END Of Nvidia! Elon Musk's TERAFAB Just Reset AI Race, Dojo 3 | Jensen Huang SHOCKED

Kind: captions Language: en Jensen Wong from Nvidia said he's never seen anything built so fast in his life before. >> What would happen if Nvidia's biggest customer decided to never buy its chips again? Tesla may have just laid the first brick for that scenario by owning one of the world's largest chip manufacturing facilities, Terapab. The project is said to span over 100 million square ft where chips as advanced as 2 nanometers will be designed, manufactured, and packaged entirely inhouse. Tesla, SpaceX, and XAI may soon no longer depend on GPUs like H100 or Blackwell, which could mean Nvidia losing billions of dollars in annual revenue. But the most controversial claim is that Elon Musk's AI5 chip could deliver performance comparable to H100 at just onetenth of the cost. A figure that if true would force the entire industry to rewrite the rules. The truth might surprise you. Nvidia does not own any chip fabrication plants. Its headquarters in Santa Clara spans only about 750,000 square ft. Rather than manufacturing its own chips, Nvidia designs them and outsources production to companies like TSMC. Tesla is choosing a harder but potentially more dangerous path, controlling the entire value chain and pursuing extreme optimization for each application from self-driving cars to robots and satellites. So, how could Terafab reshape the global financial landscape? And could this mark the end of Nvidia's era of dominance? In [snorts] today's special video, we break down Elon Musk's new nuclear weapon. Why is terrafab a real threat to Nvidia? >> We either build the terrafab or we don't have the chips. And uh we need the chips. So, we're going to build terafab. >> Yes, Tesla is still one of the world's largest buyers of AI chips from Nvidia to run massive supercomputing systems for training neural networks behind full self-driving and the humanoid robot Optimus. Tesla has invested billions of dollars in high performance chips like H100 and the latest Blackwell architecture. That's why Elon Musk has acknowledged that without the computing power provided by Nvidia's hardware, Tesla's AI development would slow significantly, especially as the global AI race intensifies by the day. However, Elon Musk is extraordinarily ambitious and Tesla's core ambition is absolute self-reliance. This is driving efforts to break free from dependence on Nvidia, particularly at a time when Tesla has never needed more AI chips. While current supply from major manufacturers may only meet about 3% of the future demand across Musk's companies, continuing to rely on external suppliers would only drive costs higher, in reality, Elon Musk has little choice but to manufacture chips in house to achieve the desired pace. Dependence on outside vendors would slow down progress, especially for Tesla. That's why Terrafab is essentially being built as a computing infrastructure empire set to produce two of the most powerful chips in history inside the Terapab facility. The scale and technology could be shocking. Even top semiconductor experts might be left in awe. Meanwhile, Nvidia today is like the world's greatest architect. It designs castles like the H100 and Blackwell chips that no one can replicate. But this architect doesn't hold a single brick. It relies entirely on contractors like TSMC in Taiwan to turn those blueprints into reality. This is exactly the critical weakness that Elon Musk has identified. While Nvidia is stuck waiting in line for manufacturing capacity from partners, Musk has played the Terapab card projected to span up to 100 million square ft. So just how big is that number? Really? According to Elon Musk, the total computing power of the billions of chips this factory could produce annually may reach 1 terowatt, while the entire United States currently generates only about 0.5 terowatts of electrical power per year. If we assume each chip operates at around 250 W, reaching 1 terowatt would require producing roughly 4 billion chips annually, manufacturing at that scale with today's semiconductor processes would require thousands of acres and potentially hundreds of millions of square feet of production space. In other words, Terraab isn't just a factory. It's an autonomous industrial nation where raw materials go in at one end and ultra intelligent AI brains come out the other. The key difference is this. Nvidia sells intelligence while Musk is building the machine that creates intelligence at an unprecedented scale in human history. So why does it matter so much in the chip race? Yeah, the answer lies in vertical integration. Take the cost structure of an Nvidia H100 GPU which sells for around $30,000 to $40,000. The actual manufacturing cost at TSMC represents only a small fraction of that price. The majority comes from Nvidia's massive profit margins along with intermediary costs such as packaging, logistics, and distribution. However, Elon Musk is aiming to eliminate all those intermediary layers. Terrafab is designed to produce AI5 chips and the next generation Dojo 3 through a fully closed loop process. When you own a 100 million square foot facility, you're not just manufacturing chips. You're optimizing entire systems on site from cooling and power grids to robotic assembly lines. Terapab is being built north of Gigafactory, Texas, and is designed to consolidate every stage of semiconductor production under one roof chip design, lithography, fabrication, memory production, advanced packaging, and testing. This is a highly strategic move to optimize development speed, operating costs, and integration across Elon Musk's companies. A factory of this scale is almost unprecedented in industrial history, a concept that goes far beyond any standard we've known. Even Tesla's existing Gigafactories begin to look modest in comparison. For example, Gigafactory Texas, one of the largest buildings in the world, has about 10 million square ft of floor space. But that's just the starting point, as Terrafab is envisioned to be nearly 10 times larger. Comparisons with iconic tech campuses make the picture even more striking. Apple's Apple Park with around 2.8 8 million square ft of workspace and Microsoft's Redmond campus spanning over 8 million square ft across multiple buildings are both dwarfed. Terraab wouldn't just surpass them. It would dominate roughly 35 times larger than Apple Park and about 12 times bigger than the entire Redmond campus. And if those numbers are still hard to grasp, imagine this terapab would be nearly three times the size of the city of London or roughly equivalent to the combined area of Vatican City, Monaco, and Gibralar. Elon Musk claims that Terrafab could cut AI computing costs down to just onetenth of Nvidia's solutions. If you were a company needing millions of GPUs to train large language models, would you choose to rent intelligence at a premium from Nvidia or buy a self-sufficient solution from Musk's empire? But Terrafab's shadow extends far beyond cost. The real threat lies in specialized architecture. Nvidia has to design generalpurpose chips powerful enough for graphics, flexible enough for crypto mining, and intelligent enough for AI. This inevitably creates inefficiencies, extra components that aren't necessary for pure AI workloads. In contrast, chips coming out of Terrafab like AI5 or Dojo 3 are customuilt, tailorade for a singlepurpose ultraast AI training and inference. Think of Nvidia as a luxury SUV that can handle any terrain. While Musk's chips are Formula 1 race cars engineered for one specific AI track. On that track, the F1 car would leave the SUV in the dust. And that's not all. While 3nanometer and 4nm chips are becoming mainstream, Terrafab is targeting the 2nanmter process, the most advanced level currently in development. Elon Musk has indicated that Terrafab will leverage Gate all-around transistor technology to create chips that are smaller, more energyefficient, and more powerful than today's common 3nanometer and 4nome designs. What worries Jensen Huang most is the potential collapse of a proprietary ecosystem. Over the past decade, Nvidia has built a fortress called CUDA, a software platform that nearly every AI developer depends on. Want to run powerful AI? You need CUDA. Want CUDA? You need Nvidia chips. However, Elon Musk is constructing a new fortress with the Optimus robot and full self-driving once Terraab reaches its projected 100 million square feet scale. Tesla could have enough capacity to supply chips across its entire ecosystem and even sell them externally at disruptive prices. If developers realize they can achieve similar performance at a much lower cost on Musk's platform, Nvidia's CUDA wall could begin to crack. Can great software really save hardware that costs 10 times more? Tech history has shown time and again performance per dollar ultimately wins. At this point, many skeptics might say chipm isn't as simple as building factories. That's true. Producing semiconductors at 2 nanome or 3 nanome nodes represents the pinnacle of human engineering, requiring extremely rare EUV lithography machines from ASML. But don't forget, Elon Musk has already redefined aerospace with SpaceX and the automotive industry with Tesla. When he talks about 100 million square ft, he's not just describing physical space. He's defining the scale of ambition. As mentioned earlier, Terrafab won't be spread thin. It's designed to focus intensely on two core chip lines, each with its own factory operating like a specialized machine. The first focus is next generation AI inference chips, AI5 and AI6, the digital brains behind Tesla's most ambitious applications. These chips will power full self-driving, the robot taxi cyber cab, and the humanoid robot Optimus. Instead of relying on expensive sensors like lidar, Tesla is pursuing a bolder path. Tesla vision, a visiononly system where AI chips must process billions of operations per second to transform raw video frames into a precise real-time 3D understanding of the world. With Robboaxi, the chip's job goes far beyond simply seeing. It must understand detect lanes, read traffic signs, and predict the behavior of pedestrians and surrounding vehicles. All backed by redundant processing systems to ensure maximum safety. But with Optimus, the complexity rises to an entirely new level. AI chips must constantly maintain balance, coordinate fluid motion across dozens of mechanical joints, and learn new skills through observation. Thanks to a unified AI architecture, every breakthrough in self-driving can be directly transferred to robotics, a rare advantage in the tech world. The key differentiator of chips developed in-house by Tesla lies in performance per watt. This isn't just a technical metric, it's a survival factor for batterypowered systems. enabling massive neural networks to run efficiently over long periods without sacrificing performance. The road map is starting to take shape. AI5 is expected to enter limited production in 2026 before scaling to mass production in 2027, while AI6 is projected to deliver double the performance at the same chip size. When you place these numbers into the bigger picture, hundreds of millions of Tesla vehicles and potentially billions of Optimus units in the future, Terraab is no longer just a factory. It becomes the foundation for an era where autonomous vehicles and humanoid robots are not just possible but ubiquitous at an unprecedented scale. Terrafab is designed to operate as a machine that builds machines using the latest generation of Optimus robots to assemble chip production lines themselves. This creates a powerful feedback loop. AI builds chips. Chips create more powerful AI and the cycle continues inside a massive facility with minimal human intervention. Nvidia may have some of the world's best software engineers, but Elon Musk is aiming to control the most efficient manufacturing machines on the planet. The second chip is D3, a specialized processor designed to power artificial intelligence in the vacuum of space. Also known as Dojo, 3D3 is engineered to overcome the limitations of Earth-based data centers and enable massive scaling of compute by moving hardware off the planet. Why does this matter? D3 represents a strategic turning point in Tesla's custom chip program, shifting the focus from competing with Nvidia on Earth to powering a significant portion of humanity's computing tasks in space. By leveraging launch capabilities from SpaceX, Tesla aims to reduce the operational cost of orbital data centers compared to groundbased ones, unlocking AI scaling to the terowatt level and beyond. According to Elon Musk, the D3 chip is purpose-built for the harsh yet liberating environment of space. On Earth, engineers spend enormous time and resources managing heat dissipation and power constraints. D3 removes those limitations. Free from traditional cooling and power bottlenecks. It can operate at significantly higher power levels and temperatures than any terrestrial processor. It is also heavily radiation hardened to survive the intense cosmic radiation found beyond Earth's magnetic field. These capabilities combined with the economic efficiency of solar energy in space and the heavy lift launch systems of SpaceX could allow Tesla to deploy massive AI computing clusters in orbit potentially at far lower cost than conventional data centers on Earth. The production targets at Terapab represent a direct challenge to the current semiconductor order, especially when compared to the dominance of Nvidia. The facility is designed to start at around 100,000 wafers per month with ambitions to scale up to 1 million wafers at full capacity. To put that into perspective, this could represent up to 70% of the current total global output of TSMC, Nvidia's primary manufacturing partner. For a company with no prior chip fabrication experience to target a single facility rivaling the scale of TSMC's network is an extraordinarily bold move toward full self-sufficiency. Elon Musk has stated that Terapab could produce between 100 and 200 billion AI and custom memory chips annually. This is a strategic push to fully replace Nvidia's high performance GPUs within Tesla's ecosystem. These in-house chips would serve as the heartpowering full self-driving, the robo taxi cyber cab network, and fleets of Optimus robots with millions of Optimus units themselves helping to build and operate the factory. Notably, instead of focusing on groundbased server infrastructure like Nvidia, about 80% of Terrafab's computing capacity is projected to serve orbital AI satellites, creating an unprecedented space-based intelligence network. Here's a final comparison to highlight the magnitude. While Nvidia's revenue depends on selling GPUs individually to thousands of customers, every step forward by Terapab directly serves some of the world's largest data consuming systems such as XXAI and millions of Tesla vehicles on the road. Elon Musk doesn't need to find customers. He is his own biggest customer. This vertical integration allows Tesla to iterate, fail, and optimize chips at a speed Nvidia cannot match given its need to serve diverse external clients. Of course, despite how formidable Terrafab may sound, Nvidia still possesses powerful modes of knowledge and ecosystem advantages that sheer ambition alone cannot overcome overnight. Elon Musk faces at least three major barriers. The first and perhaps most daunting is the CUDA ecosystem. CUDA is widely considered the secret weapon behind Nvidia's dominance in AI chips for over two decades. It's not just a platform. It's effectively the native language of most AI developers worldwide. Asking engineers to shift to a completely new chip architecture from Terrafab is like asking a global English-speaking population to suddenly adopt a newly invented language. It would require rewriting code bases, reoptimizing systems from scratch, and accepting significant risks. In tech, software inertia is often far stronger than hardware innovation. Beyond that, there's a harsh reality. Building a chip fab is far more difficult than building a gigafactory. If Tesla's gigafactories represent the pinnacle of precision manufacturing, then a 2 nanometer semiconductor fab is closer to atomic level engineering to operate at this level. Tesla would need access to extremely rare EUV lithography machines from ASML machines so scarce that even giants like Intel and Samsung electronics must wait years to acquire them. And history offers a cautionary tale. Consider Tesla's 4680 battery cells after more than 5 years. and countless promises. Actual production output remains limited and the process is still not fully mature. If manufacturing batteries a chemical engineering challenge has proven this difficult, many experts reasonably question whether Tesla can master two nanometers chip production, which demands nearperfect clean room environments and nanometer level precision. Financial pressure is another burden that cannot be ignored. According to reports from Morgan Stanley, the cost of operating Terrafab could climb to as much as $45 billion, far exceeding initial investment estimates. This represents a high-risisk gamble, especially as the EV market begins to saturate and Tesla must allocate significant resources to its roboaxi ambitions. While Nvidia sells chips across the entire global economy from healthcare to finance, allowing it to recover costs quickly, Terraab is for now primarily serving Elon Musk's internal ecosystem. This difference in economic scale could put Tesla in a position of massive cash burn before it ever sees meaningful returns. In short, over the next 1 to 3 years, Nvidia's position remains relatively secure. Tesla will still need to spend billions purchasing chips from its rival if it wants to keep projects like FSD and Optimus on track. However, in the long term, post 2027, Terapab could become a definitive statement of independence. If Elon Musk succeeds, he won't just build a chip factory. He'll set a precedent never seen before, a car company that fully controls its own silicon destiny. And in that future, Nvidia may have real reason to worry because its biggest customer could also become its most formidable competitor. So, do you think Tesla's Terrafab can actually outperform Nvidia's hardware by 2027? What's your take on AI satellites brilliant future or just space clutter? Thanks for watching. If you enjoyed this deep dive into the Tesla versus Nvidia Titan clash, don't forget to smash that like button and subscribe for more tech analysis. What's your pick for the AI future? Let us know in the comments below. See you in the next one. [music] >> [music] [music]