Blitzhiring and the Complicated Case of Windsurf, OpenAI and Google | E2151

25:59And in fact, there are startups now that are putting on glasses on workers and you can watch them work. and that's becoming the training data to take that person's job. I've seen what you're talking about and it's gotta be brutal to sit there and realize that not only are you not getting paid much to do hand assembly work that's going to be automated, but you won't be paid anything for it in the near future. All right, anyways, Devin's buying Windsurf. Devin? Wait. Yes. So Cognition AI, which is the startup behind Devin, announced right before we went to air that they're going to buy what's left of Windsurf that Google is not buying for$2.4 billion.

26:40Got it. Okay, so there's a breaking news story while we're taping. Devin is buying the remnants, but we don't know what the remnants of that carcass are. Essentially, OpenAI wanted to buy all of Windsurf. That didn't work out because Microsoft was a sticking point as they sort out some deals with IP between OpenAI and Microsoft. The period of exclusivity ended. Google swooped in with$2.4 billion, bought out most of the investors, got the founders and researchers they wanted, left behind the company. Everyone thought it was going to be a carcass. Then Cognition, who's behind Devin, showed up and bought everything that was left.

27:16Which is the ARR. And Devin is what? What does that company do? I'm curious. So Devin is the much vaunted AI engineer that you could hire to work with your company. It was essentially, going back in time, it was the early agentic developer, if you know. Got it. So the Cursor product is loved by a large number of people, right? And it's making 100 million. 500 million. Cursor's now at 500 million AR. Okay. Yeah. It's mind boggling how fast that was going. So who gets that? Who gets that? Well, so no, wait, Cursor or Windsurf, Jason? Okay. Cursor's a separate company. Windsurf is this company. Windsurf's product competes with Cursor.

27:56What is Windsurf's product called? Is it called Windsurf? It's called Windsurf. Got it. Okay. So Curse is the leader in the space. They're making 500 million. Winsurf was at 100. I've now got it in my mind. I was juxtaposing the two. So whatever's left of the Winsurf product is going to go to this Devon company. And then the talent goes to Google. But will they have a competitive product, I'm wondering. I mean, I don't think - Because Google Gemini does code. There's actually a code project. Yes. What's the name of their code project? Google has a cursor competitor, a Windsurf competitor already.

28:34So Google has a product called CLI that brings its Gemini AI models for coding into the terminal. There's a similar product from, I believe, Anthropic, which is Cloud Code, and then OpenAI's operators around the same thing. There's also Gemini Code Assist. Yeah, that's the direct competitor, I think. Do you want to make this even more hilarious? Please. So today, just because why not? Today, Amazon released a new thing called Hero, and it is a new agentic IDE that works alongside of you from, quote, prototype to production, made by some folks at the AWS team. They're kind of keeping it on a separate brand.

29:15You can pay for it, but now there's another one. So everyone's trying to solve this particular nut and they're spending a lot of money. I'm just glad that we had Varun on the show a couple months back, link to the show notes. And now he's a billionaire. So shout out Varun. Well done, man. Great. All right. So there's going to be 20 different ways for you to have a code assistant help you. It's all going to be free or close to free. And these products were greatly under, I mean, there's tons of free products and even the paid ones, which I guess GitHub's co-pilot was the original, it's like$100 a year or$100 a month.

29:50These products all seem phenomenally cheap for what they are. We're all familiar with AWS, Amazon Web Services. That's the cloud platform that powers so many of your favorite brands. But do you know about AWS Activate? That's their program for startups, where they provide up to$100 ,000 in AWS credits for all startups. whether you're backed by an investor or you're bootstrapping, money is time to keep innovating, time to delight customers, and time for you to keep gaining traction. You need runway, and AWS's Activate is going to help you with that runway. We hear this story from so many of our founding university companies.

30:29They're finding product market fit, the words getting out about their product, just a little boost, finding some savings here or there to get a little extra time can be the difference between getting traction and bringing in revenue, or hey, let's call it what it is, running out of money, okay, and shutting down. AWS knows this. That's why they've created the ultimate toolkit for early stage startups looking to boost growth. With AWS Activate, you're going to get up to$100 ,000 in AWS credits, hands-on support and training, plus exclusive discounts with some of our favorite companies and tools.

31:00So start getting the support you need at every stage of your startup journey. To learn more, visit aws.amazon.com slash startups slash credits. That's right. AWS.amazon.com slash startups slash credits. Now, the question that we often hear is, what are the margins like at a company like this? And you were mentioning earlier, Jason, the kind of last 10 % problem, last 20 % problem. That dovetails neatly with the company that I was talking to for the show today called Vellum. They're a YC winter 23 batch company. They raised a$20 million Series A. Why do we care? It's another company in the AI space raising around.

31:36Well, what I liked about them was that they were trying to work on essentially what you're describing, that the gap from prototype to production in AI is still huge. And these founders had run into this problem at their last company. They were at Dover for a while. And so they made a company to do this. And I think it's a pretty cool idea. I have a lot of notes about how they're going about helping companies kind of solve the AI pipeline. But what I care about a lot is on the pricing point, I asked them, do you have SaaS gross margins or do you have like 50 % gross margins? And this company says that they have SaaS gross margins as they're growing three to four X every year.

32:11So maybe even with relatively inexpensive pricing, a lot of AI companies will make economic sense long-term. Maybe they're not as gross margin negative as we anticipated. Yeah. So when we see the price we talked about in the last episode, the price of like tokens keeps plummeting as we have this crazy competition, then if you build a product that uses Claude, Grok, pick your open AI as your backend, and these things keep, the tokens keep collapsing, plus you have DeepSeek, plus everything else on HuggingFace available to you, yeah, then the margin might actually be growing as the cost goes down.

32:50How many of these backdoor acquisitions have there been? And I just wanted to note for the audience, I think you did a little bit of research on it. Is this the sixth or seventh one? Yeah, yeah, yeah. Go through it. And maybe we could show it on the screen. One, two, three, four, five. Five. Okay. Yeah, I'll just pull this up for everyone so they can enjoy an excellent view we have here in Notion. And you can see the docket, folks. This week in startups.com slash docket, you can see the notes we use here for the show. Yeah. So just pulling them up here. So the first one that kicked this all off was the Microsoft Inflection AI deal.

33:26That was back in March of 24,$650 million in cash. It was an IP deal. That's a common refrain in these acquisitions. Took the founders, most of the talent, and then the company pivoted to the enterprise. Go forward a couple of months. Amazon, Adept AI, June 24. We know it's in the nine figures, Jason, but no one got closer than that. Again, a licensing deal. Took the founders and some of the research team. Then number three, Google, Character AI, August 24. two and a half billion. That's getting expensive. Non-exclusive IP. So the same idea of buying intellectual property access, founders, and some researchers.

34:01Then this year we saw Meta and Scale AI. This was kind of the weird one. 14.8 billion, but they bought a stake in the company and took the talent. But then the company was poisoned because no one wants to work with Meta's subsidiary effectively on their data. So that's going to be a mess. And then most recently, Google and Windsurf, July of this year, 2.4 billion for the license. co-founders and some talent. And then as we saw later on, Cognition slash Devin is buying whatever's left. So this has now happened five times. You know, trend starts at three. This has all happened in the last year, couple of months.

34:35I'm curious to talk to an attorney. If anybody knows an attorney who structured one of these deals, we'd love to talk to them in the background. He's Alex at launch.co. I'm Jason at launch.co. Alex W at launch.co. Alex W at launch.co. So I'm, yeah. You're not JCC. You're just JCC. You were here first. I was, yeah. There's no other Alex here. All right, let's keep going. What else is in the docket? So I was particularly excited to see that we are seeing Tesla expand their geofence in Austin. Well, okay. No, I saw the image go by like everybody else. Yeah. All right. So we're going to try to do this with the minimum of jokes possible.

35:17Friends, we've been covering the expansion of robotaxis around the world from Chinese companies to Waymo to Zoox to everyone else. Of course, Tesla is now in the mix, Jason. They launched in Austin. Seemed to go pretty well. There have not been headlines of fires and explosions and crashes. There was one, a tire touched a car a little bit. I would say there were three notable instances. The tire touching, the going into the wrong lane. Oh, that's right. And then there was the phantom stop. There were three. And all of those at low speed, all of those, the car was basically in control. All of those, there were safety monitors there.

35:57So even in the worst case scenario, I think in any of those instances, it would have been a fender bender and no harm to human life. So will society be okay with this? Well, there were, remember there was a Waymo that went down an alley? So there was the Waymo that bumped into a pole in an alley. Now, it probably shouldn't have been in the alley. Alleys are probably edge cases. And it hit a pole. But it hit it at such low speed that it looks like it hit it at, I would say, that dent is a 10 to 15 mile per hour dent. Yep. And, you know, if there had been a, you know, God forbid, a person between the pole at the time, it would have stopped, I'm sure.

36:46But how did a Waymo hit a pole, you have to ask? Now, when did that happen? I think that was like two years ago. So I think you're going to see that type of mistake. Zooks had one as well. Zooks had somebody pull out. It was going extremely low speed, and it was in the right-hand lane. somebody pulled out into that right-hand lane, and it kind of moved to avoid it or slammed on the brakes or something. I think there might've been a touchy-touchy, tappy-tappy. But these little ticky-tacky, tappy-tappies, if they occur once every couple of thousand rides and nobody's hurt, I think society will accept it.

37:26But will regulators accept it becomes the issue. And will regulators accept it if you don't have LiDAR. Not that LiDAR would have solved any of the problems, but if they feel there's like a higher level of, you know, technology that could be there, whether they're right or wrong, I think regulators are going to, if these ticky-tackies happen too consistently, regulators are going to step in. This is why I'll keep saying it. There's no shame in the safety driver game. Put a safety driver in there, folks. Volkswagen's doing safety drivers. Zooks, Tesla or safety monitor in Tesla's case, I would have put them in the left-hand side.

38:06That's just me. I can disagree with Elon on stuff. He's obviously a little closer to the issue than me. I just think for optics, I would have put it in the left-hand side. But whatever it is. But here it's an expansion. It goes up, it looks like, to the northern part of the city. It does. And it hits the, I think, University of Texas, which is kind of a big deal. And this means Lon could go from the University of Texas or this could be our office could be in this zone and it could go to south of Congress. I don't know, Lon, if this could actually do that. Less people think that this expansion to the tip of the northern part of Austin.

38:46I thought we were going to pull it off. All right. Yeah. I mean, it looks like it's expanded like 5X. I don't know if that's average, but it looks like about five. It looks like, you know, when you look at the map, it looks like it got an extra three inches of expansion there. I don't know if that's average or not. All jokes aside. All jokes aside, this is the Tesla logo. Okay? You've all been trolled. Unless you're seeing something here that's not here, it's just the upside-down Tesla logo. That was my joke. I can't move it. What I have now up on the screen is the Waymo service area in Austin with the Tesla service area imposed on top of it.

39:24And according to all the data we could find, Jason, this is roughly a doubling of the Tesla geofence. And also it now has people are calculating about 40 square miles versus Waymo's 37 square miles. So it may be now that the robotaxi effort from Tesla is now larger in footprint geographically than Waymo. So, and I just, I mean, I try to never have you going up against your friends because it's not what the show is for, but you were pretty serious about like, let's go slow. Let's be moderate. Let's be safe. And this is faster than even my bullishness expected. So too fast or just better data on the operational side?

40:01You know, there's a great self-correcting mechanism here. if you go too fast and you get a tappy-tappy, touchy-touchy, regulators are going to be lobbied by haters, by concerned citizens, by people in government who might think slow is better. And in technology, you have had this position for a long time of like, hey, net-net, it's going to save lives because distracted drivers. I 100 % think statistically you're right. Then I look at regulators and knowing what I know about regulators after that cruise mishap, which really wasn't their fault. Somebody got hit by a car. The person ricocheted, as crazy as that is, into a slow-moving cruise.

40:47And the reason cruise got in big trouble in San Francisco is because they dragged the person a little bit. And then? And then edited the video and didn't show it. It's not the crime, it's the cover-up. The cover-up always worse than the crime. But even still, life happens. People will get hit by cars. People will jump in front of cars, right? People put cones on Waymos. So the self-correcting mechanism is if Tesla goes too fast, if Volkswagen, if Zooks, anybody goes too fast, then a regulator might say, hey, you're going too fast, and then correct them. There's a lot of edge cases. I've been having perfect rides in my hardware for Juniper model of the Model Y.

41:29I will say, like, you get into a construction zone on 290, which is a highway that goes east-west from, you know, hill country all the way to Bastrop and past the airport, the Austin airport. You know, my Tesla cannot figure out what they're doing with that construction zone because they are going for it. So construction zones, mislabeled signs, tricky stuff. This is the stuff, the edge case is where it really matters because I've had one or two instances, even with my hardware four, where it could have done serious property damage, serious harm to me, like life harm at 45 or 60 miles an hour.

42:10It's a lot of speed. And that's why I think staying off the highways, keeping these cars to 35 miles per hour, even if the speed limit's 40, I think they should just throttle them to 30. You know, if I was running this, minus five miles per hour, whatever the speed limit is, which I think is legal to do. And then, you know, unless it's like 25 miles an hour, just keep it slow, steady, wins the race. Because what's going to happen is somebody is going to feel like they're losing the competition. And I don't think Tesla feels this pressure. Waymo doesn't feel this pressure. Volkswagen's not going to feel this pressure.

42:46But there's going to be somebody who feels the pressure that they're falling behind. And somebody's area is faster than theirs. And it kind of feels like the Waymo, Uber, Volkswagen, Zooks, Tesla competition is now heating up. This is where regulators can do their job and say, here's the rules of the road, literally. We want you all. Well, I think the regulators should pick the zones that are. I don't know why Tesla and Waymo get to pick their area. I think the regulators should approve that. And I think local regulators have the right to do it. So I don't think this should be a federal thing or a national thing.

43:22I think local regulators should set their rules. They should have their say in this. And everybody should go nice and slow and steady. And they should go nice and slow and steady and add Providence, Rhode Island, as the next stop in their expansion. I'm going to keep saying that until it happens or until you tell me to stop. Oh, you know what the other big note here is? Very disruptive. I saw some influencers who are in the thing here because it's not open to the public yet. It's open to influencers and their friends or whatever. Oh, RoboTaxis in Austin. RoboTaxis in Austin are open. It's not open like Waymo is.

43:54It's invite only. It's invite only still. And they let their influencers in because I think they feel like they would get a bad, you know, that people have it out for them, which I get. You know, it's a politicized moment or whatever. it's still 420. I can tell you that ride from the northern tip to the base, from tip to base, that's not a 420 ride. That's like a$15 ride. So actually, if you happen to have that commute area and you happen to have this, it's kind of like the early days of Lyft, Sidecar, and Uber. You're going to be subsidized for your ride. That's a really cheap ride. Well, producer Lon just weighed in.

44:34He thinks the Omni, where a capital factory is, is inside the Tesla peen. And as is his neighborhood. So he is going to take a Tesla commute. We're going to get him to apply. And then we're going to make him do an on-screen reporting about how it goes. So producer Lon goes self-driving. He's L-O-N-S on x.com. Somebody invite him. Or his email is lon at launch.com. So if you're in it, please email Lon and invite him to go take some rides. Yeah, we want to see how it is from Lon's—I'm not going to—what is Lon? He's not the team Luddite at all, but he's like slightly more skeptical than we are. No, he's pro-tech.

45:14He's pro-tech. Yeah, he's just a little sprinkling of—no, he's just—actually, I'd say he's probably like slightly more liberal than like a Clinton Democrat, but not like an Elizabeth Warren, but he's probably got some Bernie Sanders to him a little bit. He cares about the working man, probably very Trump-esque in that way. I'll let him speak for himself. I think he believes in universal healthcare. He was a Bernie bro in 2016, but it's eight years later. I don't know if he's still a Bernie bro. Joe Rogan was a Bernie bro. All right, let's keep going. Let's move on. Okay. I've been keeping tags on open source AI because we talked about this a lot a year ago, back when Meta's LLAMA project was doing quite well.

45:55And then Meta has pretty publicly stumbled with its LLAMA 4 release. And everyone kind of went back to thinking that we're going to see essentially closed source AI run the world as XAI, Anthropic, and OpenAI kind of battle it out. Well, turns out the answer is maybe not. And there's a company in China called Moonshot AI, and they have been referred to as an AI tiger by Chinese media, essentially putting them in the deep seek kind of collection of the Chinese AI industry. they're the they're part of the mag 7 of chinese ai if you will and they dropped a couple of new models that are called the kimmy k2 models and here's the thing that's killer they're legitimately open source weights and everything and they're very very good and i just did not expect this to happen because this is a company that raised a billion dollars last year and most recently raised another 300 million with tencent participating at a three billion dollar pre-money valuation and they're dropping arse-kicking open source models.

46:53Here's some of the data, Jason, just showing off how they do. Now, keep in mind, this dropped, I believe, before anyone could get their hands on Grok 4. So there's no XAI comparisons here, but this is against DeepSeek, Quinn, which is Alibaba, OpenAI, and Anthropic. And as you can see, these models are competitive with kind of current state of the art and they're now absolutely free. Now, I don't think American companies are going to adopt open source Chinese AI models or closed source Chinese AI models, just given tensions between the two nations. But from the rest of the world's perspective, this is an enormous gift.

47:28And I find it to be honestly pretty encouraging and optimistic that we're still seeing this level of improvement. And I just can't believe what a gift it is to have open source information out there of this magnitude, Jason. I'm blown away. Well, this was what OpenAI was supposed to be doing before they closed it down and made it a for-profit company. It was supposed to be the original premise that they sold the world on with their nonprofit status was, we're going to make this free as a gift to humanity. It's too powerful for any one person to have. Therefore, if everybody has it, that's the way to go.

48:05Something changed along that way, and people changed their mind. And now we have this great paradox that China and that country is saying, you know what? We might be six months behind, 12 months behind, whatever the test might be, three months behind. Who cares? You know, what? Having an open source option that's six months behind the state of the art or 12 months behind the state of the art or maybe sometimes dips ahead of the state of the art. Yeah, it's got the great price. So for, we'll be sitting here in four or five years and it will feel very much like MySQL, right? There was a period of time where people said like, I'll never use WordPress or MySQL.

48:48How would I ever use something open source in a corporate environment? And then the White House is running on WordPress, open source, you know? Or large companies, you know, Uber, Airbnb are probably using MySQL or, you know, some version of, you know, those. So it's going to be really amazing to see. This might relate to the story we saw with Windsurf. We were trying to figure out, like, why do they not care about the product and they care about the people? Maybe everything being built right now isn't as important as what's to come. And having those people iterating at a high velocity in your organization is more important.

49:32If you don't know exactly where this innovation ends, then what's been accomplished to now is going to feel like the web pages of the 90s. So you got to get the talent, have them locked in for five, 10 years in your company. But, you know, four or five years is realistic. And if they're locked in for the next five years, you increase your chances at the$10 trillion prize. Which is – I'm writing that down because we keep going back to that. So I think that's going to be one of our trends for the back half of the year. Okay. Looping back though to AI. I want to make a point really quick because there's something interesting going on that I don't think we've actually talked about enough.

50:12So we talked about China, talked about the United States, and that's kind of where the conversation stops when it comes to foundation AI models. There's not really any other major nations out there except for Jason. Have you heard of France? because it turns out Mistral is the only other company on the open router leaderboard by token percentage of just users from that service. But Mistral, the French kind of like AI champion, is ranked sixth in the world for token usage on open router. Shockingly, look at them. They're actually pulling it off and Europe might have an actual stake in the AI future.

50:49We joked, we joked, but they're still here doing it. Yeah, I mean, it does feel like with these open source projects and talent moving around, the papers being written, we have some parity between these platforms. There's a funny meme of a circle of who's in the lead. I got it. You know, I don't know if we showed it here before or we were, did we show it last episode or we showed it in our group chat? I can't recall, we do a lot of shows, guys. Yeah, I think it was like, you are here. And it's like, it had DeepSeek, it had Grok, it had Gemini, it had Claude, it had OpenAI, like a circle. And it's like, takes the lead.

51:25You are here. And then here he goes. So yeah, Grok introduced the world's most powerful model. Next, OpenAI introduces the world's most powerful C. So you basically need to do this once a year. And then Polymarket has betting on this. And you know what? I wonder if Polymarket resolved Grok winning. Because remember, we had brought up the poly market. It was the AI. It was the leading AI model by a certain date. And it was, I wonder what the result, what it was using to resolve the bet, right? Cause all of these markets have to be resolved. And that's where this is. This is very interesting. Okay.

52:05So I found it. Juicy, juicy, juicy. Pull it up. This is the question is which AI, sorry, which company will have the best AI model at the end of July. And we talked about it probably somewhere in here. Google was storming. And then you'll note that with the release of Grok 4, XAI seems to have made a dent. And then they've kind of lost it again, which is interesting because the market is now re-betting actually more than ever that Google is going to have the best AI model by the end of July. So either Grok 4's benchmark. Maybe they pee too early because look, somewhere in the middle of July, if you hover your mouse over when the gap closed.

52:44When was that? When was the gap? That was July 8th. Right. So last week, between July 8th or whatever, the gap closed and X almost flipped it. But remember, this market is closing on July 31st. So there's no way for XAI to drop another model. No, they've done it. They dropped the numbers. Yeah. So I don't think folks are doing this based on the poly market when they're dropping it. But if we scroll down here, you'll see DeepSeek and Meta have no chance. What is the rule? Less than one, less than one. The market will resolve to the company which owns the model, which has the highest arena score based on the chat about Arena LLM leaderboard.

53:27When the table under the leaderboard tab is checked on July 31st. Okay, so this is not using the same thing that we were talking about Grok last week. We're using something else for Grok, But yeah, so this will be the arena. There's a variety of ways to look at things. And you'll note here that actually, according to LM Arena right now, Grok 4 is not really showing up yet. Number three. I don't know if it's actually been - It's number three, actually. Oh no, that GPT-4? Where is Grok? That's Grok 3. Grok 4 isn't in here yet. So Grok is three. I don't know if - No, but that's Grok 3, not Grok 4.

54:06So Grok 4 isn't on the leaderboard yet. Yeah, not yet. Okay. But that does not mean that it won't show up there. It means that right now. But all this is fun and games just to show everybody that if you were ever curious about what it's like to live during a time of great technological change, well, here we are. This is what it looks like. It's messy. It's aggressive. It's expensive. It's exciting. And at the end of the day, we get lots of cool new tools. So I just love technology. Hey, one thing, and we should end on this, I think, because we've got a lot of show here. I had asked you to look into the depreciation of servers.

54:39Did we do that? I have a lot of good data on that for you. Great. Let's start this discussion because I was told by some folks in the know three years, and then these things are going to be obsolete. You said some people had expanded it to four. This is critical because putting in 100 ,000, 200 ,000 servers at Colossus, Grox, or CoreWeave is a company that 5x their valuation or something when it went public at the peak. Their valuation is based on all this. you have open AI, incredible revenue numbers, but what is it costing them? Because they're putting a lot of servers to work. So what is the depreciation schedule?

55:22My challenge to you is to figure out the depreciation schedule that people were using for like H100s or these latest NVIDIA servers. What did you come up with? What did you find? All right, so I looked at this from two different perspectives. The first one is what are the hyperscalers doing? And when they discuss servers, They don't always break down strictly AI versus strictly not. There's some commentary in earnings reports, but the big numbers are as follows. Alphabet raised - And by the way, you can show this on the screen, I think, because it's in the docket. This week in startups.com slash docket if you want to follow along with us.

55:53And you can find today's date, July 14th. We keep all the dockets up there. So you have tons of content you can look at and you can build on this and put comments into our notion if you want to. Yeah. I'd always turn on comments so you can always drop in, leave notes. I love when people do that. Okay. So Alphabet increased its sural lifespan most recently to a six-year depreciation timeline. That was back in 23. Meta moved it to five and a half years this year. Microsoft moved from four to six years back in 2022. And then this is when it gets interesting. Amazon moved the other direction. They moved from six years depreciation to five, shortening the timeframe and therefore just increasing the effective cost of their data center investments, their CapEx, if you will, Jason.

56:33Now, the Financial Times did an analysis of this back in 2024, looking at the aggregate impact of those changes. So not an entirely up-to-date number, but they calculated that the hyperscalers alone are going to save about$10 billion by increasing the length of time they think they can use their servers. So what does this kind of look like in practice? Well, I went to the CoreWeave latest earnings report, their Q1 numbers. And as you can see in this little data set here, the depreciation and amortization cost at CoreWeave, which is a NeoCloud, rose from basically 80 million in the first quarter of 24, all the way up to 444 million in the first quarter of this year.

57:14So when you spend a lot of money, your costs go up. Now, should we worry about this? Is this a concern for the profitability of some companies? The hyperscalers, no. So I went ahead and looked through some Q1 data from both Meta and Google. And if you look at their CapEx, which is servers and so forth, they're spending$14 and$17 billion apiece. And their depreciation and amortization costs in the same quarter were basically$4 and$4.5 billion. So it's a lot of money, and those costs will go up as they invest more money, Jason. But nothing that I saw struck me as worrisome, I guess, if that makes sense.

57:51Yeah. I mean, I think the bigger issue, putting aside worrying, is how much revenue are these generating versus what they're costing? So when we looked at those first set of servers, those are Amazon Cloud, CPUs. They're not the GPUs in all likelihood. The meta Facebook ones, that's to serve up your Facebook page. That's not an NVIDIA H100 or whatever it is doing that. So I guess the question is CoreWeave might be the closest number. So how much money is CoreWeave making on those servers is the next piece of the puzzle we have to sort of tease out. Forget about worrying about it. Just does it pencil out?

58:30Does the 100 ,000 servers in Colossus, if they last for four years, and 100 ,000 servers times$100 ,000, whatever that equals in billions of dollars in servers, what does that generate in revenue? And then the next piece is what happens to those in year five or six? Do they get turned off? Do they get thrown out? Those are the two remaining questions I have. So if anybody has feedback on that, AlexW at launch.co or Alex on Twitter, you can just reply to us at Jason at Lons. Yeah, that's the piece I would like to know. What happens in year five or six of these? Do they go into like, do they get resold?

59:15Do they have some afterlife doing some other purpose, like indexing images or something for, you know, tagging of photos? Or do we just not know the value of those? We put them out to pasture slowly to do the last bit of data crunching. And Kabir did find for us, NLS Kabir, that NVIDIA has a recycling program, which is pretty cool. So I think if something gets so old that you're just going to throw it away, you can send it back to the mothership and they can do something with it. But we have seen the past cases. I wonder if that's even real. That could just be window dressing. Like they take them back, but who knows what they do with them.

59:46Do they actually take the time? Is the cost of taking them apart and melting it down and doing something with it actually, does that actually cost them money? or do they break even on that? Apple will take back your products as well. So I wonder if those are just greenwashing programs. I always suspected the Apple one was a greenwashing program for the oldest machines. I know they can resell an iPhone that's three years old or four years old in an emerging market, but is there a market for an H100 that's five years old is the question. Well, if you have a five-year-old mega NVIDIA GPU, I want to put it on my wall.

1:00:25So yeah, just shoot me an email because that'd be a fun, fun walk. I'll have all that for you, Jason, on Wednesday when we're back. Yeah, okay. You can read the docket, thisweekinstartups.com slash docket. He is at Alex, I'm at Jason, executive producer Lon at LONS, and we will see you all next time on This Week in Startups. Bye-bye. Bye-bye.