Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI

00:00:00All right, everybody. Welcome back toyour favorite podcast. It's the All-Inpodcast. It's the summer. It's August 6.Having a hard time getting a quorum hereon the podcast, but David Friedberg ishere. David Freeberg is back. Our Sultanscience. How you doing, brother?>> Great to be with you.>> It's great to be with you. And everybodyloves when Brad Gersonner is here. He'syour Bruce Wayne. If markets are yourgame, he brings that namaste to yourpayday. Yes.>> Buy his glasses at discount and he'llget you one of those fancy Trumpaccounts. All right. Welcome back to theprogram, Brad.>> I love it. I love it. You bring therhymes back.>> I bring a little intro back. We've beentrying Chimath is on the road right now.Chimath is on the road, but we will geta field report from Chimath and uh I Icall Daniel somehow Sax is going to behere, but you know how he is. He'salways late because you know get a phonecall from very important people, but hewill break in at some point. Oh, wait. Isee in the text here. Oh, there he is.He made it.>> Hey guys,>> you made it.>> How do you like my beautiful summerjallet?>> Uh, it's incredible. It fits perfectly.You look warm.>> I don't know how Jimoth does this.

00:01:14>> Let your winners ride.>> Rainman Davidand>> we open source it to the fans andthey've just gone crazy with it. Loveyou. We can walkin.>> Well, here's the report, everybody. Aseverybody knows, Chimath is on the road.He uh Oh, here he is. He um this is aphoto sax. He went he went to check hisdata center progress. Uh I think that'sin Colorado or Nevada where he wasbuilding a data center.>> I think that's on Dune.>> Oh, it's on Dune. Yes. Dune 4. Oh, yes.Here he is admiring himself. Oh, look.Here's Nat. You know when a meme hasreached its peak? when your wife startsdunking on you. There it is. And here weare. This was at the Christmas party, Ithink. Oh, you were on CNBC with AndrewRoss working. Oh, there you go.>> I wasn't sure if it was my Twitter feedthat was just selecting into it, but ithit everyone, right? This was a viralthing.>> This has hit everything. All right,listen. We got a lot to get to. Enoughwith the shenanigans and small talk.Google had uh two major shakeups to itsAI staff on Wednesday. Deis Hassabishas moved to chair of Deep Mind andchief scientist at Google. Reportsdescribe this as Demis stepping down orbeing kicked

00:02:29upstairs. Uh we'll get intothat, but Google framed it as apromotion uh and says he was steppingup. Here's Axios's quote explaining theshakeup. Quote, "Google's Gemini 3.5 Prois months behind with some companysources telling Axios that it's in partdue to low morale." Interesting. Severaltop researchers, including Gemini'sco-lead, have left the firm forcompeting AI labs. Chef Dean, plus threeother AI superstars, are leaving Googleto start a company called DiscoveryLoop. Dean is a legend. Freeberg, uh,and I think you worked with him atGoogle, one of the world's great AIengineers. He was employee number 30,joined in 1999, and has worked there forwhat I understand continuously for 27years. Discovery Loop's going to befocused on deep scientific breakthroughsin AI. Google Share is down 4% on thenews of Dean leaving. So 200 billion inlost market cap if you want to correlatethose two things. Freeberg, this is youralma mada. What are your thoughts here?Is this creative destruction? Maybethese people weren't delivering and theywanted fresh blood or is this just thesiren call of doing a startup in an ageof unlimited capital for AI andunlimited opportunity just being toomuch for the OGs at Google to not takeadvantage of. Maybe it's the thirdbucket, which is

00:03:44if you're the board andthe management, you're having a debateabout how to best deploy capital. Googlehas made a commitment to deploy $200billion in capex this year in AIinfrastructure data center buildout.Because of the capex and accelerateddepreciation,making an investment in AI compute inthe US right now is hugely taxadvantaged. And because of the extremedemand for compute, it's a prettyobvious kind of ROIC model, return oninvested capital. So if you make thissort of an investment, you havesignificant demand for that computeinfrastructure. You're very good atrunning the compute infrastructure. Thatcapital can deliver massive profitreturns for you with very highconfidence in some forecasted period.Building the most advanced frontier labdriven model also takes tens of billionsof dollars of capital. And the questionreally is can you deliver the profitsfrom the model? And in a world whereopen-source is becoming so good and openweights models are catching up soquickly and all the frontier labs arecatching up to each other so quickly,does it really make as much sense todeploy tens of billions of dollarsagainst

00:04:59building a model? And I thinkthat the scientists that we're seeingtransition out are the scientists thathave been at the core of modeldevelopment of making these frontiermodels. And they were certainly firstout the gate.>> You can look at some of the earlyinterviews with Jeff Dean from a coupleyears ago where they actually had a chatGPT equivalent internally a year beforechat GPT came out from OpenAI. Googlechose not to release it for fear ofcannibalizing search and so on. That'swhen Sergey stepped in and there wasthis whole kind of revitalization. Butas time has gone on and as everyone hascompeted on models, as we've talkedabout many times on the show, I thinkit's pretty obvious that it is very hardto get the same sort of return oncapital invested in model development asit is in capital invested on computeinfrastructure and being model agnostic.What Google has is probably one of thegreatest install enterprise bases in theworld for compute. So they have the mostenterprise customers. They have the mostconsumers. And in both cases, they don'tnecessarily need to have the best modelto make an incredible business. They canbe model agnostic. They can work withanthropic. They can work with OpenAI.They can work with SpaceX. They have asignificant ownership stake in SpaceXand in Anthropic and they can work withall the open weights models. They canhost them all.

00:06:15So now if you're one ofthe great computer scientists, you'reDemis, you're Jeff Dean, you're thiswhole crew, and you're inside at Googleand they're allocating capital not toyour models, not to the things thatyou're most interested in, but they'reallocating capital to infrastructure anddata centers and supporting the broadecosystem of models, you start to say,well, given the fact that I can go downthe road and visit Brad Gersonner and acouple other people and raise a couplebillion dollars at a multi-billiondollar pre- money with a PowerPoint deckbecause I'm the greatest in the world atdoing this, that might be a better path.for me and I think that that's themoment. So I the way I would frame it iscapex is high alpha low beta in datacenter infrastructure that capital andmodel development theoretically could behigh alpha but it's very high beta it'sa very risky way to deploy capital so soif I'm the board I'm the management I'mdeploying more capital in computinginfrastructure less capital into modeldevelopment that's what I think's goingon>> Brad what's your take on this>> I think David nails it I mean listen thesame thing's going on at Microsoft rightis out this week saying, you know,citing Morgan Stanley's report andsaying they're seeing over a 30% returnon invested capital in tokens as aservice, right? So, in theinfrastructure business. So, I thinkDavid's exactly right. Those are such

00:07:30good businesses, right? You you you youdeploy capital, everybody's running itfrom you, but the scientists who want tobe involved in super intelligence, whowant to cure cancer, who want to be onthe frontier of these models, right?They're sitting there dealing with thischannel conflict at Google because youknow uh uh Google cloud wants all of thecompute in order to rent it out toAnthropic and those those building thefrontier models internally want thatcompute in order to compete withAnthropic. So you have this inherentchannel conflict uh between thosewanting to build the models. I thinkDavid said it really well. Um and Ithink that's a that that's a bigchallenge for them. It looks like it'sbeing resolved in favor of being more ofan infrastructure company. So where doesyou know telescope out for a secondSpaceX also reported this week they alsohave channel conflict. They're rentingout their compute to anthropic at thesame time they're trying to build theirown model with Grock and Cursor. Youhave that channel conflict at Google.You have that channel conflict atMicrosoft although I don't even reallysee them pushing the frontier anymore interms of models. Meta's talking aboutgetting into the infrastructure as aservice game. And then at Anthropic andOpen AI, you don't have any of thatchannel conflict. They say we're not inthe infrastructure business. We're onlyin the model business. So,

00:08:45I think it'sa, you know, uh a a clarifying view aswe look forward that we may in fact nothave those companies on the frontier ofof model development if all these peopleleave. By the way, thanks to the lawpassed on capex depreciation, if youassume a 26% corporate tax rate, everydollar you deploy in capex, because youget to write it off in this year, you'rebasically getting 26% off. You know,that's money you get right back. Yeah.>> Yeah. Hey, uh Sax, let me have youcomment on uh this as well. Poly market,which companies will have the number oneAI model by the end of this year onDecember 31st? Um, now of course in thelast time they did this anthropic one,so they're not on the list. They're thewinner, but who will have it uh goingforward? OpenAI 32%, Google 20%, Alibaba14, and then you got Moonshot, XAI,Meta, Bite Dance, all at about 10%. SoSax, your thoughts here on what's thebetter business? Is the better businessbeing in the language model, frontiermodel, or is that getting quicklycommoditized? and really you want to bein the token sale business or is thatalso going to be a commodity and youjust need to be on the applicationlayer?>> Here's what I think is going on in termsof the the market structure is when Isaw this Google news,

00:10:00my reaction wasand then there were two because likeBrad was saying, we used to have fivemajor companies in the hunt to be theleading frontier lab, the leadingfrontier model just a year ago. Nowwe're really down to just anthropic andopen AI. So the market for frontierintelligence has become a duopoly. NowElon is still on the hunt. I'm sureGoogle would say they're still on thehunt, but like Brad is saying, they mayhave contradictory incentives therebecause they can actually do quite welljust with their compute. So I think thatthe market for frontier intelligence hasbecome a duopoly. I think it's a verypowerful duopoly. I don't think it'sbeing commoditized. I think that whatwe're evolving to is a two-tier marketstructure where there's a market forfrontier intelligence and there's amarket for let's call it kind ofcommodity or lagging intelligencewhatever you want to call it that's 6 to12 months behind. There is a market forthose tokens those models but thereality is you can't charge anything forthe weights. You can charge for thecompute you can charge for the inferencethat you're providing. You can chargefor essentially consulting services tohelp put the whole thing together.But if you're not at the frontier, youcan't charge for the model layer itself.

00:11:16If you are at the frontier, you cancharge a premium. And that's whereAnthropic and Open AI are. And I thinkthe proof for this is just you look atthe growth rates of these companies. Thelatest we heard is Enthropic is now over80 billion of ARR. Started the year at10. It had forecast 100 billion as exitARR for the year. And most people saidthat that would be impossible toachieve. Now it looks like they're goingto do it with a couple of months tospare. So their estimates are going up.I mean 110 120 or higher for end of yearARR. Open AAI seeing acceleration. So Ithink what you're seeing now is a veryclear bifurcation in the market. You'vegot a frontier model duopoly that cancharge a premium. I think of it likeApple. You know Apple's competingagainst Android. It's open source.Android actually has more users in theworld, but all the monetization goes toApple because people are willing to payfor the premium experience. I think in asimilar way, people are willing to pay apremium for true frontier intelligenceif it's really at the leading edge. Butif you're not the leading edge, there'sa huge market for that, too. But it'shighly commoditized. People are justwilling to pay you for the compute.>> So, I mean, that's what I see happeningright

00:12:31now.>> Yeah.>> Jason, what do you think? Jason, what doyou think? thing. Uh, well, if you lookat Google Cloud, they posted 82%year-over-year revenue growth, which issomething we've never seen uh, in thehistory of these cloud providers. ElonMusk and XAI just had the SpaceXearnings. We're going to get into that,but they also had massive uptick intheir Elon web services, as I've dubbedit. And if you look at Google, I stillthink Google will be the number one uhAI company because they have so manypeople using AI inside of their productsalready. They have five products nowwith over three billion monthly userseach. Android search, Gmail, Chrome,YouTube all have over three billion. Ifyou've used any of these productsrecently, uh they are becoming AI firstproducts. YouTube especially, butobviously Chrome and Gmail, you'reseeing um tools pop up there for AI. Andthen Freeberg, you kind of alluded tothis. They have 13 products total withover a billion. And that now includesGemini. In Q2, Gemini had over 950monthly active users, triplingyear-over-year. They will be the numberone AI company in terms of consumerusage by far, I think, this year. Thatdoesn't mean that

00:13:46the um frontier modelsare not great businesses. They obviouslyare. But I have been using exclusivelynon-frontier models. And for 95% of thejobs I'm doing, Sachs,it's good enough. And I just postedabout this, you know, um and Elon and Igot into it a little bit here, and Ithink you referenced this in our groupchat. Uh I I tweeted just the other day,the difference between the open sourcemodels I'm using and Frontier isnegligible already. I believe that to bea true statement for the work I'm doing.And he said, Elon responded back to me,it's actually a world of difference. Youknow, if you're doing something otherthan making a copy of a video game oryou have incredible speed needs, theFrontier models are not necessaryanymore. They're just not necessary. Thepeople using the Frontier models aredoing it because their company set it upand they it's too hard to implement opensource right now, but it's going to geteasier and easier to implement it. So,I'm still going with open source andGemini being the leaders in this.>> Look, I think it it's true for your usecases that let's say the cheapercommodity intelligence that middle ofthe market is good enough. Look, anAndroid phone would be good enough forme. I could get by on a cheap Androidphone. You know what? I still pay apremium for

00:15:01this because I use it somuch. So if you're a business that let'ssay you are a hedge fund and you're in ahighly competitive industry, you don'twant to take the chance that you're notgetting the best intelligence to poweryour models, you know, and there's a lotof industries like that where thecompetitive dynamics will drive you topay for the best intelligence. There'salso situations, this goes back to theblog post that Decagon posted, which isif you're looking for use cases, youalso want to use the true frontierbecause again, when you're dealing withimmature use cases, you don't know wherethe value is going to be and you'researching for opportunity to use AI. Youjust want to use the best because again,the return on finding those use cases isgoing to be so much greater than thesmall premium you're paying at the tokenlevel. So I think there's a lot ofexamples like that when you know the usecase is immature where you're in acompetitive industry where you're justdeploying AI you want the convenience ofthe full>> so why not go frontier model>> you know unless your employees are doingsomething stupid like you create aleaderboard and they're token maxing Idon't think the cost is that great andagain the benefit that you're getting ishuge so a lot of people are just likegive me the best I'm willing to pay apremium for the best>> I'll take a slightly different take Ithink that it's not necessarily do youtake the best model or the open

00:16:16sourcemodel. I think that there's a blendthat's happening. At least that's what Isee. For example, we'll use open- sourceopen weights for a vast majority ofsimple workflow applications. But whenit comes to specialized applicationswhere we really need to have highquality model proficiency, for example,in life sciences, in genomics modeling,I am going to go for the premium model.If I'm working at a media company andI'm trying to do AI rendering of video,I'm going to use Gemini's model thatdoes video. It is the best model or Soraor whatever the best model is for thatparticular application. So, I think theidea that there's kind of a model thatyou pick for everything, I think is thefalse assumption. On the consumer side,it is likely the case that the consumersare not going to be using someopenweight model because they can pay 2040 bucks a month and get chat GPT orGemini or Claude and be very happypaying 40 bucks a month and they'llbasically be able to minimize their costto run that for consumers. Forenterprise, I think the enterprise isgoing to be very active in selecting ablend of models that are going to makethe most sense. Very cheap openweightmodel for simple workflow applications,individual employees spinning up an app,whatever. and then more complex modelsfor those really key workflow tasks

00:17:32andthen specialized models. And I will sayit is way too early to count Gemini outon building incredible specializedmodels. They have the best video data.They have the best life sciences data.They've been working on this for farlonger than anthropic or open AI in thelife sciences side. They're very wellahead on that front. I mean Demis isstill going to be running isomeorphiclabs. So when it comes to thesespecialized models, verticalizedspecialized models like video, lifesciences, protein folding, I think theseare the things where you're really goingto see Gemini shine and then everyenterprise is going to have a mixture.But hey, if you can be the cloud serviceprovider with that mixture of models,which is what Google GCP can now be, I'mgoing to sign up for working with GCPversus working just with Anthropic. Oneof the things this has created isdownward pressure on the pricing. We sawOpenAI and Claude uh do massive pricecuts uh for tokens. So they arereacting. They're not taking it sittingdown and the orchestration between thesemodels is being built into a lot ofharnesses inside of enterprises. Sowhat's your take on the downwardpressure on token pricing or is thisjust great for consumers andenterprises? Because>> listen, we're got massive competition.>> That's the thing. America's winning.This is exactly what you want. We havemassively competitive market. We haveChinese open source, domestic opensource,

00:18:47Frontier National Labs that aredoing what they're doing. We havedownward pressure on pricing. You know,David re referenced a duopoly. You know,I think it's hard to call it a duopolywhen you're, you know, only a few yearsinto this and you have giants likeAmazon, Microsoft, and Google. I dothink he's right. I do think they'veemerged, you know, as the pure plays.Their revenues would suggest thatthey're, you know, they're gaining shareof wallet. But there are two points Iwant to make here because I thinkthey're non-conensus views that werespoken this week. One was Elon'sresponse to you, Jason. Right. Over thelast two weeks, everybody's been sayingthat the Chinese have caught up, thatopen source tokens have caught up inintelligence, that they're much cheaper,etc. And Elon comes out and says, "Notso fast. We're entering the singularityand the frontier models are way furtherahead than people think." I believe thatto be true. I think for your use casethey're very similar but I don't thinkthat's the most sophisticated use casethat people are trying to train on andtrying to experience. And then Jensencame out this week and said closedmodels are actually cheaper you know ifyou don't have to build it for yourselfif you don't have to uh you know thetraining costs and a lot of expertise tofine-tune and maintain and guard railand keep it safe. So he's basicallymaking the argument that not only are

00:20:02the the the frontier models furtherahead, but that the cost differentialbetween the two is not what everybody'smaking it out to see to to be, which Ithink explains why they continue to runaway with it on the revenue side of theequation. Um, but I think we havehealthy competition. I you're right JCLyou know for the vast majority of usecases I think token consumption is goingup for the open source guys while shareof economics is going up for thefrontier labs I think that's what wewant to see yeah and it's just Androidversus iPhone all over again oneplatform makes the profit one gets themajority of users at least globally inusage all right let's talk SpaceX hereuh they had their first earnings reportas a public company shares dropped 13%uh I I think because people were alittle concerned about the surging AIcapex. It's down 30% since going publicin June, but it's now trading at, itseems to have settled in at a 1.4trillion valuation. Went publicobviously above 2 trillion. Q2 resultswere uh spectacular is the only way toput it. 7.8 billion in revenue, up 92%year-over-year. Let that sink in. Uh and67% quarter over quarter. AI revenue.Elon Web Services more than tripledquarter over

00:21:17quarter to $2.6 billion.That's not cursor. That hasn't closedyet. Uh but that's going to be one ofthe great purchases in history. This isfrom Elon Web Services uh renting outcompute specifically to Anthropic andGoogle from the Colossus uh collectionof servers. But capex was up 18.4billion in the quarter. That's 6xyear-over-year. Obviously, you can dothe math there for a run rate of about75 billion dollar. I'll stop there andget your reaction, Brad, to the SpaceXIPO. I know you've been tracking thisand commented on it heavily.>> I mean, listen, I think that one, first,let's start off. $1.4 trillion of valuecreation for this company isextraordinary. So, the fact that frompeak to trough, it's down 40 or 50% fromthe IPO. We had that chart out a fewweeks ago. Remember that within 6 monthsof the IPO almost all these tech stocksare down 50% peak to trough. We see itagain here with SpaceX. I thought it wasa really solid quarter. I thought hisguides were pretty extraordinary. 100billion in ARR by the end of the yearand he pulled forward the $1 trilliontarget in ARR by a year from 2031 to2030. Now to just put that inperspective, Morgan Stanley's 2030revenue estimate is 325 billion

00:22:32which isalso extraordinary. Remember, thiscompany did 18 billion in revenue lastyear. So whether you're taking MorganStanley's numbers or Elon's numbers,clearly the market is not pricing thatin. At 2 trillion, we were pricing aheada couple years. I think now it's, youknow, the the value reflects kind ofwhere we are. The market has questionsabout a few things. Here's what theyare. Number one, on the rental business,the rental of compute business, herented out a huge block of compute toAnthropic. It's the question that we'vebeen talking about here. Are you goingto use the compute to build your ownfrontier model or are you going to rentit out? And if you rent it out, are yougoing to be able to find those peoplewho have the capital to offtake thatcompute? He's talking enormous numbers,10 to 20 gigs, and people are wonderinghow they're going to be able to financethat. And remember, those businesses,the GPU rental businesses tend to tradeat very low multiples. Look atCoreweave, etc. On the frontier modelbusiness, I think this is the sleeper. Ithink he said on the call that Grocktripled tokens in the month of July.That doesn't include cursor. Curser wasalready on a path to go from 3 billionto 10 billion by the end of the year.Curser plus Grock could be at 10 to 20billion by the end of the year.

00:23:48Thatwould be an extraordinarily valuableasset going to trade at a much highermultiple than the data center business.And then of course we haven't eventalked about Starlink and what he'sgoing to do uh you know I think going torun the table on mobile. So this is thenormal consolidation. We have funds likeuh across Silicon Valley that aredistributing their shares. The stock hastraded down a bit. Nothing surprising tome here. Now it's all about execution. Ithink the most important thing to watch,the two most important things to watchare number one, how do the Grock andCursor revenues end the year?>> And number two, um you know, thetraction they get on, um you know,continuing to replace traditional mobilecarriers with Starling. The distributionstarted I think today or yesterday. Igot my first distribution from a fundI'm in. I'm in a couple of funds thatare in SpaceX. Seems like everybody's inthat and that will obviously createdownward pressure if you are amongst thepeople who want to cash out and been init for a long time but I'm holding thesefor my grandkids. Saxs your take onthese spectacularI guess is the only way to describe themresults coming from a vertical thatwasn't part of SpaceX's business butnine months ago. Yeah, look, I thoughtit was a very bullish earnings call. Iwas a little bit surprised that thestock went down after the earnings call

00:25:03because not only was it a beaten raise,but also I think Elon spoke to a lot oftheir plans. The only thing I would addto to what Brad said was aroundStarship, Elon basically said, we allsaw it, right, that the Starship testflight was successful. The Starship'sfloating in the ocean, the heat shieldworked. That's going to enable moreflights of Starship now at a moreaccelerated rate. that paves the way forthe V3 satellite which enables much morebandwidth for the Starlink network whichthen powers the whole direct to cellplay. So you had that piece of it. Imean just the whole telecom aspectseemed very on track and they're verybullish about that and then you've gotthe whole AI data center play. Now onthe data centers I think what they saidis that they expected to go from 1.4gawatts of compute to about two by theend of the year. And Elon said that thespot pricer computes in the $30 to $50per watt range. So, you know, you do themath, a gigawatt is a billion watt. So30 to$50 per watt means 30 to50 billionper gawatt. And I think they're at thehigh end of that range right now. Sowhen Elon says, look, we're going to endthe year at 100 billion of ARR, all youhave to believe is that they're at 2gawatt of

00:26:18compute running for $50 a wattto hit that. That doesn't includeStarlink or the launch business or theGrock cursor piece or any of thesethings. So I think that's why they're>> multiple ways to win is what you'resaying, Sax. There's multiple ways towin with the stock.>> I think Starlink's just an unbelievablejuggernaut cash machine. If you look atthe financials, their segment reports,space, connectivity, and AI. And on theconnectivity side, the Starlink side, itthey generated $2.6 billion in adjustedEBA. You can kind of approximate that tobe kind of operating cash flow. Spacewas kind of, you know, negative 200million. So call it break even and AIwas plus 1.1 billion. But AI, to Brad'spoint, it's unclear whether the pricingthey're getting on compute rental todayis temporary and at a premium because ofthe lack of compute available in themarket today. And people that needcomput are paying Elon a premium forthat comput. So I think there's aquestion mark where that goes. But theconnectivity piece on Starlink, 4.3billion in the quarter and 2.6 billionin adjusted EBIT. He's got 12 millionsubscribers. That's doubledyear-over-year. $66 arpoo per month uhwhat people

00:27:33are paying per month. And hegrew 20% quarter over quarter. So if youextrapolate this out, he's pretty closeto being at a 24 million subscriber runrate on this multiple. Assuming thisenterprise stuff which is like airlinesand other things scale which they seemto be scaling with the consumerbusiness, Starlink alone could begenerating on the order of $40 billionof revenue topline with a huge amount ofthat flowing to free cash. That could bea $30 billion free cash flow within theyear. That alone provides the cash flowto fund much of what what Elon's doing.And if you just put a 30x multiple onthat, which I think you can becausethese subscription businesses are veryhigh renewal rate, very low capac, Ithink you could probably get a 30x juston the Starlink business. The Starlinkbusiness alone could be a trilliondollar market cap within 2 years, within18 months, let's say. That I think fundsall of the rest of this as kind ofscience projects and upside. So, I'mkind of making a bull case. It's crazyto me how well the Starlink businessperforms and you can see it in AT&T andVerizon, Huenet, Viaat. I mean, thesecompanies have been decimated. I used tohave a Huenet satellite dish on mySonoma County ranch in order to getinternet. That's

00:28:49what we had to use. Itwas like, you know, 200 bucks a month orsomethingorbit, right? And they take forever to>> terrible service. And that market gotdecimated by Starlink. And if helaunches the handset thing, thatsubscriber growth is going to go rightnow. He's adding 2 million subscriberson the consumer side a quarter. Youcould see that going to four to 5million a quarter. You could actuallysee an acceleration in the consumermillion mobile sub 400 million mobilesubs just in the United States.>> I think you can make you can make thebull case on Starlink alone and then therest of it is like, hey, is Elon goingto do well with investing the excesscapital that's spitting off of Starlink?How's Elon going to do with that money?Well, I don't know who else I give it toto like, you know, do what he's doingwith Starship and with AI compute andthe terra fab.>> Oh my god, this is a science fiction uhcounty text>> how the US gets off of this dependencywith Taiwan and China fromsemiconductors. If Elon takes this onhis shoulders and he delivers what he'sshowing as a vision here today, this isgoing to be the greatest semiconductorfabrication site on planet Earth. Well,you know, I I would say something, youknow, David, to your point, you know,how many CEOs or founders would justtake that Starlink business, which is

00:30:04such an exceptional business,trillion dollar business going to twotrillion, and they would not take any ofthese other risks. They would not doterapab. They would not try to build outthe data center. They would not try tobuild their own model. That's highlyrisky but highly important investmentsthat are being made. I mean, it isheroic and important that we have thislevel of I just think unbridledenthusiasm for innovation uh on thefrontier that Elon's doing. And I wishwe saw more CEOs, more public companieswilling to take this level of risk. Wejust got done talking about, you know,some CEOs maybe that were taking lessrisk because the safe bet was was easierto make. Elon refuses just to take thesafe bet. He's taking all the dollarsfrom this thing where he has anextraordinary business and plowing themback into these things that arecritically important to the UnitedStates. And by the way, Brad, such agood point because if you look at otherCEOs and other management teams, they'regetting in on this. They're starting torealize that buying back your shares,giving dividends is not as important asbetting on the future. Door Dash gottaken to the woodshed because they'reinvesting too much in capex. Obviously,Google got smacked with their capexspend. So, that keeps happening over andover again. And just on the headwindsthat

00:31:19SpaceX is going to face, thearguments that I think will turn out tobe wrong, but they're valid to talkabout here are, hey, is this demand fortokens and compute going to keep up? Ordoes onrem and desktops and open sourcemodels getting smaller, better, doesthat actually mute at some point demand?I don't think it does. I don't knowthere's an upper uh I don't know there'san upper bound for on demandintelligence. The second one obviouslyis Starlink is for people who are in arural neighborhood. If you've gotVerizon fiber to your building orSpectrum, you're not putting nor can youput a Starlink on your building. So, thepiece there that's going to be um uhexplained probably in the next year ortwo is every single Tesla sold is goingto have Starlink in it when they getthat merger done. What that means isyou're going to have Wi-Fi networks uhconnecting any phone to any Tesla, sayall those robo taxis out there. You'llbe able to connect also directly withthe next generation of Starlink. So yourphone will be able to direct if it's gotclear line of sight. It's going to beable to connect to any Tesla on theroad, which there are many that allfuture ones will have a Starlink builtinto them. So those are super promising.And then finally, you know, there's beena lot of speculation about thevaluation.

00:32:34Brad, you brought it up. Ithink at liquidity when people wereasking you and I heard you talk about itheyprivate companies venture capital we area voting mechanism and then when it goespublic it becomes a weighing mechanismand sometimes you'll have this moment intime uh where there's hand ringing aboutthose valuations and the hand ringingpeaked in the last quarter you had 160times uh price to sales ratio for Teslawhen it first came out 160 times rightyou take their this 2 or3 trillionmarket cap and you put it against asmaller revenue number. Well, if youlook at the revenue number increasing,now we're down to a 45 times price tosales ratio. So, some kind of uh balanceis occurring here. Yeah, Brad, betweenthese private and public markets as wellas the increase in revenue.>> Yeah, I I I mean, honestly, I think thisis all super healthy. I think the SpaceXIPO was extraordinary. I think theconsolidation here is perfectlypredictable. And now you have a companyat 1.4 trillion that I think if you takea three or a fouryear view, you can seeyourself tripling your money in thisbusiness at a very reasonable valuationon the Morgan Stanley numbers or on theElon numbers or whatever. But that'salways been the bet. Do you

00:33:49believe thatElon is the greatest innovator and agreat allocator of capital? But theprice of entry matters, right? When youget carried away on day one of an IPOand you buy this thing over twotrillion, you got to know that this isgoing to happen. And I was on CNBC theday of the IPO and I said I would wantto own this company but I'm not suretoday is the day I would buy thecompany. Right? And so I you know>> entry price matters. I mean fundamental>> but but let me give you another one. Youknow like we've talked about theanthropic IPO uh or a lot of people havetalked about it later this year. I heara lot of people saying 1.5 or$2trillion. David just talked earlier thatit's going to be run rating over ahundred billion maybe by the end of theyear. That's like 10 to 15 timesrevenue. That is not that much for acompany that just grew 10x and isrumored to be profitable in Q2. And so Ilook at the market, the consolidation wesaw in the month of July, you know, weput in the Leopold bottom hopefully inJuly that, you know, a lot of semistocks were down, you know, 30. Hey,hey, listen. The guy's doing great. He'sapparently still up 80% for the year.Just made another big privateinvestment. I I I I think he's done anextraordinarily good job building a firmin a short period of time, but themarket did panic around that. Yeah. Umas as as he

00:35:04had to cover, I think all ofthat is really good. So, as I look aheadmarching to these IPOs later in the yearon the back of the SpaceX IPO, I thinkwe're in in in really good shape. Um youknow, particularly if these revenuescontinue a pace. You know what Elon'sreally good at is just building stufflike factory specific physical physicalsites. That is such a core advantage inthis world where everyone's competingfor data centers and fabs. The softwarelayer needs hardware in the physicalworld in order to deliver their softwareservices. And there is no one betterthan Elon at actually doing that. Lookat how gigafactories have been stood uparound the world. This is his corecompetency. So Brad, like when you putElon up against Adaio and a SAM and evenan Alphabet which has 27 years of doingthis, I I mean man, Elon's got a coreadvantage if this is what this worldcomes down to.>> He says something like that on the callwhere he said, "Look, putting up datacenters is nothing compared to thedifficulty of putting up a rocket,right? It's like, you know, creatingdata centers is not rocket science." Sothey take some of those hardwareexpertise that they have from SpaceX andthey put them into data centers andthat's why they've been able to standup, you know, more data centers or orbigger data

00:36:19centers faster than all thecompetitors. A couple points there. I isit clear why Starship is so important toStarlink? Okay, let me just explain thisquickly. So basically SpaceX hasdeveloped a new V3 satellite that has10x the bandwidth of its V2 satellite.So currently the Starlink network ispowered by V2 satellites. They deploythem on the Falcon 9 rocket and theylaunch about 27 satellites per launchand that adds about 2.6 terabits persecond of total network capacity.Starship deploys 60 of these V3satellites per launch. That would add 60terabytes per second of total networkcapacity per launch. So over 20 timesmore capacity per launch. That's thepower of it. So if they get Starshipworking and by the way the last test,not only did it prove that the heatshield worked, my understanding is theyactually launched or rather theydeployed 20 V3 satellites as a test andthey were able to make connection withthose satellites and prove that itworked. They even had cameras on them.The reason we were able to see theStarship

00:37:35was because they're like,"YOLO, let's put some cameras, HDcameras on them.">> Right now, I think those satellitesbasically it was just a test and theythey burn up.>> They they burned up. So, I think thenext big milestone here will be whenthey launch Starship with, let's say, 60of these V3 satellites, put them in thecorrect orbit, make connection withthem, add the bandwidth to the network.That's going to be a big milestone. Butyou play this out to its logicalconclusion and the bandwidth availableto the Starlink network goes up 10x oreventually 100x times and that's whenthey can do all the interesting thingslike direct to cellular. There were someinteresting hints that Gwen Shaw talkedabout about with ground stations aboutwhat they could potentially do there andI think>> and they might buy T-Mobile or somethinglike that sack. it's easily within theirrange of purchases>> and and I think Elon mentioned somethingabout potentially the the Starlinknetwork could eventually handle roughlyhalf of internet traffic. So I mean thisthis thing could get so much bigger thanjust 12 million subscribers to yourpoint Freeberg. But look I I want toactually talk about the data centers fora second Brad. I do have a couple ofquestions about this. So Elon mentionedthat okay we're going to be at 2 gawattby the end of the year. He said that wewill be at five to 10 next year. Closer

00:38:50to 10 than five. So let's just sayeight. Okay. So I'm just making that up,but it's in their range.>> So let's just say that's an ad of 6gawatt. So they go from two to eight.Okay. To me, there's two questionsthere. One is how do you know that thespot price is going to stay where it is?You know, can it stay at $50 per watt?How do we know? How do we track that?How much risk is there around that? Igot the sense on the call that Elonthinks that number is going up becausethe market is memory constrained rightnow. I think he mentioned that we mightsee a 20% increase in memory productionnext year but the demand is going up200% plus. So the market is constrainedby whatever the bottleneck is at thattime. Right now the bottleneck ismemory. So where do you see the spotprice going? How do we know how muchrisk is there around that? And then theother question I would have is if you gofrom two to eight gigawatts that youhave net is six. We know that a gigawattpower data center is you know 50 billionof capex.>> So six incremental gigawatts of computewould be 300 billion of capex next yearassuming they build that right. I meanthey have optionality around that I'msure.

00:40:05So how do you finance that? Youknow what's the most non-dilutive way?They said their payback is a year orless. I'm sure that's tied to the spotprice. So, you only have to finance itfor a year. And the question, do youthink Nvidia gives them that financingor how will this play out? I guess is myquestion.>> It's a great framing, David. First, it's$50 billion per gigawatt to buildminimum. Okay. So, you're $300 billion.So, in order to finance that, it seemsto me you either have to go into themarket and borrow the money or you haveto do a dilutive equity raise. neitherof which they want to do. Um or you getNvidia to backs stop it. Um whichthey've indicated that they're going todo more of. But the problem there isNvidia shareholders don't want them backstopping unlimited because the fear inthe world is that that spot price atsome point right may go against you andwhen it does the payback period changes.Nobody thinks that the payback period isgoing to be one year even though thespot price is suggesting that it is thattoday. Right? Just a few years ago,people thought you would get paid or notfew years ago, a few months ago, peoplethought you'd get payback over fouryears. So, you basically spend 50, youthen earn 10 to 15 per year. You getpayback over four to

00:41:20five years and thenhopefully you get the sixth year, whichreally takes you up well above 20% interms of your returns. Um, right now theshortage is so acute and the willingnessto pay from the front Frontier Labs isso high because they all recognizethey're on the verge of some massivebreakthroughs that they're willing topay three, four, 5x market pricing inorder to get at scale compute. Andthat's what happened with the Anthropicdeal uh with SpaceX. I think Anthropicwould buy a lot more of that today ifthey could. Same with OpenAI. You'resaying Brad they would be willing tooverpay by a factor of up to 5x.>> Well, that's the 50 that you know that'sthe $50 per watt that David wasreferencing. They would be willing topay this 30 to 50 if they could getatscale compute that would give them acompetitive advantage over the otherpeople in the market. And remember,there aren't a lot of people who havethe offtake revenue that can afford tobuy compute at this scale, right? Itwasn't the Chinese open source companiesthat were buying, you know, SpaceX'sexcess compute or building the 10gawatt, you know, plant in in Ohio.That's open AI and anthropic. So, thevast majority of the offtake commitmentsare coming from Anthropic, Open AAI, andNVIDIA,

00:42:35right? When you hear about thehyperscalers building all of this, youknow, this compute out, they're buildingit out to sell to the people that we uh,you know, that we just mentioned. SoDavid, Netnet, if he builds 6 gigawattsnext year, and by the way, probably onlyElon um, you know, can actually stand upthat much in that time frame. LikeJensen said to me on the pod, he's like,"Nobody comes close. Microsoft doesn'tcome close. You know, Google doesn'tcome close in terms of standing it up inthat time frame." Um, I think that he'sgoing to have a challenge, you know,getting all of the componentry, right? Iknow he can stand it up, but can he getthe memory? Can he get the chips? Can heget the land powered shell all in time?I think the offtake is there, right? Butto put it in perspective, this yearAnthropic and Open AAI combined, theirstarting total compute was like 5 GW.So he's talking about incrementallyadding more than they had as combinedcompanies, right?>> That's not that much of an increase whenAnthropic is growing 10x year-over-yearand OpenAI is maybe at what 4x or maybehigher now. So>> the demand exists in the world. Thedemand exists in the world today. Ithink it will exist in the world forwell, you know, the next 12 to 24months. But there is a wall of worry inthe market.

00:43:51The reason we saw thepullback in July is Kimmy scared peopleinto thinking, oh my gosh, they're goingto undercut the Frontier's revenues. Andif they undercut the Frontier Labsrevenues, who the hell is going to payfor all this compute? That's why you sawa 40% trade down in the core weaves ofthe world and you know the the all ofthe the semiconductor stocks andsemiconductor related AI>> in some ways the fact that there's adiscussion going on that this next 10gigawatts is going to cost you knowSachs$500 billion dollars and you askedthe question where does that come from asecondary offering does Nvidia put it ontheir books do they create SPVS offtheir books like some people are doingyou know the fact that we're having thisconversation everybody's is aware of it.The market has been educated on it meansI think people will be able to change inreal time if it doesn't come to pass orif it slows down which I suspect thiscannot keep up at this pace you knowmore than another two years or so. So aas our good friend Bill Gurley likes toremind us, he's like, I can't believe>> that we're all just taking in stridethis level of seller financing, right?He would call it circular revenues,right? But the market has gottencomfortable with this. And remember,

00:45:06like we saw in July, if there is a scareabout demand, the whole sector tradesdown.>> Yeah.>> Everything will trade down, you know,together because that's just theleverage that you're pumping into thesystem. you're effectively backstoppingpeople's ability to build ahead of theirrevenue. So, it becomes much moreviolent if you ever see demand slippage.Um, you know, famous last words, I don'tsee it today over the course of the next12 to 18 months. Um, but you know, youhave these unknown unknown moments thatcertainly causes people to be fearful.>> Credit spreads are blowing, you know,have continued to to stay wide on thesedeals. So, there is fear in the marketabout them.>> All right, everybody. The fifth annualif it's September, you know it's timefor the All-In Summit. The fifth annualis happening. Yes, that's right. UhDavid Freeberg's been at work and wehave an allstar allstarlist of people joining us. Jensen Wang,founder and CEO of Nvidia. If you careabout where AI is headed, you won't wantto miss this conversation.>> The best the Oracle>> Satia Nadella, CEO of Microsoft, fan ofthe pod will be coming on for the secondtime. Jared Isaacman from NASA, the one,the only Brad Gersonner and Bill Gurley,BG2 coming back. SpaceX's Gwen Shotwell,my guy Jake Paul,

00:46:21Nick Shirley. Lot ofincredible people coming. Martin Shreleymaybe is even coming. He's that's goingto be fun. Go to the allinssummit.com toapply today. Allin.com ortheallinssummit.com.Any of those will get you there. Andwe're taking over Universal Studiosagain. We'll have our own privateplayground. Dave Friedberg, great job onthe summit. Casino night, too. I heardit's going to be a big casino night.>> Biggest yet. And the concert to beannounced who will be performing at theconcert, but it is going to beincredible. So, I'll just say one of thethings about the summit, we've hadpeople come to the summit from over 60countries. It's really incredible tomeet all these people, entrepreneurs,investors, people that are just reallyinterested in the topics that we talkabout. We try and have the world's mostimportant conversations, but it's reallythis amazing community experience.That's what brings folks back. So, wetry and invest more and more every yearin making it an amazing experience, notjust cool content on a stage, which Ithink is what a lot of these other showsreally deliver, but it's like how do youactually come and have a have anexperience for a couple days. It's goingto be awesome. So,>> it really is those three things that wefocus on. One, you're going to learnsomething, right? You got these greatpeople on stage. You're going to learnsomething from them. You're going tomeet new people. You're going tonetwork. And then you're going to havethese great experiences.

00:47:36It's thetrifecta. Folks, you're excited, Brad?You excited to be back? What? What arethe dates? What are the dates again?>> Look at your calendar. You're speeding.>> September 13th through 15th in LA. Thiscouldn't be better dates for the summit.I mean, we're we're going to be within60 days of an election, midtermelection. We're going to be within 30days of an IPO, you know, potentially ofanthropic. I mean, like, it it's goingto be heated. The SAS apocalypse, notthe Saxp apocalypse. This is the SASapocalypse is, I guess, winding its wayout. Uh, the indigestion might beclearing. Air Table just got acquiredfor less than it raised. It's aprofitable SAS company, a great product,$480 million, half a billion dollars inannual revenue, growing 20% a year,respectable if it was a public companywith almost a billion dollar in cash hasbeen sold. It's been sold for $1.28billion, about 10% of its peakvaluation, which was 11.7 billion in2021. Now, they did have a bunch ofcash. showed me include the cashposition sale was 2.25 billion. Theywere acquired by a firm called bendingspoons. This is an Italian company,Milanbased company. They buy challengedbut you know interesting businesses.AOL's legacy business, Evernote,

00:48:52Eventbrite, Vimeo, Meetup.com and theyjust went public last month. Shares uhthat is uh bending spoons went publiclast month. Shares sh 15% on the airtable news sacks. When we look at this,this was a company that had done a lotof things right, had a massive amount ofcash in their war chest, but rumors weremaybe the founders were a littleexhausted. Maybe some of the investorswere exhausted who bought in at a highlevel. What can we take away from thistransaction in bending spoons? Are theythe buyer of last resort now?>> Well, I think they're creating a greatbusiness for themselves because I thinkthis will end up being a fairlyprofitable acquisition for them. Let mejust add a piece to this which is AirTable spun out its AI agent business uhwhich is known as hyper agent into aseparate independent company prior tothis acquisition. So I think what'sgoing on here is that the founders andtalent of the company they said look wedon't want to have to make this legacyproduct work that's basically a privateequity play. I'll explain what thatmeans in a second. We want to focus onthe new thing, the AI company. That'swhere the big value creation is going tobe in the future or the potential forit. So essentially, the talent is goingto focus on the venture play and thenthey're selling the

00:50:07private equity playto Bending Spoons. Now, why do I thinkthis could be a good acquisition forBending Spoons. I think there was areally interesting data point that I sawin the commentary on this, which is only30%of Air Table sales team was makingquota. They had a 30% sales attainmentnumber and that told me a lot about thisbusiness. Okay, what it told me is, andI'm reading between the lines here, butthis was a company that had a successfulPLG motion, in other words, organicgrowth, productled growth, and they weregrowing about 20% a year. But that wasnot good enough for its board. You know,these are investors, some of whominvested at an 11 billion peakvaluation. So, they're looking for aventure type outcome. So, what happens?The board pressures the founders to dosomething that frankly is unnatural forthem, which is they say, "Look, youshould bolt on a traditional salesledmotion here to get the growth upfaster." Does that work? No. Theyprobably get a little bit of growth outof it, but they only get 30% attainment.So they've got hundreds and hundreds ofsales reps here trying to push on astring and it's not making it growfaster. So now what's the opportunityfor

00:51:22the acquirer here? Bending spoonscan go in here and do what Elon did atTwitter. Eliminate 85 90% of the coststructure. Don't do this salesledmotion. Just go back to your productledgrowth roots. You'll probably keep mostof that 20% growth and it'll be a veryprofitable company. you'll be able to>> 80% profitable probably right>> probably. I mean>> 400 million to the bottom line pays forthe acquisition in a couple years.>> People are saying they're only going togenerate 30% ebidom margin. I think likeyou're saying it could be 80 90%. Idon't think you need to keep most ofthis business or most of the coststructure associated with this business.Um Air Table is a company that has itsfans. Um I think they will probablystick with it and you know you'll you'llbe generating I don't know you couldprobably generate 300 million of Ibida ayear or 400 million uh while growing youknow 10 to 20%. So that's a play forbending spoons>> and the venture investors here Sachsthey're happy to get their money backand move on to the next thing. It's abit of a push for them, you know, interms of at the blackjack table ratherthan they've got to go 10x just to catchup and then they would have to go 10xagain to make their LPs happy. It's notgoing to happen.>> I think the question is if BendingSpoons

00:52:37can basically take this businessthat's not making money and probablygenerate 400 million a year of EBA andpay for the acquisition in just threeyears.>> Amazing.>> Why isn't that something that thecompany could do on its own? And I thinkthat's the structural problem is I thinkit's very hard for both VCs who are onthe board and the founders to shift intoprivate equity mode. Why? Becausethey're going to have to demolition whatthey've built, right? They've got allthis loyalty to the team. They don'twant to think about how do I eliminate80 90% of the cost structure. It's justnot what they do. I mean, what whatfounders want to do and and the outcomethat the board members are going for isa venturebacked outcome. And I thinkthey could have done this. They could dowhat Bending Spoons does, but>> they're not built for it, Sax.>> They're not built for it. And moreover,the structure of the cap table is allwrong because they're sitting behindthis giant liquidation preference. Allthese investors who have to get paidback who invested at this 11 billionvaluation and and you know, all the wayup>> the the incentives are broken, Brad. Andyou you yourself at your firm, Alimter,you were pretty frisky in this period.You made a lot of bets. So uh I don'tknow if air table was one of them uh butyou made some sass bets there.

00:53:52Some ofthem were at high valuations. How areyou looking back at that time period?Any lessons that you take going forward?Multiples of revenue can compress veryquickly. Right? It works great when thecompany's growing greater than 50%. Butremember it's just a heristic. It's justa very rough estimate used almostexclusively in Silicon Valley you know.So people are saying oh my god thisthing sold for two times revenue. Butwhen you actually look at it on a lookthrough basis, probably sold for maybe30 times free cash flow. I don't thinkit's easy to get it to 400 million inibita. I think if it was, the boardwould have done that. I'm on, you know,we're involved in some of thesecompanies. Once they slow down, thecompany morale goes to hell. Turnoveramong your customers, uh, you know,begins to spike. Um, it starts to feedon itself. So I sucks to go to workevery day. What do you need to keep?What do you need to think about? Verytricky. I don't know the core productand what's happening in terms ofturnover in the core product, David, butmy hunch is that the core product hasstarted uh to really fizzle as theadvances in the core product has sloweddown. You're seeing a bunch of churn outof it on the product side and now peopleare saying listen it's almost impossiblefor a software company today to keep anydecent sales people

00:55:07to keep anydifferent decent product developmentpeople because they all want to go workon AI.>> Agreed. But you don't need them for thisproduct.>> I mean the market the market's beingefficient. I mean look this is where Ithink bending spoons has an advantagethat the company's board and founderswouldn't have which is they already havean infrastructure right they have a coreteam at bending spoons that's managingnow I don't know dozens of theseproperties and so they can plug this in.I think AI in a way makes their jobeasier because in the past the reasonwhy you couldn't eliminate like all ofthe talent the infrastructure is becauseyou needed the institutional memory. Youneeded people who knew the codebase. NowAI can learn the codebase instantly.>> That's an interesting insight.>> And so yeah>> maintaining is easier with AI.>> I think maintenance mode becomes wayeasier with AI because you don't needthe historical knowledge anymore. The AIcan go in and sort of reconstitute thatthat historical knowledge. Let me getyou in here freeird uh if I may. when welook at the lessons from peak zerp andSAS and then we look at you know thismoment in time this surging AI marketany parallels that we might find here uhor lessons uh between the two>> between Zerp and AI era>> the ZER SAS era

00:56:22we had a lot of veryhigh valuations a lot of enthusiasm alot of suspending disbelief we're herein the AI era we just talked about youknow the price of compute and all thesecompanies being at a 100x uh price tosales ratio. Any parallels here or not?It's a kind of a softball question foryou.>> No, this is a very different paradigm.Uh the AI capex buildout and modeltraining which is where the predominanceof the capital is flowing is not aboutsome high multiple on revenue which iswhere capital was flowing into SAS. It'slike oh you get a 20x multiple turn adollar into 20 that's great let's do itall day long. this is a very differentstructure and strategy and capital umallocation process. So I don't thinkthat I I would look at them as beinglinked.>> It was a softball question to be honest.I was letting you hit it out of thepark.>> Look, I mean obviously SAS companieswere overvalued during the Zer era fortwo reasons. One is that we hadartificially low interest rates. So wehad a kind of a a speculative assetsuper bubble. But the other is thatpeople were treating these things likeguaranteed annuities. and actuallygrowing annuities, they'd look at it andsee, oh, 120% net dollar retention, sothis thing will just grow 20%year-over-year

00:57:38forever as a base case,right? And they were then priced thatway. But what we've seen with AI isobviouslythere's disruption and you can't to Bradsaid, I'm sure they're seeing elevatedchurn right now and it's not an annuity.Things can change. So obviously nowthese things are trading at a muchgreater discount. All of that beingsaid, let me just say I don't think youcan extrapolate to the entire SAS spacebased on this one company, Air Table. Ithink there's some things about AirTable that make it very different than,I don't know, let's say a Salesforce ora Workday is, you know, Air Table wasalways a little bit of a quirky product.I remember at the peak hype for thiscompany, people were saying like, oh,this is like a new Excel or a new GoogleSheets. It's>> new Microsoft Office. Yeah.>> Yeah. It was basically a spreadsheet forwords. That's how people were wereviewing it as this like this new kind ofspreadsheet for for words as opposed tonumbers. And it never achieved that kindof promise. It never achieved that kindof ubiquity. People understand how touse spreadsheets. Everyone uses them.Air Table never got to that point. Mostpeople still don't know what Air Tableis. Again, it had its dedicated fans,but it was a hard product to explain topeople. When do you use it?>> It had a cult following.

00:58:53It had a cultfollowing but but it never it neverachieved that sort of level ofacceptance. It was neverself-explanatory in terms of why youshould use it, what the use cases are,they never were able to kind of get themarketing right because of that.>> And to be honest, if you look at claudeco-workagents, those things are now doing whatair tableable did. So>> it never carved out, I think, a nichewhere it was super clear when you werealways supposed to use error table andand really it was part of thishodgepodge of of this grab bag youshould you could say of no code tools.This is the category it was put in andno code has to be the most impacted themost disrupted area of SAS right nowbecause I mean what is clawed codereally good at? I mean that's theultimate no lovable claw codeperplexing. The thing with yeah thething with air tableable or retoolthings like this is it's true you didn'tneed to be a coder to use them but youhad to learn how to use air tableableyou had to learn how to use retool allthese it was kind of these you knowalternative programming languages in away and you just don't need to learn anyof that anymore I mean you use claudeand you just tell it what you want it tocreate and so you know if you do want tocreate a some sort of new dashboard somesort of I don't know like a verbalspreadsheet or

01:00:08whatever you just tellClaude what you want. You don't havethis learning curve. Look, all of SAS isbeing impacted right now, but this hasgot to be the most impacted area. So, Idon't know that you can totallyextrapolate based on what's happening toAir Table. I don't necessarily thinkthat you want to replace your CRM, yourERP, your HR system with somethingthat's been vioded. You want thecertainty, you know, for anything thatinvolves compliance. I got I got to behonest. My team Sachs made I don't doyou use like a portfolio uh offthe-shelfSAS tool for managing crafts like um uhportfolios and everything.>> Well, we we vibe coded somethingactually.>> So yeah, we just did the same too. So myteam just built something that is somind-blowing that to buy it withoff-the-shelf software would have been aquart million dollars in software andlike a million dollars in integrationover two or three years. and we built itin a month and and now we have completeinsight into the whole portfolio, thecompetitive set, the founders,everything going on.>> Keep in mind that one of the reasons whyLeopold got blown out, okay? I mean, isbecause he bet on the SAS apocalypse.Remember, it wasn't just that he wassuper long these chip stocks that had acorrection.>> Oh, is that right? He was short.>> He was super he was short Adobe and awhole bunch of other SAS companies.

01:01:24Andthose trades also moved the wrong way onhim. So again, I just think that it'spainting with too broad a brush to saythat all of SAS is going to getobliterated here.>> Yeah.>> And there was a really good post aboutthis. Let me just quote from this wherethey said, "Nobody buys Microsoftbecause Microsoft writes the best code.They buy Microsoft because Microsoft isthe rail that everything else runs on.Active Directory is where your employeeidentities live. Excel is where yourboardex numbers come from. Teams iswhere the compliance recordedconversation happens. Azure holds a Fedramp high authorization and Departmentof Defense impact level 5 clearance,which means a defense contractor cannotcasually swap it out for somethingcheaper and so on down the line. So,there's a lot of really good compliancereasons why if you're a largeenterprise, you're not going to want tospend tens of millions of dollarsripping out something that costs you amillion dollars a year. It just thatjust doesn't make sense. And I noticedthat Ben off just tweeted 5 minutes agothat 15 out of 15 cabinet agencies runon Salesforce. Look, the government isnot going to rip and replace Sal withsomebody vibe coded. So look, not allassass is equal in this dimension.>> I just want some uh Figma. I just thinksome of these SAS companies with greatfounders who are in it for the long termand they have like passionate userbases.

01:02:39I think they will make the jumpto AI first products and I I put Figmain that bucket. Just to wrap this thissection, please. IGV is up 20% in thelast six months. It's up 20% in the lastfive years.>> Explain IGV, please.>> So, the high growth software stockindex,>> right? Snowflakes up 88% in the last 6months.>> That's it's an IGV is an ETF of>> IGV is an ETF of of growth softwarecompanies. Right. So, to to David'spoint, there was a panic about softwarecompanies. There was a big trade out.you know, honestly, they they performedpretty well and and as he mentioned inthe month of July, they were up when alot of the semiconductor AI stocks weredown. And some of these companies, databricks, snowflake, click house, etc. aredoing extraordinarily well. As I justmentioned, Snowflake's up 90% in thelast 6 months, which puts it in the samecategory as the semiconductor AI stocks.So, to David's point, you can't throwthem all in the same bucket. But I dothink that for these no code a lot ofthese application software companiesthey're realizing like that that youknow the game is up sell the company getwhat you can get you know importantlyhere in the air table story all thelatestage investors right we passed onthis in the last three funding roundsright which I think we're at

01:03:542 billion 5billion and 11 billion but all thoselatestage investors which were the mostvenerable of growth firms they all gottheir money back and the early stageinvestors ended up making a lot so ifthis is a failure This is a pretty goodfailure for Silicon Valley.>> This is one of the points that was madeat that time which is hey we this is astrong enough company and team andrevenue base that if we just get ourmoney back with the optionality heymaybe this would be a good investment.You could say the same thing about someAI bets.>> Well this is one of those cases wherethe liquidation preference actuallymattered. You know normally it doesn'tmatter but>> straight money here. I my understandingis this wasn't like they had like a 7%you know interest rate or they didn'thave like a participating preferredwhere you get two times your money backand then they do the trade. Does anybodyknow? Because I looked deeply into this.>> I was just saying I think that net ofcash they may have come in a little bitless than the total cash raised but itseemed like everybody got made whole.>> Yeah. But if they had the I guess sachsthey we live through moments in timewhere companies had to guarantee a 1xright you know>> 1x liquidation preference is standard itjust means you get your money backbefore other people start to profitwhich is appropriate>> but the interest rates were taken outright of these deals uh I think during

01:05:10the standard terms you know what's knownas clean terms is just a simple 1xliquidation preference the preferredjust gets their money back before thecommon starts to participate in asuccessful sale of the company. Thatjust makes sense, right?>> Participating preferred is the doubledip, right?>> Yeah. And look, we've never done that.You know, we believe in clean terms. Noone's trying to be punitive towardsfounders. It's just it doesn't makesense for some people on the cap tableto be making money while other peopleare losing money. It just doesn't makesense, right?Well, that that that's just a transferof value from some people on the captable to other people on the cap table.So, the standard thing you do is youmake sure that the investors get paidback and then everybody is participatingin the upside.>> Okay. Fourth story here. China istraining on US data from US providers.Forbes published an investigation calledthese American startups are makingChina's AI smarter. And I think thisrelates to a lot of your work in theearly part of the administration. Sachs,they claim US data labeling startups areselling valuable training data toChinese labs which in turn is helpingthem catch up with the US frontier ones.Two startups, Sergei and Meror

01:06:25are bothvalued at over $20 billion. They selltraining data sets to people like OpenAIAnthropic federal agencies. Um they allsell the same data sets to top ChineseAI companies. According to this report,like Tencent, Bance, Alibaba, Moonshot,etc. Top six AI labs in China, accordingto this report, uh are spending $500million a year buying what Forbes callssecret sauce, uh PhD written content,reinforcement learning, knowledgepipelines, uh all that kind of greatstuff. I have investments in a couple ofthese companies, including Micro One.The founder of Micro One didn'tparticipate in selling to China. He madethat decision. effects. What do youthink here about this new wrinkle interms of really the secret sauce behinda lot of these models is the data? We'verun out of uh open data on the web.Obviously, we talked last week about thebooks being uh you know having thespines taken off of them and scanned in.I mean, people are looking for data.Mira, micro one, all these companies areproviding it. Should they be providingthe same data and selling it to Chineseopen source companies or not? Well,look, I think we got to decide what ourobjective is here. Are we trying to justget in like a

01:07:40full-blown economic warwith China? Are we just trying toprevent all of our companies from doingbusiness over there? If that's ourobjective, then you can take thatposition. Historically, the rules havebeen that you want to be careful abouttechnology transfer of technology thathas a dual use, right? That it has amilitary application. My sense of datais that it's largely a commodity. Imean, data labeling certainly is. If youbasically tell them that they can't usedata labeling, I guarantee you there'sno shortage of labor in China that theycan use to do the data labeling. Infact, they probably are. What I'm sayingis there's a lot of ways to get thisdata. So look, if we basically ban thesecompanies from selling to China, weshould expect reciprocal actions takenby China to ban companies over thereselling to us. Maybe rare earths. Thesetwo countries are not completelyindependent of each other. By the way, Iwant us to be as independent andsovereign as possible. I don't want tohave any dependencies. No dependencies.>> But we still at this moment in time dohave some dependencies. So I think youhave to ask the question is this datareally proprietary? Does it have a dualuse? Does it have a militaryapplication?

01:08:55>> Yeah, I don't think it has military.It's definitely not data labeling. Thisis like hiring PhDs, hiring superprofessionals to, you know, createunique data sets. So it's>> China can do that too and I guaranteeyou they are. I don't think this isgoing to give us a decisive advantage inthe AI race. It's going to annoy it'sgoing to create annoyance. It's going tocreate friction. and how bad do you wantour relationship with them to be? Do youwant to risk starting another trade war?Look, I'm not against restrictions whenI think they're going to pack a punch.For example, I'm really glad that thefirst Trump administration limited theexport of EUV lithography machines toChina. You know, that was all the wayback, I think, in 2019.So, that was a really importantdecision. And so, look, I think targetedstrategic controls make sense. I wouldjust make sure that this one actuallymeets that bar.>> Brad, any thoughts here on thisopen-source catchup, the data being soldto China and our adversaries? Are youconcerned about these open source modelsand then us providing data to them?>> First, you know, I'm in absoluteagreement with David that we wantmaximum competition at as we sit heretoday, the US is winning. We talkedabout it at the start. Our Frontier Labsare winning. Our open source is winning.

01:10:10and we have fairly limited regulations,right? She's coming here in September ina bilateral meeting to meet with thepresident. We're advancing relations ona variety of fronts. So, I thinkeverything looks good and you want tocontinue down that path. With that said,I will tell you that this will irritatepeople in Washington who feel that thisalong with distillation and other thingsum could be the export of chips. All ofwhich at a certain level make sensecause people to wonder whether or notwe're making it too easy on the Chineselabs to catch up with American labs uhyou know in the race to frontierintelligence. So it you know it's thetype of story Jason that I think willcontinue to muddy the waters that willcontinue uh to be monitored. The reasonI don't think it will cause us to changeour stance with respect to China isbecause we're winning. But if thepresident asks his adviserss, you know,one of these days, six months down theline, are we winning against China? Andall of a sudden he gets a response, no,we're no longer winning, they've caughtup, they've passed us, etc., then thesethings will get a lot more scrutiny uhthan they're getting today. I think theonly reason they pass muster today isbecause we're still leading the race. Igot to say, using Kimmy and Quen and,you know, GLM 52 for the last

01:11:2660 days,my lord, these things are good. And Idon't think it's very patriotic to begiving them an advantage. I wouldn't doit. I'm glad the company>> Sorry. What's the advantage? What's thedata set that you you're worried aboutthat's so proprietary?>> Any of these data sets are um created byexperts here in America who are givenlike the queries that have errors inthem. So when you give um you know athumbs down to a query that's highlytechnical, it could be code, it could bebiology and science, these are you knowPhDs going in there and putting in thelatest and greatest content and thenverifying it, double verifying it. Andthat's why we're getting better andbetter results out of the LLMs. Soessentially, you're just helping themcatch up. And this could be a bigadvantage for America if we weren'tsending it there. I think a big reasonthese models are getting better isbecause data is being leaked to them.>> But what makes you think that Chinacan't do this? They have tons of PhDsover there.>> They would have to hire No. No. If theywere to do it at this scale, they wouldneed to hire the best and brightest uhscientists and experts in the west. Sobasically all the knowledge of the Westis being um you know put into packagesfor our LLMs to get better. They'resending those same packages andreselling them

01:12:41to Chinese companies,which means they catch up just as quick.I think it's a big part of why they'recatching up in line with distillation.You know, they're it's it's really verysimilar process.>> Look, if there's something trulyproprietary here, I don't want us tosell our secret sauce to China. So, youknow, I'd have to look into that and seelike is there some real secret saucehere. But this idea that it wouldseriously disadvantage China, you know,they're graduating more math and sciencegraduates every year than the rest ofthe world combined. I mean, they don'thave a shortage of smart people,especially>> and we're graduating and kicking themout of the country. That's the otherproblem. We got to get that fixed.>> Well, it's like this a lot of differentissues here. I don't know how many youwant to conflate, but I this idea thatthey can't but the this idea that theycan't recreate those data sets. I mean,look, if there's something trulyproprietary here, if it has a dual use,if it's military related, but I don'tknow that that's what this is.>> Well, they're all proprietary by design,but I don't know about the dual use cuzI don't have the data sets here. Allright, folks. That's another amazingepisode of your Allin podcast. Thank youso much, Brad, for joining us, Chimath.Good luck on your world tour. Hopeyou're enjoying a little rest and goodluck um trying to buy a white

01:13:56turtleneckthis season. that are sold outeverywhere. So, go to the all-in.comstore, allin.com/store.We have 1,000 signature Chimathautographed white sweaters coming. Youcan sign up in advance for those. Allproceeds go to charity. By charity, Imean yacht fund. All right, we'll seeyou next week, everybody. Bye-bye.>> Let your winners ride.

01:14:27We open sourced it to the fans andthey've just gone crazy with it.>> Love you. Queen of>> yours.>> Besties arethat is my dog taking a notice in yourdriveway.

01:14:47>> Oh man, my habitasher will meet me up.We should all just get a room and justhave one big huge orgy cuz they're alljust useless. It's like this like sexualtension that we just need to releaseourselves.>> We need to get mercy. I'm going all in.

01:15:12I'm going all in.