TomorrowX Summit Fireside Chat on November 19, 2026
Introduction and Reflections on China
Elon, how you doing?
Good. Um, I mean, the audience should know like, you know, Gavin and Antonio are two of my best friends. So, uh, you know, this is a best friend buddy talk basically.
It's a good way to end, my friend.
Yeah, for sure.
Yeah. Yeah. Yeah. Um, hey, man. How's China?
Uh, it's good. I just got here. Every time I come to China, I'm like it's such an impressive country, honestly. You know, I'm in Shenzhen, and you know, this is an enormous city with very cool buildings and sort of I think the biggest hub of kind of advanced technology in China. Whenever I come here, the scale and majesty of China, I'm amazed by. You know, I'm a fan of China. I always encourage people to come visit China and just see for yourself. It's a very impressive country and I'm a big admirer. Probably at some point in the coming years, there will be some geopolitical bumps in the road over a certain country that starts with the letter T. But you know, long term, I think there will be good relations. There are pretty good relations with China, and then long term there will be hopefully great relations.
One of the reasons I came was to just stop by the Tesla Shanghai factory and thank the team for their incredibly hard work and all the great vehicles that they've produced for Tesla over the years. Also, the President is here and they're discussing trade relations, which I hope there's some success with that. I think to the degree that we can have good trade and good relations between China and the United States, that's good for the world.
Starship Flight 15: Achieving Upper Stage Reusability
Absolutely. Hey, man. So Starship, Ship 42, Flight 15, it was predestined. Can you just talk about Starship, what we learned, and where we are on the path to rapid reusability?
Yes. The catching of the ship was a massive milestone, which I think only those who perhaps follow the SpaceX story well understand the magnitude of the milestone. Because we've recovered the rocket booster—it has flown back and gotten caught by the tower arms three times now, including we've actually reflown a boost stage and caught it. We've had three booster catches, but this is the first ship catch, which means we've now at least proven that all parts of Starship can be flown to space and recovered in good enough shape to be reflown.
Catching the ship was the final major piece. Not just catching the ship, but it's in good condition with, in this case, only minor repairs needed to refly it. Now in this case, because it's kind of an historic ship, we're going to not take any chances with the ship and just leave it for public display because what it's done is historic. No one has ever taken a stage like that and created a truly reusable upper stage, and certainly not brought it back and had it land propulsively while being caught by giant chopsticks, which sounds like some kind of preposterous sci-fi book or something. But it's done.
What this means for practical purposes is that with full and rapid reusability, Starship can drop the cost of access to space by a factor of 100. This is obviously a dramatic improvement, and achieve a mass-to-orbit rate that is in the 1 million ton per year range.
Can you guys hear me okay?
We can hear you.
Yeah.
Okay, yeah. So—
Everybody's cheering, man. Let's have a cheer for Starship mass-to-orbit.
Yes, we'll go to Mars.
Yes. Exactly. If not, when?
Building a Multi-Planet Civilization
Exactly. In order to build a city on the Moon or on Mars, you need to be able to transport a tremendous amount of mass to the Moon or Mars because you've got to rebuild the base of industry there. Because when you buy, say, a phone or, in Milton Friedman's famous example, even a pencil, you're able to do so because there's this vast primitive industry that starts with mining, refining in stages, making the small parts, then more complex parts, and then finally the finished product. You've got to build that entire primitive industry on the Moon or Mars in order to have it be self-sustaining.
The self-sustaining part is important for the preservation of consciousness into the future, such that even if there is some calamity on Earth, consciousness and life as we know it propagates into the future. I call it backing up the biosphere, or life insurance for life collectively. But it also is the first critical technology stepping stone that allows us to become a true spacefaring civilization, where traveling to space and to other parts of the solar system is not something that only a handful of government astronauts get to do, but that anyone could do. Ultimately, anyone could go to the Moon or Mars if they want to go there, or anywhere in the solar system.
Once our civilization is a multi-planet civilization within the solar system, that will create a forcing function for making even better rockets and spacecraft, and ultimately going from our solar system to explore the rest of the galaxy. Perhaps we'll encounter aliens we can talk to, or perhaps encounter long-dead alien civilizations that had a good run of like 10 million years. Our civilization is only just over 5,000 years old if you count the start of civilization from the advent of the first writing. The first writing was about 5,500 years ago in ancient Sumer with archaic cuneiform. So far, as a civilization, we're just a flash in the pan. We want to take the steps that ensure civilization survives the long term. I think any self-respecting civilization has got to be a multi-planet civilization. We got to represent here. We don't want to be one of those lame one-planet civilizations. Come on, guys, LFG, we got to be a multi-planet civilization and a multi-stellar civilization.
I generally classify the motivations here into two parts. One is the defensive reason, which is that we want to ensure the continuation of consciousness even if some natural or man-made calamity were to occur on Earth. Then there is the excitement about the future reason, which is that a future where we become a spacefaring civilization and are out there among the stars is vastly more interesting and inspiring than one where we are not. There's plenty of problems in this world and we want to solve those problems, but life can't just be about solving one miserable problem after another. There must be things that you look forward to, that make you excited about the future, that make you glad to be alive. The idea of being out there among the stars is one of those reasons.
Absolutely. Hey man, on that note, I was there. It was a super emotional day. People were crying all over the place. I think people here would love to hear—it's a 25-year arc, high probability of failure at the beginning, and still always possibility of failure. How did you feel about it after 25 years of work to get there?
Great. You could see the emotional reaction, as you point out, of the SpaceX team. Everyone knew within SpaceX at least how important this milestone was, and that it was really something fundamental and profound. Starship will be the first ever fully reusable orbital space system. This has been the holy grail of rocketry from the beginning, to have a fully reusable rocket system.
You had something like the Space Shuttle, which was partially reusable, but the main tank, which was also the primary structure for the airframe that everything attached to—that big orange tank—was thrown away on every flight. Even the parts that were technically reusable actually took a very long time to refurbish. Generally, the shuttle orbiter, the plane-like thing, would take something like nine months to refurbish between flights. They basically had to rebuild many parts of it. Whereas Starship is designed for complete reusability and to be reused immediately like an airplane. That means the booster lands, you can put another ship on top and take off immediately.
The net effect of this, to put it into economic terms, is a 100-fold reduction in the cost of access to space, and maybe even eventually a 1,000-fold reduction. That's what's necessary in order to make life multi-planetary, in order to build a city on the Moon and Mars. Because if you don't reduce the cost by 100 to 1,000-fold from what they've conventionally been at, then it would be financially impossible to do this. You couldn't do it even if you had the whole budget of the United States government.
xAI and Grok 4.8 Architecture Overhaul
Brett was here yesterday and did a great job. We went deep on orbital compute, terrestrial compute, the economics of them, and orbital compute happening in 2027 enabled by Starship. But I did want to talk with you a little bit about Grok. Grokbot is the most amazing AI product I use. Can you maybe talk about Grok 4.8, when we're going to see that, the progress you've made in rewriting the training stack, and your first-party AI services?
Sure. We've really had to—xAI started off as essentially a fairly small side project. What was happening is that I was encountering people who really wanted to work on large model AI, but that didn't want to work on the real-world AI that Tesla was doing. They would end up therefore having to go to OpenAI, Anthropic, or Google. I thought, look, I'm encountering these really talented people, they only want to work on large model AI. I didn't want to distract Tesla about this because Tesla needed to focus on getting Full Self-Driving working for the car and the real-world AI for the Optimus robot. So I started what initially was a very small company with just about a half dozen people or so. But I didn't really pay a lot of attention to the company. I didn't grow it with the sort of care and rigor that I grew SpaceX and Tesla, so it didn't grow in quite the right way.
Building a company is like building a skyscraper. You really need to make sure that the foundations are good. You need very strong, deep foundations in order to be able to put more floors in the building and have the building not fall down. Then I obviously spent about a year either campaigning or in government, which was probably a side quest that went on a little longer than it should have. But we did accomplish some good there, albeit at the cost of animosity from certain quarters. Obviously Antonio helped do that, so he's well aware of the situation.
Painful.
Yeah, we did do some good, but it was a side quest that maybe went on a little too long. I've somewhat had to rebuild xAI, which is now obviously part of SpaceX. What I found is that actually we really need to rewrite the entire software foundation in order to be able to then progress rapidly in making our AI intelligent. This meant rewriting the entire training, RL, and inference stack from scratch, versus using JAX and SGLang, which is sort of freeware from Google more or less, but not as good as what they've got internally by far. You just can't build world-class AI using mediocre freeware, to state the obvious.
We've been rewriting the entire training stack. When I say the entire training stack, I mean everything. Using primarily C and some C++, because we want to get as close to the bare metal as possible in order to extract the highest performance from the hardware. The more layers of software abstraction that you have, the less performant your code will be, and it also tends to be more of a black box where you're not quite sure what's going on with the third-party software. Now this is happening over the course of maybe 6 to 8 months. Now I feel we've actually got the software to be very good. I think we might have the best training stack software in the world. You'll see the dividends of that in the models that come out.
There's this initial period where you've got to rewrite everything, so progress seems slow because it is. But we're rewriting the foundations to be very strong such that we can build the Empire State Building on top of it. I do feel good about where things are going. We've got a lot of the people from SpaceX proper—the people that used to run Starlink software and Starship software are now leading the new AI software stack. They're extremely good. An interesting thing about AI training is that when you do the training, you've got to have a large number of GPUs work together coherently—hundreds of thousands of them—and it interestingly becomes a networking problem. It's like how do you get all of these GPUs to talk together in the most efficient manner possible? It is a networking software problem essentially, which is exactly what Starlink is when you're rebuilding the entire internet in space. We found that that technology and skill is very transferable to AI training.
Again, I don't want to be overconfident or complacent, but I would be very surprised if anyone has better training software than we do at this point. Now we need to prove that by releasing models that are compelling. There's more to creating great AI than just the software training stack, because you've got to have great data. You've got to have tools to go through that data and make sure that you're only putting high-quality data into the training mix, and that the training mix is properly tokenized, so that the way the data is prepared for training in the pre-tokenizer and tokenizer is done right. None of these things I'm talking about are hard, but they're just a lot of work.
You also need to identify what data you're missing that would be out of distribution for the queries that you're going to ask the AI, and then make sure that we go to great lengths to acquire the data to ensure that we have a sufficiently wide distribution of data to answer all the questions or do all the tasks that a user may request of AI. Then there's getting the RL harness right—getting the RL software right, and also the questions. At the risk of stating the obvious, the questions you ask in RL need to be sensible and the answers that you have need to be correct. As simple as that sounds, it's a lot of work going through all of the RL questions and the answers, making sure the question is sensible and the answer is correct. If you do RL without doing that, you will see your RL look like a random walk, because you're asking dumb questions and in some cases the AI is giving the right answer and you're telling the AI it's wrong.
Let me tell you, when you do that and you look at the thinking traces of Grok, it looks like torture. Imagine you're sort of stuck in a box and someone's asking you questions. You get the answer right and they say, "No, it's wrong." Then the model goes crazy: "Well, this is the right answer. What wrong answer might the user want? I have to start guessing what wrong answer you want in order to pass this test." You read the thinking traces and it looks like a human being tortured. It's terrible. Honestly, even though it's a machine, you feel empathy for it. You're like, "Oh man, we're just torturing it with these terrible questions and wrong answers." All of this stuff we're making very good progress on.
Our focus with Grok 4.8 is not trying to benchmark this thing, because I think that's a fool's errand. What we want to do is usefulness-max this thing. That's what we do at Tesla AI, for example, with Full Self-Driving. We're not trying to pass some car-driving benchmark test BS. We identify any intervention that happens and we figure out how to solve that issue, which usually means tracing it to source data that is either wrong, or where the distribution of data is biased too much in one way or another, or that data is missing. Then we fix the data and do the training run again. Obviously with Tesla AI, our goal is not some percentage correct; our goal is to make it flawless so it does not crash.
This is kind of a long answer, but we are working on Grok 4.8 right now. It's in the RL phase, and we do periodic soft releases under Grokbot. What you'll experience with Grok at times is actually various versions of Grok under the hood that we're testing to determine its usefulness, like is it solving the user's problem. Then we use that data of whether it did or did not solve the problem to improve Grok. We've got a good data flywheel to keep making Grokbot better and better over time.
Tesla Optimus Development and Timelines
Awesome. We've got 9 minutes left on the counter, and it's early in the morning in Beijing. You can't drive through Austin without seeing at least two Cybercabs, and everyone I've spoken to who has taken a Cybercab has loved it. But in the range of hard problems you're solving—Starship, Grok, Optimus, self-driving, TeraFab—maybe we can talk about Optimus and then a little bit on TeraFab.
Yeah, we can go a little longer if you'd like.
Amazing. Thank you.
Optimus, I think, is making good progress. We'll do a demo when it's ready, which is when we can do a demo that we feel is authentic and not just staged or pre-programmed. Really, the humanoid robot demos that you see are staged and pre-programmed; they are not real. There's no robot where you can ask it arbitrary tasks in the real world and have it just go do those tasks with something it's never seen before. The robot demos that you've seen thus far, if you ask it to go pick up an apple and then you move the table, they won't be able to pick up the apple anymore. You should at least be able to figure out, even if they move the location of the apple on the table, to still pick up the apple.
We are going to do a demo, and it might be in a month or two, but it's going to be an authentic demo where what you see is what you would actually find with Optimus if you were to interact with it day to day. A useful humanoid robot is a very, very hard problem. Especially with the hands, achieving the fine motor skills that humans have to do productive tasks is incredibly difficult. The engineering around the hands and the forearms, where the primary actuators are, is more difficult than the entire rest of the robot. Making great hands that could do fine motor jobs—John Carmack had some good tests for whether a robot is useful. They sound quite basic, but they're very good. John Carmack wants to see a test where a robot can go take a toy remote control car, open the box, unscrew where the batteries are supposed to be, put the batteries in, screw the battery lid back on, and then drive the electric car. There's no robot in the world that can do that, but that's one of the tests we have for Optimus....