Tesla, SpaceX & xAI Are About to Make History

Picture a giant AI computer: dozens of GPUs and CPUs, memory chips, network switches, and cooling hardware, all working together as one machine. Now picture SpaceX rebuilding that entire system as a satellite, putting it on a rocket, and launching it into orbit.

On August 4th, Elon Musk, the CEO of SpaceX, said the company had decided to build exclusively on Nvidia. He called the new Vera Rubin system the best AI computer and said the companies now have a close partnership on many levels. SpaceX also announced it is working with Nvidia to design the compute payload inside Starmind, its planned AI satellite, which will use an optimized version of Nvidia's Vera Rubin NVL72. Elon says the first launches should begin next year.

One revolution

We're not living through three separate revolutions: AI, robotics, and energy. We're living through one, where each piece amplifies the others.

Training neural networks requires massive amounts of compute. Compute requires massive amounts of energy. More energy means more compute, and more compute means smarter AI. Robots are essentially AI that can touch the physical world. And soon enough, those AI systems will be in space, which is filled with energy from the sun.

The more energy we have, the more AI and robots we can deploy. The more AI and robots we deploy, the cheaper energy becomes. That's a flywheel, and it's starting to spin. The way I see it, Tesla, SpaceX, and xAI are spinning it as one machine.

Brain, vessel, body

Framing this as a Tesla story undersells what's happening. The real convergence is happening across an ecosystem of companies that Elon Musk controls. Think of it this way. xAI, now part of SpaceX, is the brain. SpaceX is the vessel. Tesla is the body.

Grok looks like another chatbot competing with ChatGPT. That misses the point. I think Grok is the AI that will eventually run the entire ecosystem, trained on real-time X data that no competitor can access, on Tesla's billions of miles of driving data, and eventually on Optimus robotics data.

That's the data moat principle. You can hire a competitor's researchers, read their papers, and train similar models. You cannot conjure a social network with hundreds of millions of users. You cannot retroactively collect driving data from millions of vehicles. Stack those sources and they multiply.

I'm already seeing it in my own car. The Tesla Cybertruck in my driveway just got an update through Grok, which now connects to agents that do work on my behalf. While it drives itself, I talk to it, and it will upload videos, update my website, clip videos, or store a video idea and create an outline.

Catching the rocket

SpaceX is the vessel. Starship gets framed as a Mars rocket, but I think the near-term implications are even bigger: access to space so cheap that entirely new possibilities open up. AI data centers in orbit where solar is 24/7. Starlink as the nervous system connecting everything globally.

In space there is no atmosphere to block the sun. No night, no clouds, no seasons. In the right orbit, solar panels in space can collect up to eight times more energy than panels on Earth. Run AI inference in orbit and you only need to send down the results.

Getting stuff to space has always been expensive. SpaceX has already brought launch costs down by an order of magnitude, and Starship promises another order of magnitude. SpaceX now thinks it has solved the Starship heat shield problem. The heat shield is the most important piece in making a rocket fully reusable.

Without a proper heat shield, a returning rocket gets destroyed by the atmosphere, like a meteor. With one, in theory, you just refuel it and send it back up. That's why SpaceX designed the tower with the chopsticks: to catch the rocket, put it back on the pad, and launch it again shortly after.

I think that's what eventually gets SpaceX to 1,000 Starship flights per year, and eventually a cadence of a Starship flight per hour. When the cost to put stuff in space is down to the cost of fuel, the cost per kilogram in orbit plummets by 100 times or more. At that point a data center in space gets cheaper than one on Earth, with all its cooling, grid connections, and permitting.

I am not saying this happens in 2027. But in the four or five-year time frame, the lowest cost way to do AI compute might actually be with solar powered AI satellites, as Elon claims.

A brutal place for a computer

Space is a brutal place to put a computer. In a normal data center, racks sit on concrete floors, power arrives through the grid, and pumps move water through the cooling system. If a chip fails, a technician pulls it out and replaces it.

Space removes every one of those assumptions. Every kilogram has to ride a rocket. The computer has to survive vibrations during launch. Radiation can flip bits or damage electronics. Heat from the chips can't blow away into the air, because there is no air, so it has to leave through large radiators. And when a chip dies, nobody is driving a repair truck to the satellite.

It's a problem that still needs solving, and I think the economics will force the solution. It's too good to pass up.

From Cybercab to Optimus

Tesla is the body: the cars that collect data, the robots that will do physical work, the batteries, the solar panels, and the factories that build all of it.

Tesla is training a neural network to drive the way humans do: look at the world and make decisions based on what it sees. No pre-mapping required. That's the end-to-end neural network approach versus the rules-based approach, a fancy way of saying it's 100% AI. From my perspective, there's really no contest in which one scales.

The best example is Tesla's Cybercab, which started giving rides in Austin in early September. Tesla has harvested, I think, over 14 billion miles of real-world driving from customers. That camera data went into a massive AI brain, and the brain figured out how to drive, to the point where these cars no longer need steering wheels and pedals.

That same technology is now going to be used for humanoid robots. Optimus is fundamentally an AI project. It has cameras for eyes and runs visual input through neural networks, the same fundamental architecture as FSD. Instead of steering and braking, the output is arm movement, hand coordination, and walking. The underlying AI approach is identical.

The $40 trillion math

By my estimate, the global market for human labor is somewhere north of $40 trillion per year. Factory workers, warehouse workers, delivery drivers, construction workers: anything a human does in the physical world that doesn't involve a computer. I'd put the entire global automotive industry at around $3 trillion per year. The labor market is an order of magnitude bigger.

That's why I think Optimus, and humanoid robots in general, are the main event for Tesla. Everything else is secondary, and the company is already acting like it. There's no bigger signal than this: Tesla tore out the Fremont factory lines in California that made the now discontinued Model S and Model X, and it is converting that space to build humanoid robots.

I expect the cost to manufacture an Optimus unit at scale to land somewhere in the range of $20,000 to $30,000, and Tesla has explicitly stated targets in this range. The US minimum wage varies by state, so take $15 per hour as a representative number. At 40 hours a week and 50 weeks a year, that's $30,000 in wages alone. Add employment taxes, benefits, training, and turnover costs, and you're easily north of $40,000 a year.

So the robot costs roughly what a human worker costs for one year. Except the robot works for years before it needs replacing. It doesn't call in sick. Once trained, it doesn't forget. And it learns new skills through software updates, based on what the other robots in the fleet learned that day.

Elon Musk, the CEO of Tesla, has talked about a 10,000 robot warehouse. Picture a gigantic warehouse with 10,000 humanoid robots inside, trying to build a chair or fold laundry. I bet you that they should live stream this. It would be the greatest TV show of all time.

That trial and error is what teaches the robots. The second one robot learns how to do something, the whole fleet knows it, because it's just software. More robots in the world means more situations, more training data, and better AI. That's the flywheel again.

One loop

The loop starts with Terafab, the giant chip factory Tesla and SpaceX are planning. Here's how I see it. Terafab produces space-optimized chips. Earth data centers prove the architecture and train the models. Starship launches the computing satellites. Starlink links them to one another and to users. Satellite demand raises Starship's launch rate, which spreads fixed costs across more missions and makes each rocket cheaper. Lower launch costs make more orbital compute economical, which creates demand for more chips.

Tesla sits on the other side of this system, with potentially hundreds of millions, if not billions, of machines that need intelligence in the physical world.

My take: no other organization on earth has this architecture. Google has AI but no physical infrastructure. Legacy automakers have manufacturing but no AI. The Musk ecosystem has all of it, under one person who can coordinate across all of it.

Grok is already doing real work for people online. That's only the digital half. Optimus is the physical half.

Now imagine pairing your Grok Bot with your Optimus. Grok Bot can do anything for you in the digital world. Optimus will do anything for you in the physical world. I could tell Optimus to fold my clothes and clean the house, render new animations for my next video, edit two other videos I recorded, and check my email. Then it could hop in a self-driving Cybercab and bring me back a nice meal.

Where I could be wrong

I could be wrong about this. Maybe the organizational complexity becomes a liability. Maybe Elon's attention really is too split. Maybe the synergies I see are partially wishful thinking.

But if I'm right, if the ecosystem thesis plays out, then the individual company valuations we're looking at today are dramatically mispriced.

Go back to that computer headed for orbit: dozens of GPUs and CPUs, memory chips, network switches, and cooling hardware, all working together as one machine. I think the bigger machine is three companies: a brain, a vessel, and a body, spinning one flywheel. That's the bet I'm making.