ANTHROPIC DEAL: SPACEX IS THE 'AWS' OF ORBITAL AI

Kind: captions Language: en So compute power is now the number one bottleneck for the AI industry and we have just witnessed something absolutely historic. Anthropic compute rental agreement with SpaceX and XAI gives Claude AI full access to Colossus 1 supercomputer in Memphis. Now mind you this is one of the world's largest AI clusters. XAI originally built Colossus one for Grock but now it mostly uses Colossus 2 for its AI training. So what sort of an edge does this give anthropic and what does it mean for competitors like OpenAI? Also, is this the clearest sign of SpaceX's impending dominance in the orbital AI data center domain? I have Brian Wong with me and he's going to break it all down for you today. Brian is a futurist thought leader and he runs the science uh the prominent science blog nextbake future.com. Welcome Brian. Great to have you back. Great to be here. A lot of exciting news happening. >> Absolutely. So, quickly, let me just bring this deck up. Um, and if you could take us through this historic agreement. Um, >> before we begin, when do you I think I think Elon said there's going to be a new logo for a combined SpaceX and XAI entity. Is it going to be SpaceX AI? Any thoughts on the new logo? >> Um, no. I don't know what the new logo will be. Um but it you know may be focused more on the AI data center in space type thing. Um whereas basically getting both to be prominent because that will be the dominant revenue impact aspect. So I think that would probably >> in the long run right but you know like as short as four three four years it'll be a major aspect of what they're doing right. >> Yeah. Do you think it'll be bigger than than Starink? >> Oh the the AI. Oh, for sure. It It'll be way big. >> Well, so so Starlink is 30,000 satellites V3, right? They're going to go from the 10,000 11,000 now to 30,000 of the highspeed internet ones and then 15,000 of the um uh for for direct to cell phone, right? Versus 1 million for the AI. So you're looking at 20 times as many satellites for for the other things. So yes, it will be bigger and then that will scale can scale up. Um, yes. >> All right. Okay. So, back to you, Dick. >> Okay. Okay. So, yeah. So, um, for people who don't know, um, Anthropic and, uh, is granting the closest one as as you described, and that's for about 300 megawws. Um, Epic AI says there's about 445 megawws there. So um there could be still some usage of 100 megawws from XAI as needed but you know they'll have to share in some way u when they get fully ramped up. Um so this is immediate revenue for Anthropic. So they've already started um promoting their um increased service. So they increase usage limits where they're doubling the um cloud code five hour rate limits for each of their tiers of their subscription service and they remove peak hour limits for basically like if you want to spend and give them more money they will let you do it because they have more capacity. Um so they could double their revenue on the API side by increasing these limits which are already happening. But I I think they're already using the XAI uh thing. But there's some things maybe it doesn't start till end of the month. Unclear. So these are the token increases where basically the level one tier one becomes as good even better than what the the tier two was. So tier two used to be 450,000 tokens and the tier one with only 30,000 which was a very low limit. And now tier one's gone to 500,000. So basically they bumped up all tiers. So it's like what was 1 2 3 4 is now 2 3 4 5 kind of thing. Um and then increasing the maximum maximum usage per limit. So again people are paying per token. So then if I increase the usage they can make more money right away >> and this is money right away to XAI right so the 300 megawatts. So it because it's dedicated data center then that gives a premium. So Amazon charges 2.4 times the price of what it takes to build what they're doing, right? For a dedicated largecale data center, right? So there's a premium for people think that there's a a discount for the volume. It kind of works in the reverse in the situation that they have described here. So then if it was the full data center, $7 billion to build it, 2.4* 7 billion would be 16 billion. But if it's kind of only part of it, like 67% 300 megawws, then maybe it's 2 * 5 billion because not all 445 megawatts only 300 megawws. Um, but I can if the demand is off the hook, which it appears to be for anthropic, then maybe it end up going to the full 445 megawws and then they pay them more money, you know, like goes up to 16. I am anticipating that the demand being so large which is what um Dario was talking about on interviews you know one anticipating an 80 times increase in usage versus 10x right then they can use everything they that they can be given >> and then XI can make more money it will help with the IPO to say hey XI not losing money not losing a billion dollars uh a month it's actually break even or better by renting out only 15% of our capacity, right? If you do, you know, based on H100 equivalents, right? It's more it's more like one-third based upon GPUs like but the class of one has H100s and H200's versus the four times better B200 B300s, right? So we there's still some um thing we don't know which I would expect after the S1 filing to say here here's a date which we expect maybe next week. Here's a date when we're going to do this thing. Then there'll be a time when they say and here's the financials of XAI. >> And then this will say our financials are great. You thought they weren't going to be good. Our XAI financials are great. Our SpaceX financials are great. And we have a partner for the future with space-based AI. So the high end I think is 16 billion. The low end is eight. And the super low-end bare case is five. I'm thinking 8 to 10, but then if you use up the full 445 megawws, then um it could go up to 16 or even more. >> So this is so so what's the timeline for this? And and how much of this is a combination of um space-based uh and like terrestrial AI data? This is this is right now I'm just talking about Colossus one. This is like not talking about anything else. This is only Colossus one rental revenue, right? So I'm not talking about the spacebased stuff. I'm not talking about distributed AI. >> This is just I'm not even talking about if they were to rent half of Colossus 2, >> right? Like they could do that and then they would make even more money, right? Yeah. But basically they're they're getting the um so it's good for philanthropic. They're making more revenue instead of a run rate of 44 billion. and they go to like 60 billion maybe even this month or next month, >> right? So suddenly it's like they're growing fast. 30 billion run rate a few months ago, 44 billion just the you know like end of last month and then with this deal they go to 60 billion plus. >> Wow. >> Right. And that's like in May June right and then good for XAI and SpaceX IPOs because now instead of worrying about XI losing money then they have a revenue stream. I have a revenue stream that's making me money and I'm still doing the all the training and all development of the of the code. >> Yeah. >> And then so then we get both IPOs out. XAI SpaceX first, Anthropic Next. Bad for OpenAI. They're the third IPO for sure. >> Yeah. >> Right. So, >> yeah. So, this is this is interesting because I I How much of this is a 5D chess move by Elon? Mhm. Well, it it's um it's a it's a win-win for for him. It's a it's a win-winwin for for him. So, he he wins because Xi does better more more revenue on stuff that he wasn't fully utilizing. Two, it's a win for Enthropic. And then they also have a deal. I'm sure they have a better deal now with Enthropic because they needed to have the cursor data and Enthropic could try to limit them. I think even before the the cursor deal they were saying hey uh XAI you're using our anthropic stuff we're cutting you off because we don't want you learning from our models and training against our models >> right but then Elon does a cursor deal and cursor is one-third of the revenue for enthropic so anthropic can't cut off his right arm in order to cut you off right >> but then now with this deal it's like they're going from enemies to frenemies >> and then they and they have data. So they hear the data to to catch up and then they maybe even are semi-partnering for all I would think definitely they'll semi-partner all the future stuff but then they both hit their enemy total enemy open AI >> where it really hurts. >> So it really hurts them. So it's like and that's like the cherry on top. It wasn't the reason you did it but it's an extra benefit of like and I get to screw him. That's that's nice. So this is net net very positive for both the SpaceX IPO as well as the anthropic IPO that's expected maybe later this year early next year but before anthropic we were all expecting the open AAI IPO >> right >> and together with the with the case that Elon has in court currently against right >> OpenAI Sam >> and these are two really big negatives for that IPO isn't it >> right right it's a huge yeah it's it's even if Elon doesn't win that that uh the court case. Um >> I mean which I think their chances are low. No, I think the chances of them winning some part of it is high that they can get like something like you know you know but it's getting the full thing of like forcing restructuring that's low right but saying hey you wrote in your diary that you lied to me and you know you you you you said that these things would show that you you know the $40 million would was defrauded in some way right so getting some millions of dollars or $40 million $100 million that I think that the odds of that are pretty good, right? Because they clearly did bad stuff. The other thing is that >> if you allow them to freely do this um um nonprofit to profit conversion, >> right? If I say no rules, you can do whatever you want. You can change your everything you you did, right? that opens a door for $5 trillion of hospitals, universities, all these other nonprofits to go. So then there'll be this massive um um you know conversion where for where for where for where for where for where for where for the people who control the nonprofit suddenly say I'm going to you know make a bunch of for-profit stuff in a hospital and then pocket the money because they had restrictions you know in the 80s and 90s and various other things right and then they're saying oh I can just freely do that great because that is something where you would definitely make instead of 3% 5% profit because that's a nonprofit and then go to well now I'm unlimited. I can have one part that's the the forprofit, one part the nonprofit. I can make my hospital charge 20% margin or 15% margin and then I'll you know give some a cut to the nonprofit side right so basically all the hospitals in the United States would go right if you do not limit it but that doesn't mean that this current decision will be that right because um I it could be in appeal or in the Supreme Court where they would then make that limitation But there's no way they can allow the uh precedent setting thing of what opening I did to stand. They cannot just let that happen. >> That opens a whole Pandora's box here, >> right? But anyway, so so the I expect Elon to get something. The question is how much? It's unclear. >> Okay. So that's that part. Um so there was a shift of.3 gigawatts between. So we have this table of all of them. So in the middle are open AAI and XAI in the range of two 2.5 gigawatts and then.3. So XAI drops to 1.7 and then open AI goes to 2.8 right so and but then open XI still building stuff and so this is uh a significant shift in the immediate um u building of of data centers. Uh but we have a long-term game where everyone's building a lot as much as they can uh going into next year and and beyond. And the big advantage that XAI has is that they can do inference computing at the superchargers with the cars and with the power walls. So superchargers are about um gigawatts and then you're doubling every two three years, right? Based on the pace that they're trying to increase to. And then you already have the power there. So if I place a small rack at that location, then you know if I have like 10,000 racks, I can tap into those sub gigawatts and continue to build out. So it's um easier to add that. I don't have to build new grid, not to build new stuff that's already kind of like in place with the solar superchargers. But the cars once they get the digital Optimus to work, there's like one gigawatt just from the the hardware for cars lying around. So that'd be three times bigger than the 300 300 megawatt deal of OpenAI. >> You know, we we should we should do a separate deep dive on digital optimist because I know you've been doing a lot of tremendous work. >> Yeah. >> And I would love to chat with you about that se separately because that that is fascinating. Right. The other thing is power walls. I think there's about a million power walls out there. Each with about 13 kilowatts. So that's about 13 gawatts of power walls. >> And a new company, a startup, SPAN AI working with um Nvidia is putting like a dozen chips, 16 chip chips, Nvidia chips into each home with the equivalent of about a power wall. So that is another way to tap into more AI inference. So the thing is XAI has the Amazon capability to toss on a large amount of compute for AWS. So for Amazon AWS um for e-commerce Amazon built a lot of clouds and then and then they used it for themselves but then they added on AWS where they built like four or five times more and then they sold it to other people. This is the model that XAI will be showing that they can quickly tap into more inference via their energy network of superchargers, cars and power walls and then they can then rent it out to to themselves or to open AI. >> Oh, sorry. Or to or to anthropic the two. Bri Brian at this stage I want to ask you a question because then just so that our viewers understand so this >> there are two levels where or two stages where you need um AI compute right one is for training >> more powerful models and the other is for inference right >> now inference can be distributed >> but training when it comes to training you need your chips all together >> right >> in the same place right physically all together in the same place. So if you zoom out and you look at how um XAI is building its infrastructure here terrestrially on Earth and also looking to build in space. How much of um an advantage does this give quantitatively and qualitatively to XAI and SpaceX's ability to keep building new new infrastructure um and and moving its training as well as inference to the latest hardware either terrestrially or in space while at the same time giving out uh you know its previous generation of hardware to anybody like Anthropic who may want to hire it. Give us the big picture. >> So I think that the you know 13 gawatts of power wall increasing you know 20 30% every year the uh supercharge the 7 gawatts also increasing at 20 30 40% every year and then the cars also increasing you know at probably slower rate probably 10 20% every year although with um cyber cams that could also have a surge in growth. So you add those up, you're looking at like 20 30 gigawatts right there and then that will keep growing, right? So it'll take three years say to to fully tap that in. But if you have an average of 10 adding 10 gigawatts per year, that would be huge. You would because the whole world is adding maybe 20 gigawatts per year. So suddenly you can get to 50% of the additional um growth of of AI by being able to tap into these less conventional distributed sources where I can do it without you know with less grid building or with less whatever. Um, and then for AI in space, if I get to 10,000 launches per year with Starship, which I can see how that can happen by 2030, right? 10 10 launches this year if if we get the the next launches right. um 100 launches the year after because we have um three major uh launch facilities, Texas, Bokeh Chica, um two in in Canaveral, one already built another and with already with FAA approvals up to 120 and then I think that in the in 2028 by shortening, you know, getting your operations down to Falcon 9 efficiency where it's only a half an hour to get my launch off where I shut down the the the air traffic um exclusion zone 20 minutes ahead of time, get it lifted in 10 minutes, then instead of three hours, you know, then I can maybe get um you know, 500 launches, 200 per location, which is about what I'm doing for for Falcon 9. That gets to 600 launches. And then with the money I've raised this year, I would have it takes six months, nine months to build new launch towers. you know, probably another some six months to get the site locations and a bunch of other things. But by 2028, I should have new um launch towers, new sites online, and that can ramp up the number of launches toward a few thousand, right? If I had like six to 10 locations in 2028, 20 locations in 2029, >> that's I think what what you need in order to really scale to the 10,000 at 10,000 launches per year with 10 megawws. So like a 100 uh 100 kilowatt satellites. So 10 megawatts in each one. >> Then that 10,000 launches is 100 gigawatts. >> Yeah. >> Right. So that's the kind of scale that has to happen to do that. But once I'm doing 100 gigawatts, then I'm five times more than the regular 20 gawatt that everyone else is adding on Earth. And even with 10 10 extra gigawatts, 20 extra gawatt of distributed power um you know from supercharged other stuff. >> So the AI the orbital AI data data centers would they be for training or inference or both? >> Inference. Inference only >> purely inference. >> Purely inference. >> Purely inference. Okay. >> Right. Right. But thing is inference is the you know 100 times thousand times larger thing because I'm profitably you know you know running the models to make money. So it's like I'm I'm building it less I'm selling it more. So the ratio of selling more is increasing increasing which is what you know Jensen has said other people have said is going to happen. So then for training uh Grock models more more powerful Grock models um do you see like subsequent Colossus data centers being built out all across the rest of the data centers? >> Right. Right. You you would you're going towards from the 1 gawatt to 2 gawatt on the um drawing board on the plan is five gigawatt and 10 gigawatt. Right? So that was, you know, like you met Meta, Prometheus, Hyperion, they're 5 gawatt, 10 gigawatt facilities. Um, you know, Colossus 3 hadn't been announced, but I'm sure with the Reuben chips, they will h they will target a 5 gawatt facility and probably a 10 g facility. So you by 2030 you have a 10 gawatt facility which is what um uh Leapole Ashenbrer you know who runs his 5 billion10 billion fund says is needed for AGI so we will need to get to 10 yeah go ahead >> and um so I'm just by extension when you look at terapab we know that terapab is going to be building chips for different purposes like for cars for optimist for um the space um AI data centers Um will Terrafab when will Terrafab and will it also build chips for terrestrial AI data centers for training models? >> Yes, it will be building chips for everything. It'll be making AI5 chips, AI6 chips which will be you know on Earth in space wherever needed. Dojo 3 chips for training. Yes. So they will be making chips for all purposes and I have a bit um on the next slide I think that we'll discuss that where I discuss so in the Google thing in the the the column to the right it talks about the TPUs that they have and has tranium chips so each of the big players the hyperscalers are making their own chips right as much as possible because they don't want to pay the 70 80% margin to Nvidia where they can avoid it right but Nvidia is so good that they need to still have them Right. >> So then um going on to the uh next slide. So as I mentioned part of these deals are the cursor training data that is key to XAI with cursor catching up to enthropic right and enthropic will block it less because hey we're we're buddies now. You know I'm helping you make a ton of money. Get off my back about um um the fact I'm using the data you know which I should have right to anyway. um to do this thing. So they will basically make a truce on that training data thing where basically uh cursor and XI will use that data and they're going to be future partners for space AI and I think for distributed AI because in the AWS model >> I need people to buy it up and if you have an insatiable demand >> then I have I will have insatial supply that I cannot leverage as fast as I can bring it on. Basically they can bring on the distute AI. I can bring on the space as fast as possible. When I need to throttle enthropic down because my grock models have finally caught up or I made a new cursor code model, I say, "Hey, I'm going to make four times much money with my model versus only twice as much um as my cost for for the other thing." So, it's just in more margin, more money, and then having >> Can I ask you a question? Just going back to one of your earlier slides about the revenue, we're looking at as much as 16 billion a year, right? >> Mhm. Right. >> Um, how much does it offset the cash burn >> that uh >> X it could it it could offset completely? So, so it's about billion $1 billion a month is what they were doing, right, of losses. Although in the revealed uh financials for 2025, they were saying SpaceX makes $5 billion and then we go to negative $5 billion when we add in um XAI. So that would be $10 billion. It would be not quite $1 billion per month. It'd be like $800 million or something like that, right? So if I did grow some other things, X payments, ads, blah blah blah, then you know the at $5 billion it would be close to 400 to $500 million per month coming from this. If it's you know 8 n billion it fully offsets. If it's to the higher numbers 10 12 16 then they're making a profit from XAI and it's only using 15% or maybe 10% of their actual total compute build. Right? So that that means that they can you know if they were consistently doing 50% for other people 50% inference for us and then we have still have all of our own training right then they can comfortably be very profitable right with the 50% of the of their data center stuff. So that means they can um raise money, build as fast as they can, build you know big data centers, build um the distributed stuff at the cars, superchargers and homes. >> So the reason I um I'm kind of trying to zoom out and look at um SpaceX post IPO, right? One of the biggest concerns about um post IPO SpaceX was the cash burn for XAI, right? So here is a potential solution to that problem. Right. >> Right. >> Um immediately and then of course this becomes this takes on a life of its own and the more um SpaceXi scales with data centers the more revenue comes in and it's a flywheel that keeps going. Now what happens to the likes of AWS um Google um Oracle even you know and and all these major infrastructure companies supplying infrastructure for AI uh compute because the problems that exist in America in building data centers terrestrially on American soil continue to exist. there's growing opposition and communities increasingly don't want a data center near them or in their communities. So I'm wondering how you think of this dynamic playing out >> and uh what sort of lead that does give does that give SpaceX XAI? >> So Dylan Patel said uh you know in a in an interview he did recently and he tracks things closely with semi analysis where they look at all the token economics and profit things. the Anthropic has shown that they can be profitable with um with their AI inference, right? And then they've cut deals with Google, with um Amazon, all for it. So they he says any AI data center that's built first tier, second tier, everyone sells out. Everyone, whoever you want to sell, you will sell it. You will make your money. You'll be profitable. And then the ones using it and thropping whatever they will be profitable. So we've crossed over from we're building without profits to we're building with profits. >> So then whatever you can build you turn it on. You overcome the blockages to making it that new data center will make money. All your old data centers make money. So that's why you know think they're rerating where you know um all the AI players are making more money is because this is now we're past the point of will it make money to yes it makes money and the question of who makes money is everyone who can build anything >> right >> so one more question before we move on and this is about the life cycle of these chips right they're anything between three to six years >> right >> so how old is Colossus one now it's about almost three No, I think it's like two years. >> Two. Okay. So, what happens once that life cycle is over? What happens to Colossus one? Do you think is it going to be upgraded or is it going to be uh hired or the services of Colossus one going to be taken over by or hired by companies that are >> Yes. >> level two? >> Right. Right. So um the the H100s the which are H100 H200s that are in Colossus one right >> the rental rates increased you know so they went you know started a certain level they went down and now >> they're going back up because demand is higher it's because if that H100 H200 can still serve out tokens the fact that I can serve out you know a billion tokens mean I can make my million dollars Right? So that means that it will still have use and utility. If I'm constantly having a shortage of energy and compute, then I'm still paying for the old stuff, just less relative to the new stuff, but still at a profit, >> but it's still earning revenue. Yeah, >> it's still earning revenue, right? So then it won't drop per se you know like if the tokens start dropping you know five 10 times or like that then you know 10 times then the amount of money I can make from these tokens may start being less and then at some point it becomes I have my shell the building I have my energy and so then I would look at swapping out chips where I can finally take them out and then pop in the new ones right >> so Um, but that probably isn't for for 3 years. But it's it's um assuming I have an abundance of chips to do it, right? But I have a lot of chips. I can't stick them anywhere else. Well, let me just revamp this one. Which you see that with um the uh Bitcoin miners >> like um like Cororee and some of these other companies that were making, you know, $2 billion a year on Bitcoin mining. They say, you know, they say, "Okay, you Bitcoin miners, get out of my my data center. I'm putting in AI. I'm gonna make 10 times more money." Which is why, you know, some of these names are converting over, right? So, that is a conversion factor. So, it will be you can renew, but you choose the right time where it's like, "Okay, lease is over. Get out. I'm putting in the new one." Blah, blah, blah. >> You know, I love talking to you. This is why, >> you know, it's like we're building a mind map >> as we go along. And I think it's so important and that's why my viewers absolutely love you, Brian, >> because uh you know these >> you don't get these mind maps as clearly with anyone else as as I get with you. So, thank you and apologies for keep keeping peppering for I keep peppering you with these questions, but you know this >> anyway, let's get back to your day. Okay. So they already said that they want to talk about being future. It's kind of like if they if XAI, SpaceX, Tesla build all this extra inference on the ground in the distributed locations and in space then everyone comes to them. All the AI players come to them because money matters less than actual working chips and energy. So that is the short resource that they have. And if it's short for the next five years as far as the I can see then SpaceX has the dominant position. If they start making five times 10 times a hundred times more compute in space than anyone else then they're the ones with the um the monopoly on energy and power and and compute. Um, so partners on disputed AI. I think that I discussed that. I think that if they can get the superchargers, the cars and the power walls loaded up with AI, that means tens of gigawatts more more AI, which is uh if you have 30 gawatts of that, that is 100 times more than the deal we just did, right? So that's like >> So you're saying that it's inevitable that that gets leased out as well in time. >> They'll be it be leased out as well or if Grock can use it himself become dominant, you know, either or. So either I'm making twice as much money or I'm making, you know, four times, five times as much, you know, 10 times as much because I'm actually got the model, right? But the the the economics are not bad for leasing, right? So I will they will be making plenty of money from that, right? So it's like saying you only have an AWS model. Amazon doesn't have their own thing, but they're renting on AWS. Everyone loves Amazon, right? And you're saying I only have an AWS model. >> It's like so and I have a chance to to make my own model. It's like Yeah. So that's okay. So then here's a slide from um Goldman Sachs investment research where we're at the point where um the agents are taking over right in the the middle of 2026. You can see that consumer agents enterprise agents just starting to take over. They're expecting it to go up 24 times over 24 times by 2030 which is more than doubling every year. And this could be cons this could be conservative, right? But then you see that the demand is going 24 times. So that means if I can make more tokens, if I have the energy compute to do this, right, then that's where the money is. >> Okay. Can you Brian, can you for for my audience that may be wondering what enterprise agents are? Can you describe them for for us, please? >> That is the um cloud co-work. That's the cloud code where they're using it to make enterprise software or they're using it to do financial analysis that a company would do. It's not just a chat thing of like here let me ask you a question. I get a kind of so- so answer back. I I get a 10 pages. It's like I have a cycle of of work and I'm checking it, you know, every time and I'm going towards a big company related answer or or product of like a 100 slides spreadsheet, a 200page deck, a research report that's like McKenzie grade or something like that. That's enterprise caliber work, right? is not just right >> asking, you know, the the typical questions we have at the chat level. It's like I'm working towards something that's replacing consulting uh that I'm paying $100 an hour for something. >> Yeah. Right. >> Because this this also then has an impact on the the workforce of each of these companies that deploys these um >> Yeah. >> enterprise agents, right? because uh there's been so much of discussion about whether it's going to lead to um the workforce being you know reduced or whether it's going to augment the existing workflow and put it on steroids because you get so much more productivity. >> Right. Right. Yeah. So, um I think that you know in the all-in podcast they discussed that um people aren't you know aren't worried about vibe coding. It's like um the back in the day when you know like 30 years ago when Viscal Loris 123 came out. Oh, we're going to replace all the accountants. We're going to replace all the finance people because everyone can do their own spreadsheets, right? It's like no matter how easy you make it, the professional doing it will give you better product work, right? Um although you know you'll have something that becomes trivial, right? Then the the value of the trivial thing becomes nothing because you still pay for the elite stuff, right? So the I think the the the thinking that we'll replace the workers will is going away with the really good people will get productivity increased 10 times 100 times and then the quality of the work will go up that there is >> generally across the board the bar will be higher >> the bar >> quality of work will go yeah >> and then yeah consumer agent I think the enterprise stuff may end up being bigger in terms of the valuation part of it in terms of consume tokens this is a rough estimate of that. But the main thing is we're going to be using a lot more. We're going to use 25 times more tokens within four years. And this could be on the low end, but just means more inference compute is needed, more data centers are needed, and it's all profitable. And then the ROI on the hyperscalers is Google has shown that >> they're making you know $20 billion per quarter on their cloud revenue and it's increasing by 63% and they're having operating income of 6.6 billion on capex of 35. So they're getting you know like over 50% ROI from that you know based on this approximation based on cloud revenue. Azure is also doing pretty well. strong every growth reasonable ROI. AWS is showing a positive ROI trend. So again the the investments are getting money back for all the big players and then you know anthropic as well. Um this is a very uh small thing. It's just breaking down the ROI as well as the chips. So each of the players is getting the the TPUs, traniums, and then on the bottom line, you see that Tesla, SpaceX, XI, they're going to make their AF5, A6 dojo chips, which is instead of paying $50,000 for a Nvidia chip, they'll pay like $6,000 for their own chips for inference or chips for training, which is similar to what uh Google's doing with TPUs, which are cheaper than, you know, because they're making it themselves. they don't pay the Nvidia thing. So, as long as you have a competitive chip, you will use your own to improve the economics of what you're doing. >> Um, >> so I've I've I put this um on the big screen mode so that >> any of you watching can pause and you can zoom in >> because it's a lot of dense text. Um, so use this opportunity to pause, zoom in and study this. >> Yeah. And then the um money raised is like $42 billion for XAI. also got tens of billions of dollars in debt. So, a lot of money being spent, but if it's all profitable, then it's good. And then if I'm reducing the the cost of my chips, I'm, you know, instead of spending $50 billion to make a gigawatt, I may only spend $10 billion to do a gigawatt. So then, if I was already making money on $50 billion for a gigawatt, I'll be making way more money on $10 billion for a gigawatt. So the economics are improving faster and faster for this thing. So again, it's um if I can stand up another gigawatt of compute for only $10 billion or maybe even only $5 billion because I already have the energy at um uh superchargers and cars. I only need to put in you know some other layer then the like the car gigawatt in the car is nearly free because it's just chips sitting there just looping in. >> So so that means that the that Tesla Xi SpaceX have the lowest cost of adding in gigawatts of compute >> especially when they get the AI5 ships into the cars and the bots. It will be the cheapest way because the they've already been paid for. the power's basically paid for. The the chip's been paid for. It's just um you know the incremental cost is probably nothing. And then they can make you know like um you know uh 1020 billion um per per year from that from that uh compute. So the economics >> crazy because like if you it's like one of those machines those uh what do you call those machines that kind of um keep generating energy on their own. >> Oh yeah but but the flywheel that doesn't need external >> perpetual motion machines. >> Perpetual motion machines. Yeah. This is like a perpetual motion. >> Yes. >> Mini economy on its own. It's it's crazy. Anyway, so so um that goes to you know how this will be able to scale to to many tens of gigawatts. It's a matter of like how fast can they bring it online and then they know they have a customer right away who will take it and they'll make more money. So that goes to how 2026 is already good for open for Anthropic and and XA and SpaceX and Tesla and then 2027 2028 will be even better once they're tapping into the cars, the bots, superchargers, all that extra compute and energy. It's um it's insane amount of money. >> Yeah. >> And then so this is the macro hard with the data center. So that's basically it. So I think I've done my deck. Yeah. >> All right. Wow. This is like, you know, we should build a collection of these lessons, these deep dives with you, Brian, and then >> Yeah. >> colleate them and put them together and then you can, you know, weave them into a book of yours and publish it because it's brilliant. Brilliant. >> Okay. Great. So before we say class dismissed, I have just one question for you and I'd like to know your thoughts on this big um shift back towards CPUs from GPUs. >> Mh. >> And I want your thoughts on this tie up uh between Intel um and XAI for Terrafab. Do you see um SpaceXi now moving towards a balance of combination of CPUs and GPUs? So the um Blackwell chip um like when you have the rack that um of semi2 chips or whatever that um that uh Nvidia sells you, they have like you know I think 72 um GPUs and 36 CPUs. So 2 to1 if you go back to H100 I think it may have been like 8:1 >> right? >> So now Reuben which they announced which they start selling um any month now they um use one GPU and four CPUs. So the ratio CPUs to GPUs is increasing to CPUs because of the workloads uh agentic workloads that when I do agents more complex stuff the roing around is more difficult and I need more CPUs to do it. So thus there is and also I need a ton more memory. So basically based on the versions and what I have to do >> I need >> I think we should have a we should have a a a separate deep dive explanate explainer about this Brian because >> right >> um there's so much of um kind of confusion over why because we switch from CPUs to GPUs because GPUs allow parallel computing >> right >> right that was a raise on the >> raise on death of the whole reason why we move to >> AI and parallel computing. But then if you have sequential computing again and CPUs, >> it's a bit of a I'd love to wrap my head around it and if you can help you know explain this um because this is going to this is fascinating. I think this whole shift back towards CPUs is really really fascinating and we should really do a deep dive of this as well. It's yes it's basically to just to to quickly summarize which we'll go to a deep dive in the complexity of what you're doing with agents I need to do that with CPU that that type of work I need more CPUs to do and I also need more memory >> so that's the I still need the GPUs are still core but other stuff around it you know the best way I can >> so is this a training versus inference thing like training on on GPUs but more of inference on CPUs with a it it's it's um the inference is getting more complicated. >> Okay. >> So I so it's still inference in both cases but the inference got more complicated. So then I need to to have more CPUs to handle the complications. It wasn't you know the the a simple routing thing. It became um you know a lot more stuff around the parallel part. Yeah. >> All right. I'm going to look forward to this because I I really need to understand this and I need your help. Okay, we'll do. >> All right, we'll leave it at that for now. Thank you so much, Brian. It's been so wonderful going down this rabbit hole with you on uh you know this whole anthropic space XAI deal and um you you've helped us lay the foundation of our understanding for this. So, thank you so much. >> You're welcome.