Keep in mind while this is the date they officially became part of SpaceX, they have already been collaborating on building models together. So don't expect a significant bump in performance on new Grok models due to the additional data from Cursor--that bump already happened with Grok 4.5[1] and now 4.6.
There still may be Opus->Fable type bump when they finish training larger models. Remember, Grok 4.6 is still Opus-sized and they have larger models in training.
I'm skeptical of this deal. Cursor led the way on agentic coding, but Claude Code and Codex have replaced it. [I use Codex for everything now.]
Still, I'm hearing that Grok 4.6 is almost as good at coding as Codex, but lots cheaper. [My network is also down on Claude--Opus 5 seems like a regression.]
The optimistic scenario would be Grok being a fast-follower with cheaper rates, and staying ahead of open source models. If they can pull that off, and convince enough developers, then maybe they've got a chance.
But $60B seems insane to me. What's Cursor's market share relative to Codex and Claude Code? If it's not meaningful, then it's $60B for an acquihire.
I remember the days of Cursor, like early 2025. Now I use Claude Code for everything.
If the Cursor or its agent cli can offer their models as an enterprise offering (e.g even through AWS bedrock so all corporate compliance nonsense are met) there is a huge market they can capture.
You're being disingenuous. Tesla employee count, including founders, was 2 (or 3, depending on accounts) at the time that Elon "founded" Tesla. Tesla was a paper company with no assets to its name. In any normal company you would have agreed and negotiated cofounder status when coming into a company with that few employees. I don't think the lack of agreement changes whether someone is effectively a founder or not.
They didn't have any IP or assets. He did it because he and JB were already working on doing the same thing and it was better to join forces rather than set up two paper companies competing with each other.
SpaceX bought the revenue and customer relationship to pump the numbers before the IPO. They have a lot of wasted computing so cursor revenue is kinda "profit" as the cost is already paid.
And it was only a great deal because they probably paid with the hugely over-valuated stocks.
60 billion to drum up excitement as an acquisition into an overvalued megacorp that can now be dressed up as "AI". It's just numbers gaming, nothing to do with actual value.
Using grok at work and gpt at home, it's not better. Must be Elon fanboys and bots. It's what Elon does best. Subpar products with guerilla marketing campaigns.
Have you read SpaceX IPO prospectus? Their AI plans are completely nonsensical. I don’t believe there is an actual strategy here other than trying to surf on (and contribute to) the AI bubble
SpaceX is betting that the world will be compute/energy constrained by 2030 and that orbital datacenters will become cost-effective for inference. If they are right about that, they will be the largest company in the solar system. If they are wrong, they will still be one of the largest telecom companies.
To me this is a classic case of low-downside, very-high-upside.
One thing I find fascinating in this thread is how orbital datacenters bring out such visceral reactions.
I understand skepticism--maybe they will never be able to compete economically with terrestrial datacenters. But the profound certainty that detractors have is amazing to me.
There are a lot of unknowns, mostly around prices in 2030 (price to orbit, price per MW on Earth), but detractors are dead certain that not only will orbital datacenters fail, but they will never even be deployed because they violate the laws of physics. It's astounding.
> I understand skepticism--maybe they will never be able to compete economically with terrestrial datacenters. But the profound certainty that detractors have is amazing to me.
I need to finish the blog post about this. I'm around 10,000 words of confident right now about how bad an idea this is, regardless of each specific possible variation I can come up with about how to do it.
Not physically impossible, by the way; just a terrible idea in multiple different dimensions at the same time.
How is Starlink cooled? How is ISS cooled? They all use radiators, which work in a vacuum by--wait for it--radiating infrared into space.
Orbital data centers do not defy the laws of physics. TBD whether they are economical relative to terrestrial data centers, but please don't succumb to misinformation.
For some perspective, the entire heat rejection system on the ISS at peak can do a little less than one rack of dense AI compute power.
All of that plumbing, all of that area, all that stuff to put in orbit just to cool one rack of compute equipment.
How much stuff do we need to put up there for each rack? The External Active Thermal Control System on the ISS is ~20,000lbs, about 10 tons. How many racks are in a single datacenter? The Colossus is allegedly ~1,500 compute racks, ignoring other equipment. So that's like 15,000 tons for just the cooling equipment if its like the ISS. At SpaceX's current public launch costs (which makes a profit for them, so not base costs) to launch that much stuff is somewhere around $40B. Just the launch cost for the cooling equipment alone for just one datacenter building worth of pure compute capacity, that isn't even that big compared to what companies are talking about building.
All of that is very rough back of the napkin math. I don't expect that to be the actual numbers at all, but I'd imagine somewhere in that neighborhood unless they've got something really radical up their sleeve.
They better have a much better way to cool these things than what the ISS is doing.
That's ~13 tons per megawatt, or 1,330 tons for 100 MWs. That's 13 launches of Starship (at $10 million per launch, or $130 million). Of course, that's probably 600 individual satellites, at maybe $5 million per, so we're at ~$3 billion to deploy 100 MW to space.
You can sell 100 MW of compute for $3-4 billion per year, so SpaceX would make their money back in 1 year.
> Current SpaceX design is 75 kW of computer per ton, including solar panels, radiators, everything
Anyone can make a website. Let's see them actually make it and put it in orbit.
According to another Musk website we'd be riding hyperloops and drive through solo tunnels under our cities and we'd all own our own autonomous cab companies back in 2016.
Another website said we'd be able to buy a Cybertruck for $40k and the three-motor variety for $70k.
It's amazing. These guys see a conman with a history of lies and failures - hyperloop, cybertruck, full self driving teslas, robotaxi fleets - and somehow all their critical thinking goes out the window.
Let's just trust him this time guys, he's definitely going to nail it.
I mean to be fair to him, he's built an innovative car company, a pretty rocking space company, and now controls access to basically all government officials and politicians (X).
He's definitely super successful and not a complete conman.
Then again, he's kinda a conman because he's promised absolutely absurd things that haven't happened.
He's a complicated character, 13 years ago I thought he was great, and now I mostly think he's a douchebag. insert complicated shrug emoji.
Musk is a spoiled nepo-baby with rich parents and daddy issues who bought his way to success. His only skill is spewing attractive lies, and he does that quite well. In everything else he is - at best - mediocre.
I've got little doubt they can put some kind of compute in orbit. Shit, load a tiny model on the next batch of Starlink birds, boom, space AI datacenter done. Will it be anywhere close to the numbers they're claiming? Will it actually be useful and competitive to terrestrial datacenters? I guess we'll see.
Did the Cybertruck have 500mi of range? Was there a $40k model at launch?
I don't know if orbital datacenters will work out--it depends on the price of launch and the price of tokens in 2030. Maybe they'll try it and abandon it. We'll see.
But my point is that if it does work, if they can be cheaper than terrestrial tokens, then SpaceX can dominate the market because no one else is even close to being able to deploy orbital compute.
And if it doesn't work, SpaceX will be fine because they are printing money with Starlink.
Cybertruck is a good example--they never hit their sales targets, but Tesla is fine because they are making money on Model Y and they are investing in Robotaxi (which has a bigger potential than Cybertruck anyway). And if Robotaxi doesn't work out, they will pivot to a cheaper model or something.
SpaceX will be fine, not because orbital datacenters are guaranteed, but because they keep betting on things that may or may not work, but if they do work, they make a ton of money.
They do and you’re the one spreading misinformation… the amount of energy you need to radiate for inference GPUs is enormous compared to any computer currently used in space. Energy dissipation by radiating is literally the least effective method and would require insanely large radiators just to be effective on paper. You also need to mitigate bit flips, meaning you throw away around 2/3 of your compute capacity.
And that’s just the most basic things, before taking in account ALL the other problems you face doing compute in space someone with expertise in the domain could mention.
Datacenters in space is a grift that won’t happen in any meaningful capacity and can be debunked by a 15y old who followed their physics courses
1. Radiating heat needs large radiators
2. Bit flips require extra compute capacity
Neither of those violate the laws of physics. As I said, they may not work economically, but that depends on a large number of unknowns, including the cost per ton to orbit and the cost per GW of new power capacity on Earth in 2030. If you have a time machine and know those numbers, then can I please borrow your time machine?
Moreover, when you say, "ALL the other problems ... someone with expertise in the domain could mention," you're saying either:
A) SpaceX engineers do not have expertise in this domain.
B) SpaceX engineers are hiding all the problems that they know about.
B is basically a conspiracy theory. It is one YouTube video away from saying the moon landing was a hoax.
C) SpaceX engineers enjoy their cushy salary funded by a guy who won't take "that doesn't make sense" for an answer to work on rockets and neat technology, even if it may never really work out.
Do you know anyone that works at SpaceX? They get paid well, yes, but my friends who work there say they work you really hard. The bet that's being made is that there will be advancements in materials science that make data centers in space viable. If you hypothesize that there's a way to cool things down in space if you just had the right material to do it, it makes sense. Starlink satellites are already compute plus solar panels. Just scale it up! Or course "just" is doing a lot of work here, but that's the theory.
> Starlink satellites are already compute plus solar panels. Just scale it up!
Starlink satellites use probably less than 1/100th the power budget of even these earlier supposed plans. They don't use much power at all in comparison. They're also big flat pancake for their body, as they want to have a good bit of area for their antenna array. So the other side can just focus on cooling very close to the actual compute parts. It's a radically different set of engineering requirements.
For something line the AI1, you're looking at a few million in launch costs for the mass of the cooling system alone and that's assuming they get Starship flying around their projected costs. What's the install cost per rack or cooling equipment on the ground? Even say I'm overshooting it 100%, those numbers still don't sound very good!
> Neither of those violate the laws of physics. As I said, they may not work economically, but that depends on a large number of unknowns, including the cost per ton to orbit and the cost per GW of new power capacity on Earth in 2030. If you have a time machine and know those numbers, then can I please borrow your time machine?
Alphabet researchers have estimated future launch costs based on learning curves; they estimate that after Starship has launched 370,000 tons to LEO, they will have figured out how to reduce launch costs by enough to make space data centres viable.
They optimistically suggested this might take ten years.
Given the rate of development of Starship as a launch system, I suspect that Musk will reach federal pension age before this is viable.
So between the giant radiator per GPU and the diminished compute capacity, how much mass are we shooting up exactly? This just seems utterly infeasible.
You're missing C) SpaceX engineers know it's a bad idea, but can't say anything because it's the tyrant manchild's latest obsession, technology constraints be damned.
Remember the vision-only FSD? How's that coming along?
> because it's the tyrant manchild's latest obsession, technology constraints be damned.
Imho, there's only one reason that justifies Musk's space-based compute obsession: sovereignty.
Specifically, avoiding terrestrial country's.
This isn't about AI in space (though I'm sure dumb investors love that), it's about uncensorable Twitter in space, beamed down to Starlink terminals.
Which would be one of the few things you'd care about, if you were one of the richest people on the planet... but still subject to those damned nations and their meddling laws.
That's exactly it. Musk sees space as extraterritorial, which in the true sense of the word of course it absolutely is. But all those bits have to go to ground somewhere. As long as that somewhere is friendly no problem. That can change.
I suspect that simply blocking payments from locals to Musk may be enough, considering how Musk reacted when advertisers voted with their wallets to leave X.
But nations have other tools, as Brazil showed when a judge ordered a freeze on Starlink assets because X didn't have reachable assets within Brazil to be frozen.
Unreasonably expensive for pLEO volumes (same problem/advantage militaries are looking at) + collateral damage to other nations satellites
The only nations who would realistically have that as an option would have a robust scaled ASAT capability (maybe US and China) but no pLEO satellites (rules out US and China).
However, based on the rapid development of high-powered lasers, I think we're not too far from it being very cheap. Mostly an optics (adaptive and aperture) problem.
This also means it's likely to get deniable very quickly.
Yeah, I think Musk loves the idea of sticking his CSAM generators in space, where he thinks they will be beyond the reach of terrestrial laws and morals.
"Full self driving", the capability, is not the same thing as the product for which Telsa has been sued because they described it as "Full Self Driving" before it was good enough to get called that[0]; this is in much the same way that I can accurately say that "I use a Time Machine on a regular basis", but this "Time Machine" isn't suitable for the use for which you asked to borrow one in another comment, because this is the brand name of the Apple backup mechanism and not a DeLorean with a flux capacitor.
Would be nice if it was good enough that, as per what he said in Oct 2014, "I think we’ll be able to achieve true autonomous driving, where you could literally get in the car, go to sleep and wake up at your destination,"[1], or what he said in 2016 "Ultimately you'll be able to summon your car anywhere … your car can get to you. I think that within two years, you'll be able to summon your car from across the country. It will meet you wherever your phone is … and it will just automatically charge itself along the entire journey."[2]
Right now, from what I've been seeing, I think the product is roughly the level the Paint It Black video misrepresented the tech as having already reached in 2016. Which is more than nothing, but also not "full".
Maybe this is horse trading: SpaceX pays an insane price in stock for a vibe coding app, Trump will help Terafab and Starlink in any way possible. There are probably other Cursor investors among the Trump clan.
I feel like people aren’t updating to the possibility that a significant portion of software development could be under the purview of SpaceX. Whether it’s developer Twitter, Cursor’s planned GitHub competitor, or Cursor/Grok, that’s a lot of the developer mindshare. Granted, there’s a lot of competition in all of these spaces except for maybe Twitter, so it might not happen. But there’s a chance! And that possibility is fairly concerning.
Guess this means graphite is also going to be part of spaceX, what a weird turn of events here. Guess it’s good GH is actively working to support stacks
I get it and yet I don't. On one hand, my experience with Cursor (albeit limited) tells me that it's a perfect fit with Elon's other ventures. Then again, I don't see how there's anything so special about it that SpaceXAI would want to acquire it for either the software itself or the talent behind it. It seems more likely that acquiring Cursor is a way to get more users for Grok, seeing as how it still has a very small share of the market.
I've never seen the appeal of cursor either, but then I'm a decades-long vim (now neovim) guy and don't use IDEs in general. But for me claude/codex TUIs are a great way to fit agents into my workflow.
My suspicion is that it's a way to break into more developer workflows for Grok, even if the market share is relatively small. I've never used grok and don't plan to at this point, but I have highly respected friends who say it is pretty damn good, and some of them even started using Cursor because of this pairing, so anecdatally at least it does seem to be working.
Grok has started competing with frontier models on coding thanks to Cursor's data, with SpaceX's massive compute they can undercut competition on price, which makes the offering real appealing.
I've now switched to Grok 4.6, I was previously using Opus 4.8 but Grok is good enough now, and way cheaper which is the important part.
The $20 / month isn't as good as Codex or Claude. Codex reset twice this week while I'm sitting waiting for the entire month to role around with Cursor. I have all three so can have a pulse on everything.
So they can play the game that the other big AI companies are playing, where they invest in one another to make each one look bigger on the books, even though it's all the same money, likely.
The data is special. Data of programmers using coding agents in the real world is invaluable for AI training. And their customer base is useful to deploy new models into.
Why not just use VS Code which cursor (and antigravity and the one Amazon is pushing) is a fork of? It will be much closer in functionality and has been loaded up with agent features nonstop for the last 18 months.
If somebody missed that one news, one can easily get to the conclusion that SpaceX tried to manipulate index funds to buy their shares, and sold into an incredibly high valuation to keep the money. Not to reinvest it into the business.
I'm hoping Cursor becomes the third biggest (American) AI model company, replacing Google. Elon works best with top talent, and I think Cursor most kept its team together; were there any big departure from there recently?
I mean I get that we can all dislike this move but it might the best way to get into the race to become one of the top labs. Ofc Grok 4.6 is not there yet, I have seen people claim it's quite good and it is but it's a marginal improvement over the mid tier models from the top 2.
And it's certainly worse than the recent GLM 5.3 (especially once we get the weights for it).
All the models are good enough for almost all normal tech job work. If all you are doing is building a react app there's functionally no difference. You'd only want a frontier model to work on something extra complex but even that will be baked into the basic models within six months.
Maybe if you’re building something from scratch every time, but otherwise the difference becomes noticeable fast. Just today 5.6 Terra tried to silently import half my backend into the frontend bundle. After switching to Sol it quickly found and fixed that.
[1] https://cursor.com/blog/grok-4-5