The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang

00:00:00And I think we're all pretty AI pilled.And if you're AI pilled, that means wegot to build a lot more compute than theworld thinks. And that these models aregoing to be a lot more valuable thanpeople think. You combine that withtheir core business, I don't knowanother entrepreneur or another businessthat's a better bet on the future,right, than SpaceX. And so I think formost institutional investors, it's amust buy, a must own, a set it andforget it, right, in order to have a areal bet on both the space and the AIfuture.

00:00:40All right, here we go.Early morning Silicon Valley, BG2 isback. We're chopping it up on all thingstech and markets. To do that, I havenone other than GB in the house, GavinBaker from a Treaties. He's brought hismain guy, Andrew Fox. And of course, Ihad to draft Clark Tang into the mix, mypartner,um to talk to to talk about some of thebig questions of the day. You know, howshould we be thinking about the SpaceXIPO? You know, what are the big levers?There are big numbers out there forwhat's going to happen over the courseof the next few years. So, let's breakthat down a bit, help simplify it forfolks.I Mythos launched yesterday. I want totalk a little bit about like who's up,who's down in the race forsuperintelligence. Where are we? Whatdid we learn with the Mythos launch? AndClark was in uh in in in Taiwan lastweek um with Jensen at Computex and GTC.So, what was our takeaway there? What'sgoing on with GPUs, memory? Where arethe bottlenecks? And where do we go fromhere? To start everything off, um youknow, maybe just kick it over to you,Gavin, talking about the SpaceX IPO. TheIPO is in 2 days. Uh you're a bigshareholder. Congratulations. We're alsoa shareholder. Uh you know, we also weexpect to be buying in the IPO. The WallStreet Journal's reportingum you know, the Goldman Sachs are bothsaying 160 billion in

00:01:55revenue in 2028.Um we know that the IPO is $135 a share,1.77 trillion. Um sowhen we think about kind of what the biglevers are, there's so many moving partsin this IPO. Um nobody's better than youat just breaking it down, simplifyingit. What are the key levers that weought to be thinking about that you'rethinking about over the course of thenext few years?>> Sure. Um so great to be here. Thank youfor having me. I thought we're going tocall it B BGG B,but [laughter] we can stick with BG2.I've been your house.>> hey.All subject to revision.>> That's okay. That's okay. So I thinkthere's twobig levers or variables that I thinkpeople should focus on. And you know,I'm not going to comment on where Ithink um those variables go.But one is um you guys have this chart.Um did you did you post this on X?>> I did I did before and then we alsoincluded a new addition with uh XAI'snew deals as well.>> Yeah.>> Yeah.>> So Clark, who I've known for many years,umuh made a did a great analysis here.And he shows that XAI's deal um withGoogle for cloud computinguh generates more operating profit pergigawatt um than Anthropic, than

00:03:10Meta,than Google, than OpenAI. Uh their dealum>> [clears throat]>> actually with uh Anthropicalso generates probably more operatingprofit than anyone but um Anthropic.And so you know, um the your yourcolleague at Altimeter, Freida, also shecalculated a 55% IRR>> Mhm.>> on Claude's one.>> Mhm.>> You know, if you can borrow money at 6,7, 8% and invest in something with a 55%ARR,I'm not the most sophisticated thinker,but that math maths.>> Right.>> And so I think the most importantvariable, one of the two most important,is how quickly they bring on terrestrialdata centers.>> Mhm.>> We do know from Jensen that uh Elonbrings data centers up faster thananyone 122 days. Speed is literally costbecause every day you're payingelectricians and plumbers. That's cost.And they're now monetizing them atarguably the highest rate. And so Ithink, you know, everybody should runtheir own math on that, but that is amassive variable.>> Yes.>> Truly massive variable. The second thingis, you know, we have a chart and it'swildly out of date now. It's kind offreaking amazing. This chart is I thinkis this chart from 10 days ago?But it like the 10 or 12 days

00:04:26since thischart since we made this chartwhich shows the Pareto curves for Opus4.7 for codingfor Codex from OpenAI. And now we've hadOpus 4.8. It was already out of date.And now we have Fable>> Totally.>> and Mythos, which is freaking wild. In10 days>> Yes.>> like we would have had to update thechart twice.>> Right.>> But what the Pareto curve sure shows ishow much intelligence you can get for agiven amount of cost. And I do thinkbeing all revenue will accrue to thePareto curve. All at least kind offrontier model revenue will accrue tothe Pareto curve. And this is Paretocurve for code coding. And what I thinkis so impressiveis thatyou can see in the chart that Composer 2was Pareto dominantor, you know, at the lowest level ofintelligence with very little training.This just reflects, and I know you knowCursor well. I think you know Cursor a lot better than I do.A vast amount better than I do.[laughter]But my understanding is that Cursor andAnthropic have more tokens ofproprietary coding data than anyoneelse. And they have more tokens ofproprietary coding data than exist onthe public internet.And so they fed Cursor fedum used ChemK 0.25, used their ownprivate data, did some RL,

00:05:41somesupervised fine-tuningand they got a really good model. Andthen they spent 3 weeks in the Colossus2 cluster and they got a model that12 days ago was Pareto dominant withComposer 2.5. That's on their ownbenchmark.Um Cursor bench, so maybe take it with agrain of salt. But I think this justsuggests that the Cursor data is veryvaluable for coding and when it istrained, you know, to Chinchilla optimalor beyond Chinchilla optimal withreinforcement learning,you know, I think it suggests that XAIand SpaceX AI has a shot of being a realplayer in coding.>> I mean, I think one of the interestingthings is, you know, weyou know, the way he answered thequestion, right? We didn't talk aboutlaunch.Right? We didn't talk about Starlink orcommunications. Those up until really 6months ago were the business.>> Yeah.>> Right? And, you know, and then we mergedin x.ai and we merged in Cursor and thenwe announced these deals where it wasvery clear he was kind of building AWSright under our nose, you know, in in inin terms of this. But what I want to dois go to go to Fox. Give us thebreakdown. Three big lines of business,right? We've got the the thecommunication Starlink launch business,we've got the, you know, AI computebusiness and then I want to come back to

00:06:56x.ai that you were just clicking on. Butif we just go to the core business, whatdo we have to assume goes right in thecore business both with launch and withStarlink in order to achieve the numbersthat are out there?>> Yeah, sure. So, look, I thinkthe thing that's foundational toeverything is the launch business.>> Right.>> Right? This is the kind ofcrown jewel of SpaceX. Um it's somethingthat no one else really has, notablyreusability.>> Right.>> And soon rapid reusability. Right? Thisis, I think, what you need to believe into get to the economics in AI that makeorbital compute something that's veryeconomically attractive.>> Right.>> Outside of the idea that we are inshortage for power, shortage for chips.Right? Um so, I think rapid reusabilityis the main thing that we're watchingfor and I think most people should watchfor. Um,you know, Elon talks about it a lot, butgetting these rockets to fly at acadence that's comparable to an airline,right? And and Gavin has used thisanalogy before, butum, the old rocket industry was kind oflike, imagine boarding a plane, flyingto California, getting off the plane,the plane explodesafter. Um, so I think what SpaceX areultimately trying to achieve is have aStarship fly both stages, not

00:08:11just thebooster. Um,uh,30, 40, 50 timesbefore you have to retrofit that ship.Um,and when you do that, you're amortizingthe cost of the vehicle over manyflights, right? And that's what bringsthe cost down significantly.Um,>> But that's a really hard problem tosolve.>> Extremely difficult and look, I thinkthe company, you know, havebeen loud and clear, they're going toattempt to bring back the second stage>> Right.>> of Starship later this year.>> Right.>> Um, and then make it reusable, you know,re-fly the second stage next year. Um,and from there, ramp up the cadence. Butat the end of the day, driving down thecost of launchis what enables all of these otherbusinesses and is what makes them soattractive relative to incumbents.>> So how many how many times are weStarship just launched Starship 3, youknow, just launched. How many launchesare you know, do you think kind of theconsensus out there is assuming, youknow, two or three years from now? Likewhat is the launch cadence? Are welaunching one of these every day or welaunching one of these every week orevery month? Like where are we in termsof expectations?>> Yeah, so look, I think expectations fornow, you know, we're going from,

00:09:27youknow, call it 160 165 launches last yearup into the high hundreds of launches inseveral years and, you know, gettinginto the thousands of launches probablyin the next 3 years thereafter.>> Okay.>> Um I think the company have aspirations.>> Thousands of launches, you're launching,you're doing two or three launches aday.>> Right.>> Right. And then talk to us a little bit,what is this enable? Obviously, youknow, I'm here in Silicon Valley. Ican't even I can't even keep a call onSand Hill Road, two decades into themobile revolution. I mean, it's thecraziest thing. It's like a third world.>> It's It's a major business problem whenyou're freaking out here. [laughter]>> It's crazy. It's crazy. Right by theStarwood dead zone. And I'm like, howcan this possibly be? So almost likeit's a joke. It's the epicenter oftechnology in America and you can'tmaintain a call. Okay, so we're allgoing to switch to Starlink mobile whenit comes along because I don't want tolose that call on Sand Hill Road. Sowalk me through a little bit just againhigh-level. Um it's a big portion of therevenue growth expected in the businessover the course of the next two to threeyears. My hunch is a lot of this isdriven uh by uh by direct to cellconnectivity. Walk me through a littlebit those economics.>> Yeah, so look, it's actuallyinteresting. Umthe broadband business is still

00:10:42veryearly stage when you think about um thepercent of households that have actuallybeen penetrated to date. You look at thepercent of global households withStarlink, it's less than 1%. Uh andthat's the broadband, you know, you kindof have a base terminal at your house,on your car, on your boat, um and thenairlines now as well. Umso I actually think broadband can scaleto hundreds of millions of terminals,hundreds of millions of users. And todaythe subscriber base>> Hundreds of millions if if they getrapid reusability of Starship, which isreally hard.Umyou know, if if there's not competition.Um hundreds of millions, um it'spossible.Butmaybe>> I I always say around here, it's funny.I love seeing PM and kind of analysts inin this situation. It's exactly what Ido with Clark. Clark will say something,I'll say the future is a distribution ofunknown probabilities. It's either morelikely or less likely, so give me thedistribution. Are we talking 20% 30%?It's hilarious. It's the same>> Well, no, 100% same thing. And like I'vewatched Elon do many hard things andthis is a really hard thing. So, I thinkit's reasonable to think that they'regoing to succeed with rapid reusability,but just I just think it's important toacknowledge that likeorbital computeyou know, Starlinks, you

00:11:57know, StarlinkV3, Starlink direct to cell, we need weneed first reusability for Starship V3and then rapid reusability unlocks a lotof this.>> Right. When I see when I see the modelsthat the banks are putting out there,right? And Wall Street Journal,everybody's reported on these. Thesesame things have been widely leaked.They they they largely have the revenueon connectivity, so let's call itStarlink direct to cell, etc. Goingfrom, you know, let's call it 10 billionto 50 billion uh by 2028. And so, I'mnot asking you guys to react to, youknow, to tell me your specific numbers,but when I'm talking to Clark all I'mtrying to size up is order of magnitude.Do we think we can 5x the business overthe course of the next 3 years? Is thereenough TAM both in terms of broadbandand direct to consumer? And I think theanswer to that is yes.>> Yeah, here's what I just say very simplyis I have umI I travel with Starlink. UmI'm I'm a big video gamer and veryconsistently wherever I am in the world,Starlink is the best connection.>> Yes.>> It's the fastest, it's lowest latencyand I do think once they get to rapidreusability, it's also going to bethey're going to have the cheapest costper gigabyte or megabyte delivered umand better faster,

00:13:13cheaper has been awinning formula.And so, 50 billion, that's, you know,0.3%penetration of the global telecommarket. Now, maybe there's somedeflation with Starlink pricing.Um but that's the way I'd frame it up.>> Yeah. I like betting on better, faster,cheaper.>> Um, Clark, I would say probably thebiggest surprise of the last six sixweeksis that Elon, you know, we talked aboutit on all-in podcast, we called it EWS,Elon web services, right? That that hestruck these huge deals with Anthropicand Google.I don't even think people were thinkingabout SpaceX in the AI compute game,right? We If you looked at the models asof a few months ago, it wasconnectivity, so Starlink, and then itwas x.ai, the model.But this whole category of taking all ofthis compute, which he's uniquely goodat standing up, right? And thenreselling it in a way that's highlyprofitable was not in a lot of people'sforecast. Now it's a major component ofthe forecast. You know, you and I didthis podcast with Jensen, where Jensensaid Elon is an N of 1.>> What they achieved is is singular. Neverbeen done before. Just to put inperspective, 100,000 GPUs, that'syou know, easily

00:14:28the fastestsupercomputer on the planet as onecluster.Um, a supercomputer uh,that you would build would take normally3 years to plan.>> Right.>> And then they deliver the equipment, andit takes 1 yearto get it all working.Yes. We're talking about 19 days.>> Wow.>> N of 1 is right. Elon is an N of 1.>> And his ability to secure supply, standup the supply, you know, deploy it in away that's uh, you know, coherent andeffective for both himself and I guessnow for others. So, walk us through kindof that. It looks to me again like thisis a major component of the revenuestory.>> Totally. I mean, so we we were all atthe macro hard data center, and it wasjust very evident the amount ofengineering that was that had gone intobuilding these sites.Um,you people always talk about Google andtheir ability to to build a TPU and sellthe TPU to Anthropic to generaterevenues for AI.I think it's a pretty similar dynamichere with Elon able to secure power, uhbuild these sites faster than anyoneelse and also be able now to monetize itto um to

00:15:43the this massive AI marketthat's ahead of us.Umif you look at the the relationshipsthat he's forged with a lot of hissuppliers, you know, be it Jensen, beit, you know, all of these different umdifferent sites that actually want xAIas a tenant. Um his his ability tofinance these deals at at veryattractive uh financing rates relativeto a lot of the other players in in thespace, you know, these are advantagesthat compound over time. And when you'vebuilt the credibility to stand up thesesites and monetize at these levels, umyou know, it's very it's actually a veryattractive uhvery attractive proposition for for alot of folks involved. Um and actually,you know,if you look at the these deals inparticular,Gavin, you you pointed out, but, youknow, they're they're actuallymonetizing, you know, perhaps betterthanother players in the space by sellingthis infrastruc->> a lot higher.>> Um Google is obviously paying SpaceX ahuge premium for this compute. Fox, yousaid something that I thought was reallyimportant, which is, you know, it mayvery well be that in order to get, youknow, first in line on space compute,which Google certainly wants to do, thatthey're willing to pay a premium for

00:16:58their terrestrial compute. And so, tome, that's how you kind of square thecircle as to why the premium. Anythoughts?>> Yeah, look, I think there's some of thatembedded there, but um look, at the endof the day, SpaceX can stand up computequickly.They can stand it up coherently. Andthey can stand up a lot of it in oneplace and have it readily available. So,look, I think that's most of thepremium, but outside of that certainlypeople are going to space over time.>> I have to pay a little call option toget first in line for space.>> There you go. Good one.>> We've all been investing in the neocloud space. So, like there's afundamental belief around this table Ithat that we lack the compute needed tocontinue to push the frontier onintelligence. So, we have to build a lotof compute, okay? Now there's a there'scompetition going on. On one end youhave the hyper scalers who are buildingout that capability. Then we have AIdedicated clouds that are building outthat capability. And now literally in amatter of weeks, right, we have a a youknow a giant that's emerged in thiscategory which is SpaceX.The question to you Gavin is can theyconsolidate this market, right? Becauseif I think about a marketplace, Elon hasa unique ability to get the supply. Hehas a unique ability to cut deals on theother side and nobody can stand it uplike he can stand it up. So,

00:18:14I thinkthere might be a real consolidation inthe AI compute market where you have thehyper scalers on the one hand and on theother hand, you know, he may emerge asthe largest, strongest player in the AIcompute market.>> Yeah, so I think they're are they thenumber four number five hyper scalertoday after the Google deal?>> Um it will be number four.>> Kind of wild.>> Yeah.>> In 30 dayswe went from not being an AI hyperscaler to being number four. And wepassed a lot of companies includingOracle.>> Coreweave is a huge business, right?That we're we're investors in, you know,and have been investors in, right? Butthere are a lot of other players, theNebius's of the world, the Iron's of theworld. And I would say that they'reprobably 50 neo labs being funded inSilicon Valley right now as we speakbecause of the shortage in compute.>> Absolutely.So, that'skind of crazy in 30 days. That's justextraordinary.What I would say is there I think thereis a belief that these data centers arecommodities.>> Mhm.>> And I do not share that belief.Um I don't think anybody around thistable shares that belief.And in the same way that Elon was ableto re-engineer a rocket from firstprinciples and make it reusable,he engineered an electric car from firstprinciples. You know, everyone else wastrying

00:19:29to, you know, make an electriccar like an internal combustion enginecar and he thought about it differently.And umI think he looked at data center designfrom first principles and he designedsomething fundamentally different. And Idid actually ask the team. I said, "Heyguys, maybe I'd be a little less publicabout things that are very obvious toyou [laughter]>> Right.>> about how to design a data center, butare revelations to other peoplebecause I think what you're doing ismaybe ummore differentiated than you perhapsrealize cuz what you're doing is sological to you,but maybe lot not logical to everyoneelse. And that's how he was able to doit in 122 days.>> Yeah, to I mean to that point, Brad,yesterday we were meeting one of ourportfolio companies and we were talkingabout behind the meter and we're, youknow, really thinking about it. There'sonly maybe two or three two or threeplayers now that can actually reliablyengineer behind the meter data center.And you know, there's real engineeringwork that goes into all of this. So, ifyou think about this,if you're a gas combustion if you'reVernova and you say we only have acertain number of gas combustionengines. Now, we can sell them to x.aior we can sell them to one of thesestartup neo clouds. Who are you going tosell them to?>> Well, and there's another dynamic.

00:20:44Everyone starts making more money whenthe GPUs get energized and sold faster.So, literally speed is money for all ofthe suppliers. Power, land, turbines.So,I think it's we'll we'll see.>> Right.>> Hey Brad, man.>> And but but this is just we're justtalking terrestrial. I do I do want tohit on and then you can flip it back onme. Talk to me, okay, So, let's let'sassume, right, that they continue tobuild out the terrestrial landscape.They continue to find buyers for that.Um, walk us through, you know, what thisunlocks, you know, and how this isrelated to space data centers because Ithink, you know, once you start talkingterrafab capacity and beyond. So, we'retalking 1,000 gigs, right? And this yearwhat what we're doing, 25 or 30 gigsjust to put it all in perspective.>> 20, yeah.>> Right?>> 20, 25 gigs.>> Okay, so so once we start scaling up,walk us through,do we have to have space data centers inorder to get excited about buying theIPO, right? And then there's obviouslythis this debate in the world. I I heardJeff Bezos say, you know, I think it'smore like 6 years, but Elon's going tosay three because if if he says six,then it will take even longer. So, saythree and we may get it in four or five.But are space data centers

00:21:59integral andessential to, you know, the IPO? Andwhat do you think the timeline is, Erin?There are you guys.>> So, I don't think I think if you thinkabout those variables around whatCrusher could mean for XAI.And we do have an existence proof thatonce you really get on that Paretofrontier,revenue can scale rapidly and it'scalled Entropic. And there does seem tobe an exhaust There seems to be a lot ofdemand for coding. And I do think I'mJohn Massad posted something veryinteresting.>> The The founder of Replit.>> The founder of Replit. It he called itbitter lesson adjacent that coding maybe the fastest path to AGI and ASIbecause if you really go to coding, youcan write code if a model's good atcoding to do anything. So, I thinkthat's a profound point and I thinkcoding is going to continue to be veryimportant. So, I think if you thinkabout that variable,if you think about Starlink direct tocell enabled by Starlink V3, and youthink about how quickly they can orcannot bring on terrestrial compute, I Ithink orbital compute is isis necessary for the IPO valuation,but it's certainly important and it's>> Well, maybe another way to say it is youthink youyou may think we're going to get to ASI

00:23:15faster than we're going to get toorbital compute. That may take us from300 IQ to 400 IQ, 500 IQ, and beyond.Um and the ability to scale it up toconsume 10% of, you know, global GDP,but maybe maybe that's where we shouldmove next.>> No, no, I think on the orbital compute,I think Foxy would be great our Clark tolay out the math from first principleson, you know, Clark has this great charton, you know, the gigawatts it costs,you know, the dollars per gigawatt.>> Right.>> Walk us through the economic case.>> Yeah. Yeah, so I mean, on this point ofis orbital key to investing here? Idon't think it is and I'll first pointI'll make is what are the impliedmonetization ratesbased on expectations today for the AIbusiness? And you know, I think youthrew out the $160 billion number that'sbeen leaked out there that people aretalking about.The implied monetization rate on thatnumber is something like $14 billion pergigawatt per year for the AI business.They just signed Anthropic at 22 to 23.They just signed Google at 50.>> Right.>> Right. So, I I think you can investbehind the AI business terrestrially andstill be excited about it. But withorbital>> an important point. Excited about it ifthey can get the land and the power.>> Right. But but but I mean I I think formost investors, right? They

00:24:30get Theyhave an easier time getting their headaround how SpaceX wins terrestrially.Like can they go get land, power, andchips? The answer to that is highprobability yes, okay? And what we'resaying is at the rate they're monetizingthat, that gets you to the numbers thatare being leaked out there before youeven have to take the leap of faith thatthey're going to extend the lead withorbital data centers. But take us thereon that, too.>> Sure. Yeah, so so look, with orbital, Ithink the key thing is um two-stagereusability.>> Yeah.>> And beyond that, rapid two-stagereusability.>> Yeah.>> So, today with Starship, they've shownthat they can successfully reland thebooster.The second stage, we'll see what happenslater this year. I think they'reattempting to bring that back and thenmake it reusable by next year.Umbut the thing that's important abouttwo-stage reusability when it comes tothe economics for orbital compute,right, is the cost per kg comes downsignificantly. You know, we're talkingabout going from $1,500 per kg onFalcon, somewhere in that range, to 250per kg, something lower. Um and the morethat you can reuse the rocket,the more that price comes down.>> Right.>> Right, cuz you're just depreciating thecost of the launch.And eventually, you asymptote to thecost of

00:25:45the fuel.>> Right.>> Right.Assuming you can use a rocket forforever.>> Yes.>> Right, which will take a very long timefor us to to really achieve that. But,um and at that point, we're talkingabout something well south of 250 perkg.So,then you look at the specs of these AIsatellites. You know, Elon did a great>> Yeah, that pod that that pod wasincredible that he laid out the otherday, the specs on the satellites.>> It was really great because I think theyare finally showing people, here's howyou could viably design one of thesesatellites.And how heavy is the satellite? How manycould you fit into a Starship launch?And when you back into the numbers, youget to something like 5 MW of capacityper Starship launch.>> Right.>> There's 100 metric tons in one of thoseStarships. So, you can back into themath of how much will it cost pergigawattto launch these satellites into space.>> Right.>> Launch this compute into space. Umand the math that you get to before youaccount for things likebad GPUs, bad satellites, right, thesewill all be things that happen.But the math you get to is it's about $5billion per gigawatt of CapEx to putthese in space.>> Right.

00:27:00>> For comparison, terrestrially,talk about the switch gears, thegenerators, the transformers, the shell,getting the power, that today is about25 20 to 25 billion per gigawatt.So we're talking about a 5x reductionin coston half of your bill of materials>> Right.>> for the data center.>> Right.>> Which is a huge number.>> Yeah, just just very simply, I mean justto say that put it cost $60 to put agigawatton the ground today. And we'll call it35 of that is are the GPUs and thesilicon that's doing the training andthe inference.And 25 billion is the land, the shell,the power, and the cooling.I would hypothesize that those elementsare probably going to be inflationary,so that 25 billion may not go down.And because space, power, cooling areeffectively free in space, and when Isay space, I mean land. You know,there's no land in space, but there isspace.Umyou're you're talking about putting agigawatt into space for 30 billion andhaving lower operating costs. Now thedynamic versus 60 billion that'sinflationary, and that third and that 30billion, that five may be deflationary

00:28:16over time.But what we need to consideris you know, the reliability and themaintenance. And so as long as you know,everybody can do the math,but as long as these satellites in spacearen't failing at an at an astronomicalrate, the math maths. As you can see,and by the way, we know GPUs melt andlasers fail. We know this happens indata centers,particularly during big training runs.And yeah, I mean GPUs melt.Um so as long as the reliability andmaintenance is not dramatically lower,the math is there once we havereusability and then rapid reusabilityfor Starship V3.>> I when youwhen we look at this, okay, so we wewent through Starlink and we said,"Okay, like it it just stands to reasonwe're going to have direct to cell onStarlink." Like the assumptions thereare, you know, again, seem like you canget your head around. Then when it comesto building terrestrial data centers,again, not a hard one to think thatbased on these couple deals that Elon'sgoing to build a much bigger Starlink'sgoing to build or SpaceX is going tobuild a much bigger business there. Andthen you have this call option on spacethat would drop the price even further.The one thing we haven't talked about istheir model, right? And I find thissurprising, right?

00:29:31Six Six months ago,x.ai was competing, they were doingpretty well, but they've done somethingdramatic over the course of the pastcouplecouple months, which is they boughtCursor, right? Cursor is 700 800 peoplewas already doing incredibly well from arevenue perspective. Our ownprojections were that they could exitthis year at up to $10 billionof revenue, so they were growing veryfastone of the leading coding agents, butthey also had this incredible team withthe potential, right, to really build afrontier level model, but they werecompute constrained. So all of a sudden,they get bought by X. X has massivecompute that they can now train on.And when I think about the revenue in AIthat like if I look at that line item inthe models having it go from $10 billionto $150 billion, yes, a lot of that willbe the core weave type business thatthey have, but the question is how muchof that is going to be the core x.aibusiness that's really powered by thenew team from Cursor. So any thoughts onthat, Kevin?>> Right now, so Composer 2.5 was Paretodominant 12 days ago. It was trained onthe Kimmy K2.5 base model.>> Right.>> Now, what's happening is the Grok 4.31.5 trillion parameter model istraining.

00:30:46One would hypothesize based on scalinglaws that that will might be a betterbase model. And then the cursor data isbeing injected into the pre-trainingprocess, not just reinforcementlearning.And we'll see, and I think that is goingto be a very important data point whenthat comes out. And I just thinkeveryone should keep in mind that onceyou are at multiple places on thatPareto curve, if you have compute, youcan scale really rapidly.>> You know, that that to me is if I had tosay what the one piece that's being lostin the story,right? Like it's easy for everybody toget excited about the deals withAnthropic because you can put your handsaround that. You know how much revenueit is. I see debate about, you know, the90-day termination and how long theylast and what multiple do you put onthose revenues. But I think the thingthat's getting lost is I think they'vedramatically advanced their capabilitywhen it comes to building a frontiermodel. People outside Silicon Valley maynot know, you know, Michael and the teamat Cursor as well. This is anextraordinary team that he justdownloaded, right, into SpaceX. SpaceXwas already building good models. Andwhat they have isthey have this way to monetize computethat gives you this call option that youcan pull all that compute in-house,right, to train a model

00:32:02and then to runthe model. I suspect if there's anupside surprise, if we went around thetable, I'd say this is the place that'sgetting the least amount of attentionand could have the biggest upsidesurprise. Any any thoughts, Clark, onwhat you think is being overlooked orareas that you think are misunderstoodabout the business today?>> I I would say I would saywhat the last few weeks have proven isthat Elon, um,their team can stand up all thiscompute. Actually, if you just, youknow, went back 1 and 1/2 years, youknow, they were behind in the race tostand up compute. They were you knowthey they didn't have that many H100s.They brought in Colossus. Then theybrought in Colossus 2 at a scale muchlarger than anyone else. And now youknow as we gear for Vera Rubinyou knowfrom you know a lot of my conversationsit looks like they've you know securedmaybe up to 20% of Vera Rubin capacityespecially in the early days of you knowwhen you know these these chips are veryscarce that that they're going to have aaa lead on all of this because you knowpeople think that they can stand up thiscompute better. So I think they'll allyou know what what the last few weekshave actually shown is that Elon youknow Elon

00:33:17willtake you know take a shot at hitting thefrontier but if it you know if forwhatever reasonum theythey have over procured some capacitythis is a very scarce asset that they'veshown that they can monetize at actuallyyou know best in class margins andpayback periods.>> The irony is like you knowyou and I've been doing this long enoughto know I mean that's why Bezos builtAWS. Right? He had to build capacity forBlack Friday.>> Yeah.>> Right? But then the rest of the year hesat on all this capacity they had tobuild and he figured out a reallyincredible way to monetize this. And bythe way investors at the time 2009 2010when he was building out the capabilityaround AWS hated it.>> Of course.>> Because he was consuming all that freecash flow. My meanwhile he was diggingthe biggest gold mine in the history ofthe world. One of the biggest.>> One of the biggest.>> Among them among them at the time wasprobably the biggest.>> Yeah Google search might want to have awe'll have a we'll have a discussion.>> [clears throat]>> By the way I do think it is important.Grok 4.3 I think the cursor if theyacquire it that may end up being veryimportant. But Grok 4.3 was on thePareto frontier and has of 10 or 12 daysago and this these things move fast. Butmost intelligent 500 billion parametermodel in the world. And

00:34:33they were on thefrontier and there are four companies onthe frontier.xAI, SpaceX AI, Google one with Gemini3.1 Pro, and then the rest of it wasdominated by Anthropic and OpenAI. Butthey were on the Pareto frontier and nowwe'll see what they do with Cursor.>> Yeah. Um I want to come back to that ina second.>> way, man, I want to ask you somequestions.>> go go go. What do you think? So youthink the biggest source of potentialupside is the model?>> Yes.>> What do you think?>> I think that's the I think that's thething that's least talked about.>> Least talked about.>> Right? And so, listen.When I look at the bull bear case on theIPO, right? The bears are looking atlast year's revenue. Say it was $18billionand they're looking at the forecast fromthe banks of $160 billion, you know, 3years from now and they're saying,"Listen, not many companies in thehistory of the world have basically 8xtheir revenue over 3 to 4 years." Right?So that's where, you know, I think andpeople get nervous about the valuation.When I look at this, again, when youbreak it down as an analyst firstprinciples, part by part, which is whatI tried to do here, right? When you lookat Starlink, it looks totally doable.When I look at what they're building inAI compute terrestrially, looks totallydoable over the course of next 3 years.When I look at the model itself afterthe acquisition of Cursor, you know,combining those things around thecompute they have,

00:35:48that looks to me likeit could be an upside surprise. So Iwould say that I think that uh you know,in the IPO, but I think when you lookback 3 years from now, there's a decentchance that everybody's like, "Oh mygod, that was super obvious." Right?Even though today all of these thingshave risk associated and back to wherewe started. I'm not, you know, none ofus are here to pump the IPO at 1.77trillion. It's really to just break itdown as we do inside our shop and tosay, "What is that distribution offuture probabilities? What's theprobability that it's higher from here?What's the prob" And I think we're allpretty AI pilled. And if you're AIpilled, that means we got to build a lotmore compute than the world thinks andthat these models are going to be a lotmore valuable than people think. Youcombine that with their core business. Idon't know another entrepreneur oranother business that's a better bet onthe future,right, than SpaceX. And so I think formost institutional investors, it's amust buy, a must own, a set it andforget it, right, in order to have areal bet on both the space and the AIfuture.>> From your lips to God's ears.>> I mean, listen, I I again, I think thatI I I think that you're going to have towait, but you know, we had this chartlast week, right, that came out.Everybody was sending around Twitter,conveniently timed, and you know, it'slike shows the

00:37:03average max drawdown postIPO for like 20 companies from Facebook,Twitter, Alibaba, Shopify is, you know,over 50%. And so maybe that again willwill will end this section here. Youknow, Gavin, you and I've been doingthis a long time. We know it's going tobe bouncy around the IPO. Um,you know, how do you as a manager try totry to manage that? Um, do you try totrade around the IPO? Do you set it kindof and forget it? I would say from anAltimeter perspective, what we tend todo is we take a base position that weset and forget, right? And then we maysize up or size down depending upon howthe market reacts in, you know, in aparticular moment. Um, but any thoughtson on this chart oryou know, how how people you guys arethinking about it in particular. Youobviously own a lot going into it.>> First agree with absolutely everythingyou said and I actually think about itthe same way, set it and forget it.You've talked about you have ballast,you move around and you move the ballastto one side of the ship when you want tothe ship to lean into the wind to gofaster and you move it to the other sidewhen you don't want the ship to tipover. I think that's a great analogy.Think about allimportant companies in the portfolio thesame way. So 100% agree. I mean, thisthis chart is a bummer. What I would say

00:38:19is, you know, this data on IPOs, butwhat I would just say is this is areally unprecedented situation.>> Yes. We've never had an IPO this big.We've never had an IPO that's going togo into an index this quickly.Wesimply do not know how much sellingthere will be from investors.I would hazard a guess. I mean, I'm Idon't know.But Elon, I don't think he needsliquidity and I think he owns What doeshe own, Foxy?>> It's50%>> 50% of the company.>> way, he's locked up for 365 days or 366days. So, we know he's not selling,right?>> So, I just think it's an unprecedentedsituation and the right answer>> Yeah.>> is I don't know what's going to happenin the short term. And the right answerthat I would just, you know, encourageevery investor making their own decisionis to just think exactly [clears throat]the way you articulated it. We havethese different levers. We have thesedifferent variables. Think about eachone of them from first principles. Makeyour own decision.Do your own due diligence. Bethoughtful. But, there are a lot ofvariables here and that it is a littlefunny to me that uh you know, it was 100times trailing TTM revenue. Well, afterthe deals they signed, I think it's at39 times.>> That can change fast.>> So, they

00:39:34added $29 billion in a month.>> Yes. Now, it's>> [laughter]>> By the way, have you ever seen thathappen?>> Never. Never. And you know, it just goesto showfirst um Elon is not only a greatengineer.He and Gwen and the team are great atbusiness.>> And Brad,>> They they they understand what needs tobe done to raise the capital to get tothe next phase. They have a long-termmission in the business. And so, to me,again, what we saw in the course of thelast few weeks with cursor, what we sawwith these deals that they cut, I don'tknow that any of the mag seven couldhave moved that quickly to adjust thebusiness that they did. It'sexceptionally entrepreneurial at scale,which we very rarely see in businesses.Two other things I would just say>> you a hug, Brad?>> Two Two other things I I I I would justsay. Number one is people talk a lotabout the total amount of capital beingraised. If you add up the capital here,right, for Anthropic what they mayraise, what OpenAI may raise, what, youknow, SpaceX may raise, let's call it$250 billion.That's 1% of the Mag 7.Okay, it's 1% of the Mag 7.>> I Yeah. AndAnd we will as well. You know, like thatto me is like a bet on the future thatwe all believe in. And so, if I said,"Where are we out of consensus? What isour

00:40:49variant perception?" We actuallythink it's going to be bigger, faster,and we've thought that for a coupleyears. Um so, first, it's only 1% of theMag 7 market cap. And then youreferenced it, the amount of selling. UmI've got a chart we'll post here. Thisis, you know, the the dribble sharerelease for SpaceX shareholders. Youknow, so there's not a lot that can bereleased um up until after the firstearnings. This We saw this in theCerebras IPO. Um there's a version of ithere in this IPO. And so, again, I thinkthe banks have been thoughtful here,knowing that this is a very large IPO.And I'm not saying that won't tradedown. Like there's possibility, youknow, these things trade down. Butagain, for me, telescope out, is thereany company better positioned as a beton the future? I think what they'veshown over the course of last 5 weeks,they'rethey're they're probably number one. Butlet's move on.>> No, no, can I just say one thing aboutthe employees? I think another thingthat's unprecedented here is theemployees>> Yeah.>> and to a large degree the investors herehave had liquidity every 6 months.>> Exactly.>> the last 10 years.>> Yes.>> So, if you're a SpaceX employee orformer employee, and you wanted to sellyou've had whatever that is,close to 20 chances. And it is a matterof historical record that largeinvestors

00:42:05have been able to sell. SoI would think a lot of the people>> they've>> chosen to own it. Now, there's a newvaluation and we'll see what they do,but just this is utterly unprecedentedand we'll see.>> Yeah, I know. It's It's It's a greatpoint. We have in fact called thesecompanies quasi-public. Um you and Iboth know that SpaceX and I'd putAnthropic in in in this category aswell, Databricks in this category. Thesethings in many ways have been moreliquid over the course of the past 3years than some public biotech companieswe know. Right? And so there's acontinuum of liquidity here. We We treatit as a binary, private versus public,but it's really about this continuum.You know, let's keep going on models.You know, um Anthropic launched Fable 5,which you referenced um yesterday, whichis basically Mythos um with someclassifiers and safeguards um aroundcyber and biology, chemistry,um and distillation. When those thingsget triggered, it fails back [snorts] toOpus 4.8. Um you know, there was aCopart tweet about this yesterday. Hesaid, you know, it sold on all thebenchmarks, but what really makes itspecial is long-running tasks. Okay? Youretweeted our good friend, you know,Noam Brown. Um you know, ChatGPT 5.5also exhibited these capabilities.

00:43:21Um you know, it it led Noam, right, tosuggest that it's not very relevant todo these snapshot benchmarks anymore.Yeah, like the x-axis has to be time ortokens or compute because we can solvemost problems now if we just let thesefrontier models for a very long uh pointin time. So, Gavin, what is this newclass of model, right, Fable Fable 5,ChatGPT 5.5? What does it mean for therace in superintelligence? Who's up?Who's down? Who's still on the frontier?Um give us your thoughts.>> I mean, it's hard to say thatAnthropic's not up.>> Yeah.>> Like afterthe revenue numbers they've put up,after the Fable 5 release, and Mythos isevidently even better.But I just think that Gnome Brown postfrom yesterday, polynomial, is soprofound.And just the idea that we do not knowhow smart these models are.And we made>> Say more about that. Why don't we knowhow smart they are?>> Because nobody has run Mythos for a yearcontinuously. And we may never know howsmart each generation of models actuallyis or was, but because we don't havetime to appropriately evaluate theirintelligence

00:44:37before the next model comesout. I mean, this is a profoundstatement. And just just imagine, okay?So, I always say like when you thinkabout FSD,just imagine a human being who nevergets distracted, never gets tired, nevertalks on the phone in the car, neverdrinks and drives, never yells at theirkids, never has to go to the backseat togive their baby a bottle.And like of course you would think thatover time that is superior to humans whoare distracted.I don't know how long How long can youthink deeply about one topic, Brad?>> What do you Give me an hour. Give me an[laughter] hour. Give me an hour.>> A BIT.THAT MAKES me feel terrible cuz I thinkI can think deeply about one topiccontinuously before having a straythought enter my mindfor like maybe 5 minutes. Then I cancome back to that.Imagine if Albert Einsteinhad been able instead of, you know, andmaybe that maybe Imaybe he could think for 3 hours at atime. Clearly an exceptional intellect.But imagine Albert Einstein had justthought about fundamental physics24 hours a day.He doesn't have to eat, he doesn't haveto sleep, he doesn't have to relax, hedoesn't drink,>> never gets old,>> never gets old,>> never has diminished intelligence,>> and he thought for 1 year. I mean, wemight already,

00:45:52you know,>> have solved a lot of these intractableproblems.>> So, I just think that's an extraordinarythought. And just my takeaway washoweverbullish I was on compute before then,I'm just a lot more bullish.>> Right. Right. Right. So, so, so that isa, you know, we saw whenthat was probably what really unlockedOpus 4.6. It was the first reallylong-running model that could maintainthat context, maintain that memory, umsolve some of these longer-runningproblems, right? For us, the signal wasin January. We knew we felt like thatwas a big moment, but then when youstarted to see the revenue go up, weknew that lots of people were votingindependently, that that was a profoundmoment that they became much, much moreuseful. So,but one of the things that the consensusgoing into this year, right? So, the bigquestion going into this year was wasthe AI revenue going to show up? Were wegoing to get to these thresholds ofintelligence that caused enterprises andconsumers to use them more? And I thinkthe consensus at the time, at least onthis podcast, um the the the debate withwith my with with with Bill was theopen-source models, cheap tokens,

00:47:07werecatching up on the frontier, thatperhaps these models were beginning toasymptote, um that people wouldn'treally pay for premium tokens,and it seems to me that the evidence onthe field, 6 months into the year, isjust the opposite, right? That frontiertokens are capturing the vast majorityof all the revenues,and that in fact, if you believe in thelong-running capabilities and morecompute allows you to do that, they mayactually be extending their lead, right?On some of these models that were builton distillation. So, I just open it upto anyone around the table, what areyour thoughts on whether or not, youknow, have we challenged this thesisthat cheap open-source tokens are goingto always, you know, close the gap onthese frontier models, or are theyextending their leads?>> I I think this debate, like this samedebate has existed since the beginningofsince we started training these modelsto begin with, which was hey, we'realways kind of three, six months behindthe frontier. But empirically, like youcan just see all of the revenue hasactually just accrued at the frontier.And that I think that's because everytime we release the frontier,a whole new like slew of use cases>> Right.>> that that

00:48:22previously we could have nevertackled before, like coding.Um but also just, you know, you know,we've we've just been locked at our deskfor the last last day just, you know,hammering Claude because, you know, it'sjust fascinating the things that now wecan do with fable five that we couldjust couldn't do with opus 48 just a daybefore.>> So what are some of those things, man?I'm curious.>> So So I think it's really really good atmulti-agent orchestration. So they theyAnthropic released a um a blog postabout like different uh agent um sixdifferent agent like orchestrationpatterns that, you know, they've they'vetalked about. But really like once youstart being able to manage all theseagents, the harness and the model itselfis being arled with one another, they'reactually being,you know, fused closer and closertogether, but the model can understandthe, you know, the extent of your work.So, you know, one of the things, forinstance, is umI just threw in like seven of our modelsand just said, "Okay, like I want tocreate a master view of like my beliefsgiven all of these assumptions of allthese companies, TSMC capacity, like andthen and then produce me a report on allthis stuff." And you know, the the modelis able to reason through all of ourassumptions.

00:49:38Like actually, if youbelieve this>> Right. What are the contradictionsexactly?>> Yeah, it's it was fascinating. And andand you know, before we'd never do that,but but now, you know, I think we'rejust step one into multi-agentorchestration. We're going to do thiseven further and that's one example.I've also dumped all my all my notesinto it and it's reason across all mynotes from the last 3 years and said,you know, here are some of your ideasthat were consistent. Here are like, youknow, the sources that were actually thehighest signal to what actually playedout, you know, and then it is actuallyjust super fascinating what you could doand we've just blown through our blownthrough our limits.>> mean it's it's it's unlocking all this.I mean like they gave examples yesterdayand the release Anthropic did, you know,50 million line Ruby code base at Stripethat was, you know, refactored in a dayversus many weeks with many people. Youthink about where this is impactingbiology and life sciences just acrossthe spectrum.Umand to me it really gets back to thisfundamental point. Number one, if youbelieve this to be true aboutlong-running agents, then we're going toproduce and consume more tokens in thefuture as far as the eye can see. So theworld this gets me back to, you know,terrafab and space orbital and all thisbecausewe we we may in

00:50:53fact unlock realthresholds of intelligence, but we'regoing to have to let these horses runfor a long time in order to get there.Yeah, I would just say two things I twothings can be true.>> Mhm.>> The majority of economic value maycontinue to accrue to the frontier andman has it ever accrued to the frontierthus far and for sure the first 6 monthsof this year, but the majority of tokensconsumed in the world may be opensource.>> And they are>> today.>> Yes. I and I think that this currentstate is likely to persist. Harvey had agreat blog post that they put out on Xand they used and it's just amazing howeverything gets out out of date like in5 days, you know.But they used their own proprietarylegal data to do reinforcement learningand supervised fine-tuningwith Fireworks on an open source modeland then And used a router and a routerbeing something that picks which modelyou send which query to, and which modelyou use to check which model. And theygot better outcomes than Opus 4 either4.7 or 4.8 at a lower cost.>> Yes.>> And I think that is the future. And thereality isthey were still consuming a lot of Opus,but a majority of the tokens they wereprocessing probably were in their ownopen-source models.>> We heard the same

00:52:08thing. We did a We didumWe did an enterprise survey that we'llpost of 300 companies how which oneswere optimizing, so these are folks whoare kind of looking at model routing andsaying we're going to send certaintokens over here, which ones arethinking about optimizing, which onesaren't optimizing yet, and then what istheir expected use of frontier modeltokens, right? And they're all expectingto consume a lot more even thoughthey're already in the process ofoptimizing. Think of it in the in in thecontext of JP Morgan. If they're doingsome back of the house stuff, right, oncustomer service or whatever, they mayvery well use an open-source model. Now,I think they're loath to use Chineseopen-source models, so they're waitingon kind of US open-source models to, youknow, be able to really deliver the bangthat they need, but my hunch is forthese enterprises, a lot of that back ofthe house stuff will get rooted there.That will probably be a majority of thetokens, but I think the reallyhigh-value stuff, you know, coding as anexample, they don't want to writesecond-tier code. I think the vastmajority of that will continue to be onthe frontier.>> Um>> You don't need Albert Einstein to bookyou a trip. You don't need AlbertEinstein to do KYC.>> But but but this is the debate we had atliterally at this table two years ago.However, if you just look at the revenuecurves, right? What bill What what folksconcluded

00:53:24when they said that, theysaid, "Therefore, the frontier modelswill not accrue most of the revenue."And what we're seeing right now, it's90% of the>> That has been decisively wrong. Probablymore than 90%, and it may continue to bedecisively wrong. Frontier might be 90%of theeconomic value. Open-source>> might be 80% of tokens.Something that I think is very importanton open sourceis that you know, I think there's thisbelief that it's bearish for AI.It's actually it may be very bearish forthe frontier models. There's that bearcase you talked about. It's actuallyreally bullish for compute and hardwarebecause if the frontier models arecapturing less of the margin, thenyou're going to spend more on compute.So, the better open source does, thebetter it is for compute providers.>> And I yeah, I I will say it there is averyI would say between um spending time inthe heart of like the West, SiliconValley, and also spending time in Asia,there is like a verybig umlike a deep-seated belief in one versusthe other, which is like if you spend alot of time here, it's like all closedsource, cloud, every all traffic isgoing to go, you know, by way of thisdirection. And then you spend time inAsia, you know,the the overwhelming belief is thatwe're going to find the right model

00:54:40tothe right workload, and we're not goingto overspend.>> Right.>> And I think, you know, I would say Iwould saythe next year is probably going to bethe most indicative of which way thisfallsum becauseI think I think the reason whyuh closed source models have captured somuch of the value is because um themodels actually get the intention andactually carry through the work. Andthis is the first year where we actuallyhad agents that actually carried outuser intention from just answering achatbot request to actually producinguseful work.>> Right.>> Um now the the the level of thisintelligent has scaled so rapidly, andwe continue to push against like themost economically valuable tasks, whichare coding and finance and all theselike knowledge work tasks. But like forthe long tail of tasks, if open sourcecontinues to maintain a 6-month lag, wemight actually see a lot more opensource used foryou know, our everyday tasks that wemight actually>> basically Jensen's argument, right?Jensen's argument is you're going tohave model routingand we're just in a moment in time wherethe frontier models gain the advantagecan do long-running tasks that opensource models couldn't do it very welland so they're accruing

00:55:55all of thevalue, but as soon as the open sourcemodels can do the long-running tasks aswell, which is not far away that theytoo will grab a bunch a bunch of thisrevenue.>> Are you about to burst into reflection?>> I'm not.>> Okay. No, no, no, no are we, but I'mvery impressed by Misha and and the teamand what they're doing. I very much wanta frontier open source US lab to win. Weknow that, you know, I heard you sayrecently and I believe it to be trueNvidia any day that they really wantedto, right? They already have some greatopen source models. They couldabsolutely build a frontier open sourcemodel whenever they chose to do it andso it's not a question in my mind as towhether or not the US is going to have afrontier open source model. It's just aquestion about timing and then like atthat point in time is that you know,let's say let's assume they get theselong-running capabilities.Have the frontier labs now achievedsomething yet again that allows them tokeep keep the the stranglehold on therevenues?>> Yeah, and I just think it's if you'reWow, that's a cute ASIC you've builtthere. That is so cute. How would youlikeopen source to join the frontier?>> Right.>> How would you like that? How do you likethem apples? So, I mean I'm not surethat's the explicit calculation, but Ido

00:57:10think Jensen>> Say more. Just double click on that foreverybody at home.>> Yeah.>> If you were if they were to put an opensource model out there, how does thatimpact the ASIC landscape?>> Well,you might not have the revenueto fund [laughter]to fund that the revenue of the marginsto fund that ASIC.And I do think Nvidia is highly likelyto be the world's dominant provider ofopen source AI. And I do think Jensenwill bring open source,you know, right now it's whatever, 6months behind the frontier.>> Yeah.>> We might see itcreep closer and closer and closer.And I do think Jensen has a big businessdecision. I see this, you know, charthere, so let's, you know, chop it upabout Nvidia, as you say.But if all of his customersare going to compete with him,>> Yes.>> thenwhy not compete with his customers? Andwe have all these neo clouds.>> Right.>> So that's a cloud computing businessthat can compete with all these cloudcomputing businesses.He has his own models that are really,really good. Nematron 3 or 3.1 wasactually really, really cool from acomputer efficiency perspective. Andhe's always careful to release smallmodels so as to not tread on Anthropicand OpenAI,>> Right.>> Google's toes.But I do think that is a choice he ismaking.

00:58:26And just, you know, at if if theeconomics change,>> Right.>> I think Nvidia can join the frontier andbecome one of the world's largest cloudcomputing companies much faster thanpeople think.>> Interesting. Interesting. Clark, walk usthrough thisthis chart.>> Yeah, so so I think one of the takeawaysfrom spending time in Taiwan was therethere is certainly a lot of excitementaround the next wave of ASICs.Um, but I think I think it's like a veryclear moment now where Nvidiait used to be an argument of Nvidiaversus ASICs one or the other and, youknow, total domination one or the other.Now I think it increasingly every yearevery every one assumed that Nvidia wasgoing to lose share dramatically on arevenue scale, on a gigawatt scale, on aunit scale. And actually, if youactually look at the last few years, youknow, they've actually maintained theirshare very, very handsomely.Um, actually,um, if you accounted for the fact thatAnthropicwas not really using Nvidia. Theyprobably actually gain share againstif not for in 25 26. So, I think I thinkwhat was very interesting though was anew class ofaccelerators or ASICs.MediaTek

00:59:42with theirwith their new V8Tversus, you know, Broadcom's V8I forTPUsactually was a big topic of discussion.And, you know, I I think for ASICsthe argument now is that more and morewill look custom to the actual workloadand that is like one vector that peopleare moving in versus Nvidia now ishas kind of shown itself as the thepredominant provider of compute toa lot of the world and for, you know,internal internal workloads,perhaps they will go more and morecustom and more and more down the stack.And I I remember just, you know, 1 yearago when it was kind of a Broadcom orNvidia battle. It seems there's a lotmore nuance now to, you know, what typeof accelerators will fit which workloadsand fit which customers and fit whichbusiness models. Umand yeah,I thought I thought that was aa new topic.>> It's actually>> New realization though, I think we allkind of shared this view for a longtime.>> Yeah, I was just shocked. I mean, I'mI'm out here. I did a board meeting withone of our companiesand just, you know, their biggest onething they emphasized iswe thought the world

01:00:57would be have beconsuming less Nvidia than it is and ifanything, Nvidia is accelerating andthey just continue to out execute theircompetitors. And I think a lot of peopleare indexing to thisOpenAI gigawatt and you know, Nvidia has10.Broadcom has 10.Umwho has six?AMD AMD has six and they have warrants.And then Cerebras hasour shared portfolio companyhas a gigawatt.And I just that is what's on paper.>> Right.>> What actually gets deployed, let's see.I will be very surprised if you knowthat 10 out of 27, what's that math?Let's see who's best at math. Whatpercentage market share is that?>> 30% yeah.>> Yeah. I'll be very surprised if that iswhere they land. I think that is anextremely unlikely outcome.And especially as long as we're in awatt constrained world, if you can getmore tokens per watt, which is literallyrevenue with Nvidiathan a lot of alternatives just if youbuild your factory with another chipyou may save some money, but you'regoing to have less revenue and themargins may be lower and that's a pointthat Jensen keeps hammering and I thinkis a really important. And by the way,credit where credit is due

01:02:13the most important the most one of themost surprising things to me in thisASIC landscape>> I'd say Meta and Microsoft have beenprobably disappointing.>> Yes.>> You know who made a good ASIC?>> Yes.>> Well, I know you know.>> Yes.>> Jalapeno>> Yeah, exactly.>> from Open AI. They made a great chip.>> Yes.>> Now, unfortunately needs to run at amuch lower temperature than the NvidiaGPUs, which means you need to spend moremoney on cooling and that consumes morepower. They made a great chip.>> Well, we can I mean I think the questionthere and the question for everybody isgoing to be is that the highest and bestuse of your time? Right? Like I youknow, I tend to think that the frontiercompanies like there's this belief thatthey got to be vertical verticallyintegrated. But if you believe like I dothat the race to super intelligenceparticularly as we get these recursiveloops working may be over in the nexttwo to three years, then I think focusfocus focus focus. You exist to buildthe best intelligence in the world andto deliver the best intelligence in theworld and you that means you have tohave all the revenue. Because if youwant to build out the compute that'sgoing to be required to continue to pushthe frontier, you have to have therevenue in order to support it. So Ithink you know, subject to the focusquestion, I think they certainly did.This all brings me back to kind of areality check, though.Umyou know, we just got done talking abouttest time compute, inference

01:03:28timecompute, long-running agents. This isreally the thing that's unlocked therevenue this year. Um it all pushes usin the direction of more CapEx. Googlejust raised $80 billion,right? We've now taken the Mag 5 or Mag7 free cash flow, you know, downdramatically, 80% um from just a fewyears ago. Um and Morgan Stanley, you'vegot this chart in front of you, up totheir 2027 CapEx forecast from 950billion to 1.1 trillion. I mean, we weretalking about this with Jensen. That washis forecast 2 years ago. You know,obviously, this doesn't even includeSpaceX, CoreWeave, etc. So, I think thenumber on 2027 is likely closer to 1.5trillion.And if we compare this to the totalincremental inference revenue, so thething that the market gets worriedabout, you know, back to my Sam Altmanpodcast, you know, in October of lastyear, can we really afford to spend 1.5trillion of CapEx a year if we're onlygenerating X amount in inferencerevenue? The thing I think that lit thefuse this year was Anthropic showed upin a major way with revenue, right? Andso, we have, you know, the AI labrevenue everybody combined at around$300 billion next year, right?

01:04:43So, can'tyou know, and go roll that out to 2027uh or that is 2027, 300 billion. So,we're spending 1.5 trillion of CapEx on300 billion of inference revenue. Doesthat math math for you? And what wouldcause you, you know, to to get morenervous again about our ability tocontinue to make these investments?Because the second we get nervous aboutit, the entire semi complex is going tocome down a lot. Well, what do you thinkthe gross margins are on that 300billion? Yeah, let's call it 50%.>> I I would guess they're probably alittle bit higher than that. I might say60 or 70.But, I mean, that math starts to math,and what I would just say is I thinkthat 300 billion is low, man.>> Yeah. Yeah.>> I just think it's low.>> From your mouth to God's>> Yeah, yeah, exactly. I think I think weend this year well over 200 billion ininference revenue, well over.And so, I think the math really maths,and I do think we have to give uh Jensen>> Yeah, our friend.>> some credit because he said some thingsthat seemed outlandish.>> Right.>> And he was conservative. He was low. Hesaid a trillion 2 years ago.And I mean, he was really low.>> Right.>> And so, like, let's give the guy somecredit and think about what he is sayingright now.>> For sure, for sure. And and listen,I would say consistently,

01:06:00Elon's been taking the over.Sundar's been taking the over.Sam, Dario, you know, Dario did thepodcast with Dwarkesh when he wastalking about country geniuses in thedata center. He said that will be hereby 2028. He said revenues will go intothe low hundreds of billions by 2028.So, let's call that, you know, 3 400billion of revenue by 2028. And he saidthat a while ago now, so he may even berevising up his number. And he said it'shard for me to see that there won't betrillions of dollars in revenue before2030. And if you're on that revenuetrajectory, if we're on a trajectory to200 by the end of this year, let's callit 4 or 500 by next year, and a path totrillion plus by 2029, then the mathmaths.>> And we got to keep in mind that half ofthe spending is there toyou know, for training, maybe a littleless than half. What is it, Foxy?>> It's probably that depends on the lab,but it's I would say it's increasinglyless than half.>> Yes.>> Okay. So, we'll call it 35% is spendingthat's not revenue generating, but it'sgoing to kind of make the next model.So, I think the math maths.>> Right.>> And there's still this prisoner'sdilemma where if you opted out, that maybe an existential decision.>> And I think like coming into this year,going back to this kind of whatnarratives were violated, you know,

01:07:15I think into this year everyone expectedtoken pricing, uh the price of compute,it's all deflationary. And it will bekind of a smooth line deflationary overtime. But, I think this year what we'veseen is the opposite. And you know, it'sall comes back to supply-demand. Thedemand side of the equation seems to befar outstripping the supply. Right? AndI thinkyou look at the deals signed by SpaceXand others, the monetization rates perwatt are increasing.Umandlook, that is on a a pretty nascentsmall base of users, right? Like Alex atWell Rock, he has this great umway to frame it.Less than 0.2% of people on Earth areactually using AI in an agentic way.>> Right.>> Right? Like I'm not a technical person,but I'm consuming 500 CPU cores in a VMinstance, five GPUs 24/7.>> Yeah.>> I mean, if you draw that out to anymeaningful percentage of the population,I mean, we're going to be in, you know,this kind of shortage environment maybefor some time. So,I think that is all positive for thisROI question.>> Man, foxy, 100 to one CPU to GPU ratio.>> [laughter]>> Kind of agentic workflow.>> He said of course.

01:08:31>> [laughter]>> Five.>> Five, yes.>> Yeah.>> I'm being smart with my phone.>> Good, good, good. Excellent.>> I I will say also that ratio of 300 to1. You know, call it 1.2, 1.5.Um there there is also a rate that nowphysically we can only expandhow much we can produce and how much wecan actually increase that spend by,whereas we're seeing the opposite rightnow on the on the the willingness to payfor these tokens. And actually like whenthe willingness to pay for these whenthe monetization per gigawatt isactually increasing from, you know, callit like 20 20 billion um in the in thebest best of cases for at the beginningof the year to now like 30 to evenpushing 40>> per gigawatt>> per gigawatt.Um all of that is is a very heavy fixedcost base, but all of that is like puremargin flow through now. And you'reactually, you know, as we scale like thewillingness to pay for for all of thisand and now the all of this stipulatedby like, you know, everything we'retalking about of like how much is opensource versus not and all of thesedifferent flows, but really like aswe're climbing this curve, you know, thethe the revenue is might actuallyoutstrip

01:09:46our fixed cost base by bysignificant amount. And I think that'swhy all the labs are pushing, you know,the the the gas to the pedals becausethey all they all see like within if wecontinue this curve within like 3 years,you know, we're just going to be soshort on all the computer>> It's a great I'm sorry, but I mean Ilike it's a great point. Like if youthought you were getting a when you madethese decisions>> Yes.>> in November of 2025,you thought you were getting a certainreturn.>> Yeah.>> You may be getting triple that returntoday.>> At Tropic, no way no way did they thinkthey were going to be anywhere close tobreak even.>> Yeah.>> Right? And and and and in this part ofthe curve, and the reason like I I Icalled it accidental profitability that,you know, people have been talking aboutthat because they want to spend a lotmore money on computer. They just had ahard time doing it. Now maybe withSpaceX, you know, they could take someof those dollars and and and go spendthem other places. But that to me is,you know, a a fundamental change. Um thefirst argument against the frontier labswas they'll never generate revenue.Okay? And then we that got blown up.Then it was like even if they generaterevenue, it'll be really shitty grossmargins, and they'll never be able toget make money. And then kind of thatthat's blown up. And and you know, Ithink now, you know, people are fallingback and they're saying, "Well, they'reovercharging. This is token

01:11:01maxi." Mygood friend, you know, Chamath has saidthere's no ROI on any of this spend.It's all this token maxi. My bestevidence for why we all know, of course,when somebody puts on this much spendlike at Altimeter, we're not optimallyspending every single dollar. But, thequestion is, why are millions ofindependent businesses, small, medium,and large, why are millions of consumersall choosing to do the same thing?They're not dumb. These are, you know,rational economic actors that are allsimultaneously saying, "I want to dothis because it makes my life better. Itmakes my business better, etc." To me,that is the best evidence as to why Ithink this revenue can continue.>> Yeah. And Clark, I think like the pointyou made is dead on cuz I mean, you wantto own asset-heavy businesses ininflationary environments, and tokenpricing is going up, and supply anddemand is tightening, so totally agree.>> Um you know, as we begin touh find our way to the exit ramp and and[laughter] and wrap here, one of thethings I you know, you and I've beendoing this for a long time, Gavin, acouple decades. Um you may even sketchlonger than me, even though I'm a littlebit older than you. Umyou know, we have uh I always like to doa market check, because I find a lot oftime that analysts come on these things,and they talk their, you know, talktheir

01:12:16book, and you know, there are alot of people who listen to thesethings, retail investors and others.It's just kind of like, what do wereally think? And so, I alwayscharacterize as kind of small, medium,and large. Like, what am I doing? Do Ihave small exposure on? Do I have mediumexposure on? Do I have large exposureon? You know, and if you look at what'shappened in the markets, semis rippedthis year. I mean, like uh you've beendoing this a long time. I don't I'venever seen it before, right? I've neverseen, you know, the doubles and thetriples across the board like we saw.But, there's been huge dispersion,right, in the market. Internet's down16%,uh software's down 8% on the year. Youknow, spy and and Nasdaq are up, butreally up because of their componentsthat are related to AI and compute. Andso, the market itself has kind ofstruggled. Meanwhile, if you were in thestuff that we were invested in, we'veall done pretty well. I think you know,I've said it a couple times. I think ifthe Anthropic revenue had not shown upthis year, because that was the overhangon the market, I think the whole marketcould be down this year. Right? Um butthat showed up. You know, we just hadthese huge months in in in April andMay. Um for us, you know, because pricescame up so much, because I have someworry about, you know, geopolitics, themacro backdrop with, you know, with withwhat's going on with inflation

01:13:31in theshort run, and just like, you know,needing a little consolidation in thismarket to answer some of thesequestions, because now expectations arehigher. You know, we dialed back fromwhat I would call large for Altimeter tosomething kind of like medium small. Umagain, it's never all or nothing for us.It's like, what is the risk-reward at agiven price?Um and so, we think this is a, you know,maybe going to be a period ofconsolidation on way to much higherhighs. Um curious just how you run thebook, how you think about it like aportfolio manager.>> Very similarly, man. I always thinkstocks, the markets, I imagine them asrunners. Okay?And like in '22,that runner had gone downhill. It had alot of energy, man.Yeah, it was painful. It wasn't fun. Umbut coming out of that, there was a lotof kind of pent-up upside in the market.And you know, the market, particularlylast 2 months, it has run up a verysteep hill.And a lot of companies, semiconductorcompanies in particular, you know,ironically, you know, Nvidia andBroadcom, they they have been laggards.>> Totally.>> And so, but a lot of these, like I dosee a lot on X about finding the nextbottleneck. I think that was the lastgame. That game is over.You've had a lot of stocks that forgetclimbing a mountain or a hill. They've

01:14:47gone straight up a cliff, okay?>> Yes. They're tired. They need to rest.And we'll see, do they just rest at thetop of that cliff they climbed? Do theyhang out on the in their harness for awhile?We've seen some.Or do they need to go downhill for abit? We'll see, but I'm thinking verysimilarly to you.But it is and I think there's, you know,the market is seasonal. I think there'srealreal concerns around inflation andrates.>> What was CPI this morning?>> uh 4.2. I think we added core came in atlike 0.2 versus 0.3, so a little bitbetter.Um but you know, clearly we're we'reabove four again.And um and and there's short-termpressure on, you know, core PCE, etc.Um and we have some unknown unknowns,but the market, I mean, if I had toldyou the fact pattern for this year, thatwe're going to be in a war with Iran,that, you know, oil was going to be at100 bucks, that CPI was going to becreeping back up, that internet wasgoing to be down 15%. Software is goingto be down 8%. You would have said, "Iwant nothing to do with that market,right?" And here we are. The market'sdone pretty good in the stuff that wetraffic in because the worldunderestimated AI revenues andunderestimated the amount of computethat was going to be needed.>> It's odd to say you know,

01:16:02we're headinginto a seasonally weak period with allof these fears. AI has actually beenseasonal for the last three summers.Token consumption is kind of plateaued,slowed down, and that's cuz you know,college kids are big AI consumers andthey don't use as much AI, you know,hopefully they're all using it to learnand not cheat.But that may happen. It may not happenbecause of generative AI.>> is building swarms of agents, building aSpaceX model. He's going to the SpaceXIPO with me at the exchange on Friday,but I he had to build an AI model usingAI agents. He had to build a model, aDCF before we go to the exchange. He ismesmerized. He is absolutely and it'sextraordinary what he's doing.>> So he's one kid who's not easy to lesscomputer>> [laughter]>> or something.>> He's burning it. He's burning it.>> Yeah, but you know, if token consumptionplateaus, if open source takes someshare, there's a Silicon data index thathas showed, which is an index of kind ofconsumption and pricing. I think theremay have been a little bit of a shiftover the last 2 weeks to open sourcetokens that are cheaper. Like peoplelooking at that data as bearish or notunderstanding it. But nonetheless, likeI just think there's reasons, you know,to look around, be careful, bethoughtful. I always assume a bullet iscoming for me. Head on [laughter] aswivel. It's the bullet you don't seethat gets you. So, I'm trying to spin asfast as I can.

01:17:18But yeah, it's the market may need totake a breather. But man, when I thinkabout what Noam Brown saidand when I see the capabilities ofFable,it's just hard for me to get toobearish.>> I mean, like to meum and we got two, I think, of the mostextraordinary guys of, you know, thenext generation, you know, sitting inthe room. We have at Altimeter, we havedeep admiration for the work that youguys do. I always appreciate when yousend me a note about the work that we doand we publish. Um but for the guys whoare newer to the business, they mightthink this is the way that it kind ofalways was, right? And like this line,the steepening of the line of creativedestruction, the steepening of the lineof, you know, scale advantages. Um Ialways believed it was to it was goingto be true. I never thought it would betrue at this rate. I went back lastnight. In the last 7 years, we've added1 trillion of revenue to the Mag 7in the last 7 years, okay? To get to atrillion, to get to the first trillionof, you know, took over 20 years. In thelast 7, we had another tr- trillion andthat added 17 trillion in market cap.

01:18:33That trillion dollars, okay?I The forecast now that we're going toadd another trillion of revenue in justthree companies SpaceX Anthropic andopen AI over the next four to fiveyears.Okay, like not seven companies threecompanies and in half the time right andso I would say that you know, we aregoing to have bumps in the road. I knowthat it's going to be like this butwe're going to higher highs because thesize of the prize. This is going totransform five ten 15% of global GDP.There is no doubt in my mind and 10% ofglobal GDP is 10 trillion dollars. It'san exciting future to be a part of it'sfun to do it with you guys. I thinkwe're going to have to do our work to dothe things to make sure America wins andthat we evolve the social contract keepeverybody you know lift the floor takeeverybody with us on this ridebut it's a it's a it's a really excitingtime to be doing what we're doing it'sfun to be doing it with you guys.>> Yeah, I just want to say Brad thanks forhaving us and thank you for what you'vedone with the Trump accounts. I actuallythink it's super important for Americafor the world to give people an equitystake at a very young age. They theywill see it compound over theirlifetimes. This is a great thing you'vedone for the world. So thank you. I'decho all your comments like deepadmiration

01:19:48for you your team gratitudefor the collegiality and friendshipbetween our firms. I know Clark and Foxythey hang out like all the time.>> That's a people think that you know andthere are people in our business whodon't want to share anything. Our viewis like we open source itbut there are very few people who weactually call and ask their opinionbecause there are very few people who dothe thousands of hours of work that wedoyou know that are adding to that and youdo it and we appreciate that and you doas well Gavin we appreciate that. Sowith that love fest let's call it awrap.Thanks for being here.>> Thank you.

01:20:47>> Mhm.