Musk and Intel build orbital AI

Elon Musk's companies Tesla, SpaceX and XAI are partnering with Intel to build TerraFab, a 100 million square foot semiconductor factory designed to produce one terawatt of computing power per year, which is 50 times the output of the entire global chip industry combined. Yeah, and this facility in Austin, Texas is designed to consolidate the entire chip making pipeline. So, you know, design, fabrication and testing, they're all under one single roof to power billions of humanoid robots and a massive network of orbital data centers. How does putting chip manufacturing robotics and orbital rockets into a single closed loop completely alter the physical and financial limits of artificial intelligence? Well, TerraFab utilizes this like recursive loop where silicon design, lithography mask, memory production, advanced packaging and testing all happen in the exact same building. And Intel is actually bringing its 18a process node to this facility, which features ribbon FET, so gate all around transistors and power via backside power delivery. Because traditional chip manufacturing requires, you know, shipping wafers and designs across the globe between different specialists. But with Intel's power via, the power routing is hardly moved to the back of the wafer, which physically separates it from the signal wires to reduce electrical interference. And that changes the iteration speed entirely, like instead of waiting on external foundries like TSMC or Samsung, TerraFab can just manufacture, test and redesign chips continuously. Yeah, you aren't waiting on cargo ships. Exactly. It limits their reliance on vulnerable global supply chains, opening up a fully sovereign domestic manufacturing stack. I mean, it is basically like writing a book by mailing individual chapters back and forth to an editor overseas, whereas TerraFab is having the writer, the editor and the printing press sitting at the exact same desk. Hold on, let's look at what is actually coming off these assembly lines. Good point. The facility produces two highly specialized chip families. You have the AI5 and AI6 chips for terrestrial edge inference, and then the D3 radiation hardened chip for space. Yeah. And the AI5 and AI6 chips will power Tesla vehicles and the Optimus humanoid robots with a target of producing like one to 10 billion robots annually, which is just a crazy volume. It is. But 80% of TerraFab's output is actually dedicated to the D3 chip for space. And the D3 is uniquely designed to run hotter than normal chips. Right. And designing chips to run hotter changes the physical constraints of satellites. Oh, so. Because it limits the need for heavy cooling radiators, which opens up the ability to pack way more compute power into lighter payloads for orbit. Wait, back up. Yeah. Why send 80% of the computing power to space in the first place? Because SpaceX has filed an application to launch one million AI data center satellites into low Earth orbit. And Starship is designed to lift 10 million tons of payload annually just to achieve this terawatt of space-based computing. That is a staggering amount of payload. Right. But the physics behind that strategy dictate the move. Earth's power grids are severely constrained, and finding energy for terrestrial data centers is increasingly difficult. Yeah, that makes sense. But in orbit, solar panels receive five times the energy they do on Earth, unattenuated by an atmosphere or day-night cycles. Which changes the fundamental economics of energy consumption entirely. Because it limits the need for battery storage or grid approvals, opening up a future where space-based AI compute actually costs less than terrestrial compute. But I have to express some skepticism about the logistical reality of the supply chain here. Really? Yeah. I mean, there is a severe global shortage of helium, which is a critical element for chip cooling and processing, mostly due to geopolitical conflicts. So how does an operation launching 50,000 rockets per year avoid buckling under raw material shortages? Oh, you are entirely right to point that out. Acknowledging the extreme execution risks involved is completely necessary. Securing those materials is going to be a massive hurdle. Okay, we've covered the physics. Let's talk about the money. Yeah, so SpaceX acquired XAI in a $1.25 trillion all-stock merger. That unifies the Grock AI engine, Starlink, and orbital rockets. And to fund the estimated $20-25 billion cost of terrafab, SpaceX is targeting a $1.75 trillion valuation in an initial public offering. Wow. Because building 100 million square feet of factory space and securing high-NA EUV lithography machines requires unprecedented capital. And simultaneously, Intel gets a guaranteed high-volume anchor customer to validate its foundry business, which has been operating at a massive loss. And that financial convergence changes the structure of the AI industry. It limits Tesla's standalone financial exposure to these massive infrastructure costs, opening up a shared capital pool that funds the entire hardware and software stack required to pursue artificial general intelligence. Exactly. It completely distributes the financial burden across the entities. So by merging Intel's advanced 18a manufacturing with SpaceX's launch capacity and XAI software, terrafab physically bypasses Earth's energy constraints to build a vertically-integrated AI infrastructure in orbit. If human intelligence is eventually dwarfed by a network of orbital AI satellites powered by constant solar energy, who governs the physical infrastructure of that intelligence when it sits permanently outside any single country's borders? If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this. We appreciate you being here.