app for that. And I saw in the Apple App Store the Gemini promoted post is basically like VO2 is here.

And the output was so good that I showed it to somebody and they, they said he was. John had generated an image of a Lamborghini that had full TVPN branding on it. And I showed it to somebody and they were like, when did you do that? And I was like.

And they were like, that's right.

Was this how you made your promo.

Video is really just VO behind?

Yeah, no, that's our net. That's our next promo video.

Actually, also on that note, I think you probably haven't seen that part of the keynote yet, but we shipped VO3 today and it's awesome and it has audio.

Oh, that's very cool. Yeah, I was thinking about that because it was missing audio and I was like, this is the missing thing. I need to go find another model and then figure out how to inference that or whatever. But yeah, I mean in terms of, in terms of creating these like memeable moments or something that can go viral, is that something that's even Thought of or is it more just like turn it loose and hopefully.

Well yeah, I think the question is like you know, you want to distribute the capability and the question is like is that Google's job or is it the developers building on your guys platform to do that sort of last mile delivery to users?

Yeah, yeah, yeah.

I think the use case bit is, is actually really important because I think despite like we're, I feel like we're all sort of seeing everything that's happening and we're close to it but like there's so many people who actually have no idea and it's like every day I think we need like a tracker of like the number of new people entering the AI world every day. But like I imagine it's like some reasonable percentage of all people who are using these products and like you know, most products are still a blank empty chat screen somewhere. So like showing that use case of what's possible is super important. And I feel like a lot of the demos from IO today were like pushing that frontier in a, in a really meaningful way. But I think also to answer the question of like whose responsibility is it? Like I think Google, we get to do both, which is fun. Like we get to go and build a lot of these use cases for developers and make it possible for them to, you know, bring that to their consumers. But then we also get to be like dog food and like really walk the walk and build those consumer experiences ourselves and find a way to make a memeable moment.

Yeah. How much of what is announced are more like showcases that could be built into other products long term. Because I was looking at, somebody was saying like Google Assistant has less 5 star ratings than Gemini, but you could imagine Gemini being like a demo for what's coming in Google Assistant. But it seems like Gemini's just gotten so big that it's like, like bigger than Assistant now. And so you have this weird thing where you're experimenting in this fun sandbox but if you're really successful then people just start using that and then you just move off of the other thing. Is that like a deliberate part of the strategy or just like a natural outgrowth of like the modern AI development strategy that you put stuff out and if it goes really big you just let your winners ride.

I think it's a, it's a good problem to have. I think if we're in a world where the experiments we're trying or the new kind of surfaces we're building or the new types of use cases we're building Actually really resonate with users. And then we're in a place where we actually want to build those and scale those out. I think what we're trying to do. So at the end of the day, what we want to build is products that people love and really want to use. Right? And I think there's three ways you can do that. You can try to build new features into your existing products. I think in a lot of cases that makes sense. Right? So for example, like AI mode on search search today. But in some cases, if you do too much of that, you bloat the existing product, it becomes harder to use and understand versus being able to create a new surface where you actually can really rethink the ux, really rethink how users are supposed to think about the product. And then you can look at that and say, well, does this really make sense as a standalone or does it actually merge with other ways that users are using different products? Right. And it actually fits into their existing workflows in some way. And so I don't think we want to pigeonholes ourselves into just, hey, we need to consistently build on one product. Right. We want to give ourselves the flexibility to create ways for users to build new mental models and then figure out how those things come together.

And I think this search example is actually a great one because you, if you saw we announced Deep Research in the Gemini app, came out I think all the way back in December and it was like pioneered the Deep Research category and Google launched that in the Gemini app and people love it. It's, it's a crazy product experience. And then like that's now like, I think the, the, you know, the feedback was this is an incredible experience. Not everyone in the world is a Gemini app customer. There's a whole lot of people in search. And like now I think it's called Deep search is available as part of AI mode inside of just like basic Google search, which is going to reach hopefully even many, many larger, millions of more users than are using the triple.

Billion.

A couple billion, yeah.

Just a, A, just a b.

Talk about mcp. Was there, was there like a sigh of relief that Google's leaning into mcp? Was it ever a question or was it something that was like pretty easily adopted across Google?

I think it's been fascinating to see like how quickly the world has like, I feel like with standards, like people I feel like are often like very slow moving to be like, do we adopt this? There's like a whole, you know, many year battles with different standards competing. I Feel like everyone was just like MCP came out.

Everyone's like, okay, we'll use MCPCD. It's that xkcd meme. There's 10 standards. We need another. We need, just need one standard now there's 11 standards. But in this case we actually got the good ending.

It's important though, like I think developers are building stuff. I think in this moment where like you know, the world is still trying to figure out how to build robust agentic systems, like having people not spend all the time just rewriting the same framework 50 times over I think is actually a net benefit for the world. And I think the team that built MCP at Anthropic did a really good job and it was a smart move by them to actually make it an open standard which they could have easily not done. And I think it's the right thing for the ecosystem.

Yeah, it puts developers first in the right way.

Yeah. What is the overall messaging to developers? It seemed like, like Satya on stage at Microsoft's build was very much like we host every single model. We'll give you deep SEQ if you want. We're partnering with X. Google in my mind has always just been like, look, we invented the transformer. We have the biggest, best models. We're dominating the Pareto curve. Are you guys seen more as like a one stop shop or do you think there'll be more model agnosticism in the future with regard to gcp, I.

Think GCP already has this. So gcp like we do, like obviously you know, Tulsi and the team do a great job with Gemini models. We want the world to use Gemini but GCP does actually have like I think it's over 175 different models that are available through the model garden. You can get anthropics models, you can get a lot of models. So I think the, and like for customers on cloud like that makes perfect sense. They need that, like that's a requirement. Like the world doesn't want to, you know, be locked into a single model. But I think we get to see these really, really interesting experiences that can only be built on into Gemini because of the stack of Google's model innovation. So we're going to keep pushing on Gemini models too.

That's the goal.

Yeah. And then on the developers that might be working with Google, I imagine that the conversation is a little more nuanced if you're actually scaling a growing app on Google, Google infrastructure, like I ran into a rate limit as a random user, but if you're if you're actually thinking about scaling a startup, unless you're, you know, thinking about building Stargate, you're probably able to scale pretty, pretty efficiently. But have you been seeing companies run into any sort of like scaling or rate limits or challenges as they kind of grow on Google?

Email me. My email, my email's online. I'm on Twitter 14 hours a day. So whatever you need, email me. We'll make sure it doesn't happen. I think in general, like, this is like as far as like a class of problems that developers have in this moment. I think, like very real. Like I, I hear this and see people. The challenge is like compute is like a, it's like literally a currency almost. So like we have to make sure we get the right compute to the right people who are actually building stuff. And like it is a. It's not an easy process to do that. But I think if people need stuff, please reach out to the team. Actually reach out to Tulsi too.

We'll get, yeah, spread the love, you.

Know.

Don'T just create CX tickets for Logan. Talk about a couple of the other kind of product level announcements you guys had. Jules, an asynchronous coding agent. You want to speak to that a little bit? I'm curious to get your.

Yeah, we're really excited about Jules. So actually like at Juuls Google today you can sign up up for the beta. But basically the idea with Jules is we really want to start thinking through more and more use cases for how we can build essentially assistance for an agent of assistance for developers. Right. And so you can think about Juuls as trying to take aspects of the workload that are frustrating, that are time consuming and can actually operate on them in the background. Right. So the example I gave earlier today is Node JS being able to let update an outdated code base for the latest version. But I think it's really about how can you actually be able to assign tasks and let the model run and complete those tasks and come back and actually then have this relationship with the developer where these tasks can get completed, can come back and iterate together and build that collaboration setup.

And you can run multiple at once.

You can run multiple at once.

Classic Google fashion. You can sign up and use it for free right now, which I think is really cool. Like, I think part of this is if you look at what's Google's job in the world right now, like, I think part of this is like we have this stewardship role of making sure, you know, developers, people who are using search all etc. Etc. Like know what's happening with this AI technology and can actually get their hands.

I can try it.

Yeah, can try it and like know that it's available to them so anyone can sign up. I think the queue time for tasks is, is kind of high right now because everyone is flooding joules but hopefully as they scale up capacity we'll.

Yeah, I mean also and I think like what, what's been actually really exciting today is to see and get the feedback on joules because we've obviously been trying it internally and using it but actually getting it in the hands of real developers, getting that feedback, that's how we can actually a prioritize. What types of activities do we need to prioritize in joules? What use cases and features are most valuable for developers? How do we build that up? And so it's actually like a very self fulfilling prophecy I think for how this will go hopefully.

Yeah. Have either of you had a chance to actually sit and get the Project Starline demo? I remember a couple of years ago, this is the 3D. How would you describe it? 3D video conferencing.

Like video conferencing?

Yeah, yeah. And there were a whole bunch of videos from influencers who were there who got the demo and like blown away and they were like I filmed it with my camera and you can't really see the effect but you'll see my reaction. But it's exciting that it sounds like that's going into production in partnership with hp. But is there anything else to say there? Have you had a reaction or any test demos?

I'll be honest, which I had that same experience that the influencers had, which is I used the Project Starline room to take a meet call, but I was meeting with someone who wasn't in.

One of the other rooms.

So I have no idea. Like maybe I looked different or something like that, but they weren't, they were.

Like they were a normal.

So.

But it looks cool.

Like I, I, I'm hopeful.

Yeah, it's a very cool setup.

As somebody who works at home too much, I'm interested so that people can sort of feel my presence hopefully in the future.

Well, if you get in your car and drive, you're gonna miss all the DMS from developers that are, you know.

Need Logan to stay home with his Starlink.

Stay locked in.

Yeah, I mean we've been kind of in like, I mean I remember I bought a 3D TV like a, like a decade ago probably. And we've been Waiting for like, okay, yeah, 4K. We can't really see anything past 4K, but something like this would be very, very cool.

Notebook LM number four in the charts. Big launch, new mobile app did. I'm sure you guys never lost faith internally, but I think people externally were kind of had been frustrated because they liked the product so much. They felt like it just didn't feel like it wasn't getting kind of the investment. Any, any color to add there.

Lots of Notebook Outline progress. I think people sort of felt this, which is interesting. And I've had. We had a bunch of conversations around why, but it was like, the team is still cooking, like, they're still shipping a bunch of stuff.

I think that.

I think part of it was like, the expectations were just like super, super high now because a lot of the stuff they shipped was awesome, but they had a ton of new features that just landed and it's been awesome.

I think also this team is one. Even if you look at the history of Notebook LM, like, like, we actually shipped NotebookLM relatively quietly and then it got a lot of pickup because it was just such an amazing product to use. And so I think we've taken the strategy with NotebookLM where we've sort of tried to lead with the product leading almost as opposed to just trying to like, cook up a bunch of noise. And I think the NotebookLM team has just been silently making the product better and better and better.

And like, we are gonna have Tulsi cook up a bunch more noise for NotebookLM. This is a great way for you to do that.

More tweets is just.

Just firing off Notebook elements.

Well, it's good. I mean, the fact that there was like organic, real pull from the market and it wasn't just like, hey, we have a bunch of distribution. Let's make everybody aware of this right away.

Yeah. What was the original Notebook Elem? Because, I mean, there's research previews, there's developer previews, now there's alphas, betas. I feel like there's 25 different ways to say, like, don't hold us accountable to be launching this product. What is the actual, actual flow these days for just getting a product? I wasn't, Wasn't Gmail in beta for like a decade or something like that. This is kind of like in, in tech culture broadly. But I mean, a lot of things people are just like, if I can use it, I expect it to be a consumer product. But did, did, did Notebook LM go through like a particular pipeline or do you have Like a. A structure for these launches over there. Yeah.

Notebook LM came through Google Labs. Google Labs.

Makes sense.

Josh. I don't know if y' all follow Josh Woodward, but Josh Woodward runs Google Labs. And the Gemini app now is obviously generally available and not in Labs, but Labs does all the early stage bets. So actually AI Studio, the project that I work on out of Labs, same with the Gemini API, Notebook FM Jules came out today.

Some things we showcased. So we showcased on stage, we showcased Pinhole, we showcased Stitch, and with Jewels, all of these are products that kind of came out of.

And it's like break those down really quickly. Those other Labs products. Stitch and.

Yeah, so I mean, you were actually demoing Stitch earlier, right?

Yeah, I didn't. Josh demoed Stitch right before I got on stage. But Stitch, similar sort of vibe coding, vibe designing setup where you can go in and ask the models to like, go and build native mobile apps, I think is what it's actually building out of the box.

And then Pinhole basically builds off the magic of VR, sort of multimodal generation. So I think really what we're trying to do with Labs is, or Josh is trying to do with Labs, what we're trying to do with Labs is use it as a way to take these technologies and start testing what new surfaces, what new ux, what new types of products can you build. And then like, ideally, like NotebookLM, they're well loved. And we actually, like, can hit a moment in the market where we can find new product market fit. And that's really why we're trying to create this incubation engine. And then we can take those ideas and both bring them to our products, as well as actually ship them as standalone products.

And principally, they're also like very close to the model teams. So it's not this, like, you know, they're the. The, you know, they're collaborating with Tulsi's team. Like PMs are all sitting in the same places. So there is this, like, innovation at the model level turns into innovation at the product level. And I think those teams are like some. Some of the first recipients of that sort of stuff.

Yeah, it's interesting because, like, I feel like Gemini has a voice mode now that you can click over and talk to the LLM. And it's kind of like almost a separate feature path from just dictating and then having it read to you, even though that's a very similar auditory experience. You could imagine NotebookLM being like a feature inside the Gemini app, but right now it's kind of going down separate tech trees and it kind of is.

So we brought Audio Overview and Deep Research. And like, I think the interplay between these things are like all kind of connected. I think there's like some like very bespoke Notebook LLM stuff that you'll like, probably not have in the Demi app. I think the stuff that works like, there's no reason to not. It's like all, it's all under Josh's umbrella. So like he gets to easily be able to bring those.

Yeah.

And I think like the underlying technology, like Logan said, we're trying to make sure that the model teams and these products are very close together. So for example, Notebook LM has now native audio, which is one of the things we talked about today, which is really like Gemini being able to natively generate audio, which makes it more multilingual, friendly, more natural. And so that kind of technology a we can use to make NotebookLM better. We're also bringing that to Gemini live. We're bringing that to other parts of the Gemini app over the course of the next months. And I think that kind of synergy means that you're going to get get the best of these experiences depending on your use case in different places.

Yeah. It's an interesting product challenge. I've run into weird scenarios with ChatGPT deep research. I'll generate some huge research report and then I'll be like, listen to like read this to me. But it clearly gets lost once it's reading halfway through and starts hallucinating and stuff. So like there's some interesting challenge to actually make all of this. Like if you generate 5,000 words of text, actually getting it into the human's brain, it's not enough just to be like, here's your research report. Like you actually have to make it accessible.

Deep Research audio Overview, I think is that flow like.

Yeah, that makes sense work.

It's like, it's actually awesome because the Deep Research reports are like 46 pages.

Yeah, exactly. I don't have time for that.

46 whole pages.

46 pages.

Like give me a one pager or give me like three minutes of someone like, you know, having a casual conversation.

Give me a YouTube short generated by VO3. That's what I need. Yeah, that makes sense.

Eventually you two are just going to be an audio overview with your, you know, your own taste.

Yeah, yeah, yeah. We're not letting Notebook LM disrupt us. We're coming into the show completely unprepared. But we listened to a 10 minute notebook LM on the thing we're about to talk about, so we have a little bit of context. Anyways, you guys have been shipping like, like crazy. What can people expect from you now? Are you gonna rest on your laurels? Are you gonna, you know, you know, just get a little overconfident? Or is it, you know, businesses back to the grind?

A little overconfident, I think will be good because I think, like, we've done a lot of.

There's.

There's just been so much stuff that the team has done. So I think, like, people should be proud of all the work that's gone into it, but I think the work just starts now.

I was just. I woke up this morning and I.

Was like, I'm writing the doc about how we need to go heads down.

For the next six months and land 50 things that need to happen because.

There'S, like, so much stuff that's still cooking in the pipeline, and it's going to be awesome to see all that.

Yeah.

I think also, like, we're really. I really do think we're hitting our stride in so many of these areas.

It feels like it.

Even.

Even some of the posts, like, even Sundar's posts, like, over the last 24 hours, I was like, he's feeling good going into IO. You could tell. You could tell. So the energy. The energy is palpable.

Yeah. And so we want to. I think we want to use that energy. Like, I think. I think the momentum coming out from today is really about, like, okay, how do we take the excitement, how do we take the energy, how do we take the feedback and like, actually really use it to continue to just build awesome things. So I think we're just at the beginning, like, this actually really is the start of. Of what's going to be an exciting rest of the year, I think.

And we'll make sure that you two are in person for the next.

Oh, yeah, we should.

Let's do it.

More than 24 hours. Heads up if you want to come in person.

Yeah, yeah, yeah, yeah.

We calendar for these things.

Yeah. Amazing. Well, thank you both for making time to come on. It's really great to hear and congratulations on all the progress. We'll do it again soon. Cheers. All right, have fun, guys.

Yeah.

The interesting thing is they are cooking.

They're cooking. I mean, yeah, they delivered in a bunch of different ways. And I mean, obviously, like, you know what, like, tech Twitter is not real life in many ways. And you see that with the. The. The Gemini Mao's and Dows and Just five star ratings and stuff. But I mean you saw it with the, the video I generated. Like, like the models are really, really good.

They're really performing really well. And I feel like Google has been, there's so much chatter on Twitter, people making fun of, oh, I was in Gmail and I asked it to do this.

Yes, some of the products are Mrs. For sure.

Yeah. But, but clearly they're aware of deficiencies and like rapidly shipping on a bunch of different dimensions and that, you know, the energy they even are bringing to this call and the fact that they're going, even just going direct.

I think the really interesting thing is like super bullish. Marc Andreessen had that quote about like OpenAI is becoming Google. Google's becoming OpenAI. You remember this? He said this to us and I think his, his take was that OpenAI is becoming a consumer tech company and Google's most advanced in research. And I think that's going back to the Yahoo question we were talking about. If there is they invented the transformer. If there is a new architecture and it does come out of an academic lab, it's totally possible it comes out of DeepMind. Again, they did invent the transformer there and Ilya just kind of picked up the paper and was like, this is amazing. We need to implement this and go really hard at this. And then, and then everyone did. But it would be very interesting to see like we didn't even get to the diffusion language model, but what other foundational innovations they're trying because they've been ahead of the curve on getting bigger context windows, faster inference on that Pareto frontier. And if there's a new paradigm and they implement it quickly, that could be an entirely new thing. Very interesting.

Yeah. Overall, Google has been so successful for so long and so dominant that I think that it creates this effect where people almost want them to lose or want them to get taken down a notch in some way, but they're have.

To go through us because we will take a bullet for big tech.

We will. We would. We will. We would.

We won't. Because we won't have to because you will never challenge the Supreme Court.

Yeah, but Google is one of the greatest high technology companies in history and it deserves to be celebrated. And we need to just put this phrase to bed. But founder mode, like Google feels like at an individual level people are taking massive autonomy and really have a lot of momentum. And Even I'm on jules.google.com right now and they're rate limited, but it looks and feels like the product that a startup would create.

That's good.

It looks like they just raised like 300 on 3 billion. They probably did from internally. Internally. Well anyways, we should get out of here. I did get some extra context on the Coinbase issue. A friend of show said re Coinbase hack. He was sending it live while we were in that. He said the information that isn't the customer data is sold at a granular level, I. E. Info on the individual. So it'll go on a dark web market and be sold individually based on the bitcoin balance of the individual. Think of it like a standard three tier regular. Silver and gold would be a balance over 50k. And these accounts are are already vulnerable to like sim swapping and things like that. So anyways, appreciate the intel. The Lone Ranger shot that over to us.

Very helpful. But last thing, congratulations to Tyler Cowan, a friend of the show who has made the Time magazine list of the 100 most influential people in philanthropy. So congratulations to Tyler Cowan and we will see you tomorrow.

See you tomorrow folks.

We'll see you on Wednesday.

Fun show.

Bye.