Elon's Shocking SpaceX Announcement

Elon Musk just made a post on X about a massive change to Grok Bot and SpaceX. He even quoted one of my old posts. I wrote it in late August, and it talked specifically about them doing this. They ended up doing it.

I am so glad they did. This has a lot of long-term implications: for SpaceX, Grok Bot, AI, agentic processes, data center companies, Colossus, Starmind and Starship. It's like my mind's melting.

The video has the full walkthrough. Here is the same argument in writing.

What a harness does

SpaceX has a product called Grok Bot. It is an agentic harness, like OpenClaw or Hermes.

If that term is new to you, think of a harness as a program. It takes an AI and uses that AI to execute actions against your computer or against other tools.

Here is a live example. I used Grok Bot to edit the video this piece comes from. I gave it the file and told it, roughly: go ahead and edit this video. Add graphics where it makes sense. Add charts where it makes sense. Make it beautiful. Give it nice animations. Then I pushed my luck: maybe even insert a clip of my voice clone, with a British accent, saying something cheeky.

I did not know if it would do that. It did. Cheeky enough. The clone says: "I've also cut all your ums. You're welcome."

That is the tool. The harness hooks up to different ways to render the animations and edit the video. It hooks up to the voice clone program that adds a British accent to my voice. It can also check my emails. That is what an agentic harness does.

Grok Bot is not the only one. There is Meta Muse. There is GPT Dots. And there is the open source stuff, like Hermes and OpenClaw.

No model wins everything

Elon's post was about choosing the best model possible. I think this is a brilliant play by them.

Here is what SpaceX is realizing. The best possible harness, the best agent anyone could use, is one that is good at everything. But SpaceX does not have the best model at everything.

Look at Grok 4.7 right now. In my view it is very good at coding. It is very good at researching and looking stuff up. It is mediocre at other stuff. Generating 3D renders, for example, is something GPT Astra is excellent at. For unbelievably creative work, there is Opus, Anthropic's model. Opus is super good at that.

So here is my read. Instead of trying to build one model that is good at everything, SpaceX will specialize in figuring out which model is best at what. Somebody opens the harness and says write me a book. Or run my business for me. Who knows what people will ask these things in five years. The harness should be able to do it.

That becomes the background process. I think the Grok team will still work toward making the best possible AIs. But they are also going to build an AI that is very good at using other AIs. Then they put it all into Grok Bot, so it can do that automatically for everybody.

Where the money goes

There is an obvious objection. Say Grok Bot calls Opus or GPT to run a job for somebody. That money goes to GPT or to Anthropic, because you are calling their model. SpaceX has to pay to use their model in its harness.

Of course that will be part of it. Here is the key difference. SpaceX, the people who also make Grok, own massive amounts of compute. And they are building massive amounts more.

Now picture where that goes over time. This is my projection. SpaceX keeps building out compute clusters. It continues its partnership with Nvidia and builds more data centers with Nvidia chips. Terafab goes live, and they start building their own chips in partnership with Intel, TSMC and Nvidia. Fast forward more, and they start putting these data centers in Starship, sending them to space and harnessing the energy from the sun. No more fighting battles with local governments and legalities to put data centers on Earth.

What does that mean? I think SpaceX over time is going to be the largest provider of compute on the planet and off of the planet.

And that matters for Grok Bot. Whatever model somebody chooses, or Grok Bot decides to run, I think the highest likelihood is that model gets served from SpaceX data centers. If I am SpaceX and I am already serving those tokens from my own data centers, I am just going to serve them at cost. I do not need to pay another company to serve their tokens from my data centers.

The $100 math

Here is the part I would bet on. "I bet you a ton of money" that the deals SpaceX builds will say something like this: if you want to use my compute, which will be the fastest to serve the tokens, the most efficient and the most plentiful, then "you have to give me the ability to serve those models at cost for my Grok Bot."

Why is that a huge deal? Most model companies have humongous gross margins. When you call Opus 5.5 on Anthropic's API, or use Claude or ChatGPT, the company charges you what it cost them to generate those tokens, plus a big margin. My guess is something like 80 to 90% on those tokens. Reported company-wide gross margins are lower: The Information put 2025 at about 40% for Anthropic and 33% for OpenAI. The cost of each unit of intelligence is so low, and they know it is super valuable, so they can add massive margins.

SpaceX can serve it at cost instead. Same level of intelligence, roughly 80 to 90% cheaper. Over time that makes its ability to serve tokens extremely economical.

Make it concrete. Say I have a hundred bucks in Grok Bot and I want to use $100 worth of whatever to do a job. I am not spending that $100 on Anthropic or GPT. I am spending it on Grok Bot. Grok Bot is going to be very good at deciding which model is best, so it picks that model and serves it 90% cheaper. That leaves me 90% of that money to use somewhere else, because SpaceX is the one serving those models.

That is why I call it unbelievably brilliant.

Harnesses eat tokens

I think these agentic harnesses are going to explode in popularity. Grok Bot, Meta Muse, GPT Dots. They are extremely useful. You go in, tell it to do something, and it does it for you. Buy me some supplements. Clear out my junk email. Let me know when somebody reaches out to me. Text this person something. Design me a website. Run my merch store for me.

This is the most underrated part of SpaceX's long-term business. People are very focused on Starship. They are also focused on the data center part of the business. Some people are sleeping on Starlink. But the amount of tokens this agent harness is going to drive is going to be insane.

Here is why. These actions are not like going to a chatbot, asking a question and getting an answer. They require a lot of tokens, a lot of units of intelligence, to execute. That will come down. It will get more efficient over time. But it is not going to be cheap. Serving those tokens will take a lot of money, a lot of compute, a lot of chips and a lot of data centers.

Over time, I think Grok Bot and SpaceX generate a gigantic amount of revenue on that business. And all of it is digital.

Grok Bot plus Optimus

Here is the crazier part. Why do you think people keep saying Tesla and SpaceX should merge? What I just described happens in the digital world. Optimus, the robot, is what happens in the physical world.

Now imagine pairing them. Grok Bot does anything for you in the digital world. Optimus does anything for you in the physical world.

Picture the request. Fold my clothes and clean the house. While you do that, render me some new animations and 3D renders for my next video. I recorded two other videos, so edit those. Check my email and make sure everything is good. And when you are done with the laundry and the house, go to the store, pick me up a nice meal, hop in a self-driving Cybercab, drive yourself there and bring it back.

To me that is a humongous deal and super underrated. Very few people understand how big of a deal this truly is.

Cheaper means more

And this is just the beginning. As the models get better, the compute gets more efficient and the costs drop, the things these tools can do will keep growing.

There is an idea called Jevons paradox. The way I explain it: as the cost of something useful drops, demand explodes, because far more people end up using it. It happened with coal. It happened with shipping. It happened with steam engines and a bunch of other technologies. Compute and chips too.

Put that next to the at-cost math above and you can see why I keep saying this is just the beginning.

Back to the edit

Which brings me back to where this started. I handed Grok Bot the file for this video and asked it to make the thing beautiful. I asked for a British version of me to say something cheeky. I have no idea how the edit will do.

Some of you have noticed videos coming through with my clone voice. It is going to get better. I promise. There is also going to be a lot more of it. I am going to be making content that is automated, and it is still going to be unbelievably valuable. That is my goal long term. Keep sharing your feedback.

I am not going anywhere. There is just going to be a lot more high quality content. The British version of me already put it best: "I've also cut all your ums. You're welcome."