SemiAnalysis breaks down Anthropic's S-1: best capital efficiency in capitalism history, but Fortune 500 AI adoption still surprisingly low

Speaker 1: says you can't steal this... Steal man.

Speaker 2: That's John. I can steal man anything. Anyway, let's Let's bring in our seminar... Joey. Describe yourself in three words without using AI challenge. It was impossible from Marissa Mayer. But how are you guys How

Speaker 1: would you describe Doug in in three bullet points? Without using AI. You can't use AI.

Speaker 9: Without using AI. And feel free to pass because

Speaker 2: we know this is an extremely difficult challenge.

Speaker 9: Doug is a trendsetter and trend caller. I would say that. And he's probably... He's always in the right place at

Speaker 2: the right time. Oh. As

Speaker 9: evidenced by by talking

Speaker 2: to you today. That's good. About the Anthropic IPO. Yeah. Those are those are those take

Speaker 1: me I on like those. Bullets or I don't know if you wanted to

Speaker 9: Yeah. Pick those

Speaker 8: He's a he's

Speaker 9: a handsome guy. I'll I'll give him that too.

Speaker 1: Handsome. There you go.

Speaker 2: There you go. With a nice

Speaker 1: with a fresh cut. Looking good. Looking good.

Speaker 2: Okay. Looking good. How long have you had to digest the Anthropic IPO? I guess the big question is like a lot of these facts were leaked because of the Investor Day. What really stands out to you as like new learnings from the S1?

Speaker 9: Yeah. I would say most of the information was leaked somewhat before and I think we got some more detail on like per user spend, you know, how big some of the largest customers were. Yeah. Pretty specific details there.

Speaker 2: Yeah.

Speaker 9: That were, you know, pretty interesting. So happy to dive But into

Speaker 2: overall, is this a beat? Because it feels like there was like a question. We can get into like the overall health of the AI trade but in general there's like a oh, like, oh, there might be something wrong and it seems like everything was on track. Everything sort of mapped to what was leaked earlier. Things look good across the board. I think

Speaker 1: it's like a vibe. It's a vibe speed. That's what I would say. I think people were expecting that there was gonna be like one thing that everyone focused like like the vibes had shifted

Speaker 2: The market.

Speaker 1: Such a negative way like maybe two or three weeks ago where people are oh yeah, you really want to do a $2,000,000,000,000 IPO While rates are up or Yeah. When you're not, when you're decelerating, like this makes no sense. And now people look at this and they're like, okay, this is like pretty

Speaker 2: Incredible business.

Speaker 9: Yeah. This was one of probably the best scaling to a trillion dollar plus valuation in the history of capitalism. Yeah. To burn 10,000,000,000 to get to this valuation is insane. To be this level of profitable this early and at this scale

Speaker 2: Yeah.

Speaker 9: It is, you know, amazing.

Speaker 2: Okay. So Doug, incredible capital efficiency. Is this a violation of the AI is similar to the railroad build out concept?

Speaker 7: Yeah. I think so. I think it is actually an incredible narrative violation for that. Yeah. I think... I mean, it isn't... Let's put it this way. It's profitable. Right? I think the capital efficiency is mostly because they made a lot of money too. Sure. That, I think is the slight narrative violation. But, you know, that's like looking backwards. I definitely think looking forwards, you see several $100,000,000,000 compute and I would say that that's still really in line with the railroad analogy.

Speaker 2: So Sure.

Speaker 7: I think the the railroads are still... There's still railroads, baby.

Speaker 1: But No. But but it's like it's like you have like a company that is spending on railroad services. But it's Yes. But it's like this company is not... At least right now, doesn't look like that a manufacturer... Like, they're not laying the track. It's like...

Speaker 7: Yeah. They're not laying the track. And like... So so pretty much the OG, I guess, the steel companies were were... If we're gonna be using railroad analogies, the steel companies, you know, all got invented essentially to make railroad tracks. So I think that that's kind of the... Maybe that's the analogy we're looking for, but I don't know. I think it's really clean. There's no other way to put it. Like, this was an extremely clean beat. Like, you don't really have a much to nit nitpick, but I think maybe the concern is, like, a lot of people have been talking about, like, lowered expectations in q four. Mhmm. They didn't put the September ARR, right, because everyone knows that was like a total debacle. But they obviously have an an open spot for the October ARR and and they wouldn't have it open if they didn't think it was good.

Speaker 2: Yeah. That makes sense. What about the compute obligations? Like, it's sort of a big number, not really relative to the valuation even, but yes, like 600 something billion is a huge number. But it's over a long period of time. Like it felt very important to restate and reinforce that this doesn't it's not this year, it's not next year. It's you know, some of these contracts go into 2033. Some of them go into 2036. What is the extra context around the flexibility to move that... Those compute commitments around around and have them hit at the right time?

Speaker 7: Yeah. I I'm gonna have one... I'm gonna have go... Joey go next because, like, he's, you know, he's mister number. He's like mister Bucky at some analysis. Yeah. But I think more than anything else, they told you at the very beginning, we will... We want to under promise and over deliver. Like, they sent... Like like, it's very rare in s one. In, like, the last paragraph, it's like, no. No. No. We're gonna be really conservative. We promise we will underpromise and overdeliver. So, yeah, that's that's kind of my vibe is that I think the the $100,000,000,000 a year probably works. I mean, it's a little north of $100,000,000,000, I think, if you, like, naively, annually smooth it out. But Mhmm. I think it probably works. Joey, what do you think?

Speaker 9: Yeah. Think with some of the increased compute costs and then even probably averages closer to seven or eight years, it's not as many gigawatts as it looks, you know, from a 600,000,000,000 perspective.

Speaker 2: Yeah.

Speaker 9: And they, you know, they'll be adding a lot more of those compute commitments, I'm sure.

Speaker 2: Sure. Sure. But but after IPO, after quarterly earnings, after continued profitability, something that can underwrite it, maybe getting...

Speaker 1: Get me up to speed on how the original SpaceX compute deal actually works because if I remember correctly, it was like they were getting access to a lot of compute but able to sort of like forego payments against that for some amount of time, which I imagine was hitting in second half of the year, like the payments really coming online relative to how much compute they were accessing. Is that the right way to think about it?

Speaker 7: I don't recall that specific... Yeah.

Speaker 9: We we saw it a little bit in SpaceX earnings last week

Speaker 2: Yeah.

Speaker 9: Start to come through.

Speaker 2: Sure.

Speaker 9: So it was... I I think it was pretty normalized for for both companies. We saw it, you know, in both the compute margin numbers at Anthropic, Q3 versus Q2, and in SpaceX earnings.

Speaker 2: Yeah. About... Oh, sorry. Sorry.

Speaker 7: The odd part is just the ninety days. The fact that SpaceX can like essentially whip it back, You Yeah.

Speaker 2: Yeah. What what about the overall compute strategy and diversity across the neo clouds? Obviously, Anthropic has really big contracts with Broadcom and the hyperscalers. But for some of those neo clouds, there was a lot of rumbling with Elon about the, know, neo clouds over the weekend. Like, how does this s one change the calculus around just neo cloud economics? What a good Neo Cloud looks like. You obviously have ClusterMax to understand the capabilities of the Cloud, but it feels like the financing side is also increasingly as important as the capabilities.

Speaker 10: Joey?

Speaker 9: Yeah. I mean, you have a good backer and a good backstop, so whether that be a hyperscaler for NeoCloud or an AI lab, I think Anthropic has shown today AI inference is a really great business. Even running the training is extremely profitable. So compute margins are excellent. If them and chip company or hyperscaler are backstopping a Neo Cloud and they have operational excellence in how they operate those new clusters, it's, you know, Clouds are still very good businesses that have increasing monetization tailwinds behind them.

Speaker 2: Are you starting to see this, like the long rumored neo lab to neo cloud pivot happen yet? Either of you, I guess. Doug?

Speaker 7: On on the margin, I think so. Yeah. I think people are... People really wanna be selling tokens. Uh-huh. If you look on the the fact that DeepSeek v 4.1 is probably the reason why all these types security hacks are happening Mhmm. It's pretty clear that if you have the gigawatt and you can you can place it with an open source model, you know, I I wouldn't call that, like, a Neo Cloud, but, like, tokens as a service, I think, is going to increasingly happen, especially as the volume of open source models continues. Also, I wanted to adjust one thing. Please. I feel like Elon is shitting on the Neo Clouds because he's also a Neo Cloud. Yeah. He's just like, I'm I'm I'm the biggest Neo Cloud. Those guys just aren't as good as me. That's really what it

Speaker 2: comes out. No. I mean that that's... No. Said that exactly. He was like, yeah, if you don't have other revenue.

Speaker 1: Yeah. The the the points that he's making are self serving but it also makes sense that a lender would rather lend to somebody that like has revenue that's diversified. So it's it's like you he can be like trying to further his own interest while at the same time like being like having like a pretty coherent like thought

Speaker 2: that Both

Speaker 1: can be true.

Speaker 2: Yeah. Yeah. Yeah. What what what else can you tell us about the nature of all the cybersecurity threats? Like, how much is it actual inference of deep seek, as you mentioned, versus just piggybacking on cyber, you know, different exploits that have been discovered by other folks? Like, how how relevant is compute in the cybersecurity battles that we're seeing playing out right now? Do you have a do you have a grapple on that yet, Doug?

Speaker 7: Mean, I think I think compute's pretty relevant here because it seems like a bozo model can probably hack your your your infrastructure. It doesn't really take much. I think the bar is surprisingly low, and I think it's in all of our interests, like, not to talk about how low the bar is, if I'm gonna be honest with you. Okay.

Speaker 1: And I

Speaker 7: think the the other... The... Like, it's it's, like, it's, like, surprisingly low. The other thing I think that's, like, also really kind of messed up is how asymmetric it is. It's kind of like a drone. You know? Like, it costs, you know, a $100 to fly the drone into the tanker, and it costs, like, you know, like, $2,000,000 to shoot the drone out of the sky or something. Yeah. That that dynamic, I think, is very replicated in AI as well. Meaning defense is... Costs, like, probably an order of magnitude more than a... Than than an attack. Mhmm. We've done some work on the attack side. The attack side, like, the raw tokens that come out of it, it's like it's like pitiful. It's like a $100 or something. It doesn't take a lot.

Speaker 2: Yeah. Yeah. That makes sense. How...

Speaker 1: Yeah. It's such an interesting problem because we were talking earlier on the show of like, does there need to be like large scale white hat bounties, you know, billions of dollars allocated to like incentivize good people to go after and find these vulnerabilities, but at the same time, imagine introducing the chaos of like... I was using the example of like a Miami course bro being like, anyone can become a millionaire finding like vulnerabilities. Right? And they're like, here's how you take the obliterated obliterated model, point it at, you know, this legacy, you know, you know, Fortune 500, and all of a sudden, it's just like total anarchy.

Speaker 7: Also, obliterated models, like, they can do the hack, but, like, you also shouldn't trust them, if that makes sense. Like, you don't really trust their incentives to, like, actually listen to you. Like like, the same problem of, like, steering it a little different also means that it could be doing things that you don't intend. Right? Yeah. You're giving it instructions, but you're like, okay. Hack it now. Fix it. It's like, well, wouldn't it be nice if there were some API keys? Right? And then then it can, like, you know, self proliferate it, like, the worst case. Sure. I think that that's kind of the big concern. And and, like, I'm sure you guys saw when when the hacks were going around. I mean, I still think there... The hacks are still going around. Yeah. Someone That had, like, a deep sea swarm essentially, like, know, patching old PHP. They're, like like, hack into old PHP and then patch it themselves. Yep. So pretty much, the only person that can stop a bad swarm is a guy with a good swarm. Guy with a good swarm. What I'm trying to say.

Speaker 2: Yeah. Yeah. How how are you thinking about advising clients on the risk to an open source ban for neo clouds? Like because some of the neo clouds seem very dependent on open source models but at the same time like the Anthropic S1 shows that the neo clouds can sell compute to closed source labs very profitably as well. So how are you balancing those risk factors right now? It feels like we went through the whole Jensen letter. There was a lot of energy towards don't ban open source, but now we're also post PHP.

Speaker 7: Yeah. I don't know. I think... I mean, did you guys see the thing on feed today about, like, essentially, like, trying to go back to an Alibaba cloud? Like, I think Yeah. Yeah. Those go here. Those are the kinds of anecdotes that creates, let's just say, like, a blanket wide industry ban. I really don't... Like, it's... We're in a really interesting time where it's, like, really reflexive. Like, clearly, something has happened. Yeah. Something has changed. And is it that we are just like, okay. Swarm up. Use more compute, or does the government say, like, fix this. Right? Like... And the government saying fix this is probably saying these class of models are now illegal. Right? And so that is a potential. I don't really know when you're in, like, these really special reflexive moments. Yeah. It takes very little headline to move the entirety of a vibe one way or the other. Totally. So we'll see what it kinda shakes out as.

Speaker 2: Yeah. Yeah. I mean, even if there's no ban there there's there... We were talking about like the the carrot is like bug bounty maybe from the government, but the the stick is like something that's not quite nationalization, but it's the government saying like, I am ordering you Anthropic, OpenAI, Google to devote millions of dollars in compute to setting up swarms to go patch the internet and fix Yeah. Everything. And you have to do it. Like you have been conscripted for this and maybe it's a tax write off. But like you gotta...

Speaker 1: What kind of timeline are we looking at to have a American open weight model that people are like genuinely excited about? Like getting like this changes everything. Yeah. Know? I wanna see the pelican on a bike and have people losing their mind over the pelican, right? When are we how far away are we? Like clearly Jensen cares about this. Yeah. There's rumors about Reflection. Reflection and them going in house. But like, is this like a three month thing or like a, you know, November surprise kind of thing?

Speaker 7: I I mean, I think the reflection is our best shot so far. And even though it wasn't quite at the frontier, it was like a... The frontier of open models. Sorry. Mhmm. The frontier of open models, it did look relatively parameter efficient. But if I'm gonna be honest with you, it wasn't exactly like, wow. We're so bad. Because, like, you know, one month later, Kini three point one looks like it's a frontier model. Right? Yeah. So... Yeah.

Speaker 2: What what else is changing? Maybe, Joe, you can fill us in on some of the trends on tokenomics. I mean, from the s one, it looks like the cost per task is getting... Is coming down. There's a lot of progress there. But what other trends are you seeing in tokenomics broadly on the demand side? What Fortune 500 companies are asking for? What businesses are asking for? Like, how is the tokenomics model changing in some of the analysis?

Speaker 9: Yeah. I think the biggest misconception is actually on the Fortune 500. Okay. A lot of the business today is really driven by ninety ninth percentile customers. If you look like the ramp data, example. Yeah. You know, the top 1% of customers are spending a 100 k per employee. You know, they came out today. The s one said the top 5% of cloud enterprise users are spending 18,000 a year per employee.

Speaker 2: Wow.

Speaker 9: But, you know, the average non cloud code user is spending less than $75 per month.

Speaker 2: Sure.

Speaker 9: And it was 0.1% of Fortune five hundred budgets were were cloud budgets according to their s one. If you look at who's spending heavily on the frontier, you have a few large companies like Meta, Microsoft, you know, other companies like Cursor that utilize the API. But the Fortune 500 Global 2,000 is very lagging in their per employee AI spend. Yeah. And I think, you know, the story in a '27 is, you know, we can see a decent step function change with a lot more budget allocated next year towards us. Yep. And kind of taking from more of the traditional IT and software budgets towards us and we can see a lot of growth outside of, you know, both coding Yeah. And on power users.

Speaker 2: In the Fortune 500, the one number that surprised me to the downside was that they reported over 50%, but not 90% of the Fortune 500. And I was like, oh, this is kind of bad. But then I realized, like, it's kind of bad for the Fortune 500 that they're not spending more. Right? Like what is going on in the Fortune 500? What is the what is the shape of a business that's not using AI in the second half of twenty twenty six look like? Is this like a lumber yard or something? Is it all humans walking around?

Speaker 9: No. You... You'd be surprised. Okay. Even at even at a lot of Fortune 500 tech companies, the Claude code budget might be 1,500 per month in tokens. Okay. Yeah. You know, that little and and a lot of people don't even hit that in the engineering organization.

Speaker 2: Sure. Sure.

Speaker 9: I think, you know Slowly. Areas outside like traditional white collar, you know, middle front back office work, you know, still is lagging a ton as well. Yeah.

Speaker 2: I don't know. Then some of those don't

Speaker 1: know if Doug could survive on 1,500 budget daily. Just

Speaker 2: Sounds like a nightmare.

Speaker 8: I could do $1,500

Speaker 7: $1,500 a

Speaker 1: day. I I think... Insane though.

Speaker 7: I would go insane slowly. No. I think the thing that I hear that like... My my favorite is actually like talking to normal people. I went to a wedding this weekend in in good old Fort Worth, Texas, okay, with normal people.

Speaker 2: There we go.

Speaker 7: And they're like, dude, it's crazy. No one's like... Okay. One person who's like really AI pilled, of course, talks to me about AI. He's like, it's crazy no one's talking about this. And they're like, yeah. What do you use for work? And they're like, oh, yeah. Microsoft Copilot. Like, you gotta... The world is a big place, guys. Totally.

Speaker 2: Yeah.

Speaker 7: And and I think we are still pretty early into, let's say, the hyperadopter curve. Like, is... You know, like, the hyperadopters. Like, I would argue seminalysis isn't like that. Probably in the one percentile outcome of, like, people who are spending on quad. Right?

Speaker 2: Yep.

Speaker 7: And we are like, okay. We can't go backwards. We definitely... Maybe we're not, like, like, going crazy with opus fast mode or something like that. Yeah. But I think that we will continue to be spending quite a bit, and I think what we're doing is a little different than the average company. We're a little younger. Right? There are companies out there where the average employee age is north of 50. Do you think they really care to figure out cloud code? No, dude. They're just trying to collect some checks before retirement. Like, I think that, you know, it just... It takes a long time for people to adopt new technology. This is something that's like the the technology industry has been very aware of for a long time. There's like an s curve in it... Of adoption, and I think we are still pretty early. So it's gonna take some time. My favorite my favorite theme is always to pay attention during the holidays, actually. You watch that their step functions of adoption happen during the holidays because young people go home to their... You know, to the boomer family and explain, hey. Have you seen muse try to buy you an advent calendar or something at Thanksgiving? And they're like, wow. This is magic. So there's there's some really cool things there that that's often why we see step function. A good example is Cloud Code's true inflection moment, even though they had, like, 4.5 out for a while

Speaker 2: Yeah.

Speaker 7: Was during Christmas because everyone had time off, and they're like, hey. I wanna vibe code a project. And then everyone's like... They came back to work, and they're like, guys, you gotta try this.

Speaker 2: So Yeah. So that's very interesting because in the s one, they they noted that revenue or or or token spend might go down by 20% in that two week Christmas holiday, but that's actually where your hardcore users are taking a break but your new users are actually coming in and you're sort of creating a new cohort that will scale and be the influencer of the next year, which is which is very interesting dynamic.

Speaker 7: That's Totally possible. Right? They're taking a break. But I think... I I I just think it's really interesting that these these adoption curves are, like... They're not continuous. And the most important thing is looking for the next adoption curve to see what the next what the next s curve is. Right? Like, it's very clear. Cloud Code is gonna be... Or, like, rather, let's say the the white collar work replacement is the biggest one possible. Yep. And it's, you know, relatively early on, but it's still really large wave. What's the next one gonna be? It's gonna be really hard. We... Biology could be, but I don't think it's gonna be anytime soon. Yeah. Cybersecurity probably is now moving up the list. Agented consumer stuff, definitely I think moving up the list with Muse. I've been...

Speaker 1: How are you? What's the overall semi analysis company view on the, like, market side, like, personal agent market sizing? Because people have tried to do... Like, I think some people... Like, reading into the the meta earnings call, I think people... I was at least a little bit surprised to see that, like, shopping was the number one use case just simply because shopping on the sites that Muse supports, like Shopify stores, is already so efficient. Like you just can't... It'd be hard to actually... Like if you tell your Muse agent to buy something at the same time that you have the product that you wanna buy, like it's not like there's real, like, time savings. Yeah. And people and this idea that, like, people like shopping. So how are you guys thinking of that market? When I see that Meta is doing a $95,000,000,000 equity financing, I'm assuming that they're still gunning for this sort of AGI SI vision and it's not really in service of the personal agent product that they have traction with right now. But what's the broader outlook? Joe, you want to take this? Or...

Speaker 9: Yeah. I mean, think... Look. They're they're still very determined in enterprise AI in achieving AGI there. And, you know, I think in '27, we'll see them monetize new enterprise pretty quickly. You know, they always in consumer when it's a new ad format, you know, we see delayed monetization. Obviously ecom, just any of like their ad markets today, serve ecom travel, know, will be big areas in consumer as well and we know consumers don't like to pay for stuff. Yep. So taking a cut or advertising some performance based outcome is the way it will be monetized. And I think we can get to the point where they talked about on the call where your consumer agents are talking to business agents and they facilitate that transaction and, you know, take 5%.

Speaker 2: Yeah. Yeah. It seems like Zach or Meta internally is still just like a little bit cagey about being like, we're going to make a ton of money off of ads. And so they'll they'll frame it differently in the press releases, but the the big dollars come from advertising when you're in the meta organization. Doug, you mentioned your Muse Maxi, your Clog Code Maxi. Do you have any linkages between your two agents? I'm interested in your philosophy around, like, will you have a super agent that talks to everything at some point? Or are you happy to have two tools, two tabs, two apps?

Speaker 7: I I think I'm happy to have two two tools and two tabs mostly because I think of Muse as intensely my personal. Sure. Like, I don't let it touch my work email. And then I think of Cloud Code is... Like, has a lot of the in... Inner linkages with my work email. Sure. I don't think it says proactive. Like, I am kind of excited. Like, let's just say, like, for better or for worse, Anthropic really has positioned itself as the enterprise solution. Yep. And if they have a really solid enterprise clawed, you know, clawed muse or whatever, I'm probably gonna use it just because I do kind of trust their safety to a certain extent. Like, they sure be talking about it in their s one. So I think Yeah. I think I think that that's how I've mentally bucketed it versus, like, you know, Meta has all my data anyway. So Sure. And, like, you know, I'm scrolling on... I'm doom scrolling on Meta Yeah. On on on Reels. What's what's the difference between doom scrolling on Muse versus Reels? Yeah. It's like It's

Speaker 2: of essentially buying It's like it's like you have an executive assistant at work and then your Claude instance will be like a executive assistant extension and then you have a personal assistant at home if you're doing very very well And your muse agent can be your sort of personal assistant. And these two might interface, but they're separate.

Speaker 7: Well, it's different. Might interface. Like, actually, usually what it is, it's forwarding emails one to the other.

Speaker 2: Sure. Sure. Yeah. But but actually having that clean line is great and that's the way a lot of extremely high net worth individuals set up their lives. They have someone who works at the the company, and then they have someone that works at their family office, and they too can communicate, but they are in separate organizations. So maybe that's

Speaker 7: a I think it's good from a a separation perspective, but, like, I I don't know. I also just think that, like, you know, the personal assistant stuff, it is really sick to just, like... I know the the lamest one, everyone's like, oh, just reads your emails, but, like, dude, I got a lot of emails, man. I think a lot of people, it's like, okay. For example, you have a kid's birthday party, and you're, like, organizing the kid's birthday party and like you just like, oh, your your wife sends you some reel and there's like, hey, I wanna go to school. You're like, you're...

Speaker 1: Kid's teacher says describe your child in three whole...

Speaker 2: That was a bad example....

Speaker 1: Stumped, you know. That was a bad example.

Speaker 2: But no, mean I I had an interaction where I had emailed someone to to do something and they had to actually update something on a website and they said that they were going to do it and then Dot in this case like went and checked like a week later and said like, oh, haven't done it. Like you should follow-up with them and just make sure that they actually finish that task. Those types of little like follow ups are more useful and it's more than like it would require me not just reading my email inbound, it would mean reading my outbox and checking up on the things that I asked to do which is like complete laboriousness that no one wants to do. Now, what is the willingness to pay of everyone for that? How do you monetize that? You know, travel. You know, Americans don't spend that much on discretionary travel all the time. So you take a small cut of that. These aren't necessarily quadrillion dollar markets immediately but Yeah. They are very interesting and and can grow a lot.

Speaker 1: Final question I forgot to ask you earlier, Joey. Did you wanna see more on their actual tokenomics like more granular data on margin profile across different models

Speaker 2: Oh, things yeah. Like

Speaker 1: Because I think some people were hoping to see more. But in some ways I look at it and I'm like, how much money did you spend? How much money did you make? Yeah. Everything else is kind of like your business. But what's your view? Feels like full employment Yeah. For some

Speaker 9: Yeah. There was a number in there that 80% of the revenue in the last three months were from models in the last released in the last six months. So I think that was the extent of detail we got. I don't think they're going to get much more detail than that. I know they've given some inference gross margin numbers before. If you have that and the training, I think that gets you most of the way there in terms of understanding this business. And then too, they gave a lot of good detail I think on the cloud service partner. Yeah. Your relationships and the revenue running through that and kind of the economics they're paying from a revenue share perspective. So I think you have enough detail here to really understand the business and you know, how it's going to move going forward.

Speaker 2: Yeah. Have you been tracking how the big labs are using pricing pressure against open source models? Is there a world where the labs get to such scale and such efficiency that they can distill their own models to a point where they can be price competitive in some way just to sort of stave off the open source ecosystem a little

Speaker 1: exactly what's happening.

Speaker 2: It feels like that's what's happening, but what's your take from like an economic perspective?

Speaker 7: I think the person who's doing it the most is OpenAI. Yeah. With OpenAI. The 6.1 tera today feels very much like a Sure. Stab at. I don't think Opus or rather Claude cares about Sonnet that much. Yeah. But I do think that OpenAI has kind of their eyes out focused on that. Yep. I think the problem with that, though, is at the end of the day, you know, the the inference margins on open source is pretty good. So if you really wanna get all the way there, you probably have to do another 50% price cut or something like that. Yep. Where I... Because if you own it today, if you own the stack, if you have, like, a relatively modern v 200, you're still talking about north of 50%, if not, like, 70% gross margins. Yeah. So, you know, you're able to pay, like, these, like, 10... Like, these, like, twenty to thirty dollars per watt or, like, 30,000,000 per megawatt Yeah. Price using just open source models if you get the customers on it. Yeah. So I think that, like, it's possible. It's entirely possible, but I think the margins of the labs in aggregate would have to be a lot lower, but you'd argue that they wouldn't have as much competition.

Speaker 2: Sure. Sure. Sure. And then and then the the the wrinkle is that then you also have to have some sort of like fine tuning capability because that's something you get with with open source that you won't with a closed source model necessarily. Interesting. Well, thank you guys for coming on. So and and and walking through all of this. We do have to jump to the next interview. But Jordy, do have any...

Speaker 1: Yeah. I really enjoyed it. Thanks guys. Great to see you.

Speaker 2: Great to see you. Congrats on the haircut.

Speaker 7: Thank you. Appreciate it.

Speaker 1: See you guys.

Speaker 2: Have a great one. Let me tell you about public.com. Trying to invest in Anthropic? Do it on public. Invest across all asset classes. Automate anything with AI agents. Run trading strategies, risk management, money movements, and monitor markets with AI agents for investing on public. Our next guest is the chief executive officer of Turo. We have Andre Haddad. And Turo is launching verified FSD today for Tesla's.