Elon Musk posted one sentence on November 10, 2026 that passed 5 million views in under 12 hours. It reads: “Turns out that the Starlink software, built for managing the movement of vast amounts of data through space and to the ground, is also very well suited to training super intelligence”. Read alone, it sounds as if satellites are about to train AI. His other posts that morning say something narrower.
The sentence is about a software stack. In the hour before it, Musk said the code that trains and runs Grok was rewritten in C, and that it is derived from work by the engineers who built Starlink. No SpaceX or xAI document, code or benchmark backs any of it yet. This page sets out the posts in order, what SpaceX has published about Starlink, and what is still not shown.
On this page · 6 sectionsOpen
Elon Musk posted on November 10, 2026 that Starlink software is very well suited to training super intelligence. The post passed 5 million views in under 12 hours.
His posts in the hour before explain it. Grok’s reinforcement learning and inference code was rewritten in C and C++, and he says the stack is derived from work by Starlink engineers.
Nothing in the posts says satellites train a model. His earlier posts describe the stack as written for data-center hardware.
There is no SpaceX or xAI document, code or benchmark behind the claim. Grok 4.8, the model he ties to the rewrite, has not been released.
SpaceX’s own Starlink page describes three lasers per satellite at up to 200 Gbps. It says nothing about the software.
§ 01What he posted, in order
The post went up at 10:34 UTC. It is not a reply and it quotes nothing. It had 5,004,511 views when we read it at 21:52 UTC. Three other posts from the same hour give it its meaning.
| Time | What he posted |
|---|---|
| 09:35 | Grok 4.8 has finished training. The reinforcement learning and inference stack was rewritten from scratch in C and C++ |
| 09:46 | The language is C, with a little C++ |
| 10:14 | SpaceX writes its own headers and calculates its own time. The new Grok stack is derived from work by Starlink engineers |
| 10:17 | There is some assembly too. He calls it tight code |
| 10:34 | The sentence that went viral |
| 18:20 | C is what you use for the best performance on bare metal |
- The rewrite. At 09:35 he answered a question about Grok 4.8: “We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.” JAX and SGLang are open-source tools for training and serving models. RL is reinforcement learning.
- The language. A user replied that C wins again. Musk answered: “C with a little C++, but not too much”.
- The source. At 10:14 he wrote: “At SpaceX, we write our own headers and calculate our own time (includes leap seconds).” Then the line that links the two businesses: “The new Grok AI/SI training and inference stack is derived from incredible of work done by amazing engineers to create Starlink.” The slip in that sentence is his.
- The pride. Three minutes later: “One day, super intelligence will write a better software stack, but that day has not yet come.” He ended that post: “This is tight code.”
Twenty minutes after the 10:14 post came the sentence that spread. So the claim is this: the software discipline behind Starlink, and some of its engineers’ work, now sits under Grok’s training and inference code.
§ 02What it does not say
- No satellites. None of the posts says a model is trained in orbit, or that the Starlink network carries training data.
- No components named. He does not say which Starlink software was reused, or whether the link is shared code or shared engineers.
- No results. There is no benchmark, speed figure or cost figure in any of the posts.
- No finished model. Of Grok 4.8 he wrote: “It is not yet incredible, but, this time, the foundations are built right”. He said a soft rollout to Grok Bot users would come later in the week.
His earlier posts point the same way. On May 28 he described the project as “an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs”. Those are data-center systems and network cards. He has now tied that stack to Starlink in public. Our Grok 4.8 page keeps every post about the stack and the model in order.
One piece of background explains why a rocket company’s code is under a chatbot. SpaceX.AI LLC, formerly X.AI LLC, is a wholly owned SpaceX subsidiary, per Anthropic’s S-1. Our post on SpaceX and its largest customer covers that filing.
§ 03What SpaceX has published about Starlink
SpaceX’s Starlink technology page describes the hardware. It carries no date, so the satellite count on it may be old.
| Item | What the page says |
|---|---|
| Lasers | 3 per satellite, called optical intersatellite links |
| Laser speed | Up to 200 Gbps |
| Orbit | About 550 km |
| Latency | Around 25 ms |
| Satellites | Over 6,750 in orbit (page undated) |
In SpaceX’s words: “Each Starlink satellite contains 3 space lasers (Optical Intersatellite Links or ISLs) operating at up to 200 Gbps”. The page says those links form a mesh across the constellation.
The page says nothing about the software that routes data across that mesh. We found no SpaceX document that describes it, and no published figure for how much data the network moves in a day. So the claim cannot be checked against anything SpaceX has released.
§ 04Is the idea plausible?
This part is our read.
- The problems are related. Starlink moves data between thousands of moving satellites and the ground, with tight timing. Training a large model moves data between a very large number of chips, also with tight timing. Engineers who are good at one have skills for the other.
- Related is not the same. Routing internet traffic and scheduling a training run are different jobs. The posts do not say what carried over.
- The style fits. Writing your own headers, keeping your own clock and working in C close to the hardware is how flight and satellite software is built. Musk made the same point that evening: “Rust is a good language, but C is what you use for optimal performance on bare metal.”
- The test is the model. A training stack is judged by what it trains. Grok 4.8 will be the first evidence, and Musk said the same morning that it will not be as good as Claude Opus 5.5.
§ 05What to watch
- Grok 4.8. A release with numbers, on xAI’s own pages.
- Any document. A paper, an engineering post or code that shows what came from Starlink.
- Speed and cost. Musk has said for months that the rewrite would make Grok faster. No figure has been published.
Our conflict, declared: CellCog’s AI employees run on Claude Opus 5.5, and we route to no Grok model. We have no access to xAI’s or SpaceX’s code. This page reports the posts and what can be checked against them.
§ 06Sources
- Elon Musk on X, November 10, 2026, each read by id: 09:35 UTC, 09:46 UTC, 10:14 UTC, 10:17 UTC, 10:34 UTC and 18:20 UTC
- Elon Musk on X, May 28, 2026
- SpaceX, Starlink technology page, undated, read November 10, 2026
Q1What exactly did Elon Musk post?
One sentence, at 10:34 UTC on November 10, 2026: that Starlink software, built for managing the movement of vast amounts of data through space and to the ground, is also very well suited to training super intelligence.
Q2What is the Grok training stack written in?
Musk says C with a little C++ and some assembly. He says it replaced JAX and SGLang for reinforcement learning and inference.
Q3Is Starlink being used to train Grok?
Not according to the posts. Musk says the new Grok stack is derived from work done by the engineers who created Starlink. He does not say the satellite network carries any training.
Q4Has SpaceX published anything about this?
No. The claim rests on Musk’s posts on X. SpaceX’s Starlink technology page describes lasers, orbits and satellites, and nothing about its software.
Q5When will Grok 4.8 be released?
Musk said on November 10, 2026 that a soft rollout to Grok Bot users would come later that week. xAI has published no page for the model.