OpenAI’s $1T Buildout, Trump–xAI Alliance | Delian Asparouhov, Garrett Langley, Matan Grinberg, Francis Pedraza, David Paffenholz

You're watching TVPN.

Today is Thursday, September 25th, 2025. We are live from the TVPN Ultra Dome, the Temple of Technology, the Fortress of Finance, the Capital of Capital. Good morning to everyone in the chat. Good morning to Gold Rock, Ragav, John, Techno Chief, of course. The post of the day is Alex Heath in Sources. I printed it. He says, Scoops. He's scooping.

Scoop.

He says OpenAI is planning to bring ads to ChatGPT. Also, in a message to employees, Sam Altman says he wants 250 gigawatts of compute by 2033. He calls OpenAI's team behind Stargate a core bet, like research and robotics. Doing this right will cost trillions, he says. Well, hopefully he can save time and money with ramp.com. Time is money, save both. Easy to use corporate cards, bill payments, accounting, and a whole lot more all in There we go. But I was thinking about what Sam Altman's projections have to say about the build-out of AI. We're in pretty crazy times. Certainly lots of people are comping it to 1999. Lots of people are saying 1996 if they're bullish.

Absolutely.

The exact year matters a lot. I mean, even 1999, I feel like there were still—

Some people are saying this time is different.

Still 18 months. But the technology is even more magical than the internet potentially. And of course the internet already exists, so it's easier to scale revenues as we've seen. And so my take on this was the AI deals are getting crazier. Obviously this week we saw Nvidia is going to invest $100 billion in OpenAI, which in turn will be used to buy Nvidia GPUs. But we aren't fully in 1999 mode, even though a lot of people are saying that, because big tech just has so much cash to fund the build-out. And that wasn't quite what was happening in the dot-com boom. In the dot-com boom, there was a lot more leverage. The price-to-earnings ratios were crazy, 10x higher on median. And so we're not in the crazy, crazy debt era. There's a little bit of that going on with Oracle, but it's not this huge— Yeah, if you're— The debt-to-equity ratios aren't Insane yet?

Yeah, yeah. Certainly players like Meta have tapped the debt markets, right? They did an offering which blew out, but it is a small, you know, it's a, it's a some minority percentage of their earnings.

Exactly, exactly.

Yeah, yeah.

It's—

Oracle went out this week as well for a $15 billion convertible debt offering. Yeah, offering. I'm sure that'll get filled quickly again because it feels relatively low risk, even though they're forecasting well beyond what a traditional public company would be doing.

Yep. But the nature of the deals is getting more unique. It's different. I mean, we're seeing letters of intent now, which is just a very fuzzy term.

Or investments being announced. that are still in discussion.

Exactly.

Right?

Exactly.

This NVIDIA OpenAI investment, if you read the print at the bottom of the—

It was a non-binding letter of intent.

Yeah, they're basically saying, we're talking about this.

Exactly.

This is like the—

But we haven't papered it yet.

—framework for the deal.

Usually, you go the opposite direction. You get the leaks, then you get the actual announcement once the deal is fully signed. The deal signs, and then the press tour starts. But now we're putting the press tour a little bit earlier. And I keep coming back to YC Demo Day because At YC Demo Day, the letter of intent is often pulled out for companies that don't quite have the ARR or the DAUs to justify whatever they're trying to fundraise. And I wanted to revisit a story from YC Demo Day, my favorite YC letter of intent story, probably Boom Supersonic. Extremely— we've had Blake on the show multiple times— an extremely ambitious project to bring back commercial supersonic aviation, basically Bring back the Concorde. Boom was one of the first hard tech companies to go through YC. There were a few at the time. This was the Sam Altman era, again. And they didn't really fit the current mold. They couldn't show a graph of DAUs or ARR, obviously, because they hadn't even built planes yet. They were building plans to build planes. There's only so much you can do in 3 months.

It was a founder bet.

Yeah, it was a founder bet. And Blake also—

This guy was a product manager at Groupon.

Yeah, exactly.

He's got what it takes to take us back to hypersonic.

Yeah, yeah, so it wasn't—

And it sounds like I'm joking, but I'm totally not when you see Blake's level of conviction and his inability to not go supersonic.

Yeah. But it wasn't like he was like former Lockheed and was spinning out some technology that immediately could just like piggyback on the shoulders of giants immediately.

Yeah.

Like SpaceX has been out.

He had, yeah, exactly. He had to really build stuff from scratch and put the puzzle pieces together. And one of those puzzle pieces was a massive LOI. And so, He stood on stage and said, you're going to think this is a sci-fi project. You think this is a moonshot project. But the one thing that I can tell you is that I have an LOI. It's non-binding, but it's for $2 billion. And so that was what went on the screen, I think, around there. And he had gone to Virgin, talked to probably Richard Branson, and Virgin had optioned 10 airplanes in a non-binding LOI. And the deal value was roughly $2 billion. And so if it was fully exercised now—

And I think people generally believe this is the kind of business, it's like if you build it, there will be demand.

Exactly.

There will be demand from the airlines. There will be demand from consumers.

But having the LOI, like, understand—

Means something.

Yeah, it means it shows that you can go in that room and convince them that if you build it, they'll buy it. And then that leads to investors saying, okay, well, I will back you. And so Boom's had ups and downs. They had a down round. A lot of YC folks came back in, invested in the company again just to keep it going. And Blake's still at it. He's getting regulatory approvals, which we saw, and building planes one at a time. He's actually flying. And as loose as non-binding LOIs can be, they still help pull forward the future and can be a useful tool in making everyone around the table feel like there's mutual buy-in. 'Cause everyone's at least said, if this happens, I'm in. If that happens, I'm in. And if everyone commits, and if the capital partners commit, and the customers commit, and everyone commits, then the thing should actually happen. And so, the Boom story for me underscores how unclear the path of technological development can be. Sam Altman is now guiding toward 250 gigawatts of compute at OpenAI by 2033. which is a staggering number. We saw a post, which we'll get into, a single gigawatt data center is many football fields long, visible from space. It basically terraforms the Earth. And while many people are writing off Sam's projection as merely attempting to win the contest of saying the biggest number, his projections aren't entirely new to the world of AI compute forecasting. Leopold Aschenbrenner predicted a single training run cluster around 100 gigawatts coming online by 2030. So, 3 years earlier, 1/3 the size just 2 years after. And then Leopold predicted in 2028 that there would be the first 10-gigawatt cluster. 250 gigawatts by 2033 lines up with that. But wow, America will look dramatically different if that comes to fruition. In 2023, for reference, 2 years ago, the US was producing an average power in total for everything, not just AI. of 485 gigawatts. And so Sam is saying that just OpenAI, if everything else stayed the same, would be 1/3 of all power consumption in America for that entire year in 2033. Now, obviously everything else is growing and there are other labs, and you would imagine that Amazon and Anthropic keep up and Google and DeepMind keep up. And so it will be an entirely new chapter of the economic story.

What are some of Elon's Promises again, or goals.

Yeah, I was talking to Tyler about this because I— so I mapped out the power projections. This is straight lines on a log graph, of course. Leopold is the taller bars here. Jordy, you can see this. And Leopold is projecting logarithmic growth, which is exponential growth. This is every tick on here is an order of magnitude bigger for gigawatts. And you can see Leopold is projecting A straight line on a log graph, basically exponential growth continuously based on how he's mapping out the few dates that he gave. He didn't give projections for every single year. SAM is actually— it's so funny because the SAM story is framed as like 250 gigawatts by 2033. This is so insane. But he's actually sort of guiding towards deceleration. You can see SAM's Sam's darker bars here.

He's being humble.

He's being humble relative to—

He doesn't want to, you know, overshoot.

Yeah, relative to some other folks in the category. So, Tyler, yeah, what did you find? Did you see anything from the other labs?

Yeah, so I was trying to figure out like what all the other labs are doing, like what they look like on that graph. So none of them have done real kind of projections. Like no one has said like, okay, by 2035 we're gonna have 250 gigawatts, anything like that. So for Meta, Zuck, the big number that you could use for this is his, at the White House dinner, $600 billion through 2028.

Yep. So it's an aggregate number, but they're only doing 50 or 60 now. And so to get to 2028, that's only 3 years away. So you have to do 50, 100, 200, 400, something like that. It has to ramp really fast.

Yeah, I was trying to figure out what does that actually mean in gigawatts? Like, what is the actual price that you could— how many could you get out of this?

Convert the dollars to wattage.

And so basically, I used— I looked at Colossus 2, which is maybe the 1 gigawatt one. And then estimates are— this is very non-accurate. This is totally guessed. But maybe somewhere around $15 billion to build that.

Okay. Order of magnitude, $10 billion something, $10 to $20 billion. Let's keep it super vague. for a 1 gigawatt.

Yeah.

And so—

So with $600 billion, you can do like maybe like around—

60 gigawatts, 50 gigawatts.

Yeah, like something in that.

If you were to just copy and paste a bunch.

Yeah. But that's, I mean, that's like Elon, like—

Elon precision.

The most efficient, like this is the fastest of all time. Like it took them like 6 months to get to 300 megawatts, which is like double the next fastest build out, which was the Crusoe OpenAI one.

Yeah.

So I think it's probably— that's a little aggressive.

And Zuck's number was at the end of 2028?

Yeah, it was through 2028, I believe.

Yeah, I mean, the end of 2028 for Sam and Leopold, they're more thinking around 10 gigawatts, I think. So interesting to see if Zuck's projections kind of outpace what's there. I don't know.

And you guys saw today CoreWeave and OpenAI expanded their contract to a total contract value of $22.5 billion. $22.4 billion.

$22.4 billion.

Okay.

$6.5 billion. So—

It's a deal for ants.

Just praying and praying.

Yeah.

Another thing you could use besides Leopold is the AI 2027 numbers.

Yeah, I was wondering about those.

So if you look at those, like, once you get to 2027, after that, the numbers like explode. It's like, oh yeah, there you go.

That's the fast takeoff.

It's like $30 trillion on data centers.

Yeah.

But through 2026, 2027, it's like around like $500 billion a year. And that's like $500, I think it's like $650 in kind of the end of 2027. So like, you could imagine we're like reasonably on track for that. Yeah. Considering Meta alone is doing around $600 over—

Yeah.

Through 2028.

Yeah, if you aggregate all the labs.

Yeah. So I think that's like fairly reasonable. So yeah, it's hard. I mean, people are like, if you look at the timeline, everyone is like, okay, we're at the top. Like, this is so bearish.

Yep.

But if you look at a lot of people who are like super bullish on AI, we're still Broadly on track, I would say.

Yeah, this is what Rune was saying, how everyone, all the richest people have voted, and they've all said, we're going all in. We're putting all the chips on the table.

It's not about immediate ROI.

Yeah.

And I mean, honestly, if it's funded with cash and equity and not too much debt, there's a world where you take the hit of maybe some stagnation, and you get through, and you are more of the Amazon story than the— NextStar story, or I forget what it is. I always forget the names of the companies that went away.

Even in the bear case, it's like, there's the book Boom by Byrne Hobart, which is like, bubbles are good because broadly, they don't affect the normal economy because it's like all the—

It's the most risk-on. Yes, the interesting takeaway from Boom was that the common narrative is that retail gets hurt the most in a bubble, but in fact, it is not retail. It is the wealthy that get the biggest haircuts because they're the most leveraged at the peak of the bubble and they have the most wealth to compress. Whereas the average American through the number of bubbles that Byrne Hobart analyzed is not super levered long the most risk-on asset at the time of the top. Jordy, what were you gonna say?

So there's a company called iREN.

Oh, yes.

The company pivoting from Bitcoin mining to being an AI data center company. Their stock is up tremendously. Andrew Wilkinson hit the timeline a couple months ago at this point, I think, and basically said— I read it as financial advice, but he was basically making his case.

Yeah.

The stock is up massively since he came out and made the case for why he thought they were undervalued. The current share price is $48. It's up Let's see exactly. It's up from around, I guess, $17, $18, I think, when he recommended it. So more than double, more than triple. So it's up massively. He's now calling for another 7x because he says iron is about to control more power than the Hoover Dam, 2,910 megawatts, so 2.9 gigawatts. And so anyways, he's now calling for the company to— yeah, again, 7x again. He thinks that $300 per share could be a layup. He lays out a potential bear case, right?

Okay.

Given that he says, if AI turns out to be a nothing burger and nobody wants Iron's power, land, data centers, and servers, that could be bad. Unable to if they can't really capitalize on the demand because they just don't have, they're not, the data centers aren't configured to the level that they need to be to be great assets for either hyperscalers or other AI companies. So lays out some potential downside scenarios. But I do wonder if we'll see some of these companies that are trying to take a crack at actually competing with the hyperscalers as like cloud—

Mm-hmm.

serious cloud providers. I do wonder if some of these companies will get to the point where they have a lot of energy, but they basically just have to sell to another more sophisticated player that can actually take over and turn it into— help it realize its potential.

Yeah. The chat is asking for an update on TBPN merch, if it will be available for our most loyal soldiers in the chat. I think we got to set something up.

Yeah, we got to figure out how to make this happen. The first run of merch, we didn't want to sell it. We had never made merch as a company before, and it didn't feel right to sell something that is not our core competency. But we had a meeting yesterday talking about the next run of merch and getting that available by the beginning of this coming year. So stay tuned, and thank you for your patience.

We got to get Restream on some merch. One livestream, 30+ destinations. They make our stream possible every day. Multistream and reach your audience wherever they are. I had a friend that asked for some advice. If you're making money during the bull run, what should you do if you feel like you're at the top? What's the safest portfolio for a young person these days?

Rotate into watches.

Watches.

Go entirely into Swiss horology.

I mean, if your comparison is NFTs, Yes, the watch market did actually draw down post-NFT crash.

And there were people that converted NFT earnings into watches and they probably did better than—

I feel like the— what was it? The Rolex market sold off by like 30% or something. But I mean, a lot of NFTs went to zero.

Still above retail for most watches.

But yeah, I mean, it feels like there's the existence of a risk curve and The companies that are trading at the highest, you just have to go back to the basics of the Warren Buffett stuff of like company with no earnings trading at tens of billions of dollars. That's gonna be the most dangerous. They will be most punished in a down economy. Whereas a hyperscaler that still has enterprise licenses and a huge ad business is gonna fare better even if they get a haircut. We saw this with Meta during the last cycle. I mean, they traded down a huge amount, like almost 50% off the metaverse highs. But built back up pretty quickly on the next wave. And then, of course, you got gold, you got Bitcoin, you got real estate, you got plenty of things. But what would you counsel someone in their 20s making decent money to do with excess earnings at this point in time?

Excess earnings.

I don't know. I think that the young millennials, Gen Z, are overly obsessed with investing.

Mm-hmm.

It's all that a lot of people can think about, and it ends up being a massive distraction from just getting better at your craft, your career, and just increasing your earning power. So I think the number one— I mean, I talked to the guys on the team about this. We're obviously talking about different private companies, different investments all day long on the show. And so it's easy to get kind of fixated and get kind of FOMO around this stuff. But anybody knows if you've made a sizable investment or If you've made an investment ever that's like double-digit percentage of your net worth and it's trading liquid, that can be just such a—

Emotional rollercoaster.

Yeah, just such a massive distraction from what you should be focused on in your 20s, which is just increasing your skillset, building your network, figuring out how you can increase cash flow. I don't know. You had a—

Well, yeah, my take was there's huge alpha in not even picking the correct asset, but just setting up your life so that the money actually gets into the asset before you can spend it. Yeah. And so, when I started making money in my 20s, every time I would get paid, I would go to the bank. I would take out something like $1,000 of cash and put it in a safety deposit box. And that just built up over time. And I realized that if I was out on the weekends, I could not access it because you'd have to physically go to the bank to get the cash. And so, I realized that if you're out with friends and they're like, oh, should we spend all this money on whatever? If you have it in your checking account, you're like, yeah, no problem. If you have it in your savings account, it's supposed to be in savings, but you can just click one button and move it over.

I remember being at a teenager and realizing that your savings account was literally one transfer away from checking.

It's ridiculous.

Wow, this takes a lot of willpower to actually make it savings versus spending.

Yeah, exactly. So if you build up a literal hoard of treasure, A, it's incredible psychologically because you open up that safety deposit box and you can physically see the wealth growing, which is incredibly enriching. It's incredibly satisfying and it makes investing Very much addicting. But even just going into your payroll provider and route— usually you can route your paycheck to some sort of investment platform, maybe Public.com. Investing for those who take it seriously. They got multi-asset investing, industry-leading yields, trusted by millions. And then you can set up automated buys on ETFs if you like the Mag 7, if you like gold, if you like Bitcoin, whatever you want. You can have automatic investments happening so that you're not spending all your time trying to become like a hedge fund guy, unless that's actually your life's work.

Ideally, the service that you would have liked is somebody you go and you give them cash.

Yeah.

And they even— they use movie cash to give you— show you the pile. They take that, they actually invest it.

That's a good service. But to access it, you have to go—

You can go look at the cash. You can— you can—

The physical key was really, really great. It was—

it was—

I've never had one of those.

Yeah. Safety deposit box is underrated.

Underrated. Do they still make those?

They do.

Yeah.

Really?

Available at pretty much every bank.

Have a little—

no, no.

Like a mouse.

You have a little— It is somewhat rodent-coded to just be hoarding. Dragon-coded, potentially, to hoard your pile of gold.

Or could kind of cool and mysterious if it's all over the country. Like you've got a little stash in Dallas, keep some in the Bay.

What do you think, Tyler?

I actually know a lot of friends who have a Bitcoin wallet, and then they have the seed phrase in different safety deposit boxes all over the state. So they have to drive— it's like 4 hours to drive to every single bank.

That's diamond hands, baby. They're not paper-handing anything. And so if they're just continuing to contribute to that— the trick is that the cash flow typically comes in monthly or every other week installments. How do you actually set up that so that it's happening on a regular basis but still inaccessible? That's a little bit tricky. Maybe someone can build it. Maybe someone can use Privy, our wallet infrastructure partner. Wallet infrastructure for every bank. Privy makes it easy to build on crypto rails, securely spin up white label wallets, sign transactions, integrate on-chain infrastructure, all through one simple API. Barely AI says, looks like Nvidia and OpenAI will be paying zero investment banking fees for their $100 billion partnership. Summit. Hit me with the wah wah wah. So sad for the investment bankers. We talked to Kerry No Interest.

He was telling us Wall Street in, cut Wall Street in.

Come on, break us off. Break off Morgan Stanley. Break off Goldman Sachs.

Those guys need to break. It's really impressive. I mean, it's just these sound deals at this scale is very cool.

Yeah. I mean, that's what everything is. We talked about this with Nvidia spending A few million dollars lobbying in Washington, but Jensen's there. And founder to president, founder to founder, that's the nature of these deals. It's always the nature of these deals.

You got to price in the Gulfstream expenses with the lobbying fee.

You can tell more about how a company is doing in DC by where their Gulfstream is parked than their lobbying budget, for sure. Where's the Gulfstream parked? Has it been at Mar-a-Lago? Has it been in D.C.? That's what's moving the needle these days. This screenshot says Altman and Huang negotiated their pact largely through a mix of virtual discussions and one-on-one meetings in London, San Francisco, and Washington, D.C., with no bankers involved, according to people close to the talks who declined to be named because they weren't—

This was a multinational deal.

It was. They were going all over the world. London, they were probably talking about it in white tails. Remember, they were in white tie in D.C. Who knows? They might have been chatting to each other. I don't think they were both at Hill and Valley, but they might have been there overlapped at something or other. Maybe the inauguration. The arrangement calls for Nvidia to invest $10 billion at a time in OpenAI.

$10 billion slugs.

They have $60 billion in cash and cash equivalents right now. So they can, without generating any more free cash flow, they can just keep paying that.

Would, if you were Jensen, would you Would you lever up massively?

Maybe.

Become just entirely indexed to each other, right?

I mean, it's happening.

Say, if we go down, we're going down together.

There's more news about the AI keiretsu that's coming together. OpenAI partners with Oracle and SoftBank for 5 new US data centers.

7 gigawatts is And what's crazy is they're still, while they're doing this, they're doing deals with Broadcom. They're doing deals with Oracle, with CoreWeave. CoreWeave.

Microsoft still. I mean, they're buying from everyone. No one wants to be GPU poor in 2025 going into 2026. You want to be keeping the GPUs cool, not on fire. So Wall Street Journal says OpenAI laid out its vision for a vast $1 trillion—

I think it's better if— if you're scaling GPUs but keeping them on fire. If they're not on fire, partners are gonna be looking and saying, hey, got that.

Yeah, you don't want cold GPUs sitting there idle just eating a hole in your balance sheet. You wanna be maxing those tokens out, getting paid. Showcasing the development of a Central Park-sized complex 180 miles west of Dallas. This is from 2 days ago by Berber Jin in the Journal. OpenAI unveils plans for seemingly limitless expansion of computing power in Texas prairie. Startup showcases ground zero of AI boom and its plans to shepherd $1 trillion in infrastructure spending. So the trillion-dollar number is getting thrown out, thrown around for the first time because Stargate was originally $500 billion and now we're up at $1 trillion. Is that right?

Yeah.

And remember, even at $500 billion, Elon was basically saying, hey, nobody involved here actually has the cash to do this.

So wait until these stocks run up and Oracle's up $300 billion.

Larry says Hold my beer.

I'm good for my ten.

Hold my green juice.

I'm here for. I'm I'm I'm good for my half trillion.

But yeah, once you get into the trillion dollar range, you're approaching the GDP of you know some some serious countries, right?

Serious countries. OpenAI disclosed it would ultimately need more than thirteen times the computer the computing power of its first nascent site, which is raising which is rising out of the Texas brushland. Frenzied construction here has turned a sea of red. dirt into 8 hyper-futuristic data centers, bringing online roughly 900 megawatts of capacity. More than 6,000 workers labor on the project each day, including electricians, plumbers, and steel welders alternating between 2 10-hour shifts, 7 days a week.

Wow.

They are— they're— what is that? 2 to 10, 7. 2 to 10, 7. 2 a.m. to 10 p.m., 7 days a week, something.

Yeah. People in the data center business see people with computer jobs saying, oh yeah, I do 996.

Yeah, I mean, I wonder if they have 2 shifts, 10 hours, does that mean they have 2 crews that are working 70-hour weeks? Are people taking days off or are they rotating their shifts so that you get a day off every couple days?

I think they would get into pretty hot water. I don't think you could like run and say you're working 7 days a week forever. I don't think employment law works like that.

Gray towers of gas turbines have dotted the landscape since the spring, offering backup power. On a tour with reporters Tuesday, Oracle and OpenAI executives showcased the 1,100-acre site, calling it at the largest AI supercomputing complex in the world. Outside, workers wearing dust masks and sunglasses shuttled around in buggies, trying to shield themselves from the 100-degree heat. One of the buggies flew a flag that said, Jesus is the answer.

1,100 buggies.

Not very AGI-pilled. I thought we were getting God in a box. What's going on? Just Eliezer Yudkowsky shows up. He's like, No, no, no. Like, they're building God here. You're building God. You're doing the thing. Like, what are you doing here? There was literally nothing here a year ago, said Anuj, who works on the OpenAI computing team. They also announced 5 new data sites, data center sites across the US built with Oracle and Japanese tech conglomerate SoftBank. It said the new facilities would help bring online nearly 7 gigawatts of power, enough for almost 8 million homes, enough for a lot of API calls from Devin, from Cognition, the makers of Devin, the AI software engineer. Crush your backlog with your personal AI engineering team. Company executives made it clear that the Abilene site was just the beginning, noting they envisioned a need for more than 20 gigawatts of computing capacity to meet the explosive demand for ChatGPT, which now has more than 700 million weekly users. I was laughing about, you know how in ChatGPT, if you fire off Pro, like GPT-5 Pro, it kind of thinks for like 20 minutes? And everyone's saying like, oh yeah, this is good. Like, we want the agents to be reasoning for longer and longer. I feel like there's an opportunity for some startup just to be like, we've cracked the code. Our model reasons for 2 days straight. So send us your request and we'll get back to you in 2 days.

And the agent just like is pretending. So it's just like twiddling its thumb.

Yeah, it just has a progress bar.

Twiddling its thumb until last 20 minutes and then just starts going.

No, no, no. Of course what you do is you have the loading bar there and then secretly you have a human doing the work and you have a human that's like dealing with this.

Well, I think the thing is like the longer the period, the higher the expectations. And I just don't know that— I think that, I guess the thing that I would want to know is like if you're running a deep research query for 20 minutes and then if you decide to run it for 2 days, How much better is it, right?

Yeah.

If it's 10 times better, could potentially be worth it for different types of tasks. If it's 30% better, is it worth the wait?

Yeah. There is a world where you just route it to the human. There's also the world where for the really low—

Don't encourage this. People have already gotten into trouble.

Yeah, they have. They have. The other funny idea is for a lot of the There have to be a certain set of queries that just get asked the same every single day, and they should just store those in a database and cache them. Because you have to imagine that every single day someone is asking, just going to ChatGPT and lighting the GPUs on fire to just ask, what's the capital of Illinois?

Or what's the capital of California? Can you run a GPT like pro deep research report for what is the capital?

You can do it for anything. Yeah, you can, you can go to deep research and fire off 20 minutes of compute and say, what is the capital of California? Answer in just one word. And like, it will just think for 20 minutes and then spit out Sacramento.

Like, it won't think that long. It like thinks based off how hard it is.

No, deep research will—

Yeah, I just said, what is the capital of California? Answer in one word.

And it's deep research.

Yeah, it's doing a deep research query.

Let's see how long this takes. It's not going to come back in 2 minutes. It's going to cook for 10 minutes.

All right.

I was a little bit late, but I just started a timer.

Okay.

What do you think?

Also, there's a sense where the embeddings of facts is basically just the model already. You can just think of a model as a compressed version of the internet, or at least non-reasoning models.

Yeah.

So it's kind of like—

But wouldn't it be even more compute efficient to just do fuzzy lookup of— of these questions and see if they sound—

Siobhan in the chat says cache-based retrievals are—

Oh yeah.

But apparently not here because GPT-5 Pro—

It's not fully implemented.

Okay. Okay. It reasoned for 47 seconds.

Okay.

And it—

this is, this is Pro. You didn't do deep research though. I wanna see you turn on deep research and do it.

Well, this is the Pro, which is research-grade intelligence.

No, no, no. That's not the deep research.

Are you thinking?

No, no, no, no. You need to click on the little plus button. next to the chat box and then check the box on deep research. Do they still have this? I don't know. They might have deprecated this. I think it's still there.

It's still there. Still there.

Okay.

I'm doing a deep—

you're doing deep research.

No, no, this is good. This is good.

Really verify.

Just to confirm, are you asking for the current capital of California or are you interested in historical capitals as well?

Yeah.

Let's see. Current capital.

Tell it to also design a a website in Figma. Think bigger, build faster. Figma helps design and development teams build great products together. Get started for free. While we're waiting for that, let's keep reading from the Journal. Each gigawatt of capacity is expected to cost roughly $50 billion, Tyler. $50 billion per gigawatt. That's the number in the Journal, meaning the company is laying the groundwork for at least $1 trillion in infrastructure spending. Demand is likely to reach closer to 100 gigawatts, one company executive said, which would be $5 trillion.

Which is roughly the GDP of Germany.

Or Japan.

Yeah, GDP of Japan is $4 trillion. Germany is $4.6 trillion.

Good luck. Good luck, Germany and Japan. You're getting left behind. You're part of the permanent underclass now. It's over for you.

I feel bad for lighting the GPUs on fire for this. It says It's cooking. Got it. I'll confirm the current capital of California for you and get back shortly.

And it's doing people a favor. It's cooking. Like, this is the real pitch for the—

I'm getting my money's worth.

Brandon Jacoby in the chat says, breaking, Jordy discovers the UX issues of every foundation model company.

Fix it, Brandon.

Yeah.

Yeah. Aren't you doing consulting now, Brandon? You gotta get in the trenches.

Yeah.

All these folks. If the labs haven't hit you up If someone who spends 10 hours a day studying this stuff—

Just reading Wikipedia.

Just like, how many drops of water did this use? How many fish had to like go find some other home?

Extinct. Yeah, it should come back and it's like, thought for 15 minutes, put— made 2 species extinct, lit 7 gallons of diesel on fire. Okay, it's checking all the sources.

It's good. I like a detailed report. I like to know that The AI did its job. Don't just sit there, Clanker, and riff off cache-based retrieval. I want the deep research every single time.

Even for one-word facts.

Even for one-word facts. I want you to be sure. I don't want any hallucinations.

Well, I mean, this is a tough question too. A lot of people that aren't from the West Coast or the United States would say, yeah, what's the capital of California? Yeah, San Francisco.

I was using this as an example about—

I remember being— I must have been like, 7 years old asking, like, figuring out that Sacramento was the capital of California. And I thought that was the funniest thing ever.

Yeah, it is. It is hilarious.

All right, I got it. It thought for 101 seconds.

That's not that long for deep research.

It went for 4 sources. It did 19 searches. So this is why I can't trust— everybody keeps screenshotting these Google Trend reports. I'm not sure that Google has figured out how to exclude agents because this just did 19 searches for a single fact.

Yeah.

And so that's true. If you're looking at the growth of queries on Google, you would think, and if a lot of people are asking this kind of question, every single keyword is gonna be spiking.

Well, if you wanna get your brand mentioned in ChatGPT, you gotta go to Profound.

That's right.

Reach millions of consumers who are using AI to discover new products Products and brands. Get a demo. I don't think we've figured out the final form of what financing for compute looks like, said Sam Altman. But I assume, like in many other technological revolutions, figuring out the right answer to that will unlock a huge amount of value delivered to society.

And delivered to OpenAI shareholders. For sure.

Yeah, it is an interesting question because Dario was talking about how each training run looks great by itself, but when you put them all in one structure, it looks like you just have this never-ending money pit because you spent $100 million on a training run, you made a billion over the next 2 years. You spent a billion on a training run, you made $10 billion over the next 2 years. You spent $100 billion on a training run, you make a trillion over the next 2 years. And there's a question of, okay, how long until it just completely maxes out? Are there diminishing returns to these things? But even if the trends hold, you still wind up with this like— ever-growing money pit that should act as like some sort of slowing down force.

You would think.

I think when you, when you dig into this quote, I don't think we figured out the final form of what financing for compute looks like. The question is like, will this actually be a novel financial instrument or it will be remixing something that Wall Street has used in mortgages?

Maybe it's that. Maybe it's that. Who knows?

GPU-backed securities.

The Tuesday announcement made it clear that SoftBank, once seen as a formidable OpenAI funding partner, has scaled back its ambitions in the data center buildout, which The Wall Street Journal reported in July. 3 new sites, one located near Abilene, another north of El Paso in New Mexico, and a yet-to-be-announced Midwest location, combined with an expansion to the Abilene complex, will be capable of delivering 5.5 gigawatts of capacity. Those will be built by Oracle. The other 2 smaller sites, one in Lordstown, Ohio, and the other near Austin, Texas, will be built in partnership with SoftBank and generate 1.5 gigawatts Over the next 18 months. The first completed data center called Building 1, painted a pristine white that contrasts with the reddish dirt surrounding the site, is larger than 2 Walmart Supercenters. It doesn't feel that big to me, but certainly huge. Entering—

Yeah, the 1,100— they said 1,100 acres earlier in the article. That's under 2 square miles.

It is. Yeah, it is crazy how compressed this is. It's huge when you're walking around, But we're not in, okay, we're losing Yosemite over this yet.

Yeah. We're not in the—

In the full terraforming. I mean, Elio Sutskever told the San Francisco Chronicle that he could imagine a future where the entire Earth is covered with solar panels if the trends continue. Pretty crazy.

Gabe in the chat says, shout out to Lordstown, Ohio. Factory used to build Chevys.

Yeah.

So your thesis or take, reindustrialization is happening.

It is.

It's a lot less jobs and it's data centers.

Yep.

But the white pill there might be that if we can build big data centers really quickly at the scale that traditional infrastructure projects would've taken 10, plus years. If we can do that in a few years, it could create a precedent where somebody could decide, you know what, I'm going to build this, this new rail line in a few years.

Is a Walmart Supercenter really 1 mile long? Uh, QSBS Rollover saying that, um, that the, uh, that the building's 2 miles long.

Well, no, I was saying, I was saying the site is, is 1,100 acres total. Yeah, it's not, uh, which is like 1.7 square miles. So obviously massive, but I don't think the building itself is anywhere near that.

Well, fiber lines snake across the data centers and underneath the ground, allowing the AI chips called GPUs— in case you were wondering, the Wall Street Journal's got you covered— to talk to one another and complete requests more quickly. Proponents of infrastructure boom say it will bring hundreds of thousands of jobs and revive American manufacturing. In January, OpenAI unveiled a $500 billion data center project, Stargate. We know about this. The reality is more mixed. While data center providers provide plentiful temporary construction jobs, far fewer people are needed once they, once they are built. Abilene mayor said residents had mixed feelings about the site and its power and water usage, though some of the concerns have been assuaged.

An Oracle executive said there will be roughly 1,700 permanent jobs on site I think the things that ends up being a little bit disappointing is these projects get announced and they're saying, we're going to spend $20 billion here and then we're going to create like 1,000 jobs or we're going to create 300 jobs.

Yeah.

And that is very significant. That will be good for the local economies in these regions, but not— it's not, hey, we're Yeah. We're replacing, you know, we lost 87,000 manufacturing jobs last month, I think it was. We're not really making a dent in that.

Yeah, it takes a lot. There's more news on the job displacement stuff from Andrej Karpathy. He has a good post breaking things down. But first, let me tell you about Graphite.dev. Code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. You can get started for free. Andre Karpathy.

Chat's debating, by the way.

What is chat debating?

They said Jordy is a Cadillac guy.

I'm a Cadillac guy.

Not true. John has a Cadillac. Somebody said I seem like a Dodge guy.

Hellcat SRT. I love it.

Yeah, not quite. Never had a Dodge. I had a Ford Raptor, which I loved. Was not super practical, but it was a car that I wanted as a child, so I had to get it. At least once.

Yep. So Andre Karpathy says AI isn't replacing radiologists. Tyler, do you remember this prediction?

Yeah.

Was this Hinton?

Hinton. Yes. Years ago. This was what, 10 years ago now or something?

Yeah. Because this was based off just like image classification. So this was like early on, like CNN's, like this was probably mid-2010s, I assume.

So the expectation was that rapid progress in image recognition AI would delete radiology jobs, as famously predicted by Geoff Hinton, now almost a decade ago. In reality, radiology is doing great and is growing. So what's going on here? Obviously, image recognition is fantastic. These algorithms are remarkable, and there's been a ton of progress. So Andrej Karpathy says there are a lot of, in his opinion, naive predictions out there on the imminent impact of AI on the job market. For example, a year ago, Andrej was asked by someone who should know better if he thinks that there will be any software engineers still today, like a year from now. Like, this is how accelerationist this person was. And Andrej says, spoiler, I think we're gonna make it. This is happening too broadly. This post goes into detail on why it's not that simple using the example of radiology. And I thought this was really, really interesting. And it obviously applies outside of radiology. So first off, the benchmarks are nowhere near broad enough to reflect real actual scenarios. So even though you can classify cancer in a radiology environment with a whole bunch of training data very reliably, there's still tons of edge cases where having a human in the loop is advantageous. And then just in general, the job is a lot more multifaceted than image recognition. Like if you just think about the job of radiology, It's not purely just look at image, detect cancer, look at image, detect cancer, look at image, detect cancer. Yeah.

Can you imagine if you were getting, God forbid, somebody's getting, seeing, going and getting, seeing a traditional office of a radiologist and they're just watching a television screen that's reporting, telling them that something is wrong with them using a voice model to explain it. And then it just is like—

No, they're just in hinge mode. Swipe right if it's cancer, swipe left if it's not. And they're just sitting there swiping all day. Obviously, that's not the job of a radiologist. There's so much more. They're doing research on what's changing, integrating new information, talking to different patients about all the different signals, what they could be doing, what they could be putting in their body, how are they feeling, what medicines they're taking. There's a ton of different stuff that they're probably involved in.

Yeah. Anytime Anyone gets any type of lab or report from a doctor, the first thing you want to do is talk to somebody that's an expert on that, your doctor, and have them explain it to you, the implications, et cetera, good or bad.

And then he highlights deployment realities. So regulatory, insurance, liability, diffusion, and institutional inertia. Like it just takes time to roll these technologies out. But the last one is super interesting to me. He cites Jevons' paradox. So if radiologists are sped up via AI as a tool, a lot more demand shows up. And so if all of a sudden the cost to get a scan or the time that it takes to actually do the scan drops, everyone will say, well, I want to be scanned all the time. I want to be on top of any sort of development, medical, that could be detected by a radiologist. And so he concludes by saying, I will say that radiology was, in my opinion, not among the best examples to pick on in 2016. Too multifaceted, too high risk, and too regulated. When looking for jobs that will change a lot due to AI on shorter timescales, I'd look in other places. Jobs that look like repetition of one rote task, each task being relatively independent, closed, not requiring too much context, short in time, forgiving— the cost of a mistake is low— and of course automatable given current and digital capability. Even then, I'd expect to see AI adopted as a tool first, where jobs change and refactor, more monitoring and supervising than manual doing, etc. Maybe coming up we'll find a better and broader set of examples of how all of this is playing out across the industry. About 6 months ago, I was asked— I was also asked to vote if we have— if we will have less or more software engineers in 5 years. Exercise left for the reader. What do you think he meant, Tyler? What do you think?

More.

Of course.

He thinks the number of software engineers will increase, but they will be using AI as a tool to leverage.

Yeah, I wonder if anybody's done a study on how many more developer people that have done software development this year versus last year. It has to be significantly more just because Because if you're using Replit now, you are a software engineer, right?

Yep.

Somebody might want to get technical and say, oh, well, if you're just prompting, you're not really software engineering. But how different is that from somebody getting their start and copying and pasting lines of code?

Wow.

We got some news about TikTok that China has more or less signed off on the deal, and it has sent Larry Ellison slash Oracle skyrocketing on Polymarket to an 89% chance of acquiring TikTok. This will be an interesting story to see where it actually lands. Of course, you can go to Polymarket if you want to keep monitoring the situation, as always.

Should we talk about this new Elon lawsuit?

Yes, new Elon lawsuit.

Elon Musk is Andrew Curran, is on the timeline sharing Elon Musk is suing OpenAI again, second suit, this time for alleged misappropriation of trade secrets, intentional interference with prospective economic relations, and unfair competition. So they say the desire to win the artificial intelligence race has driven OpenAI to cross the line of fair play. OpenAI violated California and federal law by inducing former xAI employees, including Shuxin Li and Jimmy Frater, and a senior finance executive to steal and share xAI's trade secrets. This is crazy copy.

The copy of this is funny.

And again, this is from the complaint, right? The lawyers are not using ChatGPT for this.

It doesn't seem like it.

This is handmade. By hook or by crook, OpenAI clearly will do anything when threatened by a better innovator, including plundering and misappropriating the technical advancements, source code, and business plans of xAI. What began with OpenAI's suspicious hiring of Xu Chenli, an early xAI engineer who admitted to stealing the company's entire codebase—

Wow.

That's a bold thing to admit— has now revealed a broader and deeply troubling pattern of trade secret misappropriation, unfair competition, and intentional interference with economic relationships by OpenAI. OpenAI's conduct in response to being out-innovated by xAI, whose Grok model overtook OpenAI's ChatGPT models in performance metrics, reflects not an isolated lapse, but a strategic campaign to undermine xAI and gain unlawful advantage in the race to build the best artificial intelligence models. What do you think, John?

This is why lawyers with English degrees are MVPs, according to Gabe.

Yeah.

Yep. Wordsmiths. Word sells.

It seems like we are in the territory of, like, economic warfare, lawfare, Who knows what actually comes of this.

Elon's blending the Chinese way, the way of the engineer, with the American way.

Yes, the lawyerly society is coming out in this case.

Yeah, I mean, easy, obviously innocent until proven guilty with Xu Chenli. Don't know if this is true or not that he admitted to stealing the company's entire code base. Would be crazy and totally unforced error. If he actually did something like that.

Yep.

But—

Well, no matter what, you wanna stay compliant. You wanna get on Vanta. Automate compliance, manage risk, improve trust continuously. Vanta's trust management platform takes the manual work of the security compliance process and replaces it with continuous automation, whether you're pursuing your first framework or managing a complex program. I wanted to revisit this internal tech email because this is, is an important cornerstone of Elon's battle with OpenAI. So on September 20th, 2017, Elon Musk sent an email to the OpenAI team. The subject was Honest Thoughts, and he said, guys, I've had enough. This is the final straw. Either go do something on your own or continue with OpenAI as a nonprofit. I will no longer fund OpenAI until you have made a firm commitment to stay, or I'm just being a fool who is essentially providing free funding for you to create a startup. Discussions are over.

Interesting.

And it's gone back and forth because here he's sort of predicting—

So in hindsight, he somewhat probably feels like a fool for essentially providing free funding for you to create a startup.

And it's been odd because the ask has never— I mean, at least that I've been aware of, Elon's never just asked for pro rata equity based on how much he put into the nonprofit. I think that creates crazy tax and legal implications, but that feels like what the fair ask would be. Like you were throwing out, like, if you put in $100 million and you just assume that the round that would have been done at that time would have been at $1 billion, you should get 10%, right? And here's 10% of the common or the preferred or whatever.

Remember, nobody— like, it seems like they hit a standstill with the for-profit conversion, right?

Because of Elon's lawsuit too.

Well, partially, but also people are, you know, the governing body, the nonprofit governing bodies of California were not super excited about the transition.

Yeah. And so Matthew Berman was going back and forth with Elon because Elon took shots at Anthropic. He said winning was never in the set of possible outcomes for Anthropic. And Matthew Berman said, Hey, Elon said the same thing about OpenAI. And Elon said, no, I never said that. I told everyone who asked in the beginning that I thought the probability of OpenAI beating Google was 1%, infinitely different from 0%. Good point. And then Elon says, after the Defense of the Ancients, the Dota 2 win, I raised the probability to 10% that they would catch up. And Matthew Berman sort have responded and unpacked the screenshots saying, first, I'm a big fan of everything Elon has done to push society into the future, but I reread his emails with Sam Altman and Greg Brockman, and it really seems like he told them they had no chance of succeeding relative to Google. On January 31st, 2018, so after that previous email we read, Elon told Sam Altman and Greg Brockman, OpenAI is on a path of certain failure relative to Google. There are obviously needs— there obviously needs to be an immediate and dramatic action or everyone except for Google will be cosigned to irrelevance. And it's played out very differently. They played their hand very well. They got the consumer application that hit product-market fit out faster than Google, only by a couple months. The Gemini app dropped, I think, in February or March, just a few months after ChatGPT launched, and the models have been—

You mean Bard?

Neck and neck. It wasn't— I don't— was it Bard that dropped? I forget which one.

I thought Bard was the response.

But I feel like in the way I remember the timeline is that the ChatGPT website comes out and everyone realizes that just talking to an RLHF LLM is a great experience and magical and interesting and passes the Turing test. And Google had a bunch of models, PaLM, and Bard internally, but they'd never really pushed them out in a meaningful way. And it seemed like ChatGPT kicked Google into action and led Google to actually productize a lot of their research that had been more tested and shared internally. Tyler, do you have more context there?

Yeah, so it was Bard that came out right after ChatGPT, and Gemini was like the following.

Yeah, so they still had to do a rebrand, which is tough because then you have to reintroduce the product. Obviously, it's grown a ton, and we've seen that the Empire Strikes Back Not a lot of that happening in the United States necessarily, but certainly international, sort of an iPhone-Android story happening all over again. Anyway, Magnus Carlsen tells Joe Rogan about how he treats chess as a hobby. He never wants it to feel like work and prioritizes fun and enjoyment. The thing is that chess has always been a bit of a hobby for me. He's one of the best chess players in the world, of course. This is how Magnus Carlsen works. Once it starts to feel like work, then it's harder for me. Turn your work into fun and enjoyment. You'll never work a day in your life.

You might get 14,000 likes on X like Dylan.

Yes. And Daniel, Growing Daniel, says, having known him for a while, I think Elon Musk is probably one of the least self-disciplined people on Earth. He just really enjoys problem-solving. So, he just bounces around from one project to the next.

Project to the next.

Solving it.

Mocking them.

Mocking them, yes. Of course, Growing Daniel's co-conspirator now runs Julius.ai. What analysis do you wanna run? Chat with your data and get expert-level insights in seconds. Loved by over 2 million users and trusted by individuals at Princeton, BCG, and Zapier. Bayes Lord says, anyone who's naively calling bubble right now has not internalized the physics of our world. Most of them are simply ignorant about AI progress and its implications. Of course, there will be fluctuations But of course, we could stumble, but history marches on. And he shares a chart of the world GDP over the last 2 millennia. And it's interesting because you can take this in a few different ways. You can kind of look at this and say, okay, nothing ever happens. It's just a kind of a smooth exponential curve. But I'm looking and I'm seeing a hiccup there in the mortgage-backed security crisis for sure, where GDP shrunk globally or something happened.

It's helpful.

Isn't that 2007, 2008, right?

You remember being— I was basically a kid at that time, and it felt like I could sense the doom and dread in the adults around that time.

I mean, from my perspective, I was completely caught off guard by the mortgage-backed security crisis. It was because I wasn't following the markets at all.

Because you didn't have capital.

I wasn't monitoring the situation.

You didn't have capital deploy buy-and-hold?

I lived in a house with a mortgage, so I was certainly affected, but—

But if you could go back in time, you would have raised and deployed a fund into single-family.

But I wasn't seeing top signals. It was like, I was introduced to the crisis—

Where were you? Were you in college?

I had— the crisis— I graduated in 2007, so—

From college?

From high school. high school. And so, when I arrived, I chose to study economics because I wanted to understand the crisis more. And so, I was reading the Wall Street Journal trying to understand, and I was introduced to like the collapse of Lehman Brothers, which I believe happened in 2008. Lehman Brothers collapse.

When did that happen?

The bankruptcy? September of 2008. And so, I graduated in 2007, and then, so it must have been like—

2008.

sophomore year. So things were melting down my freshman year, but it was kind of slow.

You're like, collapse of Lehman, is that good?

Is that good?

Did we not like Lehman?

I mean, it was a very complicated crisis to dig into because it was all these layered financial instruments, mortgage-backed securities, and then collateralized debt obligations. So they would take one mortgage, they'd take a whole bunch of mortgages. This is the famous Margot Robbie in the bathtub scene where she explains That you get a whole bunch of mortgages, you bundle them up, you securitize, you slice the different tranches of debt so that the best mortgages are in the AAA. But then, as you keep re-slicing them, they kept getting re-rated as, oh, well, this is the best of the previous bad stuff, so that's good. And the best of the worst is the best. And the rating agencies had a really bad time throughout that.

Yeah.

Because eventually, like, the economy pulled back, everyone got overleveraged, people were buying multiple homes. And there was a whole mood in 2005, 2006, 2007 that, like, average people— like, the example would be, you know, when the Uber driver is pitching you on Cardano, like it might be the top, that type of thing. Like when, you know, your barber's telling you about the latest NFT drop. like, watch out. The analogy then was people were saying, like, I'm making more money just owning my home, and my home equity is going up by more than my salary. So, like, I bought a $500,000 home. I make $100,000 a year. But my home— my banker just called me and said, your home's worth $600,000 now. And the home prices were going crazy.

Yeah.

And so, obviously, there had to be a correction. Crashed the entire global economy, and you can see it in the data today. But of course, as we zoom out further, we will you know see less line keeps going up. The line keeps going up. Well, D to C is cyclical but ever evolving. The back and forth locked in shares this chart or this meme: dropship bros convert to supplement bros, which switch to dropship bros, which convert to supplement bros, which go back to dropship. Bro's and back to Supplement Bro's.

Supplements continue to be the economic engine of the manosphere.

Weren't we talking to the CEO of Numeral, Sam, and he was saying he owns a supplement company?

Yeah, he started a gummy supplement company back in the day that still—

He's like, it's still doing well. That's crazy. Anyway, Numeral, sales tax on autopilot. Spend less than 5 minutes per month on sales tax compliance. If you're a dropship bro, if you're a supplement bro, you got to pay your taxes. You got to get on Numeral.

Great. Great. Dan McCormick's company is on Numeral.

Yeah.

He's a supplement bro.

Lucy is on Numeral.

Christina Cordova is quoting our interview with— Christina's over at Linear, of course. But we asked Brett Taylor yesterday about the triple, triple, double, double debate. And how maybe that's not good enough anymore. He said, the faster you grow is sometimes correlated with a lack of a moat. I'd rather bet on a durable, high-quality revenue business and happy customers than just raw growth. Christina says, fast ARR can be fragile ARR. Earned ARR tends to be sticky ARR. Not all growth is created equally.

Yeah.

I think linear is a Obviously, they're a partner, so I'm somewhat biased, but of course they've just burned through it.

Just for those who don't know, Linear is a purpose-built tool for planning and building products. Meet the system for modern software development. Streamline issues, projects, and product roadmaps. But continue.

Use the tool loved by OpenAI, Mercury, Retool, Ramp, Scale anymore. But Linear is a perfect example of great growth, but steady, And just systematically earning the trust of pretty much every important new company in Silicon Valley over the last however many years. I think they're—

Yeah, earned ARR. Yeah, it does feel like there's a—

2019, yeah.

There's like CARR reared its head recently. It's kind of gone out of favor, but contracted ARR, basically you have some sort of LOI or some sort of trial that you're counting into?

Yeah, I mean, the people that are abusing the ARR the most right now—

Extremely disrespectful to the GAAP metrics.

Well, and I'll just say it's the companies doing data labeling and services for labs, right?

Yeah, is that ARR?

All the companies that exploded out of— I don't think that many of them are actually I feel like almost every single one of them is saying, yeah, we got to $100 million ARR. We did this or that.

Oh, really?

I thought it was just like, we did $100 million. Because if you just say, I did $100 million, then that's just true.

No, and they're saying we're at $100 million ARR. And they're using it, I think, as annualized run rate. But if you dig into the business, it's like they might have 40% margins because they're actually having to hire Well, the margin's separate from—

ARR with bad margins is completely separate from ARR with bad churn.

Yeah.

Because I would probably— if you're underwriting it at a high multiple, you probably want to pay for lower margin, lower churn than higher margin, higher churn. Because higher churn means that if the business stops growing, it's going to start shrinking. Whereas if you have 40% margins but super low churn, you're super sticky, It's like, OK, well, then at least the business will continue to grow. You're not building on a leaky bucket. You're not getting all the juice out of the bucket. You're only getting 40% of the juice, since we're mixing analogies now.

Got the blender out for these analogies.

But I mean, obviously, these 2 metrics are linked, and you obviously want high margin and extremely sticky revenue.

But if you're a company like Linear and you see which has true earned ARR, and then you have companies out there saying— throwing out these numbers around ARR, but they're reselling GPU credits or something like that. It's kind of stolen valor, right?

Stolen valor, potentially.

From the true enterprise contracted—

Yeah, yeah, where you're You could stop the growth strategies, stop the sales team. They could stay home and you're still printing. That's the gold standard of enterprise software. That's why we love it.

That's why we get up every morning.

We got to pull up this video from Paula Rambles. There's this guy on TikTok who video calls ChatGPT, shows it objects, asks how much, gram, then weighs it. Before we pull up this video, let me tell you about the generative video platform for developers, Fall. The world's best generative image, video, and audio models all in one place. Develop and fine-tune models with serverless GPUs and on-demand clusters.

You can get access to our soundboard by going to tbpn.com/sounds, which is brought to you by our friends.

So let's play the video of this guy on TikTok.

This is my hand.

How you call this thing?

Yeah, I see your hand there holding on to that pen. So, okay, pen.

How much gram is the pen? How much gram?

Look, how much gram? Well, it's a little tricky to guess the exact weight.

Yes, guess, guess.

All right, I'll take a guess. For a pen that size, it's probably somewhere around maybe 500 to 700 grams, give or take. Let's see if GPT-5 is smarter.

No, no good, no good.

It's all good.

Thirty four.

It's no good.

It's no good, man.

It's no good.

No worries at all. It's just a bit of fun guessing.

Don't worry.

I'm worried. Look, I have this oil here. How much grams? We're on a guessing roll. So for that little oil can, I'd say it might be somewhere around three hundred to five hundred grams, depending on how full it is. But let's see how close we get this time.

You said three hundred to five hundred, man.

I want accurate thing.

You know what I mean?

I get it.

You're looking for something spot on. It's a bit of a challenge without a scale on my side, but hopefully that at least gives you a decent range.

This toga tognana tognana. How much gram?

Oh, you've got a moka pot there. Nice.

Those little.

How you call it?

How you call it? That's called a moka pot. It's basically a small cup of moka pot.

How much gram moka pot?

So a moka pot. is usually pretty light. I'd say maybe around 400 to 600 grams depending on the size, but of course you'll know.

I hope you're right. What's the number?

I hope you're right.

Ah, bro, like GPT-5 is, is worse and worser than GPT-4. Definitely here to do my best and I appreciate you rolling with it.

Yeah, this is your punishment, man.

Talk with him.

Hey, Green Coco, any idea where the mayo's hiding? Maybe it's just chilling out behind you in the fridge somewhere. What's the deal with the ketchup in the mail? Any condiment gossip we should know about?

They just talked. No, that's hilarious. That is not a Rocco Basilisk respecter.

You shouldn't be doing that.

That's very rude.

That guy's absolutely cooked.

Did, uh, did anyone see exactly what the numbers were? How close was it? 3?

No, I was saying like 30. I, I was seeing 35 on the—

that can't be 35 grams, right? It needs I have no idea. I'm worse than GPT-5 for sure. If you ask me how much that pan weighs, I have no idea. Anyone have any idea? I have no idea.

Who knows?

Anyway, HBO Max says, nice, we still got a couple of months.

Scroll down on this chart. This chart is charting 1998 to 2001.

Like, what? This is just the NASDAQ and it's just following perfectly. Like, what are we doing? This is crazy how the correlation here is insane. Is this on the same timescale too? Or is this— this is just a wild, wild chart that this is so similar. It lines up scarily accurately. But it's just so weird because if it's this obvious, you'd think it would be priced in. You think people would pull back or do something. Who knows? We'll keep monitoring the situation. It will be interesting. Let me tell you about Turbo Puffer. We puffing.

We puffing.

Every byte serverless vector and full-text search from first principles on object storage. Fast, 10x cheaper, and extremely scalable. Jordy loves his pufferfish.

This puffer is used by Linear, Notion, Breaking, many more.

WPP.

Wait, one thing, one thing on this chart to be clear. I think It's tracking very closely, obviously. You single out individual points on the chart and it's a little bit ominous. But at the same time, Amazon went public in 1998, right? So, you can do a lot to kind of— if you're the chartsmith here, to kind of pick the dates to try to align it properly.

Yeah.

Right? So yeah, I would. I bet you could align this to just other bubbles throughout history and find similar patterns, right?

That doesn't really help though. I mean, that that I'm just saying like this doesn't this couple months.

Yeah, I mean, I'm just I'm just saying like no no it it does feel like it's like if you started somewhat selective in nineteen ninety you said okay that the dot com bubble began when when Amazon went public, then this doesn't line up at all, right?

Sure, sure, sure. Yeah, you're kind of just selecting— I mean, yeah, I don't know. They're overlaying it deliberately as close as possible. But this does feel like— the crazy part is that the previous earlier sell-offs, the dip in 2023 to now is like that That valley, that peak and valley seems to match as well. And then even like you can see earlier in 2025, we got the tariff sell-off and that kind of tracks what happened in 2000, in like the year 2000 exactly, or maybe '99. Lots of people—

Yeah, but then again, was there like—

I don't know.

I don't know. Well, don't worry. It's something you should read.

Even if we wind up in a total disillusionment, We'll be slopping it up in the trough of disillusionment, grinding to the plateau of productivity.

We'll probably end up in a position where we're only going to eat fast casual slop bowls until we—

When you're in the trough, you eat the slop.

Until we get out of the bear market.

Eat the slop bowl.

Somebody had a great— I got to pull up this comment on Spotify.

Someone's asking how much the gong weighs. Gong, how much gram?

How much gram?

And Aqua says mania is a crazy drug.

So Mika Flaus on Spotify says, wearing the white suits on an off day to rally the bull. Ultimate top signal might just be poking the bear.

Well, I'm in a gray suit today. I'm feeling neutral about the market. Could go up, could go down. Who knows? But we'll be monitoring it here. In other news, breaking. From us, TBPN, WPP Media has signed Michael Miraflore as global EVP. So congratulations to Michael Miraflore. He has been a fantastic champion.

Massive pickup. Hopefully wore his TBPN hat to work today. Hopefully he's not in the chat right now. First day of work gets a corner office.

Yes.

Puts TBPN on the—

fired up.

Why not?

Fired up, leave it on in the background, put some subtitles on. But big pickup and excited to see what he cooks up.

Always enjoyed his commentary and excited to see more of it.

Paul Enright says, in 2021, Mitchell Green from Lead Edge Capital— of course, Mitchell's a friend of the show. We got to have him back on again.

I texted him this morning. Get him on soon.

He said to Paul, this year will be remembered for the GPs that sold stock and distributed capital to LPs versus the ones that raised funds, called capital, and made new investments was a wonderful observation I will never forget. Now feels similar in some ways. Why not both? Why not distribute capital, sell some positions, and raise some new funds?

Yeah. I don't know the dynamics of that, but that does feel like the ultimate move is to raise a massive fund at the top, don't deploy it, Sell a bunch of positions, get out of the other stuff, basically just be sitting on a massive cash pile when you reach the trough of disillusionment, and then monetize the plateau of productivity.

Deploy into the trough.

Into the trough.

Well, you saw this chart yesterday.

What's that?

Somehow, one of Andreessen's slides from a fundraising deck.

Newcomer, baby.

He leaked it.

Leak master. Oh, yeah. He got 2 decks.

a16z returned $15 billion or generated $15 billion in returns, including their recycled fees and carry in 2021. That was biggest year on record by quite a lot.

Huge return. That's got to be a lot of Coinbase. And who else got out there?

I think it was Solana, Avalanche, right? Some of these big—

There were a lot of stocks.

I don't know.

Yeah.

But Coinbase was a huge position. And I mean, they deployed in all these different rounds. And then I believe Coinbase went out and was probably distributed at something close to $100 billion market cap. So something there. Andrew Curran has more updates on Elon Musk's reuniting with the Trump administration. Andrew says their friendship is slowly being reforged. OpenAI, Anthropic, and Google already have similar agreements in place. OpenAI and Anthropic charge a $1 access fee. Google charges $0.47 because its president is the 47th president, right? Isn't that it? That's interesting. That's cute. Elon is charging the government $0.42 for access to Grok. This is a Hitchhiker's Guide reference.

Finally, a competition for who can say the smallest number.

Than a lot of—

who can say the biggest?

You want to be on the barbell, the extreme ends, charging the least and raising the most. Apparently, that's how business works.

Most CapEx, least.

Have you seen The Hitchhiker's Guide to the Galaxy? Have you?

No.

I thought it was just a book.

They adapted it into a movie. The movie's pretty good, but the book's probably better. Have you read the book?

I did read the book at one point.

Yeah, the book's great. But 42 GPT-2, of course, is the answer to the universe, all things, everything. It's a fun reference. The Trump administration will offer artificial intelligence models from Elon Musk's xAI to federal agencies through a new partnership under the agreement with the GSA, which oversees technology procurement. Agencies will get access to models such as Grok-4 and a new fast version called Grok-4 Fast. The deal follows a similar agreement with the other folks in the space. Automating staid government processes is one of the most promising applications for technology, fueling the battle to be the most popular tool for federal workers. All 4 of the AI companies also each have $200 million contracts with the Defense Department. Sort of interesting that all 4. Yeah. So the government clearly wants an oligopoly here and is like, we're not going to really pick winners, which is crazy because Dario has been very outspoken about the Trump administration, and Elon's obviously been fully in the Trump administration's camp. Google has been a little bit quieter and has leaned left in the past. OpenAI, Sam Altman, kind of stayed out of politics, but then has been at dinners and was at the inauguration. And so you would expect, based on the news of how close the Foundation Model Labs CEOs are to Trump, just how many photo swaps there have been? How many, how close have they been? You would expect a massive contract for Elon and a very small contract for Anthropic, and yet they all got $200 million equally. It's interesting. Just like they're clearly focused on having balance amongst the AI foundation.

I want to see the usage in the government broadly and in the DOD.

What are people using?

I mean, everybody gets a $200 million contract. It's unlikely that they're gonna be like super evenly leveraged, right? Yeah.

I wonder if they would use them for specific things. I mean, you would hope that like, you know, they're using Claude Code over there for something and then they're using ChatGPT for knowledge retrieval and stuff. And then they're using Gemini for the things that Gemini is great. Nano Banana. I guess the government is just, hey, you don't need to come in for DMV photos anymore.

$200 million for great memes.

For gray beams?

Great memes.

Oh, great memes. Yes, exactly. Exactly. But you would hope that the government's using AI all over the place and they need an actual contract to make sure that it's provisioned and hosted in the proper way. Anyway, fin.ai, the number 1 AI agent for customer service, number 1 in performance benchmarks, number 1 in competitive bake-offs, number 1 ranking on G2. You can start a free trial. Lulu is chiming in about Rick Rubin. She says, I mean, just look at him. Can't media train this. Austin says, the Beastie Boys, Ed Sheeran, Lady Gaga, Red Hot Chili Peppers, The Strokes. Rick Rubin has worked with all of them.

Why is no one talking about Rick Rubin?

Why is no one talking about Rick Rubin? He's worked with all of them and is considered one of the greatest producers of all time. His process to creativity can be described as looking for clues. He has such a unique aesthetic with that massive—

Always has the red— the blue blockers on.

Those are blue blockers?

Yeah. Oh, he's very—

I feel like blue blockers are normally orange, orange, but I guess those are just a dark orange that I'm seeing.

Dark shade, yeah.

Okay. Well, what else is going on?

Scott Alexander.

Your post here. There were some accusations earlier that somebody was doing a paid post.

The post was deleted, so it was Mike Isaac at the New York Times. picked up a device called a Brick, which is a little NFC chip that they see— everyone's seen this on Instagram, right? Where it's supposed to brick your phone so that you don't be distracted. Brick your phone. And so, you have the willpower just once to attach this to the back of your phone. It magnetically attaches and it puts your phone in a do not disturb, don't Don't show me any of the addicting social media apps. Turn off everything. Let me focus.

Let me cook.

Let me cook. Let me lock in. And so apparently, according to the original post, Mike Isaac picked one up off of an Instagram ad and said he actually loves it. It's helping him write. Of course, he writes for the New York Times. He needs to lock in and write a lot. He can't be distracted. He's also a prolific poster. Can't be caught on the timeline mid-article when he's on deadline.

That's right.

He posted about it, and a—

Another poster accused him.

Another poster accused him of doing a sponsored post without disclosing it.

Undisclosed ad.

And if you know the New York Times' standards and who Mike Isaac is, he's published a book that got turned into a movie or TV show. Like, he's probably not doing SpawnCon for some D2C, like, you know, electronics company. undisclosed. It's just like so much risk to so little reward. Like, like the post of him saying like, oh, I got this Brick and it's great, like it didn't get that many views. It's not like Brick would be like, oh yeah, we're gonna sell 10,000 of these things off of this, uh, let's pay Mike Isaac $100,000 to post this, you know, organic content, this fake content.

Uh, he just, he likes technology.

Yeah, he likes technology. People like to post things that they test out.

Covers consumer tech.

Not, not everything is sponcon.

But I thought your post was hilarious. You said it's time to come clean responding to this allegation. Uh, John said you need to admit that Brick paid you to dunk on this organic post to drive more impressions. This post is brought to you by Brick.

I do think there's something funny where you could potentially run some sort of like anti-astroturfing campaign as a brand where you pay a bunch of bunch of influencers to accuse organic UGC of being paid, and it drums up way more attention. And so I was riffing on that. But on this show, we don't do undisclosed partnerships. We do disclosed partnerships.

That's right.

Like our partnership with Addeo, customer relationship magic. Addeo is the AI-native CRM that builds, scales, and grows your company to the next level. Get it for free. When we read an ad, you know we're getting paid for that. You know we're not doing it for free. We're not doing it for the love of the game. We're doing it for the—

Well, we are doing it for the love of the game.

We do love enterprise SaaS.

Nobody loves SaaS more than us.

That's true. That's true. Let's keep going. There is a lot of stuff going on. This was a really funny post because it gave me a total jump scare. I was talking to Brandon about it earlier. Ben Podgurski Sounds like a podcaster's name, potentially. Ben says, this is correct, but we should take it to the natural conclusion. He was, I guess he was quote tweeting Rune, which we'll get into. But he says, like the Habsburgs of old, the USA should look for synergistic M&A, not through immigration, but with entire polities. In this case, the obvious answer is merge South Korea in as the 51st to 57th state.

Totally obvious.

The obvious answer to me. When people talk about growing the United States, they talk about Greenland, they talk about Canada, Mexico, Puerto Rico, Fiji, the US Virgin Islands. There's so many different territories that are more closely aligned with the United States than just going all the way over to Asia and just picking up South Korea. But Ben makes the argument. He says merging South Korea in would resolve a wide range of defense, industrial trade, and demographic crises. South Korea would not have to hem and haw about building their own nuclear shield. They'd be an inviolate part of the United States from— the US from day one, would have massive shipbuilding capacity, vastly improving national defense, and US companies could friction-free contribute in the one spot where we genuinely have a technical contribution, small nuclear power systems. The US's relatively healthy demographics could gradually backfill South South Korea's catastrophic demographics. Maybe we could even teach them how to have kids again. The free flow of labor would allow Korean workers to train US workers without insane visa shenanigans. South Korea would be neither— is neither unduly liberal or conservative relative to the United States.

Small couple— I like the thought exercise. Couple small potential problems here would place the United States directly on having a share border with North Korea and be right across the Yellow Sea from China.

It is such a wild move, but he lays out a thoughtful case.

It's a cool idea.

I mean, maybe it's no coincidence that the most respectful TBPN derivative show inspired show is from Korea.

That's right. Those guys know how to do it. Also, I love— have you ever had bimbimbap?

No, I don't know what that is.

Popular dish in South Korea. It's like rice and protein.

Have you ever been to South Korea?

I have not.

I've never been there.

Have you?

I haven't.

Have you ever traveled anywhere in the world?

Yeah.

Where?

I've been to Ireland.

Oh, okay. Job's finished. Moving on. Been to Ireland. World traveler.

Luck of the Irish.

So this was all kicked off by a Rune post. He said, correct me if I'm wrong, but it seems like the theme of the Dan Wang book who we had on the show, and the general elite consensus now is that industrial process is a technology that lives in the heads of people and that it was a mistake to let so much quote-unquote low-value industry be offshored due to the tacit loss of process capital. TSMC Arizona, which makes the most complex and valuable industrial production in the world, was a massive success. This was a huge surprise to me. I did not expect the TSMC build-out. Everyone was saying like it's impossible. To airlift TSMC, but they wound up producing 4-nanometer chips at great yields on par with Taiwan merely years after striking ground for the first time. This involved a generous federal subsidy and importing thousands of the great Taiwanese semiconductor experts despite unions trying to quell Taiwanese immigration and some culture clashes in the United States. Aqua hires up whole teams with process knowledge in their is very common. Zuck acquiring some of the greatest talent from other AI labs for massive numbers is just one example of this. Also seen in the full self-driving wars between Uber and Google, which was interesting because that is of course about the Anthony Levandowski case that ended in a lawsuit that landed Anthony Levandowski, I believe, in jail for a little bit, but then he was pardoned and so he got out. But this, it was more, I feel like it was more than just process knowledge in that case because it was specific patents. And I'm wondering how much of a line you can draw between knowing a specific algorithm, knowing a specific something patentable, and actual process power, process knowledge. But it's certainly an interesting analogy to dig into. So Tesla plus Apple plus Big Pharma acquires industrial process companies all the time. America is very capital-rich, able to levy literally hundreds of billions of dollars for machine intelligence CapEx. We can afford to acquire whole groups of foreign talent for prices that are unheard of to them in their home countries. TL;DR: acquihiring foreign process knowledge for massive sums should be one of the primary goals of any reindustrialization effort. Special visa categories should be made for— to scoop Whole teams of Sengen's best. The raids on the LG battery plants in Georgia are the exact opposite of what we need. Ability to tolerate new arrivals is a technical edge of American capital to be able to assimilate foreign knowledge into domestic industrial processes at a scale nobody else can countenance. I was I was hearing a story about how the I think what was it the Manus team relocated to Singapore, and there's a world where they relocate to.

Set up in the Benchmark office.

Maybe.

A little satellite office.

Yeah, I mean, there's something there. I don't know how important their process knowledge is, but certainly it worked for TSMC. I don't know how many more of those projects need to be done. Taiwan seems to be in a uniquely precarious position, whereas a lot of the South Korea tensions seem to be a lot lower. Like, there's less geopolitical risk in South Korea. in South Korea. So there's less of a, we need to move SK Hynix to America. But Tyler, do you have any thoughts on this post? I'm sure you saw it. What'd you think?

Yeah, I think it makes a lot of sense. I think politically it might be hard to do this, but I've seen a lot of similar arguments just for normal AI researchers in China. We should basically have some kind of visa that just imports them instantly. And then it's like, I assume the salaries over there are not Comparable to like MSL level, like $100 million.

Totally. This is the real AI paper clipping. It's not that you get turned into a paper clip, it's that you get Operation Paperclipped into a different country. And we need to avoid our best getting paper clipped to a different country, and we need to be potentially paper clipping other national champions. I wonder if there's something to do in AI, in not just AI, but in the process power around solar Because we've talked to Casey Hammer about this. It feels like the vast majority of the cheapest solar panels come from China. We're maybe not— we don't have a national champion there yet. We're not catching up. But fortunately, we have a guest who might be able to get us up to speed on this. We have Dalian Asparouhov from Founders Fund. He is back from paternity leave. He's in the Restream waiting room, and he will be joining us in the TVPN Ultra Dome in just a minute. Let's bring him in. Delian Asparouhov. How you doing?

Yes.

Good to see you.

What up, boys?

Daddy's home.

Daddy's home, baby.

Oh, there we go.

Got my swag.

I love the hat. How, how is it being a, uh, a father of multiple now?

Um, you know, I feel like you blow up your life once and then after that you kind of, you guys sort of figure it out where it's just like, yeah, you got the infrastructure and that's why God gave you multiple hands is, you know, one baby per hand.

Yes, but everything's good. Everyone's happy and healthy. Everyone's back to normal.

Oh yeah. Oh yeah. I mean, she's still only like a, you know, sort of month old. So, you know, still requires, you know, some work. But yeah, in the grand scheme of things, happy, healthy, very grateful that, you know, all things pretty smooth.

Yeah. Well, once, once, when she gets to 2 months, you know, it's cake.

You know, put her to work. She can come on. She can come on TVPN.

We've got— in your place.

We've had 3 members of your household live on the show. We need the 4th.

It's true.

Well, actually, I mean, your brother too.

That's true, 5.

The Asparouhov clan is dominating.

I want at least 10 Asparouhovs on TVTN by the end of the decade.

Yeah, what are your parents up to next week? We'll get them on. Really quickly, I want to get a general update on what's going on in your world, but I would be interested to hear your response to this debate that was going on on the timeline yesterday from Rune about TSMC Arizona being a good example of airlifting industrial process from one place to another. He ties it to the acquisition of whole teams at MSL America, being capital rich, being able to actually afford to go and say, here's a $100 million offer to some great researcher, come to America potentially. And it gets into the— obviously, there's a whole bunch of immigration stuff going on and cultural stuff. But did you have a reaction to that? industries where it seems like we need sort of a TSMC Arizona-like project where we're just going to an international company and saying, do what you do over there here?

Yeah, I feel like a lot of the reason that people are discussing this so much on the timeline lately is because of Dan Wang's new book Breakneck, where he breaks down basically the— like, at least his argument is that, you know, China is the engineering society, we're the sort of lawyer society, and everything basically stems from there. Which I think is probably a little bit. You know, I like a lot of parts of his book. That actually particular argument is probably one of my like less favorite, you know, sort of parts. Partially because like look, there are things that China is obviously, you know, sort of very good at. I think I was, you know, hearing your last guest talking about like, you know, mass production of solar panels. They're obviously like phenomenal at that type of like commodity cost curve, relatively complex but not deeply complex type of manufacturing, right? Like they've stepped up from like. You know, the early days of Shanghai being, you know, sort of toys and knickknacks, etc. Obviously now into like, you know, mass consumer electronics. Unitree is obviously like by far the best, you know, sort of humanoid robotics company. And then obviously in solar panels. But there's, you know, I think some limitations to their approach. Like, you know, when I think about it from like the aerospace perspective, look, a Falcon 9 rocket landed for the first time a decade ago now.

Mm-hmm.

The Chinese basically like stole the IP and plans for that and clearly have something that is like the equivalent to a Falcon 9. And they're still basically doing the early days of Starhopper, which was back— or Grasshopper, I mean, which was like the SpaceX project in 2013, I want to say.

Yeah.

That was just doing the suborbital little hop-ups and downs to practice for Falcon 9. And so I do think they have this limitation when it comes to deep, deep systems engineering and requiring some creativity that they miss out on. And I think that's why you've seen them partially succeed in semis, right? They're good at the, like, you know, sort of last generation of semis and, you know, sort of reshoring that to China. But I don't think you've really seen any progress on them, you know, sort of, you know, domesticating cutting-edge, you know, sort of semis. TSMC Arizona, you know, is at least a very subscale version, you know, of, you know, being able to do actually some amount of cutting-edge domestically. And I do think it, you know, sort of shows that there's just some breakdowns in Dan Wang's argument where, like, we are able to import that type of, like, I forget what he calls it, like industrial process or process knowledge in the United States and even develop it ourselves, right? The other area that obviously, like, we do quite well on, you know, it's not the perfect company, but like, look, like the, you know, 99% of commercial airliners in the world are still developed by the West, right? You know, Boeing and Airbus, and continue to be. And it's not obvious that China has really been able to displace that. You know, the only other areas that I think about actively that you need to be, you know, sort of reshoring Some amount of it is like, hey, if we want to compete against China and Taiwan, it is going to be almost certainly like a mass-scale production sort of game, more so than anything else. And so when you think about mass-scale drone manufacturing, mass-scale metal cutting, mass-scale fighter jet production, that's something that we're obviously behind on. Even the munitions, right? Ukraine has shown that artillery is back to being relevant again. It was not relevant at all in any of the Middle East conflicts, but now with continuous continuous internet coverage and continuous, basically, spy satellites, you can actually justify basically just like any time that your enemy makes a move, you just artillery shell the crap out of them. And that's actually back to being sort of relevant on the battlefield. And they are far better. They have these lights-off artillery shell production facilities that basically just like are 24/7 fully roboticised, produce way more than the United States can. And so I don't know that those are going to be necessarily like the equivalent of TSMC, where you're like taking somebody's preexisting industry and like dropping it here, and that like there's not an obvious place to go import from basically for that type of stuff. So I think we kind of solve that on our own. And obviously Founders Fund, we're making some investments, you know, in that space. Obviously, you know.

Isn't it true too that in China, like, yes, there's flagship lights-out factories that are truly cutting edge, the kind of thing we just don't really have here in the United States, but yet still the vast majority of manufacturing in China is just a lot of labor, like, you know, assembling cheap components?

Hmm.

It shifted. I think that's like a decade out or a decade old sort of thinking where it's like, yeah, the reason that China won for a long time was because of just like the labor arbitrage, cheap cost of labor. When you talk to people—

But I'm just saying, like, with the true— like, we can't buy the— like, When you talk about China's advantage of just being able to produce overwhelming mass across every different sector, there's certain key areas where they are just light-years ahead. And then they're still getting this massive benefit of even the components that get made that go to the Lights Out factory to be assembled with robotics.

Yeah.

I mean, when I think about their advantages in drone manufacturing, it's mostly due to automated facilities, right? Like, it's not because they, like, you know, have a bunch of, like, you know, sort of mass assembly, super precise, small hand, you know, sort of labourers or something like that. It actually is because when you think about, like, the PCB boards, largely basically fully automated, the, like, you know, propellers that they make, largely fully automated, the battery lines, largely fully automated. And so, yeah, I just like, when I think about, like, the edge that they have in drones, yeah, there's maybe some amount that's like, you know, sort of this, like, assembly edge, but, like, I'd put that at, like, 5 to 10% of the edge. Like, 90% of the edge is that they, like, have such deep process knowledge across all these individual, like, different, you know, basically, like, subcomponent sectors. And a lot of those— Colocation.

Say again?

Like, colocation effectively. Yeah.

And the colocation is just, you know, sort of wild, right? And this is also sometimes the problem with, like, this stuff in America when we, like, try and reimport it. It's like, you know, the CHIPS Act, one of the biggest battles was like the, you know, the size of funding required support from a broad set of senators and congressmen. But then by default, they want, you know, sort of jobs everywhere. And so it's been more difficult to fully sort of centralise all of the reshoring of semis in the United States into a sort of single location. Arizona has definitely been impressive, but ideally it's not just the fab. At some point you want to do lithography there and a bunch of other parts of semis centralised there. But in the United States, we actually kind of have a bit of a chips hub, obviously still some amount in Silicon Valley. Austin's become like a really big area. Ideally, those would all be in sort of the same place rather than so spread out. And so it is just It's just harder given that we're not like China that's willing to sort of like put their finger on a scale to a single, you know, city and just be like, you are going to become like the guitar-producing city and everything is going to be, you know, sort of based there. We don't, we don't do that.

Need a new guitar executive order.

It does feel like we're starting to do that with the NeoClouds, with some of the big data center buildouts, Abilene, Texas. Like we're marshaling capital.

But that's still led by individual dealmakers going and talking to these local governments and figuring out, yeah, can we do this here? Okay, no, I'm going to go across the state line over here. I'll do it over here. Sure, they're chasing.

Yeah, we're just much less top-down coordinated. But like at the same time, I think you're starting to see like the faults and the breaks in that system, you know, in China. And Dan Ling talks about this as well, where it's like, look, you know, they clearly still excel in certain areas, but they've also like— I mean, like think about it even over the course of time that I've been at Founders Fund in 2021. like, the number of unicorns being minted in China was starting to approach, you know, the United States, and there was like some real risk and fear that it was going to supersede. Now, 4 and a half years later, you know, the Chinese venture capital ecosystem is effectively, you know, totally negligible. And so, you know, I do think that they've done a decent job on, like, you know, sort of prioritising the industries that matter the most. But also their, like, capital markets are, like, you know, sort of way worse than they were, you know, sort of 4 and a half years ago. You know, they have reimported a bunch of, you know, sort of Chinese nationals from the United States back as scientists into China, you know, pretty good. But they've also Like lost, you know, something on the order of— I forget what the number was, but like 25,000 Chinese millionaires, like, are, you know, exporting their capital.

They lost the Manus team. The Manus team's over in Singapore now. They might be working out of the Benchmark office soon.

Yeah.

Well, speaking of Benchmark, I thought it was interesting, actually. Bill Gurley posted earlier, game recognize game. And it was a quote from the CEO of Xiaomi who said, we bought 3 Tesla Model Ys for disassembly and research inside Xiaomi earlier this year. What a great And of course they're announcing like effectively a copy cap, but it's insane to just say that.

Game recognized game. It's game copies game.

Yeah. It's just, it's saying, saying the quiet part out loud is wild.

But, uh, yeah, recognize game copies game.

I think also, by the way, since I was last, I was last on was when I have been poking the Benchmark bear literally basically like since like, you know, a week or two after joining Founders Fund. And I remember early on I got a call from like one, you know, one of the partners at FF being like, hey, like, look, Look, just so you know, you're definitely making some enemies over there. And I was like, okay. And then they were like, but I wouldn't necessarily like— it doesn't feel right to tell you to like take these things down because like at the end of the day, we are a founders fund and you are critiquing them for firing founders. And that is our whole ethos. So just know that what you're doing is definitely creating enemies and you guys are creating risk. So anyways, I've been poking the bear for like 5 and a half years.

5 and a half years.

And then it was just so glorious to have like finally an official response from Teton. And it was just like, it was an alley-oop. It was the perfect response. It was like, how dare you not basically like, you know, you know, control, you know, sort of a company's choice and, you know, sort of, you know, capital provider. Like, why are you letting them take Chinese capital? And I was like, oh my God, this is perfect. This is like the exact— you don't even realize how wrong you are because you don't even realize that what you're saying is, why aren't you as a venture investor telling the founder to run their fucking company? And that's your critique. And I remember reading it and I was technically supposed to be on this like one week, like Twitter or whatever, like, you know, detox.

Um, but then like, you know, you're on, on Twitter being like, I'm supposed to be on a detox right now, but I'm about to go on a bender.

I know, but I was like, I got like 5 texts from friends being like, how have you not responded to this yet? And I read it and I was like, oh my God. Yeah, fuck this detox, bro. Like, definitely fucking going in. And it was just like, it is, I think either my top or second best day on Twitter in terms of joy provided for me. Like, Like true, like existential, like identity level, like as much as like a child's birth. I mean, like that is a level of joy.

Zero user regretted seconds. Like the least regret.

You would compare it to, yeah, holding your daughter or something.

It is like, it is. Yeah, I think roughly equivalent. The only other day that I've had that was that good was the, like the day after the FTX blowout. And I was the first one to really find and screenshot the Sequoia blog post about him playing League of Legends and getting that out. And it was just, that was just, that was incredible. It was just, it was, it was like I got to break news. Yeah, I got to like poke fun at a competitor. I forced them to like have a response to all their LPs about it. They had to like take down that blog post. It was just like, oh, that was like a truly— Yeah.

Weren't you also better than SBF at, at, at, uh, League of Legends.

Okay.

I mean, like, AOC is better. Yeah. Like League of Legends. So like the bar is very low. I just like—

but at least, at least you got to respect some of the highly convicted investments in Anthropic, Robinhood. You know, he was a gambler, but, you know.

Yeah, dude, he was a good gambler. I mean, like, he just, you know, played the line a little too hard. But like, yeah, there's a world where like, yeah, I forget who was telling me this. It was somebody that was basically responsible for like brokering the like FTX like bankruptcy, etc. So they're like a Wall Street, whatever, you know, sort of type. And they basically were like, yeah, like if he had just lasted another like 6 months, the book like would have been totally fine. He wouldn't have got liquidated. And today, if he just held these investments that he made from like back then, he would be richer than Elon. Like he would actually just be the richest person on the planet.

Wait, what?

And so it's just, it's wild to think that he was like on a knife's edge.

Richer than Elon.

I don't know.

That's crazy.

Putting up—

I haven't seen that actual number, but I mean, I guess it makes sense with like the various positions he had.

Yeah, I think it was like assuming that FTX continued to go, etc.

And Solana.

Best possible scenario.

There were like, I mean, again, some of those moves were just like, it's like the Anthropic position, I forget, is like now like $25 billion.

I mean, the guy was insane. I don't know if you remember, but he went on Tyler Cowen's show, Conversations with Tyler, and Tyler asks him, would you flip a coin where there was a 50% chance of humanity being completely wiped out or 50% chance that you double the prosperity of humanity? But it's 51-49, the good outcome? And he was like, I would flip it 100 times in a row. And it's like, that's insane. Like, that's not how those things work. But he was like, the expected value is positive, so I must take the EV-positive bet. And it's like, no, no, no, no, no, please do not flip that coin. Like, you're going to wipe us all out. It's like 99%.

The best part of SBF having to go to jail is Dustin Moskovitz finally shutting the fuck up about EA.

Thank God.

We can stop talking about the dude who only met EV Positive things.

Yeah, it's wild. Well, speaking of net EV positive things, is it no longer net EV positive to invest in a company that's growing revenue triple, triple, double, double, double? Because I have a lot of friends that have been doing that. They previously they were saying I'm about to IPO my company, but now they're thinking of shutting down because Delian's according to the best venture capitalists.

Delian's a fan of zero zero zero zero zero thousand X thousand X.

But yes, I want your reaction to the state of the software markets. It's sort of an ev take a little.

Have you have you done a pure SaaS deal this year?

Um, I did just do like a like basically pure vertical AI SaaS. Let's go.

Don't even do it. I couldn't resist. He had to do one.

He had to do.

Can't help myself.

Vertical SaaS underrated criminally underrated. That's literally. I talk to every VC. It's like, you just go invest in the foundation model layer, go invest in hard tech, leave the vertical SaaS for me.

That's great.

Yeah.

But yeah, overall thoughts on the triple, triple, double, double, double being out of date with how fast the foundation model companies are growing. I had this interesting thought that it's like you could be running a janitorial company, and if you have a foundation model as a client and they're growing 10x, you're going to be cleaning 10 times as many toilets. Your revenue is going to be 10xing because they're 10xing. And if you just are a barnacle on the side of the massive whale, you're going to see massive growth. And I wonder what that means because—

We're investing in the barnacle economy.

The barnacle economy. But in general, what are your thoughts?

Yeah, I think with this stuff all the time from the space perspective too, where you can get deluded into this world where it's like the space economy is growing, but if it's only a bunch of other space startups, like, you know, trading revenue back and forth, you can, like, prop that up for, like, quite some time trading a dollar back and forth. But at some point, it needs to connect into, like, the rest of the ecosystem, right? And so, you know, I think there was, you know, some, you know, memes and jokes going around, you know, sort of yesterday about the, like, Nvidia, Oracle, OpenAI, whatever, $100 billion investment, et cetera. And it's literally just, like, the same dollar effectively going, you know, sort of in circles over and over again.

Yes.

And that can, you know, sort of work, but, like, that can also blow up. And we just— and I think we talked about this in maybe even one of the last interviews, It's like you have just this crazy dynamic where, you know, the percent of GDP being invested in this category is like, you know, the equivalent of like at the railroad bubble, etc. But a lot of the companies that are doing this have like these cash flows from other very large businesses that are driving towards this, right? You know, Zuckerberg, I think, gave the quote where he's like, look, I would prefer to like, you know, waste hundreds of billions of dollars.

Misspend a couple hundred billion.

Yeah, spend a couple hundred billion rather than lose the machine god race. And so it's just like the level of— it's just like, it's like nothing we've ever seen before in terms of like the supercycle of, you know, sort of capital going into this. Maybe that has the ROI, maybe it doesn't. When I think about it from like the investor perspective, I mostly think about it twofold, which is like how you should analyse any business, which is basically like quality of revenue and like, you know, in terms of like durability of that revenue. And then, you know, the competition, e.g., the margins, you know, sort of on that revenue.

Right.

And so I'll maybe provide like one example that I think is very strong and one that I think is like a little weaker. You guys have had, you know, sort of Melissa, you know, John Lita's, you know, sort of wife on for Cybernetic Labs. Labs. I think of that as a great example where, you know, she has really great revenue growth. And I believe it to be very durable in that most of her revenue growth is coming from things that are totally decoupled from the, like, insanity of, like, you know, Nvidia, OpenAI, et cetera, circles. Like, it literally is janitorial companies, right? Effectively, HVAC, plumbers, et cetera. Those people's revenue, their growth, et cetera, like, is completely, like, economically decoupled And so in some ways, then Melissa's revenue is completely economically decoupled from the entire AI hype wave. And so when I think about, like, oh, okay, let's say imagine a world where all of a sudden, like, you know, the Mag 7 decide to, like, you know, reduce deployment into data centres and, like, foundation model training by 100x, you know, sort of next year, which companies are affected by that? Okay, well, like, CoreWeave, I imagine, is, like, going to be affected by that. Oracle, definitely going to be affected by that. OpenAI, Anthropic, definitely going to be, you know, sort of affected by that.

Yeah, you look at, you look at how many companies are now indexed to OpenAI, right? You have Broadcom, Oracle, SoftBank, CoreWeave. Like these are—

There's a lot of them.

Yeah.

A lot of them.

And I think as an investor, you have to take a little bit of like a bifurcated strategy where it's like, it's, I think it's a generational index. And this is, you know, if there's one thing that I admire about Peter the most is, um, he just does such a great job of, even though he has lots of biases, particular, you know, preferences, etc., on one sector or another, founder or another. He's very willing to very, I feel like, soberly analyze the just like broad macro world and identify, hey, even if like I don't love, you know, whatever boring SaaS software, etc., or like the AI stuff, you know, maybe might not be real, still be willing to like analyze it from a purely investment perspective and like take it on. And obviously we've done decent checks into this field versus like I literally like find my biases so strong that I can't even like, I can't even motivate myself to even spend time on it, to analyse it as an investor. So that's where I think about the durability perspective. And then there's the margin perspective, which is how much competition is there, right? You're seeing, obviously, in these foundation models where just the token cost continues to go down, the margin never seems to really be improving. And a part of it is like, there's a bunch of companies and capital that's going into this. And some of the, let's say, AI for whatever coding has maybe a bit of a similar dynamic where it is like, you know, a bunch of companies that are working on it, the foundation models working on it, et cetera. I don't think there's going to be like 15 AI for plumber, you know, sort of companies, right? And so this is where Melissa both has this like great durability and great margins. And so I think you have to take this kind of bifurcated approach as an investor, which is both like invest into the like index effectively. Now, our approach has generally been invest mostly into the thing driving the index, e.g., OpenAI. And then you're like, you know, sort of counter strategy is invest in things that are totally decoupled from the index and like super far away. And then basically, like, nothing that much in the sort of messy middle in between. And to go back to, like, the whatever, 3-3-2 or whatever it was, I do think that in both categories, yeah, I do think there's some merit to it where it's like the growth rates and EV growth in these companies just do look very fundamentally different, and the bar is definitely higher. It's just been like, it's been very remarkable to watch Melissa and her company grow at a rate that just Yeah, it's— I just like, I, you know, if you had told me, you know, right to distribution of where you think Melissa's revenue growth is going to be, like, once you, like, started the company, this is in, like, the 99th percentile, you know? And I think it just speaks to how much of a wave this can be. And I think Peter talks about this where it's like the internet wave happened, but then it still took, like, 20 years to get the integrate— internet integrated into the rest of the economy. I think you're gonna see the same thing, you know, sort of here where, like, there might be not as bad of, like, an AI bubble crash, But the, like, implementation period still may take— might be more accelerated because just generally society's moving faster, but it may take 10 years. And it's companies like Melissa that are getting it integrated into, like, the broader, broader society.

Have you seen any companies where they've done a good job? They're still in founder mode. They're, they're, they've grown like a base of business that then when the AI story came, they could actually act, they could act on that in a way that it wasn't like, They're a public company and they need to go through this massive transformation of how they bill, but they can just add AI on top and then they're actually growing new quality revenue. Are you feeling that at all? What are you pointing to?

Let me figure this out. Sorry, I'm having to—

Oh, yes, yes, yes. Ramp.

Okay, there we go.

There you go.

Yes, yes, yes.

There we go. Ramp.com. There you go.

Ramp.com. You know, great company that has done a great job of like, look, they obviously like were growing very well before any of the, you know, the AI boom.

Yep.

But like, I just remember this very particular— I think they've announced this now, so I'm not like revealing anything, right? They have their like, you know, AI agent, you know, basically product. But I just remember the board meeting where Kareem and crew presented like the vision for it. And I was like, oh, this is like what a very competent but like larger startup that is still fast moving, knows how to adopt things, things does in this world, where it's basically like, yeah, look, we have a bunch of people that obviously use what is a very beautiful interface that automates a lot of finance work, but it still requires people to go in, set up these rules engines, do some amount of tagging, et cetera. We're just going to basically start to offer for some of our beta customers that are going to try this out. We're just going to effectively screen record what they're doing and start training AIs on that. And it's like, I just can't imagine both with if you're a net new company being like, I'm going to make AI agents for the CFO suite. It's like, well, but there's like a whole set of things you have to do. You have to have like the corporate card, you have to have travel expenses, you have to have this expense policy, accounting integrations into the ERPs, integrations into like their tax. It's like all these systems that you need to build that are like more Web 2.0 or whatever, you know, sort of systems.

Yep.

And only when you have that and you have a tonne of usage on it, like, can you then actually start to train the like AI agents on top of that usage, right? Instead So I do really, I've admired that a lot in Ramp. And so that's probably the one portfolio company that I've seen it the strongest in. And then the second is this other portfolio company of mine, Sword Health, that, um, you know, had been doing this like, uh, basically, you know, AI sort of physical, you know, sort of therapist. But most of it was that basically you had these like sensors that were strapped on your body. You would go through your physical therapy routine and then a like PT would basically then review your movement data, et cetera, and like message back and forth with you and adjust it. That was a very obvious area where like they literally had years and years of training data sets on how their digital PTs were basically interfacing with patients, they basically significantly upped the ratio. I forget the exact number, but my guess would be it was something on the order of, like, I remember when the company started, it was like 12 patients per PT. I remember by 2022 it was like 1 to 100, and I think now it's something on the order of like 1 to 1,000. And it is largely like the AI wave that has enabled them to do that. And especially because it's like this text back and forth, literally, with a physical therapist. But again, if you were starting from scratch being like, I'm going to make AI agents for physical therapy, It's kind of hard to do just that. The physical sensors, the distribution with Fortune 500s, health systems, et cetera.

Yeah, we had a company on, I think, a couple of days ago called Filevine that raised $400 million. They were, for 10 years, creating just workflows for different law firms.

Just SaaS for law firms.

SaaS for law firms. They have all the workflows, time tracking, conflict checks, et cetera, document storage. Now they can You can just add AI in a bunch of different places. And maybe they don't have their own foundation model, but maybe Claude or OpenAI models become good enough, right?

Yeah. Give us an update on space defense tech, stuff that's happening in DC. It seems like there's a ton of deals there, but they're 2 orders of magnitude smaller than the AI sometimes. It's like a billion-dollar deal for Palantir in the UK. And people are like, I just heard about a $100 billion deal over here from Nvidia. But there's some amazing stuff happening. What's been top of mind for you in space or defense or D.C. generally?

Yeah, I mean, you know, sort of big macro story for the next 5 days is it seems like we're very likely headed towards a government shutdown. I think we've talked about this in some of our prior appearances. Yeah, there's gonna be a lot of people furloughed. RIP. You know, we've talked about this in prior—

I think, can you imagine if we had a a precedent like that in the private markets where like companies just routinely like, yeah, we're out of cash, we're talking with the board, but everyone's furloughed. Everyone's furloughed. Just probably it'll be like 5 days, maybe like 40 days, you know, whatever.

Imagine, by the way, imagine like Xi Jinping coming out and be like, oh yeah, sorry, like we couldn't, like, you know, with my finance minister agree on a budget. And so we're shutting down the CCP for like, you know, it's just like, it's so funny to me that like Like, they so badly want to usurp America, and our system still has us regularly shutting down for extended periods of time, unable to agree on the most basic facts of like, what should we be even like prioritizing and allocating towards? And we still kick their ass. We have the best capital markets. We have the best nation, the best freedom, the best fucking technology in almost every single area.

If you're a career politician, it must feel so good to shut the government down for a week.

I just, what I feel bad is like, now that we're so like, Avarta is so deeply integrated in all these government programs, like, there's a lot of people that just like end up having to like continue to do their job, but like they just don't get paychecks. And it's like, and some of these people are like friends of mine now that like have worked in technology and are over there. And it's like, yeah, like you just kind of have to make sure that you have savings because like rent still keeps going. You're kind of still expected to work, but you literally just like don't get paychecks. And at some point you get some amount of back pay.

Wow.

But, but yeah, the reason we're marching towards it is like, look, we've gotten into this world where like Like, you know, I think it's something like 4 of the last 20 years we've actually passed, like, our budget on time. So the fiscal year for the government basically ends, you know, basically end of September every year.

Yeah.

We effectively, for many, many of the last 20 years, we never end up, you know, basically actually figuring out what the budget should be. So we go into these continuing resolutions. But when you're in that state and you're doing everything so last minute, it means much more regularly we have shutdowns. So shutdowns are happening, you know, sort of more often. And so this year it's the Dems and Republicans going back and forth on, like, you know, basically How much should be based just off the big beautiful bill? You know, how much, you know, should the, you know, sort of Dems have to, you know, sort of fold on, you know, some of their asks around, you know, a lot of it is like healthcare spending, you know, basically related. And at least right now, the like tone in DC is like there isn't an obvious, like, you know, you know, going to be a path forward over the next week or so. And then Speaker Johnson isn't even necessarily like releasing a schedule on when he's going to call everybody back to even, you know, sort of fix it. So it's an interesting place where You know, CRs and shutdowns generally favor incumbents, right? And so, but what's interesting though this year is like now a bunch of the net new tech companies are kind of the incumbents, right? And so, you know, if you were Anduril in 20-whatever-18, government shutdown, pretty painful. If you're like Anduril in 2025, yeah, I mean, you'd prefer for budgets to keep going, but like you can actually probably even continue to beat plan, et cetera, even with sort of government shutdown. And so mostly what it makes it painful for is some of the next-gen companies that have been started in the last year or two, really sort of painful. So at least as somebody that's closer to an incumbent now rather than a 2-year-old company, I don't mind the government shutdown that much. If anything, it kind of gives me more room to get away from some of the sort of seed, Series A, Series B companies. So, Mr. Politician, shut down away, and we'll keep flying capsules.

It's also, if the government shuts down, then if it restarts, it'd be a bullish catalyst.

We need it.

Send us to new all-time highs.

Close us out with a white pill. What's the best new development happening in defense tech space? Any positive news aside from the government stuff? The shutdown's disappointing.

Will you be buying the Unitree IPO?

That's depressing.

Wait, no, here's what's depressing is it's going out at $7 billion. What does that say if they're the leader and the market says this is a $7 billion company and we have companies that are worth $40 billion? If you look at Tesla, what premium does Optimus give to Tesla? We're clearly valuing the potential of humanoids more than at least the Hong Kong Hong Kong Stock Exchange.

This is kind of related to like how I think you could use Unitree for peace, where if I were to provide like, you know, sort of the analogy, what Apple is to Foxconn, you ideally need some US company to be to Unitree, where it's like both the US government and the CCP are not happy with the Apple-Foxconn relationship, right? The US government is not happy that Apple does so much of its manufacturing with our primary adversary and would like them to ideally really pushed to relocate more of that to more allied nations like Vietnam, India, etc., which they're starting to do, and do more of that in the United States. The CCP is not particularly happy that Foxconn puts in all this work, technology, process, etc., ships out these iPhones, and then that just ships a ton of profit dollars over to the largest company that's owned by their adversary. And so both sides are super unhappy with the relationship. And yet the relationship has now persisted for like over 20-plus years. Right. So I think there is the downside case of Unitree succeeds a lot in China and just makes humanoid soldiers that end up invading Taiwan. That's the downside scenario. The upside scenario is Unitree becomes a contract manufacturer, but an American company has way better foundation model, AI controls, sort of robot, whatever operating system and design, and uses Unitree basically as the manufacturer. And then you basically get into this unhappy wedding on both sides in a way that actually makes it even less likely that basically China—

Yeah. Only downside is the sci-fi scenario where there's a backdoor into the millions of humanoids that get deployed into the US. And yeah, I mean, yeah, but my iPhone's not going to get up at night and stab me in the chest.

I mean, it could explode and it's in your fucking, you know, I mean, like, look, you thought the Hamas fucking, like, you know, pagers were bad. Let me fucking tell you, there's like a little bomb in each of these and like you can explode like just government officials, you know. So could be could be tough.

Okay, I asked for a white pill.

That was, and I got a black pill.

Close, but we'll get you next time, Delian. Thank you so much for hopping on.

Great to see you. Congratulations. Congrats on the little one.

We will talk to you soon.

Later, boys.

You heard him talk about eight sleep. You can enjoy an eight sleep if you go to eightsleep.com. Get a pod five, a five-year warranty, thirty-night risk-free trial, free returns, free shipping.

How did we do last night?

I got cooked. I got a seventy-eight.

I got a seventy-seven.

I got a 73, so you beat me. Okay.

73.

Our next guest is from Flock Safety, a Founders Fund portfolio company, actually. We have Garrett Langley. I'm very excited to bring him in from the Roost Room waiting room. Garrett, how you doing?

What's happening?

I'll let you take the intro.

Welcome to the show.

Good. How are you?

Doing great. We've been super excited to have you on first time.

Why don't you give a quick Yeah, or hiding under a rock, maybe. Um, but, uh, yeah, Garrett Langley started a company called Flock Safety with a couple friends 8 years ago, and we catch bad guys. Um, we will help local law enforcement make just around 700,000 arrests across the country this year, um, all violent, nonviolent crimes.

What was the, uh, what year did you guys start? How quickly? What was the time to first arrest?

Yeah, just under 60 days. So we were in YC, and it was horrible because we had built this product. We had a couple neighborhoods using it, and we were like, Demo Day is going to be a bust because like, yeah, we built a camera, but like, what else? And I kid you not, the week before, this guy broke into someone's home, stole a nice road bike. DeKalb County, Georgia made the arrest. And so we went on Demo Day, we had one slide and it was this mugshot, and we were like, we build cameras that do this, please invest.

Wow, you already had the mugshot. Yeah, yeah, that's better than like, you know, we got $2,000 MRR. Yeah, real-world results. That's great and feels more critical than ever with everything going on in the world. What have you guys been up to lately?

Yeah, I mean, so we're live in just over like 6,000 cities now, so pretty well deployed, um, and building a lot. So we got into the drone business earlier this year, kind of at the end of last year, early this year. We've now got drones flying all over the country, um, just kind of launched that also for the private sector. So it might surprise you that Private enterprise spends north of $30 billion a year on unarmed guards. I think drones are a pretty good alternative. They're cheaper, they're more reliable. And so now, like, if you're—

And unarmed guards are effectively just providing like a deterrence, right? They're not meant to engage, but they're just trying to show that there's— that you're not, you know, you're not alone, basically, right? And so, yeah, drone can, can play a similar role. And potentially even a better role if it can follow a bad actor or something along those lines.

Yeah, I know, exactly. So I mean, you think about like a retail example, which if you're in the Bay Area, this happens a lot. Stolen car, car gets stolen, they drive it to a Home Depot, to a Target. We know that car is stolen today. We just notify 911, and, you know, depending on what city you're in, a couple minutes or 20, 30 minutes later, 911 shows up. Now, with our drone, that drone can get automatically dispatched off a rooftop, out of a box, completely remotely kind of teleoperated. You've now got live video streaming to local law enforcement inside their vehicle. This is the vehicle. This is the person. They're going into the store. They're leaving the store, whatever it may be. And so it's kind of the way to think about it is we just treat drones like a camera that can fly. So we're pretty good at building cameras, and now these cameras can fly and chase bad guys.

Can you give me some timelines on when the average American would be seeing just drones flying around, what scale we're going to see that? Are you bullish or excited about humanoid robots or wheeled robots? We've seen what Starship Technologies has done. There's a whole bunch of companies that are putting more simpler robots on the ground.

Yeah.

What's the actual path and what kind of What kind of timelines are you kind of operating against?

So, I mean, the good news is that these drones fly high. So we fly 400 feet up in the air.

Okay.

So if you're not looking up, you're not going to hear it, you're not going to see it.

Wow.

The other thing is our drones can cover up to probably 30 square miles effectively. And so the sheer number of drones you need actually isn't that many.

Sure.

So, if I think about a county near me, it's the largest county in Georgia. I think it's just over 500 square miles. They're going to get 20 drones to cover the entire county.

Wow.

And that's a couple million people. Like, you're never going to see these things.

So, you would probably wind up either partnering or building yourself, like, something that looks kind of like a cell phone tower, has charging infrastructure. Maybe there's a human that goes out there and fixes parts if there's problems.

Does it even need to look like that, or can you just put it on a roof of, like, a Walmart?

Even better, fire department. Fire departments are legally required to be equideployed in a— there's a federal regulation of the density of fire departments, which makes them a perfect place to also put drones.

For sure.

So you put them on top of a fire department, uh, they're fully autonomous, they live up there, it's got HVAC, it's a pretty cozy place to live as a drone. And then someone clicks a button, uh, from a computer just like you guys are on, drone flies, and then it can track the car, locks in, tracks the car, the human, It's locked in. You can do multiple drones to one operator. So even in a county like the one I'm describing with 20 drones, you know, they probably need 3 or 4 operators at all times, but like they're covering 500 square miles, like insane coverage.

Yeah. Talk to me about some of the other robots, robotic dogs, wheeled, 4-wheeled robots, humanoid robots. Like when does this happen?

Well, and before that, like talk about autonomy timelines because if you're flying 400 feet in the air, at what point, like, there's not a lot else going on up there. I imagine you could build these to be pretty fully autonomous pretty fast.

Yeah. I mean, the autonomy is already there. The issue is the FAA regulation. It's like right now, we are legally required to have a human click a button that says launch drone. And then, they need to have a Part 107 license that says they can fly.

And they're like an actual—

that's a pilot slice.

Uh, it's a, it's an easier version. It's maybe 20 or 30 hours of, you know, studying.

Okay. Specifically for commercial drones.

Yeah.

Yeah. And I think it's good, but the autonomy is there. Um, and the FAA has done a great job over the last few years of kind of like trying to catch up. They were a little slow for a while. They're catching up now. I mean, they're making better decisions. Like they're now allowing us to fly multiple drones with one operator. It used to be it had to be 1-to-1. It used to be you had to be able to see the drone. Now you don't have to be able to see the drone. So in the case of that big county, that person sitting downtown, they're flying a drone 50 miles away.

Yeah.

And I think that makes a lot of sense. So the autonomy is there. It can autonomously track the car, stick to it, so you can avoid a high-speed pursuit. But the FAA is going to make sure that it's safe. And I agree, like, safety is the most important feature we sell.

Yeah.

How much of a— It feels like high-speed chases are like, uh, definitely a bug, not, not a feature, right? Because it creates an entirely new danger, right? Of, of, hey, let's get— we've got one guy that stole a car, let's get 6 cops driving super fast.

What can go wrong? Yeah, it's probably thrilling though.

Yeah, I think it's like probably one of the highlights of the job, um, if you're like an adrenaline junkie, because you're driving 90 miles an hour trying to, like, chase someone. No, it's definitely a bug. And I think most elected officials would agree, as they've started to ban it, which creates all types of new conflicts. So I think high-speed pursuits are a really good one. I think the other thing that we've been surprised by is the reduction in service calls. And let me try to expand on that.

Please.

When you call 911, the average response time might be 30 minutes. And a part of the problem is it's just slow to drive. So we'll get there faster because of 'Cause we don't have to wait in traffic. The other issue is that a lot of the times you'll call 911 and say, hey, these 2 guys are in a street fight, and then 20 minutes later those guys are gone. We still had to send an officer. They still had to look around. Maybe they go into the gas station and grab a Gatorade 'cause they're thirsty.

Mm-hmm.

And all of a sudden, like, we've wasted hours of this officer's day. So one of the beautiful parts of the drone is actually it gets there faster and it can also remove that call for service from the backlog. so that for actual critical incidents, we get humans there faster as well. I think the other one that I'm pretty excited about is we have a number of colleges that we work with that are using drones to help escort females at night back home. So you can call 911 and be like, hey, I don't feel safe, like it's really dark, I just left this party, and you don't feel great, I think, as like a 19-year-old maybe doing that because you're like, now I'm gonna get in the back of a patrol car.

Yeah.

Now, in a couple colleges we work with, they're pushing this out like, hey, you can always call 911 and and a police drone will escort you home.

Whoa.

And we'll have video, we'll have eyes on you so that if anything starts to look suspicious, like we're on top of it. And I think there's just a ton of use cases that we're just uncovering now to make drones just a more bigger part of our daily life.

If you run the, the like economic calculation on the 30 square miles for one drone and there's cost to service that and batteries versus do some sort of deal where every stoplight So, you know, every streetlight has a camera on it and you have sort of coverage, 360 cameras on every single corner. Is there an economic calculation here that makes sense where the drone's more efficient, or is it just more flexible? Like, what are all the trade-offs that a city or you think of?

Yeah, so the way we think about it is there's a certain cost per citizen that a city is comfortable paying to eliminate crime. And we do think you can eliminate it. Like, you can't eliminate emotional crime, like crimes of hate, crimes of passion, but the capitalistic crime. Like, there was a really funny No Jumper podcast a couple of weeks ago where the guy was like, those effing flockers, they're making it too hard to commit crime in San Francisco. And I'm like, that's the goal. Because like, that guy knows—

On a podcast?

He's super smart.

On a podcast?

Wow, he called you flockers. That's hilarious. It's brutal.

That's so funny.

Yeah, it's just about like, it's about making it so that, you know, you're— somebody could still do it, but it's not economically viable.

Random podcast. That's amazing. Yes.

That's product market fit. Getting called out on No Jumper for, for stopping crime.

For stopping crime. So when we look at that though, and like, you know, the, the, the probably the most deployed city we have is spending about $22 a citizen per year with Flock. I think that's pretty reasonable. That's not actually not that much money. Um, it's a lot for that, that city, but I mean, as an individual, I believe that city's gonna solve every single crime that happens. I don't think that's too much of a burden, but I do think it's about, you know, think about building a cake. You got to have different layers. You don't want just all cake, all icing. You want cameras, you want cameras that fly. You also need software, you need trailers because there's more of a deterrent. So I don't think it's a one-size-fits-all, but for us right now, we kind of see that deployments where you're going to cover your perimeter with cameras that track cars, you're going to have kind of PTZs, pan-tilt-zooms, for your major intersections, And they're going to use drones to kind of COVID the whole city or county as an overlay.

How how do you back to humanoids?

Sorry.

Oh yeah, I know we're gonna get there.

Just yeah, I really want to know because cost per citizen is probably really expensive. We're seeing stuff from Unitree. There's companies in America. Optimus is far away. I just give me your humanoid take.

Sorry. So so my so my humanoid take is like as it stands today, they're just way too expensive. Um, so if you think about like an average retailer. some of them make a lot of money, right? If you think about like a big box retailer, they might, they might generate a couple million in EBITDA per location. But then you go to— do you kind of go downmarket to, you know, an Ulta, Sephora, a Dollar General? Their box profit is just like not very high. They have a lot of locations.

Yeah.

But their individual stores are actually not very profitable. They can't spend more than a couple thousand dollars a year per store in safety. I just don't see a world where either humanoids or, or dogs work there. And we're seeing that unfold as well, where bigger footprint locations just have both more assets and more dollars. In my conversations with retailers in particular, the kind of two-wheeled approaches have been laughed off. They get stuck in a corner.

Yeah.

They get kicked over. The dogs are seeing more efficacy. And I gotta imagine like a biped, like humanoid, is gonna be the most effective because that just sounds really scary.

Yeah.

Well, now, and you see some of them, you can't kick, you can't kick them over.

They come right back up. It's, it's crazy.

Talk about a scary situation. You're trying to rob a place, you try to kick it over and it just stands back up and just looking at you just like, what next?

There's gonna be like dark web tutorials of like how to get it stopped. Yeah.

All of a sudden, the bippers bring EMF jammers.

It's so sci-fi. You have to cut these cables on the back.

It's going to be wild. Sorry, Jordy, you had another question.

Yeah, I was just curious about how you think about product prioritization, because I imagine when you get in working with these cities and police departments, and they start using your products and getting value out of them, they just start coming to you with more problems.

Yeah.

Both like hardware, software, but it feels like the drone opportunity's big enough that at some, you know, it's, it's probably hard to, that you have too many opportunities, uh, than you could probably pursue all at once.

Yeah. I mean, I think we took the company from 1 SKU to 8 SKUs in the last 12 months that all have clear line of sight to 9 figures of ARR.

Crazy.

Amazing. Yeah, I agree. Uh, and like the pipeline Yeah, the pipeline is there. The ARR is growth. Like, it's kind of wild. This will be our first quarter, I think, where, like, our core business, like, won't be the biggest product line, which is kind of weird. I think that's a good thing for us because we're not decelerating on that side. It's just other products are really attractive. But I think it is, you're right. Like, our bigger concern is the market adoption, not our ability to build.

Mm-hmm.

I mean, this is a kind of enterprisey type buyer where they're just only going to adopt things systems so quickly. So even like on the AI front, we have a lot of ideas. We've built a lot of interesting products. They're just, they're really nervous to adopt too quickly. And I think the example I always give is maybe you're trying to book a hotel and you call the hotel and you realize you're talking to an AI agent, but you're kind of like, I don't know, this is better than the alternative because I'm just trying to book a hotel. I don't know, man. I think when you like call 911 and you're in the middle of a violent situation, I think there's something warming about knowing there's another human on the other side trying to help you. Yeah. Not to say they can't be augmented with AI, but I think that's like a much more challenging societal question of what do we want humans to do and what do we want AI to do?

Yeah.

And we are being conscious to not make that decision for the hundreds of millions of Americans we help keep safe.

Yeah. Just as a taxpayer, I would imagine that I would want the person to pick up, but I'd love for the calls to be transcribed and then AI to find patterns between what's happening. And insights and analytics and all of that.

Yeah, there's also something strange. Like, if somebody's calling 911 to report, like, a drunk— somebody that's obviously drunk driving, it's like you're taking up bandwidth from somebody that is a human that could be going elsewhere. It's like, for— maybe it's like you should— there should be no dial times. A Clanker immediately picks up and says, what's the problem? And then routes it either to a human or If it's less like, you know, if it's not like immediate, you know, some violent act is taking place, it can actually be solved by a voice, you know, voice model.

And I think that's the right approach. But yeah, I think to answer your question, like, we're pretty focused on doing 3 things for our customers. Solving more crime, whereas right now it's about a 40% clearance rate for killing someone. So like, you got almost a 50/50 chance of getting away with homicide. That should be zero. You should, every single person that kills someone should go to jail. Nonviolent crime is even worse. We want to solve cases faster. It takes way too long to solve crime, and we want to do more with less. Like most people, I don't know if y'all remember our grandparents, our parents, it was an admirable job to go into policing, to fire. Yeah, this generation doesn't view it that way, which means 80-plus percent of police departments are understaffed against their budget, and it's not getting better, and technology is going to fill the gap.

Yeah. People have compared you to the Anderol or Palantir, kind of like, I don't know, American dynamism companies or whatever.

People have compared you to good businesses.

Yeah, sounds good. But I'm interested about— we talked to Shyam Senkar, CTO of Palantir, about some of the difficulties of deploying software into governments. And he gave the example that one time they built a beautiful piece of software, exquisite system, and they went to deploy it. the team on the other side of the government that was firing it up was trying to run it on a computer that had like 128 megs of RAM instead of like a gig. And so that was kind of the dawn of the forward-deployed engineer. And I'm wondering if you have any unique solves on how to deploy systems, how to deploy software, if you like the forward-deployed model or something else. Like, what's working for you at actually getting your solutions out there? Yeah.

So really similar example. I remember distinctly we were like launching this product and the designer has a beautiful, you know, MacBook and this HD Retina display. It's beautiful, right? Super expensive setup. And she's designed this like super slick thing. And I was like, great, Mariana, like you should go like field test this. So I was like, this looks dope. Like let's launch it.

Yeah.

And she gets into a patrol car and it's a Panasonic Toughbook. It's a 13-inch screen. It's 1280 by 1024 resolution, and he's using it while he's driving 40 miles an hour. And I'm like, this product's not going to work. Like, we got to go back. But that is literally our— we've got, I don't know, hundreds of thousands of DAUs using this product now, and they're very happy, but that's their normal use case. They're driving.

Yep.

They're on a Toughbook. Like, it's a tiny screen.

Yep.

It's low-res. It's old. And so like you're pretty much treating it almost more like a mobile app, but like with a distracted teenager as your customer. And that's been pretty difficult because we want to build these like super high fidelity, you're just sitting at a desk all day doing your job, and like our customers are in the field all day. So that one's been pretty, pretty tough. And then I also think the other burden we've done a good job of going around is just IT in general.

Yeah.

Like city IT is really tough to work with. They're really tough to kind of just like get going. So we tend to build everything we can to avoid IT. There's some pros and cons to that, but so far I think it's been mostly pros.

How does that actually work? I imagine, is that just like you're only looking for like user licensing or like OAuth or something? Like, how do you— at some point you have to plug into the IT systems a little bit, I imagine.

Well, so as a good example, historically everything in our world is on-prem. Like, we're one of the first cloud-only solutions in this market.

Yeah.

And early on, that was painful because people are used to buying a piece of software, they install a desktop application, that installation requires IT support. And so building everything in the cloud, which for this, you know, for y'all, it's like, well, obviously, yeah, let's say a very— it was a very contrarian take 7 years ago, and now it's moving. But I mean, we have a customer That just a few years ago moved off of paper records. They had filing cabinets. Like, imagine if you report a crime, like, hold on, we gotta go find that file.

Wow.

And that— this is a big city. This is a city with millions of people in it.

Yeah.

Um, just got off of paper. And I think in Maryland in 2023, the cloud became legal.

It was legalized cloud.

This is—

are there people marching?

This is— we're gonna protest for legalized cloud.

Yes, we love the cloud. If there's any cities that have still banned cloud, we will—

And we hate paper. The show is presented by Ramp, of course, and we are strong in favor of ending paper and going to the cloud.

Agree. But literally, we didn't do business in Maryland until 2023 because it was illegal to be in the cloud.

Yeah. Did you ever get a test environment with a Toughbook and a car that a designer could actually go and and drive around the parking lot of your office or anything like that?

No, we make all of our employees go do ride-alongs.

Okay, yeah.

It's so much fun.

Yeah.

'Cause then you get to see the whole product in action, you get to meet the customer, you get to go catch some bad guys. Like, it's a pretty fun way to spend 8 hours.

Yeah. How are you thinking about data privacy, where data lives? There's obviously so many advantages to being cloud-based, but then you have some citizens who might say, wait, well, like, why is my image with this private company? I'm a taxpayer, I don't have a vote at your board meeting. I have a vote at my city council. How do you grapple with all that? What's the mood on the ground in various cities? What are the dividing lines?

Yeah, I'm glad you mentioned cities. So, our general stance is we don't write the rules. We create the levers for local politics to dictate. And so, data retention is the easiest one. In some cities we work with, data retention is 7 days, and every single data we capture is purged after 7 days.

Mm-hmm.

But then in some cities like Dallas, it's a year. In New Jersey, it's 5 years. State legislation for 5 years.

Wow.

I don't care. I live in Atlanta.

Yeah.

Uh, I think we're like at 30 or 60 days here in Atlanta.

Yeah.

And that's up to the local politicians to decide like what makes sense for them. Uh, the other thing that we do that's pretty unique is every single action in our system is audited in perpetuity. And so whether you're downloading, you're searching, you're doing something, that's stored.

Yep.

So the city manager, the internal affairs bureau, those groups can actually say, hey, has anything nefarious been done?

Yeah.

So that's a big concern of people. So those 2 things tend to be pretty good levers for letting local politics dictate versus, to your point, no one elected me the police chief of America.

Yep.

I just built a camera at my house.

I want to pitch you—

Great line.

A startup idea that Blake Scholl, the founder of of Boom Supersonics?

Boom, yeah, sure.

He was thinking about doing this. Have you heard this story? He was thinking about building something before he started Boom, probably right around the time you were starting Flock. He wanted to make a smart stoplight that would have a camera on it, and it would see, hey, there's no cars here. Why do I have a red light? Switch them. Feels like it would improve traffic flow. It's a situation we've all been in. You're sitting there at the red light. There's no one coming the other way. Flip the switch. Why is that? Is that a good idea? Is that something you could do? Is that something any company could do? Like, what's the market structure of the stoplight industry? Do you need to do some sort of private equity roll-up? Walk me through your— if you were putting on a VC hat, how would you interpret that pitch?

My pushback on that pitch is to show me the incentive and I'll show you the behavior. Who's getting promoted? Who's making more money if we fix traffic? I don't think anyone.

The opportunity cost of the workers, it's all diffuse in the economy.

No, but this is the question we—

Don't you think a mayor could run on, hey, we have a lot of congestion here, we have a technology— you'd have to do a study or something to show that you could reduce congestion if you had smarter traffic lights. It feels like someone—

Mayors don't— mayors get elected on 2 things. Well paved roads and low crime. That's it. That's the only things that matter. Like so, I mean, I think the product idea makes a ton of sense. Yeah. But I just go back to like one of the only reasons why because you mentioned Andral.

Yeah.

When we were pivoting the business into local government, Trey Stevens was like, "Do not go sell to police. It's a horrible market." We tried it at Palantir. It's not going to work. It's not going to work. I was like, "Well." I don't know, like, I'm a little bullish. He was an investor. He's an investor in the company. It's kind of nice. It might work. And what I didn't realize is systems like Palantir are too many steps removed away from solving crime at a local level. And like, Flock had coincidentally built a product that solved crime right away.

Mm-hmm.

And so for a police department, crime is equivalent to revenue. So we weren't helping them save money. We were helping them generate revenue. And like, that does get a sergeant, lieutenant, a police chief promoted, that gets a mayor reelected. And that's— that would be my pitch to Blake is like, we got to find a way to make this relevant to a mayor. Yeah. And like, just a little bit less traffic isn't going to move the needle.

Yeah.

And have them move off from having no technology and no cost structure.

Well, he's busy building supersonic jets. He's got a good thing going for now. But maybe, maybe that's the next thing he works on. I have a second startup to pitch you based on your answer to that question. Automated road paving startup. I go to the mayors and I say, road paving gets you paid and gets you in the job. Would that work? Is that a good idea? How would that play out if you thought that process through? What advice would you give to a road paving entrepreneur?

Yeah, there was a YC company doing the little road paving.

I remember this.

Yeah, I remember this.

But yeah, walk me through your thought process on this.

So I actually think that's a great idea. We're working on something Tangential to that. So one of the biggest issues is that to do what you're describing, you have to know where the potholes are. And cities have no idea where potholes are. And they rely on you calling 311, which I'm sure you both have done many times and said, hey, I'm just calling in to report a pothole.

Yep.

No, you don't. And that also costs the city on average like $8 per call on the 311.

Whoa.

So it's a really inefficient system.

Okay.

And so what we've been able to do is train our cameras to look for potholes. And so now we can actually report back general road conditions. And that is the beginning of them being able to say, hey, now we can actually be intelligent about where we should send people to go repave. And I think an automated robot that does the paving would be even better. But yeah, I love that idea. I think that I'm in.

That's amazing.

$25,000 check.

Fantastic. Well, you know where to reach him. He is, of course—

Oh, last review. Wanted your reactions. We had a guy named Riley on yesterday who made a Find My Parking Cops. Did you see that?

Oh, that, that park was awesome.

Yeah.

Yeah.

He's a little— did it get shut down or did it back up?

It got shut down.

Shut down.

He's—

Very quickly. I was impressed with San Francisco's reaction time on that. They shut it down like very quickly. But it's good fun.

Yeah. I'm a fan of it. I think most people don't pay for parking because most of the time you don't get a ticket. Yeah.

Well, thank you so much for hopping on the show.

Just gonna hang out. It's great to get the update.

Yeah. Congratulations on all the progress, and thank you for everything that you do.

Yeah. Come back on anytime.

Yeah. We'd love to talk to you soon.

Yes, sir.

Have a great rest of your day. Let me tell you about—

Man with 10 9-figure businesses.

Yeah. If you want to reach people in cities, you gotta go to adquick.com. Out-of-home advertising made easy and measurable. Say goodbye to the headaches of out-of-home advertising. Only AdQuick combines technology, out-of-home expertise, and data to enable efficient, seamless ad buying across the globe.

Before our next guest joins, we did need to acknowledge that Mark Leonard has stepped down from Constellation Software.

Very sad news, for health reasons.

We hope he's okay. Interesting timing, basically called the top on—

Yeah, we talked to Carrie No Interest on the show. You can go pull that up if you want to hear our conversation. With him about these private equity software enterprise software roll-ups. Mark Leonard was mixed on AI, sort of unclear on how it would affect the business long term. But the market did not like it and the stock fell off and the stock has fallen again on this very sad news. So we are, of course, sending him our best. The health reasons that he is stepping down for are unspecified at this time. So but please send your thoughts and prayers. Thoughts and prayers, Mark Leonard. And we have our next guest in the waiting room. TBN Ultra Dome.

There he is, suited up.

Matan, how are you doing? Hello, I'm doing well.

How are you guys doing?

It's been too long. Yeah, we've seen you before in the TBN Ultra Dome. Welcome back.

Pleasure.

Give us the news. Give us the give us the 140 characters and the company refresh everyone, and then give us the news.

Absolutely. So I'm Matan. CEO at Factory. At Factory, our mission is to bring autonomy to software engineering. What that means more concretely, we have built droids, which are autonomous software development agents. And more importantly, I'm here to tell you that they are the number one agent in the world.

Okay.

As we just released today in the Terminalbench benchmark, which evaluates tools like Claude Code, OpenAI's Codex CLI, Mm-hmm. Factory is number 1.

Congratulations.

And also number 3 and also number 5.

Wow.

Pick any model you want. Factory's still number 1, no matter how you slice it. Um, our droids are simply the best agents in the world.

So excited to share that.

Yeah.

What's the secret? What, what, what got you there? Is it, uh, pre-training, post-training, a bunch of RL stuff? Is it a data flywheel? Like what's the, what's driving the growth?

Yeah, great question. So I think one of the biggest things is that most, if not all of the agents out there, are built for one model in particular. So, you know, Claude Code is built to work with the Claude models. OpenAI's Codex is built for the GPT models. There are some, uh, other tools out there that aren't from the research labs that are focused on really like fine-tuning their agent harness for one given model for any given step. But what we've done with our droids is make them fully model agnostic, which actually makes them more performant in the long run. It's kind of similar to if you were a human engineer and you only, let's say, studied one coding language, you would actually be a weaker engineer than if you studied all of them. It's somewhat analogous there in terms of how we've built these droids that now allows them to be the most performant with any model that you, that you put in under the hood.

How are you viewing the market and the customer landscape? Uh, automated software engineering, that feels niched down from years ago, which was just, uh, the, the, the transformer-based LLMs can write code. Uh, but as we actually dig in, we see that the needs of a Fortune 100 customer may be different than a mid-market company, which is different than a startup, which is different than an SMB, um, which is different than a solo vibe coder, solopreneur. Uh, where are you seeing opportunity in the market? How bifurcated is the market? It seems like you've wanted to be very general on the model side, Are you also trying to be general on the customer size side?

Yeah, great question. I mean, I think the thing that a lot of the other tools have missed is the fact that the further you go into the large enterprise, the less overlap there is between software development and coding. So, for example, if you're a solo developer, your software development is basically just coding. You're not really doing code review. You're not doing that much in the way of documentation or testing or design docs And so, where we focus and why we call Droids software development agents as opposed to coding agents is that we focus on that whole end-to-end software development lifecycle, which is really where a lot of enterprises are missing. Like, there's the famous study that came out from MIT that said something along the lines of 95% of AI adoption efforts are failing in the enterprise.

I remember.

Yeah, exactly. And so, I mean, it kind of makes sense. Like, if you think about software development as a pipeline—

Sorry, I love hitting you with the surprise sound effect.

That was so good.

We'll get some more positive.

We're not bearish on you.

It's just—

No, no, no, no, that was good.

It was perfect.

It was perfect. It was just too good. But, you know, software development is a pipeline, right? And if you focus on only one part of that pipeline, namely code generation, and you expand that— I used to be a physicist. I don't know if you guys are big on fluid mechanics, But opening one part of a pipe and doing nothing to the others, you don't actually increase the throughput.

Sure.

You just create new bottlenecks.

Interesting.

And so, the reason why a lot of these efforts are failing for software development is they're focusing on just coding, but then that just punts the problem down to testing or code review or documentation. And that's why a lot of the other tools out there aren't seeing kind of the adoption and the real business outcomes that we are with the droids.

How do you see coding agents fitting into the consumer world? This might not be your business at all, but I noticed that the, you know, I've seen incredible results just from asking a question that you could hit deep research with. But if I ask it to Claude Code or Codex, it can build an entire HTML web page with JavaScript widgets and bar charts, and It just produces, it instantiates the information in a much cooler, it's just interesting way. And I'm wondering how you're seeing crossover from advances in automated AI coding into a consumer world where they might not even know that they asked for code, but they got code. It's already happening a little bit when you hit O3 Pro, it'll write some code to give you the answer. You don't even see it unless you unfurl the— the reasoning tokens. But how do you see that playing out in like kind of the mid to long term?

Great question. I think this is something that people aren't yet fully aware of, but this was actually the very first statement of our announcement, which is the best agents for software development are becoming the best agents for everything. And the reality is because basically every problem can be broken down into some sort of software development problem. And you can actually take it from Anthropic themselves. Literally yesterday, Alex Albert, the guy who runs DevRel at Anthropic, tweets out, you know, the best coding model will be the best model for many types of knowledge work. Code is how computers operate. Anything you do on a computer can be done through code. And this is really why Factory being the number one software development agent is really going to then lead to a lot of other tasks.

Yeah.

And we're already seeing this in the enterprise. So, we sell explicitly to developers, but already PMs, designers, even people on the operations side, so like finance, Biz Ops, that sort of thing.

Yeah.

We find that they're like sneaking their way in even though they didn't actually, you know, put any budget towards it. And they're starting to use Droids kind of behind the back of some of the VPs of engineering. And this is really something that we've expected because of this fact that code is really the language that computers speak. And if you want a computer to do something for you, you will eventually through some means need code.

Couple questions. What models are you guys getting the most leverage out of today? Is it a mix? Are you focused on— you said earlier, I think you're focused on a variety, trying to be model agnostic, but be curious where, who you're paying out on the back Yeah.

So right now we achieved the number 1 score across the board on Terminal Bench with Claude Opus.

Mm-hmm.

But again, we are, you know, really focused on having the model agnostic stance because org by org, they might have different preferences. Also, what the best model is depends, you know, what day it is. You know, tomorrow there could be another new best. And it's important that we have that ability to quickly swap them in. Um, so that's one thing that I think is important. It's also even task by task. There are certain models that are better, and we wanna make sure that the user has that familiarity, uh, with the model so that they can go in and, you know, swap it out if there's a task that they're doing that GPT-5 might be better for or Grok might be better for.

Mm-hmm.

Uh, you forgot to mention you raised some new money today.

Give us the news.

We got a gong here.

That is right. Uh, we have raised $50 million. From NEA, JP Morgan, and Nvidia.

JP Morgan.

Jamie Dimon's getting in.

He couldn't help himself.

Couldn't help himself.

Couldn't help himself. I love it.

Congratulations.

That's our deposits. Everybody, the world's deposits at work.

Talk to me about that last question. We'll let you get back to your busy day. Talk to me about the branding. It feels like we just got the idea of You were talking about the death of an agent recently. Now you're kind of pitching droids. It's clearly your term for—

Droids you're looking for.

Yeah, yeah.

People always talk about, you know, being— we're not the droids.

Yeah.

How much of that is like a differentiated brand that you want to be specific to your company versus like a new coinage that you want to describe a different way of working that you'd actually like to see other companies adopt? seeing the entire industry standardize around that particular piece of language, you would be happy, or would you be like, hey, they stole our brand?

Yeah, I mean, I think generally, like, agent as a term is here to stay.

Yep.

But the point is, as a kind of general purpose term, it's often synonymous with like poor quality or like a while loop wrapped around LLM calls.

Yeah.

And I think what we want to do is to make good on our promise to customers, which is giving them that best agentic experience, and we do that through droids. At the end of the day, when you have a cold and you're, you know, your nose is running, you don't want a tissue, you want a Kleenex. Similarly with software development, you want the droids to come in there, right?

Okay, I love it, I love it. Well, congratulations on the round. Thank you so much for hopping by.

We'll talk to you soon.

Awesome progress.

Thank you guys for having me.

We'll talk to you soon. Cheers. Um, let me tell you about Bezel. Getbezel.com, your Bezel concierge, available now to source you any watch watch on the planet. Seriously, any watch. And we have our next guest in the green room.

Also, I guess we gotta wait to talk about this. 2 new launches, one from Meta. Alexander announced Vibes, which is a Meta AI app for short-form AI-generated videos.

Okay, yeah, YouTube launched something with AI-generated Shorts leveraging VO3 recently. What else?

And then there's a new ChatGPT product called Pulse, but pulse.

We will dig into that.

Get into it with Francis over at Invisible.

Welcome to the TV. Francis, how are you doing?

What's happening? Welcome to the show.

Doing great.

Excited to be here. Thanks so much for hopping on the show. Kick us off with an introduction on you, the company, any news you got to share. What's good? What's new in your world?

Yeah.

Well, as you saw on Bloomberg, we've raised $100 million this year, which is a big turn of events.

There we go.

$2 billion valuation.

Yeah.

Crazy thing that, you know, when you saw Meta acquiring Scale, um, they had raised $1.8 billion and we had only deployed $6 million. We scaled profitably to $134 million of revenue last year.

That is wild. I remember I heard about you guys. I forget what context I heard about you guys for the first time, but, uh, even this was probably 2 years ago. Even back then I was like, Like you guys were somewhat under the radar, I felt like, but I heard some of your revenue numbers back then. I was like, what? How have I not heard about this company? But yeah, we wanted to—

Invisible for not much longer.

Yeah, no longer.

Now very visible.

No longer.

Now we're visible.

Buy the new domain. You're visible.

Visible Technologies.

Yes.

Yes.

Yes. Give us an— I mean, it'd be helpful to hear like where you guys have an edge in the market broadly? And then I, I, yeah, I want to kind of, um, yeah, understand kind of where you're getting—

Yeah, problem set, core, uh, like what you're actually replacing, where customers are getting the most value, anything like that would be super helpful.

So 10 years ago when we founded the company, there was a question, which is, if there's an app for everything, why isn't everything perfect yet? And Salesforce was the first SaaS company in 1999. For the last 25 years, every enterprise software company has been a SaaS company. And this has actually put customers in a pretty absurd situation where if the customer wants a cake, Silicon Valley will not sell them a cake. They will sell them tools to make a cake. And the customer then has to hire a systems integrator like Accenture to stitch together all these SaaS applications into an end-to-end solution that actually works for the business. And so that is what we were disrupting. And Palantir was not public yet, so they weren't a well-understood comp. We were kind of like an ugly duckling because we were talking about AI services. And, and so it's just a fundamentally disruptive business model.

So you were, you were competing initially with someone like an Accenture to integrate existing software systems.

Is that right?

Yeah. We build custom AI applications for enterprises and governments. And so you can think of this very different than SaaS, like a triangle. We have the horizontal platform, which is very powerful, but you have field engineers and field CTOs in a forward-deployed motion that builds custom AI applications. that actually solved the problem for the enterprise and the government. And our insight was that even though every solution would be custom, the horizontal platform would allow you to sort of build infinitely customizable software over time.

That makes sense. We've been seeing Marc Benioff and Dr. Karp going at it over various contracts, Salesforce, which you already mentioned, Palantir, which you also mentioned. Are you in that knockout, drag-out fight? Have you found a differentiator around industry or size of problem, or, or are you just going in the arena?

So Salesforce was founded in 1999. Palantir was founded in 2003. So way before the GenAI era. So we're in a sense like the GenAI native way of doing this. We hired Matt Fitzpatrick, who previously ran McKinsey's Quantum Black Labs, and he's now our CEO. And after years of being capital efficient, this round sort of enables us to go into pre-IPO motion. And yeah, if you're a public markets investor, you don't really have that many options for buying AI today. So, and this is generally true for the enterprises, the customers themselves. If you want to build custom AI solutions, Palantir is like a data integration and decision-making support company. So they really, they really focus on that. But if you have a problem anywhere else in your business, how do you build a custom solution with world-class field engineers? It's a very different type of engineer, very different type of go-to-market person, very different type of business that is familiar building custom solutions on the enterprise. You know, enterprises do not want their data to leave their systems. So they need their— they need all the solutions to be on-prem and containerized. And we have the ability to build in that way.

Do you find— do you think that there's a shift in the AI era to large enterprises doing more of these sort of like semi-custom solutions? Is that going to be a continuing trend? Because it feels like—

Yes.

The last era was very much like rip out the old custom on-prem solutions and go to some sort of like one-size-fits-all cloud solution.

Exactly.

Yes. And this is a reversal.

We're going all the way the other way.

And I think that's because enterprises ultimately need proprietary and custom solutions. So in the SaaS era, like I said, they were stitching together things and they had to do that with some combination of like their internal engineering function and systems integrators. And systems integrators are not tech companies. They're gonna overcharge, underdeliver. You're gonna pay them by the hour. So their incentive is to bill you as many hours as possible without getting fired. And that's not really aligned with efficiency. And so in this era, I think, you know, enterprises are realizing it's going to be very, very difficult for them to build world-class engineering teams and to build a platform like the one we have where we can take the best models, the latest models which we train— we have an AI training business— and deploy them in the enterprise to actually solve the business use cases that they, that they Yeah.

So, so how much should we read into you guys transitioning away from training? Is that a business that is—

We're not transitioning away. We are fully committed to training. So actually, the move of Meta acquiring Scale basically cleared the field. And so it's an inherently oligopolistic market structure where there are a few players that have invested as much as we've invested over as many years as we've invested. And so, we have like 20,000 experts, and that number is scaling very quickly. We can hire like 1,000 experts a week. So, when you're asking a model a question about the chemical properties of silver or the history of Sweden or whatever you're asking a model, we have PhDs, master's experts in every subject that are training these models. And there's evals and lots of inputs into making these models great.

Got it. So leveraging the experience on the training side in order to help the—

To deploy in the enterprise.

Got it. Got it. And then the business model for the deployment side is— I'm just going to repeat it back to you so I understand— is there's some period that looks more like services, but it becomes a piece of durable software that they can use for a long time.

Just like Palantir. Exactly.

Yes. Got it.

Cool. AI services is exactly that way.

Yeah.

And that puts us in an investing mode. We literally invest in our customers. So, we put field engineers, field CTOs in with our clients, and there's an investment period to build the solution. But once the solution's in place, that is very durable revenue.

And so, you're like a nightmare for someone like an Accenture or some of these big consulting companies that don't really have the true expertise around training these models and you get to— and have a lot of organizational bloat. Is that right?

Yes. So, Matt, who's our new CEO, we bonded over ancient Greek and Roman history, and he calls it the Battle of Thermopylae because even if Accenture has like a million people worldwide, has some huge headcount. If there's 3— if the pass is only yay wide and there's 300 people at the pass and we're always winning at the pass, then we're going to win every battle. In other words, if, you know, I believe our technology is superior, dramatically superior already to traditional systems integrators. So if you're a buyer and you're hearing Accenture's pitch and Invisible's pitch, which one are you going to pick? The Battle of Thermopylae, Spartans win.

It's fantastic.

Spartans are coming. Coming for it.

What a market. This is just a fantastic market.

Awesome.

Well, uh, congratulations.

Congrats on all the progress.

Great to be on the show. Big fan.

Appreciate it.

Thanks for having me. What's the— you said prepare to go public. What timeline are we thinking about? Is that multiple years out? Is it—

We think in terms of, you know, how many hundreds of millions in revenue do you need to have before the story is just obvious to a public markets investor. So we really want to build a business for all markets. So the intrinsic value of the business, the ability to survive in any condition is key, and then we'll take it public.

Awesome.

All right, we'll come back on anytime. This was fun.

We'd love to ring the bell with you. Cheers.

Great to have you on. Talk to you soon. Thanks, man.

Bye. Up next, we have Jacob from— sorry, David from JuiceBox coming into the studio.

Juicebox.

I believe he's in the Restream waiting room.

Amen.

Juicebox has some fundraising news. The gong's already warmed up. How you doing?

Welcome.

Thank you. Thanks for having me on. David.

Big day today.

Introduce yourself. Tell us.

You picked kind of a good day to drop a cinematic launch video.

This is a good day for it. Uh, but take us through it. Uh, introduce yourself, the company, and the news.

Yeah, uh, I'm David, co-founder of Juicebox. Juicebox is an AI recruiting platform. platform. We help our customers win the talent war, and we do that by helping them find and engage the best talent. And yeah, today we're announcing our $30 million Series A and previously unannounced $6 million seed.

Boom.

Hit it a few more times. A few more times. There we go. There we go. Congrats. Big day. Okay. AI winning recruiting wars. What does that mean?

And how many, how long until you're out of business because no one has a job anymore? I thought AI was going to get rid of all the jobs.

No silly questions. Dave, are you working more with early stage, growth stage, public companies, all the above?

High schoolers.

Where are you kind of focused right now? Because I saw— I admit I only saw half the launch video, so you had me up until whatever, 50%.

Nice. Did you see the part with the Zuck impression?

I did see the Zuck impression. Yeah.

You made it through the important part. Yeah. We work with all kinds of companies ranging from founders doing their first hire all the way to kind of growth stage companies, a few Fortune 500 customers as well. Some of those companies are Perplexity, Ramp, Quora, and a bunch more. I think the common thread for where we're able to add the most value is if the right person for the role is actually really hard to find. And that can be because they have a specific skill set, a specific set of experiences, something that makes them unique and uniquely positioned to really excel in that role.

I think of hiring as kind of like some sort of chain of events with different point solutions. There's actually finding talent, job platforms, LinkedIn, Indeed. Then there's, you know, the, you know, you might want to send someone a test or have them fill out a form or process resumes. You need an applicant tracking system. I've seen other point solutions for send a video and then ask some questions and then process the video and let the user or the hiring manager kind of review videos. There's a whole bunch of different things. Do you see yourself as kind of a cross-journey platform? Are you focused on a particular landing zone? Where's the product position today?

Yeah, it's a great question. I think of the recruiting space as really having Um, 2 different ways in which you could add value. One is, um, finding net new candidates, people that you wouldn't have discovered or wouldn't have gotten into your pipeline in the first place. Uh, and then everything that's carrying through the pipeline and optimizing that pipeline. And the latter is usually about like saving time and optimizing conversion rates. And the former is about, um, making sure the right talent is speaking to you in the first place. Um, we're focused on that part. And so our goal is to help companies discover talent that they wouldn't otherwise. And we do that by combining a bunch of different data sources and then using large language models to run the search and surface those profiles for you. And so if you think about what a recruiter does today on, say, LinkedIn Recruiter or like a typical search solution, they set a bunch of filters and then start reviewing profiles one by one. It's super manual, requires a lot of thought still to like think of what could make that profile a fit. fit. And that's exactly the process that we're able to do with LLMs.

So is this a direct replacement for LinkedIn Recruiter? Are you trying to get people to just be able to—

Or do you want to puppeteer LinkedIn with a screen recorder or API or something like widget?

I don't know if they'd like that.

Something like that. I don't know. It depends. I mean, there's been some companies that have done it successfully. Some of them got acquired by LinkedIn, but I don't know.

Yeah, no puppeteering of LinkedIn. Most of our customers are existing LinkedIn customers. Our goal is to help you find net new talent and do so in a more efficient way. So it's really about kind of enhancing the workflow of the recruiter and enabling them to do that search that would just be too manual or too time-consuming to do otherwise. Like looking through 50,000 software engineering profiles is just not something that's like feasible. But for an LLM, it's very feasible and can do it in a couple minutes.

So there are, I mean, there are different pools of talent. For— I mean, we do media video production here. We might want to hire somebody who doesn't really have a LinkedIn presence, but they might have an Instagram account, or they might have a personal website where they've done some— they put up their showreels or what they've worked on. Do you have crawlers that could go out and find those profiles on the internet? Like, how could I think about the surface area of you going out and finding me candidates and sourcing?

That's right. So we look at a bunch of different data sources for different types of roles. So I'll give 2 examples, one on the engineering side, GitHub profiles, a ton of value on those, especially where traditional LinkedIn profiles or resumes might be a bit more sparse, especially for engineers who don't put a ton of info on there. GitHub can be really rich and a really interesting data source. Meanwhile, on the sales side, often what we see is customers want to find sellers who have experience selling to a specific type of buyer persona. And so there it's actually a bit more about the company data. You know, what type of companies does their current company sell into? What type of a platform are they selling? Is it a SaaS product? Is it a usage-based product and more? And so all of that information we enrich and we try to aggregate upfront so we can make those decisions for you and help surface the right profiles there.

There's another, they put an example, they have a tag on their site called likely to switch, which would be like you of a round of layoffs, funding or vesting timelines. People don't really think about that. You can see like, okay, this person's been here for 4 years. They're probably getting a refresher, but they're vested out. That's pretty cool. What do you think? Do you have a strong thesis yourself around employment over the next next 10 years, I'm assuming Sequoia leading this round. I think that's—

Sequoia confirmed not AGI-pilled.

Confirmed.

No, but it says that we're still going to have talent wars in a decade, right? That's what that tells me. But how do you think about it?

Yeah, I think if you're AGI-pilled, it actually means hiring the right talent is even more important. And the reason for that is that if you think of like the output of a given person is usually constrained by like their time, their previous knowledge and context, and then whatever tooling they're using. And so if the tooling that they're using is getting like exponentially better, they use LLMs to automate a bunch of the things they'd otherwise do. They're much more productive than they would usually do. And that's all a multiple of the person or like the initial person and their kind of raw capacity. Having the right person in that role becomes even more important. important because if AI basically 10x's whatever you do, you want to have the highest baseline that you can add that 10x to. And so it gets even more important to have the right talent on your team. And if you think of that in like the context of a software engineer, if a 10x software engineer becomes 100x, you want to have as many 10x software engineers as you can. And so my view is that the race for talent will become more and more competitive as AI gets better and better because having the right people on your team has that outsized impact. And I think we're already seeing some of that in like, you know, the war for talent in the AI labs. But I think more and more of that will trickle down into the rest of the economy as well.

Sure.

What do you think about outcome-based pricing famously used in the recruiting industry? Most human recruiters, even if they work for a firm, they might get one month salary as a bonus if they place someone successfully versus LinkedIn, which is subscription software, completely not outcome-based. What model do you think will dominate in the future? And what are you doing?

Yeah, it's a really good question. Um, we, so we have like a per-seat pricing, normal SaaS model, but then we also have an agent product, which is priced per role that the agent works on. And so it's not quite outcome-based, but it's more like usage-based. How many roles are you deploying the agent for? Over time, we want to get closer and closer to outcome-based pricing as, as is possible. You know, if you think of recruiting agencies, it's pretty common to do like 20 to 25% fee of first-year salary, which on like a dollar value is, is much, much more than one could probably charge for a SaaS subscription. And so there's definitely a lot of things that are attractive there. At the same time, I think, The market's maybe not quite there yet or not quite ready to pay for software on like a, say, percentage of hire model, though I hope we'll be able to push in that direction over time as well.

Fantastic. Well, congratulations.

Super cool. Really insane progress. Yeah. And congrats on the new round.

We will talk to you soon. Thanks for having me on.

Cheers, guys.

Have a great rest of your day.

Thanks for coming on.

Sam Altman launched a new feature today called called Pulse. It's in ChatGPT, initially available to Pro subscribers. That's the $200 a month tier, right?

Yeah.

Pulse works for you overnight and keeps thinking about your interests, your connected data, your recent chats, and more. Every morning you get a custom-generated set of stuff you might be interested in. It performs super well if you tell ChatGPT more about what's important to you. In regular chat, you could mention, I like to go visit Bora Bora. Or my kid is 6 months old and I'm interested in developmental milestones, and in the future you might get useful updates. Think of treating ChatGPT like a super competent personal assistant. Sometimes you ask for things you need in the moment, but if you share general preferences, it will do a good job for you proactively. This also points to what I believe is the future of ChatGPT, a shift from being all reactive to being significantly proactive. I completely agree. agree with this concept, uh, and extremely personalized. Uh, this is an early look and right now only available to Pro subscribers. We will work hard to improve the quality over time and find a way to bring it to Plus subscribers too. That sounds compute intensive. That sounds like they're going to be running essentially an O3 Pro level query, a ChatGPT-5 Pro level query every night, asking based on what's going on in the news, based on what's— what this person searched for, based done all the chats recently. Let's summarize and put together some sort of dossier for them and surface that. So that's basically one big, you know, fire up the GPUs every single night. You gotta get the customers to pay for it, at least in the short term. But yeah, what do you think, Tyler? You think you'll use this?

Uh, it's pretty interesting. I, I'm wondering, is this the thing that Sam Altman was talking about, about the compute-intensive thing they're rolling out?

It's— he said, he said this week we will be launching a couple compute-intensive things. Some of those will probably be announced at Dev Day. This is one that he's just announcing this week.

This feels like it's effectively doing something like just kind of—

Hanging out in the background.

Hanging out in the background somewhat actively, somewhat passively, just gauging what your interests are. I can see this as almost like having a Google alert set up.

Amjad from Replit says, reminds me of Google But how do I get started? I don't see it in the app. So it's obviously rolling out slowly.

Yeah, you can definitely see like an agent being built into this. Like, that seems very natural.

Totally. And I feel like ChatGPT was already starting to sort of surface this concept of like, hey, do you want me to do this every week for you? Do you want me to do this every month for you? But I haven't actually set up any of those cron jobs and had it stay stick around and become part of my workflow, but I like the idea of it. And I think that this— that the big thing is just like for retention. If every time you open the ChatGPT app in the morning, instead of just an empty box, you're greeted with like, here's some interesting stuff.

Or you're getting a push notification.

Or push notification. That's going to be huge for retention.

Or it's like, hey, this product launched. Do you want to buy this? You want to buy this? Let's pull up this video from Alex Alexander Wang, just share Vibes, a new feed in the Meta AI app. Let's see if they beat the slop allegations. I think it's powered by Midjourney, so it looks pretty cool.

Yeah, I mean, Midjourney, fantastic model. Looks pretty good. Yeah, I wonder how this is actually going to— so separate app is what he's saying?

It's a new feed in the Meta app.

Okay, in the Meta AI app. So not, not, you would expect this to be baked directly into Instagram. That's probably too much compute out of the gate to just drop it in as a filter or something, or a button in Instagram. Because if you have a billion people pushing that button, you're just gonna get, uh, get swamped. I think this beats the, uh, the slop allegations. There's some cool stuff you could do with this. I mean, video generation looks good. Looks, looks VO3 level to me.

I don't know.

Obviously these were curated and selected, but we'll have to get in the app.

Uh, Tyler, before we do the show tomorrow, please use the app for like, I don't know, like 12 or 14 hours, something like that. Just straight. No, no breaks.

Just get wireheaded.

Just, just lock in.

Just lock in.

Um, Max, Max, uh, Conrad in the chat says, I'm tired of hearing about fundraisers without business metrics, top line, bottom line, value created for customers. Good point. We want to push people harder to say, if you want to come on, you got to give us some numbers. And if you don't got numbers to share other than the headline, we should have a very small call. Credit to Juice. I think they've been actually crushing it raising money.

Yeah, I mean, that is a market where money flows very freely because Because as a business, you're ready to sign a check for $200,000 to hire someone as a software engineer.

$150,000.

You're like, yeah, I'll pay $10,000 to get the job done, $20,000. So the money flows quite freely, and that's why there's so many executive recruiters, software engineering recruiters that make a ton of money just working basically for themselves, just as talent agents moving things around. Well, Meek Mill and Elizabeth Holmes were going back and forth Elizabeth says, thanks, Meek. Currently serving a 137-month sentence in federal prison. Would love to work with you on reform. Here's some legislation I have drafted. Your story inspired me. And Meek Mill says, let me check your story out. And Will Brown says, Elizabeth Holmes, this year is our year. Elizabeth Holmes, we can never forget this, tiger. Meek Mill, I used to pray for times like this to rhyme like this.

I know, focus on it too.

It's a great one.

Yeah, I think they're thinking about different, uh, aspects of reform.

Yeah. Um, well, we don't, we don't rap.

Maybe Meek Mill would want YC arrest where she has to lock in and just repeatedly build the, the device forever.

I like that. Uh, we don't rap, but we do sing.

Find your happy place.

Find your happy place. Tyler missed it.

All right, give me a second to kind of like—

Yeah, you wanna try again?

Find your happy place. Find your happy place.

Book a wander with inspiring views, hotel-grade amenities, dreamy beds, top-tier cleaning, and 24/7 concierge service.

It's well, but we got breaking news: Trump signs order to approve the TikTok deal and avoid the U.S. ban. The chat was on it way before we were, and let's see. TikTok is poised to be spun off into a separate U.S. entity to comply with a 2024 law requiring the China parent-based parent company ByteDance. Divest or face a US ban.

Mm-hmm.

Trump said, I had a very good talk with President Xi, a lot of respect for him. Hopefully he has a lot of respect for me too. And we talked about TikTok and other things. We talked, but we talked about TikTok and he gave us the go-ahead. Under the deal, a group of US investors including Oracle and Silver Lake are set to take a majority stake in the new TikTok entity. Let's give it up for Silver Lake. Good to see them throwing some size around, as well as Oracle. ByteDance will maintain less than 20% in equity to comply with the divest or ban law. Vice President Vance said the company will be valued around $14 billion. Feels weirdly low.

Low.

Just the US business?

I don't know.

I felt like— I mean, we've talked to Sean Franken. It sounds like they were losing a lot of money on TikTok Shop. And some of the— I mean, the growth had probably plateaued with—

Well, they were just losing money in general.

Yeah, but you got to imagine that that monetizes pretty well. It's such a big network, so much time on site. But yeah, it's trading like a Snapchat or a Pinterest, basically, like a non-Meta property, a non-Google YouTube property. Oh well.

Yeah, it's crazy.

Well, keep checking in on it.

So TikTok, Valued less than perplexity.

Okay.

Honestly, now I can see why Arvin was trying to get a bid in.

True. Yeah.

Everyone's like, "I've already got so much money." But yeah, I'm already at merger vehicles basically. Brett Adcock should have should have taken a crack. He could he could absorb easily absorb his market cap easily absorbed TikTok. Anderall could have.

True.

There's a lot of companies that could have. absorbed.

Uh, I'd like to see Palantir pick them up, run, run the database.

Yeah, $14 billion is crazy.

That is crazy low.

Crazy low. Uh, that's gonna—

I, I'm surprised, uh, the CCP said yes to this.

Yeah, what's the, what's the catch? I mean, this feels like it's worth like—

it feels like there's other chips on the table that are being traded around, whether it's like rare earths or, you know, some sort of of deal on tariffs, or this is such a multi-pronged deal between different countries and stakeholders. The UN thing just happened. And so there's a whole bunch of different pieces. And the art of the deal is not a single iterated price, a single asset in the moment type thing. Could Truth Social have picked it up? Truth Social's right around there too.

No, they've—

They've fallen. They're in the single digits, still a unicorn, right? Right? Our president is still a unicorn tech founder, I believe.

Yeah, $4.7 billion.

There you go. He's a unicorn founder, folks.

When's earnings? On November 7th. So tune in. We gotta lock it down.

I don't think DJT gives a speech. He should.

That would be crazy.

What you got, Tyler?

Do we know? I think there are rumors that the TikTok algorithm would still be governed by the Chinese, but then it would run on like American hardware or something?

We're going to do the inference.

Yeah.

We're going to do the inference. They'll do the training. We'll do the inference.

It's like the algorithm with Chinese characteristics.

Yes. But I am somewhat bullish on the idea of doing some post-training on top of their algorithm or doing some secondary post-inference. You see the ranking of the feed. You're doing analytics on that and seeing, OK, there's still some propaganda stuff here.

Yeah.

But is that confirmed that that is Is that what's going to happen, though?

I don't know. I don't know if that was confirmed. But it's not a total black pill if the training still happens there because you can do so much post-training, I think, and monitoring. And even just having the data, you can start running analytics and understanding what type of content is showing in what context. Are there actual— are certain trends or keywords, or is there certain censorship of certain ideas or promotion of certain ideas. At least if you have access to the data of what's being served, what TikTok users are watching, you can then understand, and then you can apply pressure or, you know, post-training or anything else. Anyway, in other news, the Lucy nicotine vending machine has arrived at Palantir. This was a journey. I worked with Eliano on this, and John says Palantir has a climate-controlled Lucy vending machine for their employees in their and you're bearish. It was a lot of fun. So we'll be sending out a few more of those, hopefully getting one in the TBP and Ultradome soon. The last thing was the Financial Times has an article about America's biggest corporations keep talking about AI but struggle to explain the upsides. We'll dig into this another day, but they analyzed hundreds of filings that suggest S&P 500 constituents are clearer about the technology's risks than its benefits. And another one of these data points of Fortune 500 companies being reticent to adopt AI. I'm not sure if it's just blanket bearish for Fortune 500 companies. If you're a big enterprise and you're not able to figure out a way to deliver positive ROI when a magical new technology comes along, it might not be a good sign for you. Jordy, what else do you want to talk about today?

Yeah, I was just confused. I was trying I'm going to dig in and see. I didn't see a16z in the announcement anywhere.

Oh, yeah, they were supposed to be, right? Maybe they pulled out or something.

I don't know.

I like the idea of the Ben and Marc Show just being hardcoded into the algorithm. They got Rick Rubin on there. Show me some Rick Rubin TikToks, some Torenberg, little Chris Dixon in there, some KB. This would be a good TikTok algorithm.

They still could be in it, but we'll see. Well, we can ask.

It's I hope the thumb is on the scale regardless of their ownership position.

Putting thumbs on scales.

I'm a big fan of thumbs on scales.

Anyways, without further ado, we, I think, got to get on with Taipei.

I think we do. I think we do.

We will see you folks tomorrow. Thank you for tuning in.

Thank you for tuning in.

Leave us a review on your favorite podcast app if you're listening there. And we love you.

We will talk to you soon.

See you tomorrow.

Goodbye.