TERAFAB: ELON'S MOVING FASTER THAN EXPECTED

Added: 2026-05-07

[00:00:00]SpaceX strikes a deal with Cursa AI.

[00:00:03]Tesla invests $2 billion in a mystery AI hardware company. Tesla and SpaceX tie up with Intel for Intel's 14 Amstrom chip technology. That's just what we know. Elon has been furiously working behind the scenes to make the Teraf project a reality as soon as physically possible. I have Brian Wong with me today for a deep dive into everything that we have learned so far and then some. Brian is a futurist thought leader and he runs a prominent science blog next bigfuture.com. Welcome Brian. Great to have you back. Lots to talk about today.

[00:00:36]>> Lot to talk about. Great to be here.

[00:00:37]>> Yeah. Yeah. It's it's I mean all of this is evolving at such a rapid pace. It's going to be so difficult to kind of stop midstream and take stock and assess where we've come because we've come so far already. We're only getting a few little, you know, slivers of information through the 10Q filing or through, you know, all I I suspect we'll know a lot more when the SpaceX IPO uh documents come out, right?

[00:01:08]>> Mhm. Yes, it will. I'm looking forward to that.

[00:01:11]>> Yeah, that's going to be fun. When are you expecting it? I mean, because that again >> expecting in May. Expecting May.

[00:01:16]>> All right. The the second half towards the end. Um, my overunder is is mid May.

[00:01:22]So, it could be like, you know, second week, it could be third week. My overunder is expecting it in May. About the middle of May.

[00:01:29]>> Middle of May. So, that that's just a couple of weeks away. Wow.

[00:01:31]>> Right.

[00:01:32]>> Exciting times.

[00:01:34]>> Exciting times indeed.

[00:01:35]>> Okay. So, um, where do we begin? Let's let's kind of set the lay of the land first, right? So, you've had these three or four deals and tie-ups and agreements that we know about. first the overarching Tesla SpaceX deal with Intel. Then you've had the SpaceX deal with Cursor AI. Um, and then you've got this $2 billion investment in a mystery AI infrastructure company that Tesla made, which we talked about in one of our earlier episodes. And for those watching, if you want a deep dive on that, please go and watch that. I'll also link it below in the description so you can take it take a look at that and also at the end. So, um, that's just what has surfaced, the tips of these icebergs, but there's a lot of work that's going be going on furiously behind the scenes, right? What is your sense of that?

[00:02:25]>> Um, my my sense is that um, activated chips um, you know, infrastructure, activated infrastructure is everything and the only way to activate infrastructure is with energy.

[00:02:38]So the critical components are I have to have enough chips, I have to assemble them into a data center and I have to have enough energy to run it. So those are the critical pieces of it and having that capability is more valuable than money because I also need money to do this thing, right? But if I only have money and I don't have the other two, I do not have um a a built data center. I only have the money to build a data center. I'm still two years away from data center if I can solve all the supply chain problems. So navigating through all supply chain problems to get to the end result of that is what you need. And then if you have a a spare gigawatt of data center then you have leverage to do deals like the cursor deal.

[00:03:28]>> So what's the cursor deal really about?

[00:03:30]Is it about providing the AI compute infrastructure or is it about an aqua hire hiring you know absorbing the the the team and the u you know the main guys involved there or is it about jointly working towards like a a coding solution that's a one-stop solution for everybody >> right so the end goal is to have the best AI with the best uh compute AI coding, right? That's, you know, you want to have all that, but the only way that deal comes together is if you have a bunch of extra compute to entice the um the cursor team to come to you, right? Because they didn't have enough compute. But Anthropic and OpenAI are also short compute. Open AAIt >> everybody's everybody's going to be short computed. It's only a matter of time, >> right? But XAI is less short compute right now. They have a a gigawatt to spare. They have two gigawatts. They have a gigawatt to spare. They have 200 300,000 GPUs to run stuff. Right? When OpenAI offered Cursor 50 58 billion a few months ago, Cursor the leader said, "No, we don't want to do it." But why?

[00:04:50]You know, OpenAI is one of the leaders.

[00:04:52]They they have all this money. Why would they do it? It's because they were short compute. they weren't going to run or train the uh cursor models as well as as they would want or the cursor team wants. So they weren't going to offer you know here have you know abundant cap here,000 GPUs 2,000 GPUs not enough for them to be the thing more likely they would just suck in all of the intelligence all the developer capability and then put it into their codeex or whatever right they wouldn't really make them the the thing but they come to XAI X is further behind the coding side you know they could make the grock but they will give them all the compute they need to properly train their models and if the the models are good and they're making their own dedicated coding model. Then they would have the compute to run it, the inference part. So there's the training >> and the inference.

[00:05:46]>> XAI is getting cursor hooked onto their in AI compute power.

[00:05:51]>> That's right. That's right. It's it's more valuable than money itself because they could have some you know openi could have arranged to say I'll borrow more money you know meta here $100 billion right the cash was not enough they needed to say I need to be able to run my models and and and run my agents to do things so thus you need to have the compute I also want need to train my models so having the compute there to train and also a large data center where the training of it you want to have you know 500,000 a million chips all in one location to train the model better and again Xia has that they have the biggest one so you can do a full training run on 200,000 chips so that you can have a smarter model right so they can have the compute it's all co coherent shared memory so they can make their models a lot better and when it's done you can then run the model I have the compute I have the energy I can run your model right and and any model so having that extra capability is good and also going forward they XA will build faster continue to build faster on more uh data center and >> just as everybody else is going to be as well. Yeah. Yeah.

[00:07:20]>> So Right. Right. But then they got some numbers.

[00:07:23]>> Yeah.

[00:07:23]>> Yeah.

[00:07:23]>> No, complete that thought out please.

[00:07:25]>> They go next level with space AI where the entire world will be struggling to make more than 20 gigawatts per year and then they'll go to 100 gigawatts, 200 gawatts, 10 times more than anyone else can do all by themselves. So >> their their lead on um data center will be you know dominant 10x more than Let me bring up this this little chart with numbers that you sent me. This is fascinating. And before we go into that, um why explain for the audience why Elon's comment about Google winning the compute race in America, China, China winning uh winning it for in the world and XAI or SpaceX winning it in space.

[00:08:16]How does that now make sense when you look at these numbers?

[00:08:20]>> So Google has um the most AI data center you know not one location X actually has the biggest one location but Google has many many locations all across the US in the world right 20 30 data centers. They were data centers to run the search because they had the biggest profitable you know use case where search was supported by advertising and then they convert all those regular data centers over to AI data centers. you have to like you know get the right chips in there, the right memory you know to do things to make a regular data center into an AI data center right and they have about 9 gawatt of that data center now thus you have Google when you do a search they come up also with the AI response as well right so they have the most and they can even add more and they convert more over to AI data center thus you're looking at about like 20 24 gigawatts at the end of next year going up from nine. So they're, you know, doubling from a large base, right? And they're four or five times ahead of open xanthropic individually, right? So they have the most and because they're already built out everywhere. The issue is that we're hitting limits. It's tougher to add each new gigawatt, each 10 gawatts, right? So the fact that they already have the locations they have the advantage to you know convert what they already have into AI data center it's tougher for anyone else to keep building because it will take you years to do it them they will get to 20 40 gigawatts by converting what they already got slightly beefing it up and then everyone else has to build net new. If you have to build net new because they already have the contract with utilities to get the power then you have to um build a new power plant new power plant could take 5 10 years >> right if you build a new nuclear reactor >> right or and sevenyear gaps delays on the natural gas >> right so there's all these limits that stop you from building more on the ground >> right but then >> as Elon explained on the dwarves podcast the problem with the gas it's the problem with the gas turbines for for gas powered plants.

[00:10:32]That's the that's the bottleneck and it's massive, >> right? But if I go to 100 gawatt in 2030 and Google's gone to 40, now I'm ahead of Google. I won in space. I have more than Google. For China, China is right now building about 300 gawatt of power.

[00:10:50]Um 330 gawatt I think of solar, wind, and coal and whatever. and they have two to three times as much power as the entire United States, right? But it's running their economy, but they can still add another, you know, half the United States in power, right? Um >> easily >> and and and run it for AI, right? They can do it easily. Yes.

[00:11:13]>> China doesn't have these problems like America. Yeah. where you've got to worry about um you know local communities um kind of giving it the greenlighting these projects or environmental clearances >> or funds or anything the CCP says okay we need this here and it's there they just pump it >> right right but when when SpaceX goes to terowatts per year then even China fall behind because you have if I toss in three terowatts that's more power the entire world right so then I've thus gotten you know um three times more than China if I had more than China each year right so thus space is the ultimate winner on this power chip front and um deep analysis by um Dylan Patel who spent $7 million with his company semi analysis which is equivalent like a hundred or a thousand researchers confirmed that no AI bubble that not only that that the demand is like five times what we can currently Right. So the margins will go up and you know we will need all these power and chips.

[00:12:22]>> I I mean I guess like most of the fears about the CI bubble were centered around open AI >> but as you see the field is far wider and larger and the players are far more numerous and important than open AI. Now the landscape has changed right dependencies have changed. So I think even if open AI goes public um and their returns aren't what are expected them to be. Right. I mean sure there's going to be a bit of a dampener but the picture is a lot bigger.

[00:12:58]>> Mhm. Right. Right.

[00:13:01]>> And also uh the fear of open AAI like let's say they fall behind you know they're falling a bit behind anthropic now. Um, so from number one to number two, although you know they'll they're trying to leaprog back with, you know, GPT 5.5 and do other things, you know, they'll they'll struggle for this year, but I think the writings to me seem to be on the wall that they'll fall to number two and then if XAI and Google can can um execute, you know, they can go to number three or number four, right? That it's like um a shark, you must always be moving forward, otherwise you die, >> right? Yeah. So if they stop moving forward as as fast as the others, then they will die. They'll be starved of money, starved of compute, and then they fall to the wayside. But for people who say that's a doom scenario for AI, they're only like 5 10% of the compute for AI, right? So as big as their 2 g, you know, like Google at 9 gigawatts, all these other things, Amazon at large numbers, and then as things move forward, you know, it'd be a bump on the road. And also it's not like they sell and then oh we destroy all the value in OBI. All the pieces will get ripped apart. You'll probably even sell out at at um at a decent number where basically I will buy your unused lease on this data center. I'll take it over >> because you're you know you can't use it, right? I'll take all your people the people already going, right? So then it you know the like the shark will rip a aart the carcass and then it'll all Everything's got to be recycled.

[00:14:37]>> Everything can be recycled. That's right.

[00:14:40]>> What else can we take away from these numbers?

[00:14:42]>> So, so the main thing is that um the the fresh companies in AI, OpenAI, XAI, anthropic, they're all all about two and in meta there too, all in the two gigawattish type range. You know, bit ahead, bit bit behind. uh te Xi's advantage is that one of their gigawatts, half of their compute is not um fully utilized because they don't have enough um um users on their thing.

[00:15:10]But that's that negative can be turned to an advantage because that can then do deals. The other thing is to look at Q4 and basically see that um all the main players are trying to doubleish or more obviously trying to do more to get to about four to five gigawatts you know for you know open anthropic Google gets up about 50 60% up to about 14 gawatt Microsoft gets to 10 you know uh Amazon gets about 10 so the other guys are moving faster but the traditional big players still have advantage for a while right and then you get to the 20 gawattish type range 2027 7 or 2028, XAI can leap ahead from its peers by going toward 20 20 gawatt, 30 gigawatts by converting superchargers because superchargers, there's 80,000 now and then they're, you know, doubling every 3 years or or maybe even doubling every two years and then they have 7 gigawatts there now and then that's going to go to 1420. that amount of inference if you add AI to every location, right? Then it starts to get about as big as the Google number. So they, you know, XI can quickly move toward the lead of the pack on Earth with with their um with their compute. And it's all >> here here's here's the opportune moment for me to ask you a dumb question, but I'm I'm doing it because I think a lot of people will need um a bit of clarity >> when you're talking about computing at or distributed computing at superchargers and even the cars. Uh Elon has talked about that multiple times.

[00:16:48]>> This is for inference not for training.

[00:16:51]Right.

[00:16:51]>> Right. Right.

[00:16:52]>> Can you explain why?

[00:16:55]So in order to do training, I need my million or 10 million chips all in one spot ideally. Um if not like super close together so that I can have my whole brain train up and learn. It's like you know the AI data center is like training thinking learning what to do making yourself smarter right if it's all separate if I take my brain and split it up into a thousand pieces I can't learn efficiently right so based on the nature of the models I cannot do the training and learning by having things all separate it's it's tougher so based on how the models currently work you need large data centers to do it um And then for the inference answering a question, I don't need to to know all world knowledge to answer your question. I just need one piece of it, right? And I also only need to think about it for a few seconds with maybe a 100 chips at most or you know sometimes you can even have a local model where it's only on your laptop or only on your Mac Mini and have like 100 gigabytes. So it doesn't take that much to answer a question once the intelligence has been made, right?

[00:18:05]But we still you need to have a certain amount. But it's plenty at a one megawatt or 100 kilowatt supercharger, right? I have a a half rack, a full rack of 100 GPUs there. So I got 100 GPUs. I got like a terabyte of memory, right? I can answer pretty much anything, right?

[00:18:29]So thus inference which will be 90 99% of all the data centers because people aren't paying you to train to get smarter. They're training you to get the answer and to do things. So that's the inference the doing things answering questions solving problems. That's the inference side.

[00:18:52]And so if I get you right, it's all about coding right now.

[00:19:00]>> Yes.

[00:19:01]>> If you look at whether it's Codeex or whether it's Cursor or Anthropic or even XAI, what they're trying to do with coding, why what gives why is coding so important? Is it a means to an end or what's going on? Explain it please in layman's terms.

[00:19:19]>> Sure. Um so one um when you look at um Gartner which uh tallies up how much the um computer what they call the information technology industry is the information technology industry according to Gartner is $5 trillion. So the whole world economy is about a hundred $120 trillion right $5 trillion of it is IT or or the computer software and consulting industries right so roughly one-third is your hardware you know all the the data centers we've been talking about also the the PCs on our tables or in our pockets and then onethird is software and then one-third is consulting right or developers or whatever, right? So that's the rough numbers. So then if I solve compute, my addressable market is at least the software and consulting portion about $3 trillion, right? And maybe some of the hardware and that's if it doesn't grow.

[00:20:24]But if I make it grow double or something like that, it's goes from three trillion to $6 trillion. So the $30 billion that that um that uh Anthropic is making now on a run rate, right? That is only 1% of the $3 trillion market, right? So and it's half a percent of a doubled market. So we know that is a huge huge market more than the answering questions market that kind of corresponds to search. search Google is like hundred billion dollars a year, >> right? So then if I'm making $20 billion a year >> on my hundred billion, I'm already at 20% of that market, but I got cap out at about 100 billion. Let's say it goes double two, you know, two 200 billion, right? It's still less than 10 times, you know, it's a little less than 10% of what the IT industry is. So it the compute side is a bigger industry. it's more valuable also more margin and stuff like that. So one just by the market >> where does knowledge AI fit into all of this? I would imagine it's somewhere in the middle because it's more a lot more valuable than just search >> but not as much as as coding.

[00:21:43]>> Yeah. So that goes >> for research or drug discovery or discovering new physics and all of that.

[00:21:48]>> Right. Right. So that's that goes into part of it is R&D. You know R&D is again a few hundred billion dollar market but it's more valuable because it has a multiply effect. When I give you the answer of an ordinary question that's it you don't pay for anything else. If I give you the answer in R&D then that unlocks other things right more spending happens more work happens you know you generate more more value right so >> so the reason I ask you this is because a lot of of what SpaceX AI is going to be doing with AI is R&D as well >> right right so the other thing is the the ultimate tight loop of value creation is AI research if I can create a faster faster AI research a better AI research right so yes there's a general research but not all things are equal answers and value creation and productivity around AI research that is the the main flywheel if I make that faster I win at everything else it's like I'm you know creating this nuclear chain reaction of stuff where it's just going to explode and take over right so the more I can create on that particular piece so that that goes to your point of The most valuable piece if I speed up my innovation because my research is faster as Elon says innovation speed of innovation is everything. So if I directly impact my speed of AI innovation so that goes to the cursor team as well. Yeah, that brings us full circle and I was just about to ask you so how much of >> this innovation and R&D and these factors play into this aqua hire of cursor by SpaceX.

[00:23:34]>> Yes. So that is ultimately everything like one there's the tactical thing about making more money and doing all that kind of stuff because if I can't keep moving that goes to the the shark part of it. I can't keep my shark moving, keep moving faster. But the whole game is won by going exponential where I have self-improving AI, right? So if I make self-improving AI, I make the self-improvement faster and faster, then you win. It's basically is the speed of my explosion, the energy of my explosion faster than that. So it's like a chemical bomb versus a nuclear bomb.

[00:24:14]chemical bomb maxes out at, you know, 3,000 degrees. It go only goes so fast on terms of the explosion. The nuclear bomb, it can go to a million degrees, 5 million degrees. It can be a thousand times, 10,000 times more powerful.

[00:24:30]Right? So then the speed of the explosion is what m so it's goes to physics of like how fast can I make this innovation speed of intelligence explosion happen. So they're working towards speed of intelligence explosion, right? And thus you end up winning. So you know the amount of computing stuff is kind of like how much burnable material do I have? I'm making a giant forest fire. I make more forest. I have more stuff, right? But the intensity of the fire is also there. So then if I can set the entire in analogous way solar system on fire with intelligence where I make all this a million a billion times more compute that would be a more powerful thing. It it goes to >> you know foundational physics of what I'm trying to do but it's like looking at it at an equation of speed intensity and size. Yeah, that's where Elon's reference to Kardashev level civilizational scale comes in.

[00:25:35]>> That's right. That's right.

[00:25:37]>> And to that and that kind of brings us back to this graph or this chart that you have.

[00:25:42]>> Yeah.

[00:25:43]>> This is all Earth.

[00:25:45]>> This is all terrestrial.

[00:25:47]>> The game completely changes when you go into space. And that's where XAI or SpaceX AI will really flex its muscle, >> right? But it, you know, like if you simplify the numbers, you know, the regular players at two gigawatts, end of the year they're at four or five, they're you know, a little faster than doubling and then try to double again.

[00:26:07]So that's the exponential double double double, you know, in one year, nine months double, right?

[00:26:12]>> Yeah.

[00:26:12]>> So then if I go off Earth, then I can, you know, 5x 10x every year, right? And and that, you know, going 10x 10x 10x is like, you know, 10 times, 100 times, a thousand times. just like it just goes really insane really really fast and it's it's taking that you know exponential curve to the next level and if you can't go to space you know you're at a slower um pace >> yeah because the limiting factor is launch capability >> right launch capability that's right and also you know building the chips which is a teraf app right >> sure but but then that's >> yeah it's all it's all same thing yes >> yeah it's all the same thing that's what we were talking about. Okay. So, so, okay, that that that explains, I think, pretty much what the cursor deal is all about. Then you've got this new AI infrastructure u startup or company, Stealth. It's a mystery one. $2 billion uh spent on spent on it and acquiring by Tesla, came out in the 10Q.

[00:27:20]What is this all about and why? because you have the deal the tie up with Intel for the gallium nitride gallium nitride technology the 14st strong unit chip tech that Intel has.

[00:27:33]What could this mystery company be or startup be bringing to the table?

[00:27:40]>> So it could be bringing some new technology to make the AI data centers um more efficient or faster to build. So or just is a team you know a more mundane thing would be just be a team of people that Elon needs is something that's removing a blockage of some kind right if it is um say photonix related right >> then that could reduce the power you need by half where basically you move um bits and bites and and and numbers around within your model but you do it more efficiently. So instead of over copper and getting heat and losses, you do it over optics and fiber and then you can get >> you're using light basically, >> right? Using light >> and then you're only limited by the physics of light.

[00:28:29]>> And then there's the there also the advantage of heat management.

[00:28:32]>> Right. Right. Because because it's more less losses losses are heat. So that's a problem because it it slows things down.

[00:28:40]It also wasting energy. So it's bad on multiple levels. So you don't want to do that. You want more efficiency and less heat uh to do things.

[00:28:49]>> Okay. So, Brian, can you just maybe backtrack a bit, explain what photonics is all about and why it is so promising because it's nothing new. I mean, we we've seen photonics uh chip technology being developed for the past few years, right? But it seems to be maturing now.

[00:29:07]So, it's it's become really important.

[00:29:08]Everybody's talking about it.

[00:29:09]Everybody's interested in it. What is it all about? Why is it so transformational?

[00:29:15]So if I can go from electrons inside of a chip, you know, semiconductor, moving electrons and doing stuff there, the charges and all that kind of stuff, and then I go to photonix um to move the data around, it um I can move it faster. So instead of, you know, 100 gigabits per second, I could go to 10 terabits per second. So I can go 100 times faster, right? Um, and then you can send stuff down the fiber and it's just it uses less energy because it's it's again just photons instead of electrons. So all the physics of it is potentially 100 times faster um um lower energy um less less heat. So you know among the various things it's better and we've been trying for decades to do more and more with photonics. So if you can make an all photonic computer and people are companies start to try to do that then theoretically you could get to this 100 a thousand times 10,000 times better computer but that has been too hard right we have not been able to get that transfer over to happen right uh so then they do is they they want to get as close to the edge of the chip as possible where instead of sending out your signal from your from your chip into another chip to convert it over to photon I want to do it right from the edge of the chip. So I get closer and closer to the edge of the chip. If I can't do that, I'm getting, you know, photonics, you know, communication like we have the network communication between racks and between data centers, that's all fiber.

[00:30:53]That's all photonics, right? But the more I can add closer to go within my rack and have communication within the rack because basically about 80% of the rack is networking and heat management and all that kind of stuff.

[00:31:10]>> That's what CUDA is, right?

[00:31:12]>> CUDA is a software for for running the parallel um Nvidia software to >> what's the Nvidia the uh the the tech Nvidia technology that links >> there's Yeah, the NV. Sorry. Yes. Envy link >> envy link right and they're converting that more and more so that's you see that oh look instead of the whole thing I I got this big pizza thing you know craft America shield so that was them converting things components that were electronics into photonics or into condensed things to make it smaller more compact less heat >> so you're essentially going from electron to photon to electron right now >> right right >> but if you can just photonix right through that just completely changes the game.

[00:31:57]>> You can go 100x. So right now we're getting about 2x doing you know all these photonix companies going $10 billion or whatever like that. Those dozen companies are helping to get us to this 2x gain, right? But there is a potential another 100x if you can go all the way but it's hard. So you bite it off step by step and then there are people trying to go the long shot of like doing the whole thing >> because Elon has spoke to spoken about um exploring new physics with the terrafab new physics for chips right other than photonix what else qualifies as new physics for chips.

[00:32:38]So um for um 30 years I I've been involved you know full closely following molecular nanotechnology. So molecular nanote technology you know in the ' 90s was really hard >> by the way you know for everybody watching the the I think he Brian is just one of the very very very very few guys who I can ask these questions to okay and he can just explain it so beautifully. So I'm so grateful to have this chat with you Brian. Okay, please.

[00:33:07]>> Yeah, no worries. Right. So, so molecular nanote techchnology is moving atoms uh molecules exactly where you want. So, and then people talk about, oh, I want to assemble at molecules exactly where I want. If you actually can do that, then you could if I can move carbon atom exactly where I want, I could make um diamond everything. So I can make a diamond chip that could have you know 10 times the performance, 20 times the performance and it be the transistors would be you know some quantum effect thing that would also be highly precise. So if you can get to that then you know the level of technology goes up by about that but it is physically possible. it's within the realm of physics to do this thing. And I have um tracked companies that um have continued to do this work. There was a big controversy back early 2000s like 2000 2001 where a Nobel Prize winning chemist said that this could not be done. But actually that whole thing and I you know I've talked to the people who who who were involved in it was just that um the the lead guy who was the face of milk technology Eric Drexler uh had said things that was pissing off the Nobel Prize winner basically saying hey your work is no good everyone should stop doing what what you're doing there all your researchers can come over and do what I think is should be doing. So there's a bunch of egos involved and then they said children so it >> children yes uh it was also there was money that was supposed to go into technology and that got diverted over into Intel and other regular technologies because the politicians don't understand real groundbreaking versus you know not so groundbreaking they understand the difference right um because if I can control things at the molecular level a molecule versus only um a um 10 nanometer structure, you know, or 100 nanometer.

[00:35:18]100 nanome smaller than 100 nanome is what their definition was. If a a nanometer, if things were a billion times bigger, your 100 nanometer cube is a size of a soccer stadium that is 100 stories tall. Um your 1 nanometer structure is a size of a of a dishwasher. And then your molecule is a size of a baseball or softball, right?

[00:35:42]So it's if I can assemble everything precisely within my um the volume of my um my football stadium very precisely, you can do what appears to be magical things, right? So um by going toward that you can get to the limits of physics and the limits of physics are more like a million times better than where we are now at least and then if you're really clever maybe a billion times better right so there is a lot more capability that can happen more than just only a hundred times gain by perfecting photonics right so the thing is these aren't new physics it's just mastering to the level of known physics right if you can actually then get beyond that then you know things like you know there's certain violations say if I could violate here then I could have faster than light travel I can or I can get to easily get to light speed so the level of technological magic that is possible by going up to physics going beyond physics is is huge and I happen to have you know studied you know going up to the level of the physics where you get the most bang bang for your buck and then going if you could go beyond or certain areas where it seems like we might be able to go beyond that just in those areas there's there's huge gains to be made. Um and then also >> when you have the most powerful AI, you can do all of this R&D and you can break down barriers and invent new physics or again discover new physics or invent ways to discover new physics >> and and for for in terms of the lack of possibility. I know that that um research teams have moved um carbon atoms and carbon dimers. So the people who said that it can't be done hard physics it has been done and they can do it like a hundred times a month right where basically you can move these molecules around but right now it's on a 2D plane uh and they have to go to 3D and stuff like that so it's still hard but the thing is there is significant level of capable work that has already been achieved it was hard one work over 20 years right but um but they've kept they they were kept it stealth and secret and blah blah blah right but in terms of like the capabilities there's there there's also capabilities around you know the carbon nanot tubes that we make those are also molecularly precise you know where you have carbon atoms into a long fiber so >> um >> and if I can make everything out of carbon nanot tubes >> the space elevator going from the ground >> to to space you could do that right so there's >> um and materials become like 100 times stronger you know 50 times stronger so so it's just um the what we can do just goes crazy once we can push up and actually make progress on some areas that have been difficult up to up to now.

[00:38:42]>> I see so many parallels and similarity with battery technology and the evolution of battery technology.

[00:38:48]>> Right.

[00:38:49]>> Cutting edge. Yeah.

[00:38:51]>> Well, the other thing is that like if I make a a perfect battery molecularly precise, right, I can make it 100 times better, right? and all the fuel cells. I could if you were to make a fractal fuel cell where because right now you have you know your anode and cathode separate, right? Just a simple separate thing.

[00:39:09]>> Yeah.

[00:39:10]>> But if I was to make them into some kind of fractal thing where they're kind of like all interweaving and stuff like that.

[00:39:15]>> Yeah.

[00:39:15]>> That the the power efficiency of transferring it all around because all these contact zones where it's happening.

[00:39:22]>> Yeah.

[00:39:22]>> It it could release you know 100 times the power, right? So just by getting control at a far more precise level and be able to assemble it means that you can get these 10x 100x gains in power and efficiency. So then instead of needing 200 kg of batteries to drive my car, I drop down to 10 kg, >> right? And I can, you know, so it's >> it's just and all the weight of the car >> bananas five times 10 times lighter, >> right? So then my car might and and then with that much power I can make a car-sized vehicle that could fly into into space, >> right? Because with that much efficiency, you know, the space flying personal go into space vehicle becomes >> what was what was the number that Elon thought would be necessary for battery uh energy density to the threshold to cross for it for uh electronics or for planes. Yeah, >> planes. How about 500 watt hours?

[00:40:25]>> 500 watt hours. Yeah, >> we we can roughly do that now. But um that would be a version where you have a bunch of small engines around it. So you have like an array of small engines to do it. Um if you have even more um you know a thousand watt hours per kilogram then you can make a far more elegant far more powerful design. So kind of like barely possible at >> we're already seeing it with uh with Archer and Joby Aviation. Right.

[00:40:53]>> Right. Right. Right. But you can have >> they they haven't got to >> Yeah. They have to get to scale and and you really want to not be just barely good enough. You want it to be like, >> you know.

[00:41:06]>> Yeah, that's what I was I was about to say. You could have larger aircraft with a longer range.

[00:41:11]>> Right. That's right.

[00:41:12]>> And that transforms aviation as well.

[00:41:15]>> Right.

[00:41:17]>> Fascinating.

[00:41:18]>> Mhm.

[00:41:19]>> What else? Um so when I look at at the hardware section um so you have this overarching deal uh with Intel and then you've got this uh AI stealth startup that was bought.

[00:41:33]>> These are both on the um the chip technology hardware infrastructure stack side, right? what what sort of um ho how does how do these two this deal and this agreement with Intel feed into each other because they're on the same infrastructure stack side >> um >> you see any >> the Intel side of thing like um speeding up the um um and having alternative to the um advanced fab the terra fab lab Right.

[00:42:09]So, the Terapab Lab, that's what's being built in um in uh Texas now, right?

[00:42:16]They're breaking ground. You know, they have it leveled out. They're going to build that as fast as they can. So, Intel is the closest one to having the same vertically integrated lab, right?

[00:42:26]They have a pretty integrated lab already. So, they can work uh XAI can work with Intel at their lab for fabricating and designing chips, right?

[00:42:38]So they can go to them right now and start designing chips better by by working with Intel and Intel has people the team the top researchers who can who been there done that know what to do right so tell the team so basically it's a ready to go let's get going as best as we can now and so then that that speeds up the lab work and then for designing the lab they can talk to Intel and say here we're looking to do this here we're looking to innovate they can say that will definitely not work. Here's what you have to do. And then they say that might work. We can try and test that. So again, all of the the learnings is is you know, you want to talk to the experienced people.

[00:43:23]>> So they they get a lot of that stuff.

[00:43:25]And then also knowing who else to talk to. Oh, you should really talk to whatever, you know, then they they can bring in the rolodex of people. I have a friend with so- and so. You need to talk to him. You need to hire this guy. Um, and then they're also talking to all of the suppliers, Tokyo Electron, ASML, um, you know, the entire >> Yeah. So, with regards to ASML, there was a, excuse me, there was a bit of confusion over. So, ASML makes a lithography machines, right?

[00:43:56]>> Right.

[00:43:56]>> And these are whatund 150 billion dollars a machine a pop each, right?

[00:44:03]>> And they produce a limited number of machines every year. So there was a confusion over just how far down the line ASML is production is completely booked out. But I think there's now uh a new kind of sense that they booked out to 27 but not 28 and there's still room to grow. But then when Elon pumps money into it and says as you rightly said in the past, hey just take this these truckloads of money and give me what I want. That's possible.

[00:44:34]>> Right. Right. Yeah. it's it's booked out, but you know, you pay an extra 20 30%, you jump the queue, right? You know, I I overpay for the ticket and and I get get the cut in line, right? And if I also do stuff like I'm going to >> fund your ability to make even more, >> then, you know, give me special treatment and let's work together and then um we're going to make an extra line. They'll start producing some stuff. we'll speed up this line and then now but I want to have half of the the gains of the line because I helped you make that faster and better. So I >> So Elon had it Yeah, sorry. Continue.

[00:45:10]>> I made the pie bigger. So give me two two slice of the pie.

[00:45:14]>> Right. So Elon had this deal with Samsung with where Samsung said Elon could walk the line. Right. Right.

[00:45:20]>> And learn a lot. What do you think he's gained from that?

[00:45:25]Um it it's he and his team are learning everything about what exactly works and and they're getting a um a boot camp um in you know crash course in all the details of what actually happening there. The reason I ask you this is because when you look at the terapab and Elon's plans for the terra fab, he famously said that he's going to completely revolutionize and rebuild or redesign the terapab and what a fab should be from the ground up from first principles for from a first principles approach and in a way that you could have a cheeseburger and smoke a cigar there inside.

[00:46:08]>> Right. Right. So all of this when you when you when it feeds so the experience of walking the the Samsung um fab line Intel's experience this AI startup all of it feed feeding into this innovation engine uh that Elon's running for the Terrafab. How does that play out?

[00:46:32]So, um, Elon's getting, you know, I think the Intel CEO said, um, it's the best thing that ever happened to the semiconductor industry for Elon to get involved in it.

[00:46:44]is because you know Elon got involved in the space industry and then he revolutionized it with you know reusable rockets and will be able to he's already 10 times bigger than anyone else in terms of launch and capacity and everything and it's going to 100 to a thousand times more right because he took what was there and pushed it where it could be pushed so it can go exponential and go way more so he's going to take the apply the same thing and he's doing it to Boring Company also did to electric cars cars, you know, electric cars, the whole industry was 100,000 times smaller than it was today before Elon got involved, right? So, he can take it and make it successful and push it. And semic industry is already foundational to all of our civilization, but he's going to take it and push it to extreme levels. Um, so one is just doing the boring company, just doing the electric car thing of like taking all the stuff that works now and making each equipment in the fab line faster and better to stress everything and push it and see where the weakest link because you have like, you know, 100 different machines inside your fab and he's going to say, "What's slowing me down with bottlenecking? How much can I push it?"

[00:47:59]really this current piece. Can I dial it to 11 and not break it? Can I dial it to 13 and not break it? Okay, all these are running at 13. We figured out how to get them to run at 13. This other one thing is running at 9ine. I got to figure out how to make redesign it to make it go 13 or put two there. So then my overall production can go up by 50%. So you can apply that everywhere to to boost the the capacity of everything. But then he also has unique um use case problems where he has to redesign everything where it's like I must redesign my chips so they run at 300°.

[00:48:39]I'm not doing that now. I have to do it when I put in space so I can make the radiators more more effective. Right?

[00:48:45]There's a physics thing that when I change from here to there where I need to um be able to handle the high temperatures. So that means they need to redesign all the chips by default. I cannot just take your regular chip. I need to make it run hot as close as possible to hot so that uh how I need so that it can work better in the environment I'm going to put it in.

[00:49:09]Right? So that's another thing that has to be done. The other thing in terms of chips in the design is that when I stick it on the moon and accelerate off, I probably want to accelerate at 100 times 100 gs, maybe a thousand gs. So basically firing out of a big cannon, right? So again, I need to change the chips so they can handle that kind of stress because I need to make that happen, right? Higher temperatures, higher G's, right? Those two big changes. The other thing is that I need to take my my EUV machine, my fab, all my production equipment, and I have to stick it on the moon. Not right away, but in, you know, 10 years, right? I need to stick it on the moon because I could be mining, I'd be making all my chips, all my solar on the moon, right?

[00:49:54]Because in order for me to go to a pedawatt, a thousand times more than a terowatt, everything has to be made on the moon.

[00:50:01]Everything related to the satellite and the chips and the air data centers has to be on the moon. Otherwise, if I'm sending it up from Earth, I only get a 10x 20x bump. I make the aluminum. I make um uh solar cells on on the moon. I can make the silicon on the moon, but I don't if I don't complete it, I'm still taking 10% to the moon. I can only get a 10 10x gain, right? Because I'm only making 90% there. So, I have to make 100% on the moon. So that means that I make my supply chain for these things not for the entire world just for these specific things of solar power, silicon chips, satellite bodies and mass drivers. I need to make that on the moon. So if I simplify that as much as possible, I make the the problem easier.

[00:50:51]But we will make a hundred teraps on a thousand teraps on the moon in order for me to get to pedawatts. That is a thousand times more. Yeah, >> I will need to make 1,000 fabs on the moon and then make 1,000 times the number of satellite factories on the moon, right? In order to get to this larger goal of making as much um solar and chips in mass as we make steel. We make a billion two billion tons of steel per year in the world. We make need to make that in chips and satellite from and solar panels from space on the moon to get to that next level. So that is um means that we need to take all and that also means we'll need a thousand times you know 10,000 times more equipment than from ASML from Tokyo Electron the entire uh semiconductor industry must be a thousand times bigger if Elon >> here's here's here's a thought if Elon takes and as Elon is want to do and he's done on so many occasions including Neurolink and transform the way you know everything happens. He it's I mean I expect him to take these lithography machines from as ASML and also redesign them from the ground up. So then um like uh the Giga Press that came out of the factories in Italy, right? Um nobody wanted them. Uh you had this brilliant technology till Tesla just went headlong into it and now all other car auto manufacturers including the Chinese are following. Elon could very well just reinvent the lithography machine based on his use case and then you could have I mean I don't know who within SpaceX/Tesla in the merged entity would be manufacturing these or would ASML then be given the job of manufacturing these >> will be both. It'll be both.

[00:52:56]>> It's both because you have um who's making the um all the batteries, right?

[00:53:03]CL makes a lot of batteries. You buy more batteries from them, but then he also makes 4680, right? And then where if the company, you know, starts not being able to hit the goals that Elon needs, then he'll say, "Okay, I have to replace it. I have to do it by myself."

[00:53:20]which he did more in the satellite world because he makes, you know, the most satellites, 4,000 a year, 4,000 satellites a year. No other company could do that, right? So then they had to vertically integrate and make more supply chain by themselves where they could buy from someone else where they could keep up. They Okay, fine. I'm buying from you. You're doing it. It's good. It's quality is good. Volume's good. I'm keep working with you. But then you can't do it. I got to do it myself. I I wanted to work with you. I mean just with the terapab the terapab is a case in point >> you can't get enough chips so he has to build it a fab himself right >> right if you could do it yourself I would let you do it yourself because I still got a lot of other problems to solve right yeah but he can't do it why got to do it myself but you know if you can >> let's let's look at the big picture now Elon vertically he loves to vertically integrate everything right >> and because of the developing or evolving geopolitical situation across the world onshoring or slash French shoring but when it comes to such a critical um you know due technology and of strategic interest to America I suppose he would onshore as much of the entire vertical stack as possible >> start to finish >> but the thing is >> so then it's possible that ASML would be forced to come to America and build these lithography machines on American soil as Right. They would have >> Yeah.

[00:54:51]>> Yes. Everything would the supply chain shorten up and put it all in one spot.

[00:54:55]Why is that critical, not optional for uh for Elon and SpaceX and Terafab is because I need to put it all in one spot, shrink it down as much as I can because I'm gonna launch it to the moon.

[00:55:11]>> Exactly. So, if it's all over the world, I can't launch all that stuff to the moon because it's like here and there and all these pieces, supply chains all huge. I need 5,000 different factories.

[00:55:22]I need to condense it into 10 factories, one factory, and all the spots. So, I can just shove them into my starships, take a thousand starships and put it on the moon, right? So, because end goal 10 years, 15 years, I want to have these fabs on the moon. So that means the entire supply chain has to be nice and compact, loadable onto a starship to take it to the moon because that's where it needs to go. It can't be here. It's got to be there. So I need to make it smaller.

[00:55:53]>> Like the boring the Boring Company drilling machines, >> right? That's right.

[00:55:57]>> You need to get into the starships, >> right? Not only the equipment, but the factory has to be movable.

[00:56:03]>> Yeah.

[00:56:05]And also I need to make the factory movable so that I have fewer things to mine. If I have a thousand different things to mine, I need this element, this element, that element. If I can shrink that down to 10 things to mine, I only make 10 mines on the moon. Yeah.

[00:56:22]>> Or only um you know, instead of 100 mines or maybe I have to go find 10 other things that aren't even on the moon. I have to go go to the asteroid. I could go other things. That makes it way more complicated. Not just I'm shipping it in from um from Saudi Arabia to to Dallas. No, I got to like I gota ship it from um you know asteroid um you know series over over to to me and I got to like travel twice as far as as from Earth to Mars, right? I got to go like 200 million miles to do it. So supply chain problems are really bad in the solar system. I need to like concentrate it up.

[00:57:06]No >> Oh boy. Okay. So I this has been a brilliant chat but >> I'm hoping to so we've for our audience we've laid out um the the highlevel view the 35,000 foot view of the terrafab and everything that we know about it but I'm hoping Brian that the next chat the next deep dive we do on the terab will be a lot more technical >> I think we should work together and and layer it and because this is a great primer for everybody watching. It gives them the lay of the land. It explains all the moving parts, why they need to come together, in what way, the geopolitics of it, how everything ties up into each other. Um, and then this is it's not a pipe dream. It's reality and it's already in motion.

[00:57:58]>> That's right. That's right.

[00:58:00]>> Next time we got to do a deeper, more technical dive, and I'm looking forward with you on that.

[00:58:06]>> Thank you so much, Brian. Always a pleasure. Till next time. Bye.

[00:58:09]>> Bye.