.....what in the Atlas Shrugged is going on here. There is not a market for either xAI alone nor for reusable rockets at present, and data centers in space is an idea that won't even see the first attempted launch for years. Yet it's so important that it must be the pursuit of anti trust legislation?? Like the other rail/utilities examples mentioned let's look to Congress to pass a law, there is literally negative proof of anti trust enforcement here. There isn't even a market!! Bad take.
To my knowledge, antitrust law doesn’t necessarily require an existing monopoly or even a fully formed market. Obviously this scenario isn’t happening right now, but it could and the risk of giving Musk exclusive control over effectively unlimited data center compute seems important enough to act on
The risk of unlimited data center compute in space is plausible only if you ignore extreme temperature swings, space debris, gamma rays, solar flares, coronal mass ejections, and the Kessler effect, and pretend lift costs are zero.
This idea is about stock price, not plausible uses of engineering.
Musk fanboi: "I am sure it's good to put all the compute beyond the protective cradle of the ionosphere. Radiation will make our AI overlords stronger."
The idea is more to put it in a sun synchronous orbit which will stay within the Van Allen belts. Going over the poles mean more charged particles than you get on the ISS but not a huge difference. Nothing like the vulnerability satellites in geostationary orbit have.
I didn’t know about the Kessler effect until Sabine told me about it the other day. There are concerns that it is closer than previously believed https://youtu.be/8ag6gSzsGbc
You don’t have to ignore these things - you just have to think that the value of data centers is getting high enough that the costs for mitigating these things become affordable.
More do the benefits of being in orbit (cheap energy, potentially cheap space) outweigh the costs (more radiation, harder heat dissipation and launch costs), of the three costs one of them is economic and two of them are arguably mitigatable as technology advances.
This is as ridiculous as creating anti trust laws against Apple for inventing the iPhone in the Jobs era. There is competition even in the notional world in which you're living.
Antitrust enforcement applies to prospective markets if a combination could eliminate a potential market entrant. There is DOJ Antitrust Division guidance on this.
There are other entrants but at time of print, no one is within 2x SpaceX's price in $/kg. Here's ChatGPT:
SpaceX: ~$1,500–$3,000/kg
Blue Origin: ~$2,000–$4,000/kg (target, contingent on New Glenn reuse/cadence)
ULA: ~$4,000–$7,000+ /kg
Arianespace: ~$6,000–$10,000+ /kg
Firefly Aerospace: ~$15,000–$25,000/kg
Rocket Lab: ~$20,000–$35,000/kg
Other small-launch startups: ~$20,000–$40,000/kg
So maybe Blue Origin gets close if they can match SpaceX's technology? Even with competition, I see a strong case for the type of rules MY is advocating. I'm not even clear what the argument against is.
Even if they raise prices, they have massively more capacity than anyone else. And that gap will increase even more dramatically if they get starship working.
Uncharacteristically underbaked take from Matt here. Common carrier regulation could certainly make sense, but the idea that the only reason to vertically integrate is to monopolize a business that doesn’t exist is not well thought through. Vertical integration can drive coordinated investment to innovate and open new markets - exactly what seems to be going on here. If there is an anti-trust issue later it can be addressed in due course.
Unless you think that business is truly transformative. If you work backwards from a future where AI is as central to civilization as something like rail freight was in the 19th century or telecommunications in the early 20th or oil now, and Elon sits at the center of an oligopoly of space-based data centers that power them, the clear forward path is leveraging his space company to give himself an unfair advantage deploying those data centers. In that future, Elon is a modern Cornelius Vanderbilt with all the negative downstream effects of concentrating to much wealth and power in one family. Except Elon is also an emotionally unstable megalomaniac.
My point is that it will take 5 or 10 years before what's he's proposing could even work. If he actually succeeds in starting to build AI in space THEN we can look at regulating it.
When you're talking about the world's richest man who controls (and manipulates) a major information channel and is willing to spend huge amounts on election influence, starting earlier is better than waiting until the problem appears. Building momentum toward legislation against a motivated lobbyist takes time and by the time the problem is obviously a problem, it will be much harder to achieve.
Most of our intuitions about how politics work was informed by an era where news media sources had fairly strong internal cultures of adherence to facts. We no live in a world where the hyper-wealthy control information media with a declining or non-existent culture of truthfulness and neutrality. Legislation restricting abuses by those hyper-weathy people is going to become much harder in the future.
Yep. Vertical vs Horizontal integration is common pain point in working through antitrust law as well discussed by Ben Thompson and Byrne Hobart in a Nov 2024 interview. [1] Examples include various phases of the industrial and the information revolutions, particularly with respect to international politics, energy, and coordinated violence including war and terrorism.
There are more details in Hobart's book, "Bubbles and the End of Stagnation". [2] I only made it a third through before realizing that this could feel like preaching to the choir. Moreover, I have some professional relations with Thompson and Hobart that further compelled me to consider alternative narrative structures for these patterns. Eg, the origins of the progressive movement under Teddy Roosevelt that include antitrust regulation against Rockefeller and other turn-of-the-century baron of industry.
Unless "due course" is the elimination of prospective markets entrants due to the monopoly leverage of the tied market. There is nothing the prevents two separate companies from partnering to "coordinate" investment in an enterprise on standard, non-discriminatory commercial terms.
I disagree, but only because "We should have common carrier regulations on space launches" is such a benign conclusion regardless of whether you think the premise is absurd.
I can 100% get behind common-carrier regulations and anti-trust in general, but credibility matters; examples matter. If you'd proposed that Pets.com was about to have a monopoly on dog food in February 2000, we'd still be mocking you to this day.
Agreed. I just posted my critique of this ... his plan creates the problem he claims to want to avoid. Here is my comment above: Matt, you've diagnosed a real tension but prescribed a Maladaptive response that would guarantee the exact outcome you fear.
Your core insight is correct: SpaceX's dominance in launch capacity could theoretically extend into AI compute. But your solution—common-carrier regulation preventing competitive advantages—reveals Victim identity thinking that treats success as a threat requiring control.
Consider the actual pattern: SpaceX achieved launch dominance precisely because NASA's cost-plus contracting created the dysfunction you now want to replicate. Boeing and Lockheed had regulatory protection and guaranteed access. Result? $400M per launch vs SpaceX's $67M. The "fair access" you're proposing is the same thinking that gave us the Space Shuttle disaster.
Your railroad analogy backfires spectacularly. The ICC didn't create competition—it calcified incumbents and prevented innovation for nearly a century until deregulation in 1980 finally allowed trucking to displace protected rail monopolies.
The Creative response you're missing: If orbital data centers prove viable, the threat of SpaceX dominance will attract massive capital to competing launch systems (Blue Origin, Rocket Lab, Relativity Space). Your own logic admits "space launch does not strike me as inherently monopolistic"—so why preemptively regulate as if it is?
You correctly note billionaires fighting each other (Bezos vs Musk, Altman vs Musk). That competitive dynamic is exactly what prevents monopoly—unless we regulate it into existence by imposing common-carrier rules that reduce the returns to competition.
Don't use antitrust to prevent competitive advantages. Use it to prevent anticompetitive behavior. There's a massive difference, and conflating them turns potential innovation into guaranteed stagnation.
I mostly disagree with the characterization that there's not a market for reusable rockets (there's a market for launch and SpaceX's ability to reuse hardware can lower its costs), but otherwise... yea, when I saw the title this morning I had a nice "WTF" moment.
There's a very strong market case for Falcon 9 and its brethren, but the case for (ugh) "Starship" is a lot more tenuous. There's fairly widespread agreement that you need Starship to work to get the Starlink math to pencil out, and you need a full Starlink network's immense payload needs for a system as large as Starship to make sense. Whether the ouroboros so formed works out long-term depends on a lot of assumptions standing in for questions nobody has yet answered.
The trap here is that there is a market for launches. And there are obvious benefits to cheaper launches. The conceit is that reusable is automatically cheaper. But, that is not necessarily the case.
Falcon 9 suggests this is not the case. The trouble with past reusability has been that it hasn’t truly been so reusable. Many elements of that are now a solved problem.
Apologies, I didn't mean for that to be my implication. I meant that it is the trap many people fall into when talking about the market for "reusable rockets." My assertion is that the "reusable" is catching a free ride in that description.
And really, I just meant my comment to be an addition to yours. You correctly state there is a market for launches. Which decouples those.
No, not at all. This was exactly the industry's widespread conclusion at the end of the Shuttle program. The so-called "big dumb booster" was probably the most cost-efficient way to meet market needs as they were then understood. Computers had gotten small, light, and most importantly cheap. Shuttle was a do-it-all space truck that was predicated in part on "go get a malfunctioning satellite and bring it home for repair and relaunch". In the 1960s and 1970s, this seemed like it made a lot of sense, but the microprocessor revolution caught everyone by surprise.
It turns out that partial reuse certainly appears to be the most cost-efficient and industrially sustainable way to meet launch needs of the late 2000s on into today, but this was only shown in practice by taking a big risk, then going out and doing it (for which SpaceX absolutely deserve credit).
It's legitimately shocking that anybody with a remotely technical background is taking datacenters in spaaaace! seriously. Matt doesn't have that background, so I'm 2/3 inclined to let him off the hook today (the 1/3 remaining is to say 100 rosaries of "I will keep in mind what I'm qualified to talk about), but we are long beyond the point that we should take everything Musk et al. say with a gigantic grain of salt.
From what I’ve seen thus far, media coveage on this topic has been pretty poor. No one’s actually saying there will be giant data centers assembled in orbit that will rival a Google 800K sq. ft. facility on the Columbia River. I assume what’s being considered is a large collection of data centers that are each the size of a basketball, and maybe the size of a small car with its solar panels are furled for launch. But I don’t know because everybody wants to write about AI and Elon.
DC-X, Roton, Armadillo never went to orbit. They were just testbeds for vertical launch and landing. Afaik, SpaceX was genuinely the first to do vertical landing of an orbital launcher. Generally, prior to F9, the perception was that it was probably possible to do, but didn't make sense financially, particularly if you were solely talking about the boosters. It's interesting how wrong this position was, and I recall how many experienced space industry vets loudly said this at the time. And this is with F9 being only partially reusable. Starship is projected to bring down launch costs another order of magnitude, and is fully reusable. I see people pooh poohing this notion and mocking Starship regularly, in much the same way F9 was pooh poohed and mocked prior to its success.
As far as whether data centers in space makes sense, it really just boils down to whether Starship pans out. No one thinks they are viable at today's launch costs. At about $200/kg to orbit they start to make sense. Obviously, Musk is bullish on Starship's prospects, hence bullish on this market in general.
The DC-X was an exciting experimental test vehicle, but it never went to orbit. There have been a lot of other non orbital powered landings over the years too. Roton ATV, Armadillo's test vehicle, Masten's test launchers etc. All cool systems in their own right but even their creators would never claim that they solved reusable boosters to orbit. It's just a fundamentally different thing.
I hear that claim so often (same about electric cars and self driving cars) but who in tech thought these things were impossible?
The need to reuse more of a rocket was what lead to the Space Shuttle for instance. Everyone knew re-usability was needed.
As far as I know noone thought reusable rockets were impossible, at most unfeasible with current technology and financial use cases. And SpaceX did nearly fail to get the tech working and for finances they became their own largest customer through Starlink.
One of Musk's singular talents is convincing people only he thought XYZ was possible, when in fact a great many smart people and talented organizations have studied XYZ for decades.
But this is true of every engineering accomplishment. No matter what it is or how impressive it is. It's always the case that it's widely understood to be physically possible before it's actually built.
I think when people say this wrt SpaceX specifically, no one is saying that reusable rockets etc were thought to violate the laws of physics. No one believed that. But it is true that, before F9 revolutionized the launch market, most space industry folks (who imo learned all the wrong lessons from the Shuttle's "reusability") believed that VTVL was impractical for orbital launches - both in the sense of being too technically demanding and not likely to be a big source of cost savings. But they were wrong, and now SpaceX absolutely dominates the global launch market.
You can say the same thing about Starlink. Again, many of the same jaded industry types thought Starlink was ridiculous, drawing (as with Shuttle) all the wrong lessons from the failure of Iridium. As with reusable boosters, no one was claiming that low latency LEO internet was physically impossible - rather, the claim was that it was a boondoggle, no one would use it, it would be a black hole for money, etc. But again this was all incorrect. Starlink works extremely well today, so much so that access to it has become a geopolitical flash point, and SpaceX makes lots of money off of it.
Same with creating a new car manufacturer in the US, for that matter. Again, obviously physically possible to build EVs etc - but doing it at scale in the US? Before Tesla actually did this, almost everyone considered the idea to be absurd.
For people to now say, oh all of this stuff was no big deal. Everyone knew that reusable boosters were the way to go. It is frankly just incorrect, and I think that (surprisingly widespread) perception contributes a lot to the present day underestimation of Starship.
Does this mean that data centers in space is a slam dunk? No, not necessarily. But imo at some point you have to update your priors re: "Elon Musk is proposing something that at first glance looks like a total boondoggle, could it actually work as a business?'
EVs clearly weren't technically impossible; hell ... the first one was built in 1888. But I think it's easy to forget that in the early Tesla Roadster days it was general industry consensus that the major auto OEMs were un-disruptable because starting a new car company was financially and operationally impossible. Also battery costs were $500/kWh. They're nearly $100 today.
This is Matt's worst take ever, possibly. Essentially seizing SpaceX as a resource for the government.
This is very different from railroads where there were many companies that needed rules of the road to compete. Its just targeting of one company because of their advanced capabilities.
Where did he mention seizing SpaceX? I thought he said to treat SpaceX the way we treat AT&T and Union Pacific and Western Union, none of which were ever seized by the government.
Matt is saying : SpaceX is too important to be allowed to operate independently. They should be forcibly turned into a utility.
This proposal from Matt is not regulating an entire industry - just this one company that might one day become an industry. Matt's proposal is turning a private company into a government utility, with no vote from the owners of this private property. It would arguably be a wholesale theft of value by the government from private citizens.
Treating the entire industry as utilities doesn’t mean nationalizing it! There are some government-run entities in gas, electric, and water, but the intercity railroads have (almost?) always been entirely private despite being regulated as utilities for about a century.
This is true, but if Musk’s plans come to fruition the Common Carrier regulations seem well suited for this. And you can’t wait for the commissioning of massive data center capacity in space. By that time it would be too late. Musk would have used his space capabilities to launch data centers, at scale, while potentially blocking access to space for AI competitors. We should not allow that last part.
Congress can pass a law very easily. Anti trust is not the vehicle for this in any way at this stage. There are other orbital carriers with more capability growing daily - Boeing+Lockheed have a partnership. Northrup buying Orbital ATK. French company Arianespace. Japans Mitsubishi. Rocket Lab has launched more than anyone except SpaceX. This whole discussion reeks of simple ignorance over the state of the industry imo.
You’re vastly understating SpaceX’s market position. Comparing Rocketlab to SpaceX? Their launch capacity is tiny. Right now, SpaceX has delivered an order of magnitude more mass to orbit than any other company besides Roscosmos (which has been operating Soyuz since the 60s). Blue Origin’s New Glenn rocket will be able to compete with the Falcon 9 and Falcon Heavy (assuming they can manufacture them at the same scale) but Starship will immediately leapfrog the competition again. None of this is to say it’s a non-competitive industry, but in terms of kilograms to orbit, SpaceX owns a huge amount of the market share, and they’re already converting it into a near monopoly on satellite internet.
And? Theres obvious competition and a market leader due to being a first mover. Comparing the two could be simply like comparing the iPhone to Android. Apples market lead led to a huge company that a decade or two later has less distribution and a dwindling market share in top markets with less differentiation over time. I dont understand why we need to pretend that LEO is an impossible task esp with the investment in it larger than anything else besides data centers atm
I mostly just took offense at you suggesting someone else was ignorant of the state of the launch industry while in the same breath implying Rocket Lab (largest current rocket: 300 kg payload capacity) had a comparable launch record to SpaceX (smallest current rocket: 25,000 kg payload capacity). I don’t strongly disagree with your conclusion but that doesn’t excuse the rhetoric.
Just want to point out that RocketLab is specifically targeting the smallsat market, so it's not a knock on them that they have relatively small payload rockets.
Rocket Lab has launched more than anyone but SpaceX. I dont believe I said they were comparable? Just that there are clear other options and rising competition in pretty much every area that SpaceX and/or xAI develops. Who knows where it will be in the future, first movers do not always equal winners, and more often than not are not the long term industry winners
Isn't the main problem that merging is largely a lifeline from unrelated markets to prop up gambles that are currently losing?
I confess my main gripe would be as a shareholder of whichever is doing better. But I also don't see the reason that the market should allow unrelated merges on a regular basis.
Stopping a merger because it's a fake cash infusion can be stopped on “brother-sister” grounds if it's truly a breach of one company's fiduciary duties, but I have a hard time seeing being the case here. Even then if the two companies are entirely privately owned by the same person it's usually allowed even then. It's probably why it's happening in advance of any IPO
Largely agreed. I would argue for more scrutiny in the merger than a ban on monopolistic grounds. In particular, I would take a fine tooth comb to make sure they aren't trying to shed any liabilities in transferring any ownership.
Totally Agree! Matt's judgement is being clouded here. Even if somehow this idea works, there will still be plenty of competition by earth-based computing centers, plenty of competition for moving things to space, if this idea holds lets just implement this for all data centers already! Lets implement this for anyone who wants to just build a really big datacenter.
It’s not about the data centers - it’s about the space launch. And he agrees - his main demand is that space launch should be treated as a common carrier utility!
This article presupposes that there are other AI companies out there with bold plans to put AI data centers in space but who will be blocked by SpaceX boxing them out on behalf of x.AI. Is that really a thing though? this seems like another Moon shot that Elon wants to try, and we should let him, and we should get out of his way. Just like with Tesla or SpaceX, he's doing something so bold and authentically innovative, that we should not preemptively put up roadblocks and create more regulatory uncertainity here--at least not for the reasons stated.
Let him prove his concept, like he did with rockets and electric cars. Then we can think about regulations after it hurts consumers or competition.
Why SpaceX and not the other half dozen providers? Why is anti trust the mechanism? Why ignore the competition from other tech companies and other heavy industry?
In terms of why making SpaceX do it it's because they are already close to being a monopoly. 85% of global launch capacity is controlled by SpaceX. A fraction that is likely to increase in the near future.
At the moment the only competition is ULA, who aren't competitive and kept afloat by government reluctance to be wholly dependent on SpaceX, Blue Origin which remains a vanity project for Bezos more than a successful company, and Rocket Lab who currently competes in the light launch market that SpaceX isn't interested in, although their Neutron rocket may make them competitive in medium launch if they can scale.
It's difficult to overstate just how far ahead of all their competition SpaceX is.
It's because they did it first. Should we have regulated the iPhone for being first? Would that have changed the dynamics of today's smartphone industry for the better? This is pure punishment politics imo
It's better to think of these anti-"monopolists" as anti-capitalist pursuing their dream for people to owns the means of production. Innovation be damned.
My point was that the downstream market dynamics are not predictive based on first mover status in any industry, and so doing regulation before the industry even functionally exists is forced regulatory capture. If Elon gets a windfall for a few years for being first it doesn't seem like such a disaster when there is now a global race on. Most of the first attempts will be in LEO which is a far smaller lift for competitors than further out GEO
I have shitty broadband at my house because cable companies were the first to run data cables into homes and then they successfully lobbied to escape common carrier laws. I think that's really all Matt is arguing. Companies that launch things into space should be treated as common carriers and SpaceX shouldn't get a pass just for being first.
I don't know that the political solution Matt is pushing for is warranted. That's because I'm fairly skeptical of the economics of space data centers. I do agree with Matt though that SpaceX is the only company that could plausibly do it if it can be done at all.
Maybe that will change in the near future, but Blue Origin has been trying to be SpaceX for longer than SpaceX has existed, so I think we should be skeptical that catching up will be easy.
Econ 101 says that natural monopolies should be regulated as utilities, while monopolies formed through market manipulation and cartel behavior should be broken up. It's two sides of the same coin. And utility regulations pretty much always apply to all companies in a sector, not just one.
I thought the issue was if you don’t treat the industry as a common carrier industry, then Elon Musk will monopolize it. He never suggested special treatment for Musk.
What is the evidence they are discriminating on launch services? They are currently helping Starlink rivals get their constellations orbited. I work in an industry adjacent to this and I have not heard any claims that SpaceX is refusing launch services.
I think the claim is that this sort of discrimination is the only explanation for the potential synergies of merging the two companies, which is the official reason for the merger.
Now if, as everyone believes, the actual reason for the merger is just to get a cash infusion into his floundering social media and AI businesses, then that’s a different story. But I believe he’s not allowed to say that’s the reason they’re merging, since SpaceX shareholders would get nothing out of that.
Potential synergy within the combined company is not evidence of real or potential discrimination though. And by the time, if ever, there are AI satellites ready for launch, Blue Origin and other launch providers will likely have closed the gap.
So Bezos will only launch AWS datacentres and Musk will only launch xAi openAi/Microsoft/Google/Oracle etc get squeezed out. All Matt wants is for all those companies to get the same access to space at the same prices as AWS/xAi so the ai field remains as competitive as it is now
Launching stuff into space isn't that hard (we've been doing it for 70 years), if SpaceX starts denying its services to other companies they'd be doing a massive favor to its existing competitors.
And any large tech company (or country for that matter) could launch new competitors if they wanted; Aerospace tech is a huge existing industry with many expertise and capabilities out there to be bought.
That doesn’t seem to be true, given the long struggles SpaceX had and Blue Origin is still having. If Amazon is having this much trouble a decade or two in, why would Google or Apple or Microsoft be able to do better?
Compared to, for example, AI investment in these companies is actually pretty low. Also, SpaceX and Blue Origin both wanted to start from scratch partly because they're both billionaire vanity projects.
Blue Origin costs 2 billion a year right now when they're weirdly huge, which isn't a lot compared to what companies are spending on Datacenters for AI right now.
If you could get a functioning rocket company up for say 10 billion, when the yearly capex investment for AI is 100s of billions; That means that if datacenters in space ever become a big thing for datacenters building your own launch corp is not a big investment.
Also, they'd be able to just hire people away from the existing rocket corps and use their suppliers. It won't be that hard to catch up.
For one thing, they face nation-state competitors who are not profit-motivated and are willing to endure large losses for long periods of time- that's a pretty fundamental difference
Before we know if it works? What if he responds by saying “forget it, I am not going to spend a fortune on this borderline impossible thing if you are already trying to get your grubby regulatory paws on it”. Now we have no space compute at all!
Musk seems like the only person for whom it might plausibly work because he's the only one facing the problem of having a huge excess of launch capacity.
I see space data centers as an answer to the question of what to do with Starship more than anything.
Leading with that specific link was a poor choice.
But you also need to read these announcements with a Silicon Valley hype decoder. Google is already working on orbital prototype launches. This is not just a thought exercise.
Catching up on other comments, I now see that you're coming at it from a position of technically-informed skepticism. I'm not arguing against that, and I even upvoted lhe top-level comment, because I agree with the main point.
I'm only responding to the first sentence:
> This presupposes ... there are other AI companies out there with bold plans to put AI data centers in space ... is that really a thing though?"
Companies are investing! That doesn't mean it will succeed, but it is a real thing.
I think your take is about spot on. It bears watching, for the reasons M.Y. stated, but the time to impose regulations is when, and if, Musk starts massive data center launches.
“Musk is trying to do something here that is bad for the country and the world and that is also bad for various non-Musk billionaires and corporate actors.”
I don’t think this is ever really justified? What reason is there to believe that space data centres are a crucial piece of infrastructure? What reason is there to believe Musk wants to monopolise the industry, let alone reason to believe he could?
There's currently 1 AI model running in space launched by the Chinese. No one even knows if running data centers in space is feasible, cooling alone is going to be very difficult to solve. Do you know how much heat an nVidia B200 generates?
I don’t think putting your data centres in space will do anything to stop government regulation. Governments regulate foreign companies with no domestic footprint all the time when they do business with domestic customers.
Unless Musk actually succeeds in making his city on Mars (or now apparently the Moon), any AI generated child porn or other illegal material made by space AI data centers will be transmitted to users who live on countries on Earth, hence giving those countries at least some oversight of the company which is sending the data (SpaceX).
How lucky we are to live in such an age of technological wonders in which we must contemplate the legal questions posed by having a drug-addled billionaire using space data centers to beam kiddie porn created by artificial intelligence down upon us from the heavens.
Technically, he can already do that in the US -- virtual child pornography is potentially constitutionally protected speech per Ashcroft v. Free Speech Coalition, 535 U.S. 234 (2002) (https://en.wikipedia.org/wiki/Ashcroft_v._Free_Speech_Coalition). Prosecutors have to show that the content *also* satisfies the Miller test for obscenity for it to be criminally prosecutable.
Fundamentally they want engagement and are apparently now using a paywall to limit the explicit image function of Grok. It’s a complete disregard for the adverse costs their actions impose on others.
Data centers are clearly a crucial piece of infrastructure already, as we see by the chaos whenever cloudflare goes out for a few hours or whatever. It’s quite plausible that space based data centers will be more reliable than ground based ones, and will grow in importance as land use and electricity production become more binding constraints.
Musk obviously wants to monopolize anything he can.
There are MULTIPLE major issues with putting compute in space, including extreme thermal differentials between sun and shade (500 degrees or more), issues with shielding electronics against gamma rays, huge issues with solar activity / solar flares, huge issues with space debris, and hanging over it all, the potential for a cascading Kessler incident.
And that is on top of lift costs, which are not small.
Data centers in space is kind of the space equivalent of “we don’t need to cook when there’s Door Dash.” Sure, if you ignore all practical constraints.
We are not likely to need anti trust for this. Meat processing and PE consolidation of service providers, yes, anti trust enforcement would be a Godsend. This, not so much.
1. AI inference is extremely robust to occasional bitflips. If a single cell in a matrix goes even from 0.0 to 1.0 (the worst case scenario for a single flip), the resulting effect is basically zero in a high parameter model. Just reload the model from central store once daily, and you're fine. This is a non-issue in my opinion.
2. Solar flares: they will very occasionally take out a small portion of your cluster, which is equally true of various natural disasters effecting your network of data centers on earth.
3. You would want these satellites in quite wide orbits, above the belt where ex: Starlink flies today. At that layer, the odds of debris hitting your satellite in any meaningful way are quite low.
I think the right way to think about this is "what percent of my satellites could I lost per year due to solar flares, various equipment failures, etc, over the expected ~5 year useful life of the satellite?" Even with a ~10% annual loss rate, I think the math pans out fine.
Heat dissipation feels like the biggest issue to me. Though I am far from an expert on that particular topic, I feel bullish on us being able to make the marginal (sub-linear) gains needed in that domain to make this project viable.
I don't know how you can do easily dismiss Kessler concerns. The cross sectional area of a single gigawatt scale solar panel array + radiators is going to be literally thousands of times the area of all satellites hitherto launched into orbit. And this all only makes sense if you can theoretically go from gigawatt to terawatt.
Demand a bottle deposit on everything that goes into space that gets refunded when it's confirmed to burn up on re-entry. Go with the Michigan model and make it 10 cents. If it's just 5 cents no one will care.
1. That’s soft errors, not hard. Hard errors result in permanent damage. Reloading the model does not fix that.
“Radiation is undoubtedly the main threat to electronic components in space. Cosmic radiation interacting with microchips can cause memory errors. High-energy ions or protons passing through a transistor can lead to short circuits, electron leakage, and, as a result, irreversible damage that can jeopardize an entire mission. While problems sometimes arise at launch, many manifest once the spacecraft leaves low Earth orbit, since closer to Earth, it is still shielded by the atmosphere and magnetic field. That’s why astronauts aboard the ISS often use electronics based on conventional “terrestrial” microchips. But GPS and GLONASS satellites are not so fortunate: at an altitude of about 20,000 km, the penetrating power of energetic particles is so high that the use of standard microchips is out of the question.”
2. Loss of entire data centers due to terrestrial disasters is quite rare.
3. High orbits are actually more risky for Kessler effect, as if you do get a cascading series of collisions, clearing those orbits could take thousands of years.
Again, the math is "what percent of the cluster can I lose per year and have this be viable?" I think it could be made to work re: shielding compute and accounting for the occasional solar storm intersecting your orbit, but I admit I haven't done any deep research on this.
I agree with #3, though the trade there is that high orbits make collisions themselves super linearly less likely. But yes, this is a risk. I do think that this could be a blessing in disguise though, in that it will constrain the AI build out to only terrestrial energy sources, which will perhaps put enough of a natural brake on it to slow down an intelligence explosion enough for us to figure out what to do about it.
I mean NASA and DARPA have toyed around the idea of putting computers or networks in space for decades. I still think it is more difficult than Elon thinks it is from a compute and maintenance perspective. I'm bullish on this happening in a decade or so. But with energy costs potentially going down and compute power continuing to get more efficient as chips do over time, is this worth the squeeze for him now? Remains to be seen. It doesn't help that he overpromise's a lot.
"what percent of my satellites could I lost per year due to solar flares, various equipment failures, etc, over the expected ~5 year useful life of the satellite?"
This sentence alone makes me think its not feasible. Who is going to spend the enormous amounts of money to build and launch something significant into space with a **5 year expected useful life cycle!!! when you could just build it on the ground and it will work for way longer than that!
People are taking a bunch of half-baked complaints and putting them together and thinking they have a fully baked complaint.
I think heat dissipation will end up ruining this idea on its own. But that one is pretty obvious to falsify: launch it and it either fries to death or doesn't, not much middle ground where we're not sure.
Most other tech issues people bring up simply aren't. They are things people think they saw in a sci-fi movie once.
"Even if you build a solar-powered GPU data center here on earth, 80% of you CapEx is worthless well within 5 years."
1) Its not worthless, its just not as good and likely has a decline that you can coax value out of for a longer period of time - making it more profitable.
2) But even if you have to replace all the GPUs, much of the rest of the complex is still in place and usable. That's 20% you don't have to replace compared to something in space. 20% is a large margin!
3) You could do a mix, upgrading over time instead of having a CapEx cliff. Or the reverse, Intel comes out with a chip that is 500% better than the current Nvidia model (not likely, but still). You can begin upgrading something on the ground, while the thing in space is stuck at its built performance.
1. A used A100 (top-tier chip from 5 years ago) is ~$4k in mint condition, less than 10% of the spot price of a top-tier trip from today (GB200). Good luck finding another capital asset that depreciates so quickly. There are few data center workloads, except maybe crypto mining for some reason, where it is useful to run a fleet of these.
2. Yes, but after 5 years on the ground you are also replacing batteries, which are very expensive and prone to failure. "The rest of the complex" is stuff (batteries, buildings, high power cabling, plumbing, etc etc) is stuff that just straight up doesn't exist in the space based deployment. Your coat of re-using it in space is 0, because it does not exist. Think of it this way: both modes of deployment have an AI chip and solar array that are near worthless in 5 years. The ground setup has all the stuff I mentioned above; the satellite setup has a chassis and radiator instead. It's not immediately obvious that at mass produced scales the former is costlier than the latter. It may be! But I feel like it will be pretty close.
I am not saying it's a slam dunk. I get that the benefits are marginal or nonexistent at the moment. I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space.
1) I agree that they degrade in cost quickly, but that doesn't mean they don't work. The chip industry is full of examples where less powerful chips are still useful in some capacity. But even if its the reverse, that just underlines my point which is that if chips are depreciating faster, than having to relaunch a whole new data center every 3 years instead of upgrading a groundbound center is worse!
2) Industrial batters don't fail in 5 years...most are targeting 20+. Beyond that, the reason that all that stuff exists on earth and not in space is because its solving for problems that a space center hasn't figured out yet! You have confirmed energy delivery, cooling understood, data transmission addressed, replacement/upgrade cycles understood and addressed, transport understood and addressed. Meanwhile in space, none of these are even understood much less addressed!
3) Let's talk about just that last one, transport. A server rack weighs in at 2-4 thousand lbs. Data center has 1000+ server racks. That's 2-4 million lbs, or 1000-2000 tons. The Falcon heavy can lift 64 tons and the Starship is posited to do 150 tons. So that works out to at best 7 and probably closer to 30 launches to put one data center in space. At 31 launches it would currently costs about $70+ million to get a Falcon launched. So current price just to transport a data center into space based on weight of server racks would be $2 billion dollars. Even if the price for weight falls dramatically, the transport costs would huge. And if AI really takes off, and we are needing to build thousands of data centers, it would take 10s of thousands of launches when SpaceX is doing a total of 150 a year. Not only do you need to figure out the engineering in space for a data center, you need to scale up launching so that you are doing hundreds a day!
"I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space."
I think that if this is true, its because AI has reached AGI and we live in a very different place than we do now.
1. A used A100 (top-tier chip from 5 years ago) is ~$4k in mint condition, less than 10% of the spot price of a top-tier trip from today (GB200). Good luck finding another capital asset that depreciates so quickly. There are few data center workloads, except maybe crypto mining for some reason, where it is useful to run a fleet of these.
2. Yes, but after 5 years on the ground you are also replacing batteries, which are very expensive and prone to failure. "The rest of the complex" is stuff (batteries, buildings, high power cabling, plumbing, etc etc) is stuff that just straight up doesn't exist in the space based deployment. Your coat of re-using it in space is 0, because it does not exist. Think of it this way: both modes of deployment have an AI chip and solar array that are near worthless in 5 years. The ground setup has all the stuff I mentioned above; the satellite setup has a chassis and radiator instead. It's not immediately obvious that at mass produced scales the former is costlier than the latter. It may be! But I feel like it will be pretty close.
I am not saying it's a slam dunk. I get that the benefits are marginal or nonexistent at the moment. I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space.
" just build it on the ground" isn't so easy when there's such massive NIMBY backlash to datacenters these days. No matter how expensive it is to launch stuff into space, it's even more expensive to fight a never-ending legal battle against every single local government in the world.
> AI inference is extremely robust to occasional bitflips
This is goofy. I question whether you understand what a bitflip is. A single bitflip in firmware can cause corruption that can brick the hardware. In software it can crash the program. It doesn't just make a parameter wrong. You're thinking at the absolutely wrong abstraction level here.
> If a single cell in a matrix goes even from 0.0 to 1.0 (the worst case scenario for a single flip)
This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0, it makes the value completely unreadable. What are you even talking about? Datacenter servers don't have ECC memory just for fun.
I understand perfectly well. By mass, 99+% of the silicon being sent up is GPUs and ECC'ed memory. Only a tiny , isolated portion is dedicated to firmware, microcontroller, etc, which can be radiation shielded at relatively low (in terms of weight) costs.
> This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0
I did not say that it would. I was specifically enumerating the worst thing that could happen from a single bit flip of working memory neural net weights, not claiming that this specific switch would necessarily happen. Last I checked (a while ago, I don't know what the state of the art for vector dbs is), neural net weights are generally stored as matrices of small width floats (fp16?), clamped at 0.0 <= x <= 1.0. The worst case scenario, in terms of weighting, is that a single bit flip maximally reverses the weight: makes a near 0 weight go to near 1. Obviously a bit flip could also turn the float into an out of bounds value or NaN, but this would be an easily detectable error, and hence not the worst case scenario.
> What are you even talking about?
Why be mean and condescending? Why not be charitable and curious instead? I'd offer you the same courtesy.
I understand perfectly well. By mass, 99+% of the silicon being sent up is GPUs and ECC'ed memory. Only a tiny , isolated portion is dedicated to firmware, microcontroller, etc, which can be radiation shielded at relatively low (in terms of weight) costs.
> This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0
I did not say that it would. I was specifically enumerating the worst thing that could happen from a single bit flip of working memory neural net weights, not claiming that this specific switch would necessarily happen. Last I checked (a while ago, I don't know what the state of the art for vector dbs is), neural net weights are generally stored as matrices of small width floats (fp16?), clamped at 0.0 <= x <= 1.0. The worst case scenario, in terms of weighting, is that a single bit flip maximally reverses the weight: makes a near 0 weight go to near 1. Obviously a bit flip could also turn the float into an out of bounds value or NaN, but this would be an easily detectable error, and hence not the worst case scenario.
> What are you even talking about?
Why be mean and condescending? Why not be charitable and curious instead? I'd offer you the same courtesy.
Yeah, I'm involved in amateur radio satellites (https://www.amsat.org/) and it has taught me that space is a very harsh environment for electronics, even relatively simple stuff like UHF/VHF analog radio repeaters.
Whatever NIMBY obstacles might exist, it just seems far easier on basically every front to just build data centers on Earth. For the cost of getting a bunch of electronics to orbit just buy land and pay off the local government to let you build a gigantic solar array.
The thing is, earth based solar panels are in the Earth’s shade exactly half the time, and are often in the shade of other things, like big clouds of water vapor. If you have the right sort of orbit, space based solar panels will never be in the shade of anything.
I am against hobbling our only successful space company- and a world leading one- because the owner has bad politics and is intemperate on Twitter. Who among us, etc.
The extent to which "maybe Elon will do it" is still the official industrial policy of the United States is ridiculous. He's going to drag us back into PV manufacturing as a part of this, which should be a giant emergency that Congress wants to solve.
I don’t think it’s hobbling the company to say that it has to offer the same deals to all customers, including subsidiaries. We already do that for other important infrastructure companies, like Warren Buffet’s BNSF railroad.
I was under the impression that, as a result of measures taken under Obama and Trump I, we are already self-sufficient in PV cell output?
Anyway, if Musk's engineers can make it dirt-cheap via process improvement and automation, so as to compete with China's "subsidies, no environmental controls, and scale" model, all to the good.
“Musk is trying to do something here that is bad for the country and the world”
Hold on there. Musk is trying to do something to solve one of the world’s most pressing problems - AGAIN! Many of these problems, that Musk attacks, are left-wing priorities (municipal traffic, practical EVs).
He is now focusing his remarkable talents on using space, which he has made accessible, to mitigate our power and climate concerns. This is what you characterize as “bad for humanity”. It’s GREAT for humanity. It’s also true that we should not let Musk exploit his space leadership to block other AI companies from innovating space AI solutions. Your argument for Common Carrier regulations is sound, and maybe what that cures is all you meant with your “bad for humanity” comment. If Musk voted for Kamala I believe you’d have made that clear.
“As a result, this merger poses the very real risk that the combined company will be able to leverage a dominant market position in the space launch industry into a dominant market position in A.I. If nothing else, that is clearly what Musk wants to achieve with this merger.”
I’m pretty sure what Musk wants to achieve is a lower cost of capital for xAI. Turns out there are plenty of people that still take his word at face value though.
Applying common carrier rules to space launch makes sense, but the rest of this seems a bit tenuous. Why does the merger matter if Musk already directly controls both companies? What about the fact that SpaceX currently launches Project Kuiper satellites that directly compete with Starlink? Is there any reason to believe they wouldn’t launch rival data centers?
Agreed, but common carrier regs for commercial space seem like a good idea in general (not just for space data centers), and blocking the merger seems bad unconditionally
I've been in the datacenter and telecom space for 20 years and the general experience of the industry is that anti-trust/common carrier legislation is good in theory but bad in practice. It is *theoretically* good to enable competition with a low capital barrier to entry. In reality, this has disincentivized significant capital investment in digital infrastructure outside of the two boom/bubble cycles unless directly subsidized by the government (which in itself is a whole other saga) - the competitive phone carriers that emerged in the 80s solely as a service on top of Bell infrastructure all failed relatively quickly, did nothing to improve quality of service, and pushed consumer prices so low (good) that there was no incentive for operators to reinvest (bad). There are other drivers at play here, of course, but the right way to deal with the AT&T monopoly in the late 70s would have been to make local exclusivity agreements illegal (something that is less relevant today but still needs to happen) and force consolidation of authority around permitting (something that is more relevant today and still needs to happen). Relative to what actually happened, this would have dramatically increased the competitive barrier of entry, which is bad (in theory), but winnowing the competitors only to well-capitalized players would have pushed competitive infrastructure buildouts earlier, increased mean quality of service and (theoretically) improved consumer pricing more quickly. Since SpaceX has neither a regulatory nor a natural monopoly, there is no good reason *in practice* to enforce the same kind of structure just because they are perceived to be clearly ahead (which is true but we are far enough away from datacenters in space being practical and there are enough competitors trying to solve the cooling issues at scale that it is plausible if not even likely that they won't be by the time this is relevant).
The telephone story is interesting. But can you explain why the 3 cell carriers and the handful of cable companies offer such low speeds at such high prices compared to other countries?
The premise of this question isn't entirely correct, but nonetheless the regulatory environment around common carriers has nothing to do with the answer:
1. The median broadband/5G speed in the US generally better than in most other countries, and the exceptions to this are almost universally wealthy petrostates with nationalized access providers (i.e. a lot of capital available to solve the problem via enforced monopoly).
1a. Above 100 Mbps, speed isn't the best determinant of service quality for consumer internet anyway, median latency and lack of bufferbloat are, and by these metrics US ISPs are generally light-years ahead of peers elsewhere, although much of this is due to work by backbone and content providers, not the ISPs themselves.
2a. Median broadband pricing is somewhat higher than in many peer countries. Controlling for service quality in a quantitative way is difficult, but this is almost certainly less true or untrue once you control for service quality.
2b. Median wireless data pricing is dramatically higher than in many peer countries.
3. Density (low) and scale (high) are huge problems in the US that don't exist in most other peer nations in a substantial way. The capital required to launch an infrastructure-owning competitor at national scale is dramatically different here than in an EU nation.
3a. A byproduct of this is that the major MNOs in Europe own a majority of their own towers. This is not the case in the US and dramatically increases US MNO OpEx on a unit basis.
3b. The cost to build infrastructure of any kind is dramatically higher in the US for reasons beyond scale (queue Abundance argument, etc.). This dramatically increases US CapEx on a unit basis.
3c. Government subsidies for broadband/cell access in the US are narrowly tailored in a way that they are not in most peer countries.
3d. All of the above lead unit pricing higher but also create other perverse incentives for operators that are bad for consumers in attempts to grow profits.
"2b. Median wireless data pricing is dramatically higher than in many peer countries."
It seems to me that our wireless providers are much better at price discrimination than peers elsewhere. Bundle everything, especially including devices and help getting new ones up and running, into a single contract for well-off but uninquisitive consumers and sweep a bunch of the surplus off the table, while offering cheaper "utility-type" plans for more sophisticated mid-market consumers and cheap MNVOs for poorer or more budget-conscious ones.
I pay a comparable amount in 2026 for four basic unlimited data lines as I did for 2 separate plans with capped data in China a decade ago, and that's without having gone down to an MVNO. Maybe China has fallen a ton in price since then, but I doubt it's gone so far as to put the shoe clearly on the other foot.
I generally agree (although the proper comparable would be Vodafone/DT in the EU, the Chinese market is too dramatically different IMO). What I am trying to get at in 3d above is, in part, this type of behavior. Of course the Big 3 here are allowed to do a lot of things they would not be able to in the EU, but I don't think the degree to which their behavior is anti-consumer would be as significant if they could make decent margin at what we would consider a promotional service base rate.
Yep. I don't even know if the bundles are really anti-consumer, though the case could be made either way.
I was, however, under the impression that US telecoms infrastructure gets basically a blanket exemption from most of the usual processes that get leveraged by NIMBY nutjobs, no?
In some respects, yes. You don't need zoning approval to build most infrastructure except macro towers and small cells (and even then not in all situations), but there is still local permitting required which has its own sets of challenges and restrictions depending on what and where - and really the issue is the lack of uniformity in what you're going to encounter from jurisdiction to jurisdiction, or in many cases within a jurisdiction depending on what individual in a permitting office you may be dealing with! A lot of success comes down to relationship building with those in control of permitting, which is a great business practice but not exactly how government is supposed to work.
Amusingly this discussion reminded me that I have not yelled at Verizon's Fios people lately and, sure enough, my bill had crept up to their listed long-term rate. A brief chat and it's now chopped in half for the next twelve months and I have an annual calendar reminder in to do this in perpetuity.
Wait, what? My experience of US 5G has been leaps and bounds and climbed-cliffs ahead of my experience in Europe or Japan, let alone China.
The last time I had a fiber connection abroad was in Beijing in 2016, but it was rated at 500 Mbps and it still sucked. My current 300 Mbps connection here is comparable in price and vastly, vastly better, and has been since we bought our house in 2017.
The plural of anecdote is not data, blah blah, but it's enough to induce a deep sense of skepticism any time someone uses rated speeds to tell me that US telecoms suck.
Having a vibrant competition in space launch is a good thing. Therefore, in the present regulatory environment (i.e., based on whatever is bugging Donald Trump), the best way to keep at least two companies in that competition is to cheer on Jeff Bezos's destruction of the Washington Post.
Full disclosure: I work in the industry, but not for SpaceX. Thoughts are my own etc etc.
I think the bull case here is quite strong. If you look at where the puck is going on two fronts (launch cost per kg, and AI uptake and power consumption growth), this seems like the only long term viable solution. Three variables to consider:
1. Do you think AI demand will continue to grow at least linearly (if not exponentially) basically forever, a la railroads, the internet, or energy production? Set aside short term shocks like bubble pops, and think in 10+ year timelines.
2. Do you think the cost of a kg to orbit is going to go down with scale (of both individual vehicles, and in number of vehicles) in the usual "learning curve" that has applied to every industry from air travel to solar panel production?
3. Do you think we can develop slightly more efficient methods of shedding heat in space?
My personal answer to these is "definitely, definitely, and probably". If yours is too, then the long term trend of putting compute in space is obvious; and the only question becomes "how soon do I need to do it?" If you are a country like Saudi Arabia or China and have basically no barriers to "tiling the desert in solar panels", you can wait a decade before this becomes the only viable path. If you are the US or Europe, where the idea of doubling domestic electricity production is laughably off the table, the time to start doing this is "oh shit, right now". To be clear: the China/Saudi terrestrial approach while your rocket industry takes a decade to catch up is clearly the safer play, but that play is closed to the West for sadly self-inflicted reasons, so it seems like the space long shot is our best bet.
What in the name of all that is holy breeds this weird sense that AI computing is "*the* very lifeblood of the future and the sole terrain on which the New Cold War will be fought," as opposed to "just another technology whose continued development and successful integration into knowledge work will raise productivity growth by a percentage point or two a year until the S-curve flattens at some point in the future?"
The argument of "is AI like any other technology (see power loom, or flight, or nuclear power, or whatever) or qualitatively different in ways that make it hard to reason about?" has been utterly beaten to death elsewhere. Either you buy that thesis, or you don't. I personally do, you seem not to, c'est la vie I suppose.
Edit: just a thought exercise, if you look at America's 10 most valuable companies, roughly 100% of their workers' output is knowledge work. Seems like valuable stuff to me.
Unless the theory is the immensely half-baked, technobabble-filled "recursive self-improvement" stuff, there's simply no reason to think we aren't headed for a capex bust in a few years followed by a long period in which white-collar work digests what LLMs can do well and people experiment with other architectures or use compute for other machine learning tools with more niche/technical applications.
The amount of economically useful work that LLMs can do at present, even if they were *snaps fingers* perfectly integrated into every white-collar field today, isn't revolutionary enough to justify either planned capex or valuations, and capability growth is already slowing down on the way to plateauing.
And when we wake up in a few years, China will still conduct 30% of global manufacturing value-add, against like 15% for the US, a figure which is in some ways overstated by exchange rates.
EDIT: Having just caught your edit... sure, if LLMs or any on-the-horizon architecture were actually even theoretically capable of doing a large fraction of that knowledge work, it'd be revolutionary. LLMs specifically are not and there are very good reasons to believe that this will always be the case, and there are to my knowledge no successor architectures posed that could overcome such limitations.
LLMs already do a large fraction of what used to be my job. The latest versions of the frontier models were mostly coded by their predecessors. I am not sure your argument about knowledge work holds up to sustained scrutiny today, let alone if we extrapolate linear improvement. And most evidence indicates that improvement is still exponential.
The fraction of my job they can do rounds to 0, which isn't surprising as it's a mainly human-facing role. Wringing value from our corporate LLM as a search assistant is, in the main, harder than simply filing and documenting things well to begin with. Wringing value from it as a copywriter is harder than writing things well in the first place, at least for me. Wringing value from it as a sounding board is functionally impossible, worse in every way than finding 20 minutes to shoot the shit with a trusted coworker. It just spits out nonsense, on par with the worst corporate buzzword presentations by MBAs with no industry knowledge. Our developers have said they find it useful for hacking stuff together but it requires a lot of policing and checking before integrating any coding outputs into work product.
What's worse is that they can do very little better when used by the engineering firms and infrastructure owners my company serves and pointed squarely at technical tasks with short horizons. The most promising applications we've got at this point, which are the same ones we had last year, and mostly the same ones we posed the year prior, are "code research assistant" and "drawing output checking engine." There's some hope for "automatic annotation engine" but at present they underperform older, purely rules-based, tools on this front. For all three applications, we're very, very deep into "trust but verify" territory as the accuracy rate is in no case over 90%.
All this adds up to an impression, not of complete uselessness, but of "another automation tool to be integrated into white collar work for some marginal productivity gains spaced over a couple decades." It's Excel-All-Over-Again, not Skynet.
"The latest versions of the frontier models were mostly coded by their predecessors."
This... does not comport with what folks in the industry have told me, at all.
The state of the art has changed dramatically in the last three months (like December 2025 is literally the inflection point). Most of the prominent professionals (like, well known legends with major success under their belt) in my field are producing most of their output via LLMs, and even the ones who are not refrain for mostly aesthetic or ethical reasons, not because they doubt the capability. I would check back in with that developer team in a month or two and see what’s changed.
Coding is a closed environment and thus LLMs show better progress. In lots of other fields, that closedness does not exist. And because it does not exist, the applicability is limited. For instance in law, some tasks like first line discovery and review of contracts were being automated and now LLMs can make that more capable.
So I am not doubting that there are some good advances but to extrapolate from programming to other knowledge work is a category error.
Everything except high level "advice and experience" tasks are immediately in the cross hairs. With software, this may at least keep many engineers employed because demand for software has (so far) been quite inelastic. Unless the amount of legal work explodes at a similar scale (for all of our sanity, I hope not...), you're just going to need many fewer lawyers to do the same amount of high-level legal work.
Now, lawyers as a class in particular may succeed in insulating themselves via regulation, since they write all of the regulation, but ex: accountants may well be cooked. I am sure that doing RL on accounting is, if anything, easier than doing it for coding.
I don’t see either of the ideas you mention discussed here. It seems that what this person is claiming is that AI computing is somewhat more important than railroads or telecom infrastructure, and will likely yield several percentage points of annual growth.
This doesn't make sense to me at all. If AI becomes the thing, I don't understand why we wouldn't simply build a bunch of them in Arizona and Nevada under massive solar panel farms. It would be far cheaper to build, maintain, cool, power, and upgrade than launching things into space.
Suppose you build a few, giant plants in Arizona and Nevada. First, those states aren't just Terra Nulla, they have their own political and environmental groups which will fight you on this. Then you have to find a way to concentrate all that solar power into one plant, which is difficult and eventually becomes impossible. Gonna need massive battery backup for nightfall, too. Then you have to distribute all that data back to the rest of the country (or world) which is another incredibly hard engineering problem.
Putting them in space, you completely sidestep all of those issues. Of course, it comes with new challenges, but perhaps those challenges will be less intractable.
"it comes with new challenges, but perhaps those challenges will be less intractable."
There's a lot of work being done in that "perhaps...less intractable." There were over 500 data centers built in 2025 alone. SpaceX did less than 200 launches in 2025. It would take multiple *Starship* much less Falcon launches to put the equivalent of one data center in the space. That's not including dealing with all the other engineer challenges that would have to be overcome simply to build what can already be done on Earth. I wouldn't expect that to be cost competitive for at least a decade and probably two - if it ever does.
You write "at least a decade or two" as if that's some unimaginably long amount of time. That's roughly how long it took Tesla from its founding to where it is now. Companies make 20 yeah investments all the time, nations even more so. The need for data centers isn't going away, and its only getting harder to find places to put them.
This imagines a world where we need to ex: double energy capacity in the next decade or less to support the AI buildout. This would require tiling nearly the entirety of Nevada in solar. By comparison, going to space on an hourly cadence is merely an engineering and capital allocation challenge. It's not obvious whether the regulator and political obstacles of the former are harder to surmount that the lure-space option.
Yeah, I think all of that will be easier to overcome than building AI data centers in space - at least for the next decade or two. Especially at that large of a scale, just how many rockets do you think SpaceX can launch that it could outpace building on the ground?
I guess we’re sort of agreeing then? I am saying this will probably become viable in about a decade, you’re saying it will be a decade or two… seems like a minute difference in timelines, mostly.
Its very hard to think about where we are in 20 years with advancements to AI. Perhaps we've achieved AGI by then! More likely IMO, we get some significant advancements but its not nearly the breakthrough that people imagine now, but more like computer chips have been. Massive changes in many ways, but not transforming everything as many claim. Or something in the middle. AI could help us figure out fusion or help build a space elevator. In which case all these discussion are missing key parts. Or perhaps it leads to nuclear war and we're building stuff in space because half the world in is ruins.
The "slightly" in point 3 is doing a lot of work. Currently radiators would be heavier/larger than the panels.
And in point 2, you mention "normal learning curve" cost reduction for space launches. But that's not what we need. We would need exponential reduction across a few orders of magnitude. That's only ever happened with compute (Moore's law) and perhaps solar.
Space launch is currently a mechanical/chemical pursuit, not an electrical/digital one.
Long story short, you can start studying it "right now", but it won't be a viable industry for a few decades at least.
Would you be able to point me towards anyone who has done an analysis of what the cost breakdown would need to be for things like kg/orbit for data centers in space to be feasible? My prior is that it's 10x the price it would need to be for it to be reasonable, so even if it drops by half that wouldn't be enough. But I'm a lay person in this field and would be super interested to understand how the optimists do their math.
If the value of data centers is also growing (and their other costs are also falling) then the lines cross sooner. What is the halving time for the cost of space launch? Three of those gets you almost to 10x.
Keep in mind that I work in the industry, so am perhaps over bullish relative to a totally dispassionate analysis. Here's a report that predicts a drop from the current customer-facing frontier on F9 (roughly $1.5k/kg, almost certainly lower if done at-cost) to <$100/kg and as low as $33/kg in a little over a decade (warning: PDF): https://ir.citi.com/gps/kdhSENV4r6W%2BZfP44EmqY4zHu%2BDy0vMIZnLqk4CrvkaSl1RIJ943g%2FrFEnNLiT1jB%2BjLJV4P9JM%3D
We're pricing in a NIMBY tax that is so high that it's preferable to pack up your entire physical infrastructure and send it to space, adding a host of design challenges around heat dissipation, rad hardening, surviving the launch environment?
This is insanity. A world in which space data centers make economic sense is a world in which property rights have degraded to the point to make any gains from building space centers infeasible.
I didn't buy "serious people are considering this" as an argument for it being feasible. No company wants to be the one who admits that they aren't looking into something new and exciting so there's hype reinforcement.
Elon has confidently written off space based solar power, space data centers are only slightly more reasonable.
It's so bizarre to me. Massive data centers are being built today without significant pushback. Any scaling issue with building data centers on terra firma would be even worse in space.
Some data centers are being built without any pushback, but a lot of data centers are getting pushback now. If this gets worse, and the best sites without pushback get used up, then space starts to look better, because all “sites” in space are equal cost and value (at least, for a long long time, since the surface area of orbit is much higher than the entire surface area of earth, let alone the land area that doesn’t have too many disasters).
I get the argument but the case seems vastly overstated. There are 1.5B acres of undeveloped land in the US alone. There's no political crisis in finding land for applications with much less capital and much worse externalities, like industrial meat processing or hog rendering or whatever. There seems to be a huge jump from "we can't build our data center in our first location, downtown Los Angeles" and "I guess we should send it to space!" You can't put these things in Mexico or wherever?
Yeah, it’s hard to tell how significant the ground based obstacles become. But the big thing for space based systems is that if they work once, then you can really get in on the economies of scale, whereas each ground based site has some site specific differences, and there’s more unpredictable weather, legal regulations, and geology on earth on a weekly basis (though perhaps solar flares become a similar issue in space on a yearly basis?)
Solar flares, rockets exploding, etc. etc. I think my main problem is that the solution fast forwards through technological leaps that would make equivalent construction on land much easier. Like this idea is dependent on new ways of cooling GPUs that don't require water or air. Wouldn't that make data centers much more palatable to build on earth?
It reminds me a lot of the basis for terraforming Mars - if there's some ecological disaster on earth we would need to create a habitable environment on Mars. But any technology you create for that would be able to fix the ecological issues here!
I definitely think that terraforming Mars makes less sense than “terraforming” Antarctica!
But space is different enough that it’s not quite so obvious to me that whatever works in space would work on the ground.
I realized after I wrote those two paragraphs that I should ask my partner, since we were having lunch together, and his research is in the optical properties of nano materials. He said that most purely radiative things that could work in space would work on the ground. But he’s been working on materials that take incoming light and re-emit it at higher energy, so that they use incoming light for cooling. If that sort of system could work, then it would work better in space, with 24/7 sunlight, than on the ground. But he did say that cooling in space is one of the big problems, and that putting something that depends on cooling in space is not likely a good idea any time soon.
As a matter of principle, I absolutely agree that if SpaceX is a de facto launch monopoly imposing common carrier-type requirements and other anti-trust enforcement is the correct policy move.
All that said, I think it’s worth noting that Elon is engaging in pre-IPO hucksterism here and that the economic case for data centers in space requires the exhaustion of current cheap energy sources on Earth, a very dramatic decrease in the costs of orbital launch (probably requiring a different mechanism than rocketry— maybe some kind of space elevator), and likely also significant advances in materials science (no convection in space so cooling requires massive radiators)
Right now, there are places on Earth and in developed countries with abundant cheap energy, many of which are in rule of law friendly jurisdictions. In the US, one of these is the Permian Basin in Texas and New Mexico, which has tons of access to cheap wind and solar power and also so much “waste” natural gas from oil production that local methane production often exceeds takeaway capacity and the price at the local gas hub goes negative.
But despite this availability— and land being extremely cheap— hyperscalers have been very cautious about building new data centers in the Permian. This is largely because of a combination of latency issues (proximity to users is super important) and maintenance challenges (Permian is sparsely populated so getting people out to fix problems or update hardware is costly).
Space is uh, considerably worse than the Permian basin on both of these axes. It really, really does not make sense except for say, defense applications where compute collocated with satellites is unusually useful. (A real market but much more limited TAM)
Again, we’re not actually seeing this at scale in terrestrial cheap energy locations like the Permian, Scandinavia (cheap hydro), interior British Columbia (cheap hydro plus cheap gas from the Montney Basin), the Arabian peninsula (abundant solar + gas), etc
I understand that part is the issue is that training is still pretty high-touch and requires close human monitoring (very plausible that this becomes automatable but we’re not there yet), and another part is that any model that’s economically worth its training costs will ultimately generate orders of magnitude more compute in inference use than it did in training (so training-only clusters would only be a small share of the full data center stack)
I think this significantly misses the point Hayes is making. Even though all of those billionaires have divergent views and interests, they are all united (along with other billionaires like Ellison and Benioff and lots of non-tech investors) begin the idea that we're close to building a machine that will replace all white collar workers, and we should built it as fast as possible.
My issue with Hayes statement isn't just that it flattens the divergent views of billionaires, I think it wrongly assumes that every billionaire on earth wants to immiserate white collar workers for their own bottom lines. Billionaires have a lot of different political beliefs and some might actually want a world with mass unemployment because they're bad people, but my suspicion is that many are smart enough to see that that is a bad future for society.
- basically all billionaires are all in on developing AI and a positive about doing it
- many people involved in AI, including some of those same billionaires, think that AI will soon be able to replace a huge swath of the workforce
There are lots of differences of opinion about what to do about that and what the implications are, but I take the conjunction of those two claims to be Hayes' point.
"basically all billionaires are all in on developing AI and a positive about doing it"
I don't think this is true at all. Is the Arnault family heavily investing in AI, or Amancio Ortega? I haven't heard Warren Buffet is deeply invested in it.
I think there are many who are, but many others who are focused on the business that got them there and it has nothing to do with AI.
I can't read Hayes mind obviously but I don't think this is what he means at all; it seems much more like a general position-taking statement in which enemies are identified (in the left's current idiom, they are all "billionaires"). The tweet doesn't even make sense as a causal statement: "billionaires" are a group of people and "globalization" and "neoliberalism" are abstract concepts. This strikes me as a thrown-off tweet, a vent, rather than some through-through claim.
You're taking two steps here, interpreting the tweet as "LLM-based AI will do to white collar workers what globalization and neoliberalism did to (American) blue collar workers" and then imputing the spread of LLM-based AI to "billionaires." Fine (and we can argue about that), but highly doubt that's what Hayes means.
It doesn't take very many if them with their hands in the right pies to make things very bad indeed for the rest of us. The ones helming the big AI companies are not particularly pro-human people.
Dude it's like three real billionaires and a couple paper billionaires that we're actually taking about. Buffett has been anti-tech for 50 years now. The collective Walmart heirs don't give a fuck. Gates is buying farmland. Bloomberg's wealth, outside of the Co., is mostly in real estate. The Koch's are still OG tied to energy infrastructure (a friend ran their venture arm). I can't think of a more diverse "class" than billionaires. Doesn't matter to Hayes though - he'll just keep making shit up.
I’m not sure that it makes a practical difference, but some people think it makes a moral difference that the billionaires aren’t particularly interested in the immiseration of the white collar class - it’s just a potential foreseeable side effect of the thing they are in fact interested in (along with the side effect of cheap white collar work now available to the working class).
I agree that that's true, but I don't think it makes an important difference. This was true for international trade as well -- the people who advocated for it (including me, still!) didn't want to immiserate anyone.
It is just not remotely true that Claude Code could easily replicate the Oracle database, or PeopleSoft, or any of the other important Oracle products.
It is somewhat laughable that anyone believes they can shape someone else's AI. The government failed to shape social media, which has helped to crater birth rates and raise a generation of nihilists.
Of course, the flip side is that we have a known history of men building the railroads. Without Vanderbilt and others, we wouldn’t have had trains, which opened up interstate trade along with the entire country. While that time was the movement of goods and people, today we are moving information.
The same millionaires of that day, private wealth and industry built it, and the country benefited. Elon Musk is a jerk with an imagination. I worked 35 years in the automotive industry. I can give a number of reasons why Musk's little electric car business was going to fail. Nobody believed he would deliver.
So, no, I would do much to harm Altman, Bezos, or Musk. Bezos and Boeing, and likely some other global rocket companies, can compete against Musk. People can pay to have a data center built in space. It has been called the commercialization of space. That is exactly what is happening.
I would suggest Matt and the Progressives let it happen, just like the railroads happened.
Matt is not trying to stop it; he specifically said to treat it the way we treated the railroads and impose common carrier rules. Matt has a touch of Elon Derangement Syndrome, which is probably why he is jumping the gun on this issue, but his point seems valid.
Common Carrier mostly means that the federal regulators have wide latitude to set arbitrary rules. If you're a business this is bad for the same reason Trump is bad- you are at risk of getting jerked around by capricious officials.
Given Elon's status, it also seems like a guaranteed outcome, and is a backdoor to let populist dems cripple the company.
I think for this, you need a tort first, you can't just do it preemptively.
The relevant standard under Section 7 (the typical statute for challenging a merger) is "substantially to lessen competition, or to tend to create a monopoly" in a relevant antitrust market.
I don't think there's a serious argument presented here that that standard is met.
First, which market? The market for LLMs? The market for space datacenters? The latter doesn't even exist, it can hardly be said to have a substantial reduction in competition.
I've worked on a several of anti-trust investigations and litigations. This post has the same problem you critique anti-trust maximalists for: it's all vibes. If you want to win anti-trust cases in court, you need sound theories, backed by specific evidence, and economics that supports the case.
Frankly, this comes off as "you don't like Musk, therefore we must find a reason to object to the deal".
Wow -- preemptively constrain a risky venture to create a new industry, just in case it works, for nakedly partisan reasons, on transparently spurious grounds of anti-trust? Not what I would expect from you.
It’s not for partisan reasons - it’s for the clear threat to competitiveness of vertically integrating a non-competitive utility with a single participant in the downstream market.
.....what in the Atlas Shrugged is going on here. There is not a market for either xAI alone nor for reusable rockets at present, and data centers in space is an idea that won't even see the first attempted launch for years. Yet it's so important that it must be the pursuit of anti trust legislation?? Like the other rail/utilities examples mentioned let's look to Congress to pass a law, there is literally negative proof of anti trust enforcement here. There isn't even a market!! Bad take.
To my knowledge, antitrust law doesn’t necessarily require an existing monopoly or even a fully formed market. Obviously this scenario isn’t happening right now, but it could and the risk of giving Musk exclusive control over effectively unlimited data center compute seems important enough to act on
The risk of unlimited data center compute in space is plausible only if you ignore extreme temperature swings, space debris, gamma rays, solar flares, coronal mass ejections, and the Kessler effect, and pretend lift costs are zero.
This idea is about stock price, not plausible uses of engineering.
Musk fanboi: "I am sure it's good to put all the compute beyond the protective cradle of the ionosphere. Radiation will make our AI overlords stronger."
A masterful plan, sir, truly quite visionary!
The idea is more to put it in a sun synchronous orbit which will stay within the Van Allen belts. Going over the poles mean more charged particles than you get on the ISS but not a huge difference. Nothing like the vulnerability satellites in geostationary orbit have.
I didn’t know about the Kessler effect until Sabine told me about it the other day. There are concerns that it is closer than previously believed https://youtu.be/8ag6gSzsGbc
See also pretty good discussion at https://aerospaceamerica.aiaa.org/features/understanding-the-misunderstood-kessler-syndrome/?utm_medium=email&utm_source=rasa_io&utm_campaign=newsletter
Neal Stephenson wrote a whole book (Seveneves) based on this premise!
Low key one of the most stressful and claustrophobic books I've ever read.
It's all very difficult to remove waste heat in a vacuum. You have to radiate it away.
Huh, very interesting and unintuitive (for me)!
You don’t have to ignore these things - you just have to think that the value of data centers is getting high enough that the costs for mitigating these things become affordable.
More do the benefits of being in orbit (cheap energy, potentially cheap space) outweigh the costs (more radiation, harder heat dissipation and launch costs), of the three costs one of them is economic and two of them are arguably mitigatable as technology advances.
Thank you Lisa!
This is as ridiculous as creating anti trust laws against Apple for inventing the iPhone in the Jobs era. There is competition even in the notional world in which you're living.
It reminds me of attempts to block the Sirius and XM merger, as though "satellite radio" was a market that needed to be managed.
We just covered this week + last week in ECON 3339 how you define a market for the purposes of bringing an antitrust case https://en.wikipedia.org/wiki/Small_but_significant_and_non-transitory_increase_in_price
In this example Matt wants to define a market that does not exist.
Antitrust enforcement applies to prospective markets if a combination could eliminate a potential market entrant. There is DOJ Antitrust Division guidance on this.
That was the stuff invented from whole cloth by a biased federal judge, right?
If by "whole cloth" you mean the Clayton and Sherman Acts.
There are quite a few competitors for this notional market and 0 evidence musk controls pricing power in any way
There are other entrants but at time of print, no one is within 2x SpaceX's price in $/kg. Here's ChatGPT:
SpaceX: ~$1,500–$3,000/kg
Blue Origin: ~$2,000–$4,000/kg (target, contingent on New Glenn reuse/cadence)
ULA: ~$4,000–$7,000+ /kg
Arianespace: ~$6,000–$10,000+ /kg
Firefly Aerospace: ~$15,000–$25,000/kg
Rocket Lab: ~$20,000–$35,000/kg
Other small-launch startups: ~$20,000–$40,000/kg
So maybe Blue Origin gets close if they can match SpaceX's technology? Even with competition, I see a strong case for the type of rules MY is advocating. I'm not even clear what the argument against is.
According to this theory, the best solution is for SpaceX to simply raise it prices so that its not cheaper than its competitors...?
The theory is essentially:
1. AI is critical infrastructure
2. Data centers in space will outcompete data centers on the ground
3. Musk's near monopoly on launch will give him a near monopoly on space data centers.
Point 1 is questionable and point 2 is extremely questionable. Point 3 is fairly defensible, but kinda meaningless without the other 2.
Matt is jumping the gun here, but the people saying space launch is a highly competitive market are missing the reason why in my opinion.
Even if they raise prices, they have massively more capacity than anyone else. And that gap will increase even more dramatically if they get starship working.
Uncharacteristically underbaked take from Matt here. Common carrier regulation could certainly make sense, but the idea that the only reason to vertically integrate is to monopolize a business that doesn’t exist is not well thought through. Vertical integration can drive coordinated investment to innovate and open new markets - exactly what seems to be going on here. If there is an anti-trust issue later it can be addressed in due course.
Unless you think that business is truly transformative. If you work backwards from a future where AI is as central to civilization as something like rail freight was in the 19th century or telecommunications in the early 20th or oil now, and Elon sits at the center of an oligopoly of space-based data centers that power them, the clear forward path is leveraging his space company to give himself an unfair advantage deploying those data centers. In that future, Elon is a modern Cornelius Vanderbilt with all the negative downstream effects of concentrating to much wealth and power in one family. Except Elon is also an emotionally unstable megalomaniac.
This is the sort of fan fiction I only dream of writing
I think this is the better plan
And if in ten or fifteen years, we have a problem.Then, we can try and pass some legislation
I think you need the regulation *before* you have ten years of one company monopolizing the effective way to expand data centers.
My point is that it will take 5 or 10 years before what's he's proposing could even work. If he actually succeeds in starting to build AI in space THEN we can look at regulating it.
But let him at least try
When you're talking about the world's richest man who controls (and manipulates) a major information channel and is willing to spend huge amounts on election influence, starting earlier is better than waiting until the problem appears. Building momentum toward legislation against a motivated lobbyist takes time and by the time the problem is obviously a problem, it will be much harder to achieve.
Most of our intuitions about how politics work was informed by an era where news media sources had fairly strong internal cultures of adherence to facts. We no live in a world where the hyper-wealthy control information media with a declining or non-existent culture of truthfulness and neutrality. Legislation restricting abuses by those hyper-weathy people is going to become much harder in the future.
Yep. Vertical vs Horizontal integration is common pain point in working through antitrust law as well discussed by Ben Thompson and Byrne Hobart in a Nov 2024 interview. [1] Examples include various phases of the industrial and the information revolutions, particularly with respect to international politics, energy, and coordinated violence including war and terrorism.
There are more details in Hobart's book, "Bubbles and the End of Stagnation". [2] I only made it a third through before realizing that this could feel like preaching to the choir. Moreover, I have some professional relations with Thompson and Hobart that further compelled me to consider alternative narrative structures for these patterns. Eg, the origins of the progressive movement under Teddy Roosevelt that include antitrust regulation against Rockefeller and other turn-of-the-century baron of industry.
[1] https://stratechery.com/2024/an-interview-with-byrne-hobart-about-bubbles-and-escaping-stagnation/
[2] https://www.stripe.press/boom
Unless "due course" is the elimination of prospective markets entrants due to the monopoly leverage of the tied market. There is nothing the prevents two separate companies from partnering to "coordinate" investment in an enterprise on standard, non-discriminatory commercial terms.
This piece is going to become a classic of “when Matt lost his mind” satire pieces. Like the Slate pitch or the million David Brooks parodies.
I disagree, but only because "We should have common carrier regulations on space launches" is such a benign conclusion regardless of whether you think the premise is absurd.
I can 100% get behind common-carrier regulations and anti-trust in general, but credibility matters; examples matter. If you'd proposed that Pets.com was about to have a monopoly on dog food in February 2000, we'd still be mocking you to this day.
That's not the correct analogy -- it would be a monopoly of the online dog food sales market based on a pre-existing monopoly of the internet.
It is not benign if it is targeted to just one company as opposed to an industry wide regulation.
Agreed. I just posted my critique of this ... his plan creates the problem he claims to want to avoid. Here is my comment above: Matt, you've diagnosed a real tension but prescribed a Maladaptive response that would guarantee the exact outcome you fear.
Your core insight is correct: SpaceX's dominance in launch capacity could theoretically extend into AI compute. But your solution—common-carrier regulation preventing competitive advantages—reveals Victim identity thinking that treats success as a threat requiring control.
Consider the actual pattern: SpaceX achieved launch dominance precisely because NASA's cost-plus contracting created the dysfunction you now want to replicate. Boeing and Lockheed had regulatory protection and guaranteed access. Result? $400M per launch vs SpaceX's $67M. The "fair access" you're proposing is the same thinking that gave us the Space Shuttle disaster.
Your railroad analogy backfires spectacularly. The ICC didn't create competition—it calcified incumbents and prevented innovation for nearly a century until deregulation in 1980 finally allowed trucking to displace protected rail monopolies.
The Creative response you're missing: If orbital data centers prove viable, the threat of SpaceX dominance will attract massive capital to competing launch systems (Blue Origin, Rocket Lab, Relativity Space). Your own logic admits "space launch does not strike me as inherently monopolistic"—so why preemptively regulate as if it is?
You correctly note billionaires fighting each other (Bezos vs Musk, Altman vs Musk). That competitive dynamic is exactly what prevents monopoly—unless we regulate it into existence by imposing common-carrier rules that reduce the returns to competition.
Don't use antitrust to prevent competitive advantages. Use it to prevent anticompetitive behavior. There's a massive difference, and conflating them turns potential innovation into guaranteed stagnation.
I mostly disagree with the characterization that there's not a market for reusable rockets (there's a market for launch and SpaceX's ability to reuse hardware can lower its costs), but otherwise... yea, when I saw the title this morning I had a nice "WTF" moment.
There's a very strong market case for Falcon 9 and its brethren, but the case for (ugh) "Starship" is a lot more tenuous. There's fairly widespread agreement that you need Starship to work to get the Starlink math to pencil out, and you need a full Starlink network's immense payload needs for a system as large as Starship to make sense. Whether the ouroboros so formed works out long-term depends on a lot of assumptions standing in for questions nobody has yet answered.
The trap here is that there is a market for launches. And there are obvious benefits to cheaper launches. The conceit is that reusable is automatically cheaper. But, that is not necessarily the case.
Falcon 9 suggests this is not the case. The trouble with past reusability has been that it hasn’t truly been so reusable. Many elements of that are now a solved problem.
I didn't make the claim that "reusable is automatically cheaper", I made the claim that SpaceX in particular can lower its costs by reusing hardware.
There are other reasons why they have a cost advantage but this is certainly one of them.
Apologies, I didn't mean for that to be my implication. I meant that it is the trap many people fall into when talking about the market for "reusable rockets." My assertion is that the "reusable" is catching a free ride in that description.
And really, I just meant my comment to be an addition to yours. You correctly state there is a market for launches. Which decouples those.
Three generations of Shuttle vets laugh bitterly.
Genuinely asking if I am misrepresenting anything?
No, not at all. This was exactly the industry's widespread conclusion at the end of the Shuttle program. The so-called "big dumb booster" was probably the most cost-efficient way to meet market needs as they were then understood. Computers had gotten small, light, and most importantly cheap. Shuttle was a do-it-all space truck that was predicated in part on "go get a malfunctioning satellite and bring it home for repair and relaunch". In the 1960s and 1970s, this seemed like it made a lot of sense, but the microprocessor revolution caught everyone by surprise.
It turns out that partial reuse certainly appears to be the most cost-efficient and industrially sustainable way to meet launch needs of the late 2000s on into today, but this was only shown in practice by taking a big risk, then going out and doing it (for which SpaceX absolutely deserve credit).
It's legitimately shocking that anybody with a remotely technical background is taking datacenters in spaaaace! seriously. Matt doesn't have that background, so I'm 2/3 inclined to let him off the hook today (the 1/3 remaining is to say 100 rosaries of "I will keep in mind what I'm qualified to talk about), but we are long beyond the point that we should take everything Musk et al. say with a gigantic grain of salt.
From what I’ve seen thus far, media coveage on this topic has been pretty poor. No one’s actually saying there will be giant data centers assembled in orbit that will rival a Google 800K sq. ft. facility on the Columbia River. I assume what’s being considered is a large collection of data centers that are each the size of a basketball, and maybe the size of a small car with its solar panels are furled for launch. But I don’t know because everybody wants to write about AI and Elon.
On the other hand a lot of people didn't think reusable rockets were possible
This is a popular notion that is completely false--at least, nobody qualified thought they were impossible. There was some question about whether they made financial sense, which is obviously different. But vertical-launch, vertical-land launch vehicles were very seriously studied as early as the 1960s (the High Frontier blog has some great light reading: https://thehighfrontier.blog/2016/03/20/straight-back-down-to-earth-a-history-of-the-vertical-takeoffvertical-landing-rocket-part-1/), and McDonnell Douglas literally flew the DC-X tailsitter testbed in the 1990s, making a number of successful transition test flights. Here is one: https://www.youtube.com/watch?v=ZL9cLvYIDPE
DC-X, Roton, Armadillo never went to orbit. They were just testbeds for vertical launch and landing. Afaik, SpaceX was genuinely the first to do vertical landing of an orbital launcher. Generally, prior to F9, the perception was that it was probably possible to do, but didn't make sense financially, particularly if you were solely talking about the boosters. It's interesting how wrong this position was, and I recall how many experienced space industry vets loudly said this at the time. And this is with F9 being only partially reusable. Starship is projected to bring down launch costs another order of magnitude, and is fully reusable. I see people pooh poohing this notion and mocking Starship regularly, in much the same way F9 was pooh poohed and mocked prior to its success.
As far as whether data centers in space makes sense, it really just boils down to whether Starship pans out. No one thinks they are viable at today's launch costs. At about $200/kg to orbit they start to make sense. Obviously, Musk is bullish on Starship's prospects, hence bullish on this market in general.
https://archive.is/ZTBAn
Studied and possible are not the same thing. Again and again, the conclusion was it was just too technically hard.
The fuck are you talking about? They flew test articles! I linked the video!
The DC-X was an exciting experimental test vehicle, but it never went to orbit. There have been a lot of other non orbital powered landings over the years too. Roton ATV, Armadillo's test vehicle, Masten's test launchers etc. All cool systems in their own right but even their creators would never claim that they solved reusable boosters to orbit. It's just a fundamentally different thing.
I hear that claim so often (same about electric cars and self driving cars) but who in tech thought these things were impossible?
The need to reuse more of a rocket was what lead to the Space Shuttle for instance. Everyone knew re-usability was needed.
As far as I know noone thought reusable rockets were impossible, at most unfeasible with current technology and financial use cases. And SpaceX did nearly fail to get the tech working and for finances they became their own largest customer through Starlink.
One of Musk's singular talents is convincing people only he thought XYZ was possible, when in fact a great many smart people and talented organizations have studied XYZ for decades.
But this is true of every engineering accomplishment. No matter what it is or how impressive it is. It's always the case that it's widely understood to be physically possible before it's actually built.
I think when people say this wrt SpaceX specifically, no one is saying that reusable rockets etc were thought to violate the laws of physics. No one believed that. But it is true that, before F9 revolutionized the launch market, most space industry folks (who imo learned all the wrong lessons from the Shuttle's "reusability") believed that VTVL was impractical for orbital launches - both in the sense of being too technically demanding and not likely to be a big source of cost savings. But they were wrong, and now SpaceX absolutely dominates the global launch market.
You can say the same thing about Starlink. Again, many of the same jaded industry types thought Starlink was ridiculous, drawing (as with Shuttle) all the wrong lessons from the failure of Iridium. As with reusable boosters, no one was claiming that low latency LEO internet was physically impossible - rather, the claim was that it was a boondoggle, no one would use it, it would be a black hole for money, etc. But again this was all incorrect. Starlink works extremely well today, so much so that access to it has become a geopolitical flash point, and SpaceX makes lots of money off of it.
Same with creating a new car manufacturer in the US, for that matter. Again, obviously physically possible to build EVs etc - but doing it at scale in the US? Before Tesla actually did this, almost everyone considered the idea to be absurd.
For people to now say, oh all of this stuff was no big deal. Everyone knew that reusable boosters were the way to go. It is frankly just incorrect, and I think that (surprisingly widespread) perception contributes a lot to the present day underestimation of Starship.
Does this mean that data centers in space is a slam dunk? No, not necessarily. But imo at some point you have to update your priors re: "Elon Musk is proposing something that at first glance looks like a total boondoggle, could it actually work as a business?'
EVs clearly weren't technically impossible; hell ... the first one was built in 1888. But I think it's easy to forget that in the early Tesla Roadster days it was general industry consensus that the major auto OEMs were un-disruptable because starting a new car company was financially and operationally impossible. Also battery costs were $500/kWh. They're nearly $100 today.
Good history lesson here:
https://archive.nytimes.com/bits.blogs.nytimes.com/2008/10/24/musk-unplugged-tesla-ceo-discusses-car-troubles/
[BART_SIMPSON_CHALKBOARD.GIF]
This is Matt's worst take ever, possibly. Essentially seizing SpaceX as a resource for the government.
This is very different from railroads where there were many companies that needed rules of the road to compete. Its just targeting of one company because of their advanced capabilities.
Where did he mention seizing SpaceX? I thought he said to treat SpaceX the way we treat AT&T and Union Pacific and Western Union, none of which were ever seized by the government.
Matt is saying : SpaceX is too important to be allowed to operate independently. They should be forcibly turned into a utility.
This proposal from Matt is not regulating an entire industry - just this one company that might one day become an industry. Matt's proposal is turning a private company into a government utility, with no vote from the owners of this private property. It would arguably be a wholesale theft of value by the government from private citizens.
Treating the entire industry as utilities doesn’t mean nationalizing it! There are some government-run entities in gas, electric, and water, but the intercity railroads have (almost?) always been entirely private despite being regulated as utilities for about a century.
This is true, but if Musk’s plans come to fruition the Common Carrier regulations seem well suited for this. And you can’t wait for the commissioning of massive data center capacity in space. By that time it would be too late. Musk would have used his space capabilities to launch data centers, at scale, while potentially blocking access to space for AI competitors. We should not allow that last part.
Congress can pass a law very easily. Anti trust is not the vehicle for this in any way at this stage. There are other orbital carriers with more capability growing daily - Boeing+Lockheed have a partnership. Northrup buying Orbital ATK. French company Arianespace. Japans Mitsubishi. Rocket Lab has launched more than anyone except SpaceX. This whole discussion reeks of simple ignorance over the state of the industry imo.
You’re vastly understating SpaceX’s market position. Comparing Rocketlab to SpaceX? Their launch capacity is tiny. Right now, SpaceX has delivered an order of magnitude more mass to orbit than any other company besides Roscosmos (which has been operating Soyuz since the 60s). Blue Origin’s New Glenn rocket will be able to compete with the Falcon 9 and Falcon Heavy (assuming they can manufacture them at the same scale) but Starship will immediately leapfrog the competition again. None of this is to say it’s a non-competitive industry, but in terms of kilograms to orbit, SpaceX owns a huge amount of the market share, and they’re already converting it into a near monopoly on satellite internet.
And? Theres obvious competition and a market leader due to being a first mover. Comparing the two could be simply like comparing the iPhone to Android. Apples market lead led to a huge company that a decade or two later has less distribution and a dwindling market share in top markets with less differentiation over time. I dont understand why we need to pretend that LEO is an impossible task esp with the investment in it larger than anything else besides data centers atm
I mostly just took offense at you suggesting someone else was ignorant of the state of the launch industry while in the same breath implying Rocket Lab (largest current rocket: 300 kg payload capacity) had a comparable launch record to SpaceX (smallest current rocket: 25,000 kg payload capacity). I don’t strongly disagree with your conclusion but that doesn’t excuse the rhetoric.
Just want to point out that RocketLab is specifically targeting the smallsat market, so it's not a knock on them that they have relatively small payload rockets.
Rocket Lab has launched more than anyone but SpaceX. I dont believe I said they were comparable? Just that there are clear other options and rising competition in pretty much every area that SpaceX and/or xAI develops. Who knows where it will be in the future, first movers do not always equal winners, and more often than not are not the long term industry winners
It's not like Ai competitors couldn't still have plenty of capacity here on earth
Completely agree. See my comment above. Matt's solution would paradoxically create the very problem he is trying to avoid.
Isn't the main problem that merging is largely a lifeline from unrelated markets to prop up gambles that are currently losing?
I confess my main gripe would be as a shareholder of whichever is doing better. But I also don't see the reason that the market should allow unrelated merges on a regular basis.
Stopping a merger because it's a fake cash infusion can be stopped on “brother-sister” grounds if it's truly a breach of one company's fiduciary duties, but I have a hard time seeing being the case here. Even then if the two companies are entirely privately owned by the same person it's usually allowed even then. It's probably why it's happening in advance of any IPO
Largely agreed. I would argue for more scrutiny in the merger than a ban on monopolistic grounds. In particular, I would take a fine tooth comb to make sure they aren't trying to shed any liabilities in transferring any ownership.
Yeah, this take relies entirely on worst case hypotheticals, which wouldn’t win any antitrust case at this point.
Plus it would be the wasted effort that the net neutrality fight or the Internet Explorer fight were.
I think there is a case to apply common carrier rules to space launch, but that has to apply to the whole industry, not a single firm.
Totally Agree! Matt's judgement is being clouded here. Even if somehow this idea works, there will still be plenty of competition by earth-based computing centers, plenty of competition for moving things to space, if this idea holds lets just implement this for all data centers already! Lets implement this for anyone who wants to just build a really big datacenter.
It’s not about the data centers - it’s about the space launch. And he agrees - his main demand is that space launch should be treated as a common carrier utility!
This article presupposes that there are other AI companies out there with bold plans to put AI data centers in space but who will be blocked by SpaceX boxing them out on behalf of x.AI. Is that really a thing though? this seems like another Moon shot that Elon wants to try, and we should let him, and we should get out of his way. Just like with Tesla or SpaceX, he's doing something so bold and authentically innovative, that we should not preemptively put up roadblocks and create more regulatory uncertainity here--at least not for the reasons stated.
Let him prove his concept, like he did with rockets and electric cars. Then we can think about regulations after it hurts consumers or competition.
I’m not saying we shouldn’t let Elon try, I’m saying we should require Space X to offer its services for sale on a non-discriminatory basis
Why SpaceX and not the other half dozen providers? Why is anti trust the mechanism? Why ignore the competition from other tech companies and other heavy industry?
In terms of why making SpaceX do it it's because they are already close to being a monopoly. 85% of global launch capacity is controlled by SpaceX. A fraction that is likely to increase in the near future.
At the moment the only competition is ULA, who aren't competitive and kept afloat by government reluctance to be wholly dependent on SpaceX, Blue Origin which remains a vanity project for Bezos more than a successful company, and Rocket Lab who currently competes in the light launch market that SpaceX isn't interested in, although their Neutron rocket may make them competitive in medium launch if they can scale.
It's difficult to overstate just how far ahead of all their competition SpaceX is.
It's because they did it first. Should we have regulated the iPhone for being first? Would that have changed the dynamics of today's smartphone industry for the better? This is pure punishment politics imo
It's better to think of these anti-"monopolists" as anti-capitalist pursuing their dream for people to owns the means of production. Innovation be damned.
The iPhone isn’t infrastructure. Do you mean the App Store?
I feel like you are confusing railroads (which should be subject to common carrier concerns) with rail cars (which should not)
My point was that the downstream market dynamics are not predictive based on first mover status in any industry, and so doing regulation before the industry even functionally exists is forced regulatory capture. If Elon gets a windfall for a few years for being first it doesn't seem like such a disaster when there is now a global race on. Most of the first attempts will be in LEO which is a far smaller lift for competitors than further out GEO
I have shitty broadband at my house because cable companies were the first to run data cables into homes and then they successfully lobbied to escape common carrier laws. I think that's really all Matt is arguing. Companies that launch things into space should be treated as common carriers and SpaceX shouldn't get a pass just for being first.
I don't know that the political solution Matt is pushing for is warranted. That's because I'm fairly skeptical of the economics of space data centers. I do agree with Matt though that SpaceX is the only company that could plausibly do it if it can be done at all.
Maybe that will change in the near future, but Blue Origin has been trying to be SpaceX for longer than SpaceX has existed, so I think we should be skeptical that catching up will be easy.
Be as skeptical as you want, that's no reason to use government intervention to punish first movers because you don't like them
If it strangled smartphones in their crib? Quite possibly...
Econ 101 says that natural monopolies should be regulated as utilities, while monopolies formed through market manipulation and cartel behavior should be broken up. It's two sides of the same coin. And utility regulations pretty much always apply to all companies in a sector, not just one.
But rockets are not a natural monopoly are they?
Industries with really high R&D costs are in a grey area. They have high barriers to entry and are capital intensive, but switching costs are low
I believe the proposal in the piece above was to make all space launch subject to "common carrier" rules.
Who says the others shouldn’t have this treatment? Matt certainly never said that.
The article is called “dont let ELON MUSK monopolize space compute” its not about the industry
I thought the issue was if you don’t treat the industry as a common carrier industry, then Elon Musk will monopolize it. He never suggested special treatment for Musk.
What is the evidence they are discriminating on launch services? They are currently helping Starlink rivals get their constellations orbited. I work in an industry adjacent to this and I have not heard any claims that SpaceX is refusing launch services.
I think the claim is that this sort of discrimination is the only explanation for the potential synergies of merging the two companies, which is the official reason for the merger.
Now if, as everyone believes, the actual reason for the merger is just to get a cash infusion into his floundering social media and AI businesses, then that’s a different story. But I believe he’s not allowed to say that’s the reason they’re merging, since SpaceX shareholders would get nothing out of that.
Potential synergy within the combined company is not evidence of real or potential discrimination though. And by the time, if ever, there are AI satellites ready for launch, Blue Origin and other launch providers will likely have closed the gap.
So Bezos will only launch AWS datacentres and Musk will only launch xAi openAi/Microsoft/Google/Oracle etc get squeezed out. All Matt wants is for all those companies to get the same access to space at the same prices as AWS/xAi so the ai field remains as competitive as it is now
They do. See the Amazon Leo launch last October and the next one later this year. Nobody wants a new United Launch Alliance.
The stakes here seem low because the market will probably never exist.
Why did you chose this subject rather than more Habsburg content? You have the best Habsburg content in English!
Launching stuff into space isn't that hard (we've been doing it for 70 years), if SpaceX starts denying its services to other companies they'd be doing a massive favor to its existing competitors.
And any large tech company (or country for that matter) could launch new competitors if they wanted; Aerospace tech is a huge existing industry with many expertise and capabilities out there to be bought.
That doesn’t seem to be true, given the long struggles SpaceX had and Blue Origin is still having. If Amazon is having this much trouble a decade or two in, why would Google or Apple or Microsoft be able to do better?
Compared to, for example, AI investment in these companies is actually pretty low. Also, SpaceX and Blue Origin both wanted to start from scratch partly because they're both billionaire vanity projects.
Blue Origin costs 2 billion a year right now when they're weirdly huge, which isn't a lot compared to what companies are spending on Datacenters for AI right now.
If you could get a functioning rocket company up for say 10 billion, when the yearly capex investment for AI is 100s of billions; That means that if datacenters in space ever become a big thing for datacenters building your own launch corp is not a big investment.
Also, they'd be able to just hire people away from the existing rocket corps and use their suppliers. It won't be that hard to catch up.
For one thing, they face nation-state competitors who are not profit-motivated and are willing to endure large losses for long periods of time- that's a pretty fundamental difference
Loved this column’s focus on specific actions that can make a difference. Thank you!
Before we know if it works? What if he responds by saying “forget it, I am not going to spend a fortune on this borderline impossible thing if you are already trying to get your grubby regulatory paws on it”. Now we have no space compute at all!
https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
https://www.pcmag.com/news/google-eyes-space-based-data-centers-with-project-suncatcher?test_uuid=04IpBmWGZleS0I0J3epvMrC&test_variant=B
Google is certainly investigating the possibility, and per the second link, Amazon as well.
You did notice they described it as a moonshot thought exercise? Not something currently plausible.
Musk seems like the only person for whom it might plausibly work because he's the only one facing the problem of having a huge excess of launch capacity.
I see space data centers as an answer to the question of what to do with Starship more than anything.
Leading with that specific link was a poor choice.
But you also need to read these announcements with a Silicon Valley hype decoder. Google is already working on orbital prototype launches. This is not just a thought exercise.
The “prototype” is literally launching satellites to evaluate TPU degradation in orbit.
That is not what most people would consider a prototype.
Catching up on other comments, I now see that you're coming at it from a position of technically-informed skepticism. I'm not arguing against that, and I even upvoted lhe top-level comment, because I agree with the main point.
I'm only responding to the first sentence:
> This presupposes ... there are other AI companies out there with bold plans to put AI data centers in space ... is that really a thing though?"
Companies are investing! That doesn't mean it will succeed, but it is a real thing.
Most importantly though I can't imagine scrappy little Google and Amazon need the U.S. government to protect them from Elon Musk.
I think your take is about spot on. It bears watching, for the reasons M.Y. stated, but the time to impose regulations is when, and if, Musk starts massive data center launches.
“Musk is trying to do something here that is bad for the country and the world and that is also bad for various non-Musk billionaires and corporate actors.”
I don’t think this is ever really justified? What reason is there to believe that space data centres are a crucial piece of infrastructure? What reason is there to believe Musk wants to monopolise the industry, let alone reason to believe he could?
There's currently 1 AI model running in space launched by the Chinese. No one even knows if running data centers in space is feasible, cooling alone is going to be very difficult to solve. Do you know how much heat an nVidia B200 generates?
I learned about how big radiators have to be from playing Terra Invicta.
I learned about cooling in space from the legendary "Stealth Spaceships" flame wars.
The only stealth spaceships are no-ships and the use Frank Herbert’s magic.
And yet, 15k posts later....
So his company can create copious amounts of AI generated child pornography in a place outside the legal jurisdictions of governments?
I don’t think putting your data centres in space will do anything to stop government regulation. Governments regulate foreign companies with no domestic footprint all the time when they do business with domestic customers.
Unless Musk actually succeeds in making his city on Mars (or now apparently the Moon), any AI generated child porn or other illegal material made by space AI data centers will be transmitted to users who live on countries on Earth, hence giving those countries at least some oversight of the company which is sending the data (SpaceX).
How lucky we are to live in such an age of technological wonders in which we must contemplate the legal questions posed by having a drug-addled billionaire using space data centers to beam kiddie porn created by artificial intelligence down upon us from the heavens.
Technically, he can already do that in the US -- virtual child pornography is potentially constitutionally protected speech per Ashcroft v. Free Speech Coalition, 535 U.S. 234 (2002) (https://en.wikipedia.org/wiki/Ashcroft_v._Free_Speech_Coalition). Prosecutors have to show that the content *also* satisfies the Miller test for obscenity for it to be criminally prosecutable.
all child porn should qualify as obscenity
Dan, thank you lol. Your comments bring me joy.
Seriously though, I don't know why he didn't just host his own fetish AI LLM for his own purposes, and keep the public one normal...
Because he thought OpenAI , Gemini, etc. were too woke/
Fundamentally they want engagement and are apparently now using a paywall to limit the explicit image function of Grok. It’s a complete disregard for the adverse costs their actions impose on others.
Data centers are clearly a crucial piece of infrastructure already, as we see by the chaos whenever cloudflare goes out for a few hours or whatever. It’s quite plausible that space based data centers will be more reliable than ground based ones, and will grow in importance as land use and electricity production become more binding constraints.
Musk obviously wants to monopolize anything he can.
There are MULTIPLE major issues with putting compute in space, including extreme thermal differentials between sun and shade (500 degrees or more), issues with shielding electronics against gamma rays, huge issues with solar activity / solar flares, huge issues with space debris, and hanging over it all, the potential for a cascading Kessler incident.
And that is on top of lift costs, which are not small.
Data centers in space is kind of the space equivalent of “we don’t need to cook when there’s Door Dash.” Sure, if you ignore all practical constraints.
We are not likely to need anti trust for this. Meat processing and PE consolidation of service providers, yes, anti trust enforcement would be a Godsend. This, not so much.
I think a couple of these are overblown.
1. AI inference is extremely robust to occasional bitflips. If a single cell in a matrix goes even from 0.0 to 1.0 (the worst case scenario for a single flip), the resulting effect is basically zero in a high parameter model. Just reload the model from central store once daily, and you're fine. This is a non-issue in my opinion.
2. Solar flares: they will very occasionally take out a small portion of your cluster, which is equally true of various natural disasters effecting your network of data centers on earth.
3. You would want these satellites in quite wide orbits, above the belt where ex: Starlink flies today. At that layer, the odds of debris hitting your satellite in any meaningful way are quite low.
I think the right way to think about this is "what percent of my satellites could I lost per year due to solar flares, various equipment failures, etc, over the expected ~5 year useful life of the satellite?" Even with a ~10% annual loss rate, I think the math pans out fine.
Heat dissipation feels like the biggest issue to me. Though I am far from an expert on that particular topic, I feel bullish on us being able to make the marginal (sub-linear) gains needed in that domain to make this project viable.
I don't know how you can do easily dismiss Kessler concerns. The cross sectional area of a single gigawatt scale solar panel array + radiators is going to be literally thousands of times the area of all satellites hitherto launched into orbit. And this all only makes sense if you can theoretically go from gigawatt to terawatt.
Demand a bottle deposit on everything that goes into space that gets refunded when it's confirmed to burn up on re-entry. Go with the Michigan model and make it 10 cents. If it's just 5 cents no one will care.
1. That’s soft errors, not hard. Hard errors result in permanent damage. Reloading the model does not fix that.
“Radiation is undoubtedly the main threat to electronic components in space. Cosmic radiation interacting with microchips can cause memory errors. High-energy ions or protons passing through a transistor can lead to short circuits, electron leakage, and, as a result, irreversible damage that can jeopardize an entire mission. While problems sometimes arise at launch, many manifest once the spacecraft leaves low Earth orbit, since closer to Earth, it is still shielded by the atmosphere and magnetic field. That’s why astronauts aboard the ISS often use electronics based on conventional “terrestrial” microchips. But GPS and GLONASS satellites are not so fortunate: at an altitude of about 20,000 km, the penetrating power of energetic particles is so high that the use of standard microchips is out of the question.”
https://maxpolyakov.com/electronics-in-space-operating-in-the-harshest-conditions/#:~:text=Radiation%20is%20undoubtedly%20the%20main,on%20conventional%20“terrestrial”%20microchips.
2. Loss of entire data centers due to terrestrial disasters is quite rare.
3. High orbits are actually more risky for Kessler effect, as if you do get a cascading series of collisions, clearing those orbits could take thousands of years.
Again, the math is "what percent of the cluster can I lose per year and have this be viable?" I think it could be made to work re: shielding compute and accounting for the occasional solar storm intersecting your orbit, but I admit I haven't done any deep research on this.
I agree with #3, though the trade there is that high orbits make collisions themselves super linearly less likely. But yes, this is a risk. I do think that this could be a blessing in disguise though, in that it will constrain the AI build out to only terrestrial energy sources, which will perhaps put enough of a natural brake on it to slow down an intelligence explosion enough for us to figure out what to do about it.
I mean NASA and DARPA have toyed around the idea of putting computers or networks in space for decades. I still think it is more difficult than Elon thinks it is from a compute and maintenance perspective. I'm bullish on this happening in a decade or so. But with energy costs potentially going down and compute power continuing to get more efficient as chips do over time, is this worth the squeeze for him now? Remains to be seen. It doesn't help that he overpromise's a lot.
"what percent of my satellites could I lost per year due to solar flares, various equipment failures, etc, over the expected ~5 year useful life of the satellite?"
This sentence alone makes me think its not feasible. Who is going to spend the enormous amounts of money to build and launch something significant into space with a **5 year expected useful life cycle!!! when you could just build it on the ground and it will work for way longer than that!
I picked 5 years for a specific reason: all of the major components (except maybe the radiators?) already have a 5-year depreciation cycle:
1. Starlink satellites expect to be de-orbited after 3-5 years, though I've read the latest models are targeting longer missions.
2. GPUs are useless junk for frontier AI purposes well within 5-years.
3. Thin-film photovoltaics generally maintain good efficiencies for a bit over 10 years, but they are also by far the cheapest part of the assembly.
Even if you build a solar-powered GPU data center here on earth, 80% of you CapEx is worthless well within 5 years.
People are taking a bunch of half-baked complaints and putting them together and thinking they have a fully baked complaint.
I think heat dissipation will end up ruining this idea on its own. But that one is pretty obvious to falsify: launch it and it either fries to death or doesn't, not much middle ground where we're not sure.
Most other tech issues people bring up simply aren't. They are things people think they saw in a sci-fi movie once.
"Even if you build a solar-powered GPU data center here on earth, 80% of you CapEx is worthless well within 5 years."
1) Its not worthless, its just not as good and likely has a decline that you can coax value out of for a longer period of time - making it more profitable.
2) But even if you have to replace all the GPUs, much of the rest of the complex is still in place and usable. That's 20% you don't have to replace compared to something in space. 20% is a large margin!
3) You could do a mix, upgrading over time instead of having a CapEx cliff. Or the reverse, Intel comes out with a chip that is 500% better than the current Nvidia model (not likely, but still). You can begin upgrading something on the ground, while the thing in space is stuck at its built performance.
1. A used A100 (top-tier chip from 5 years ago) is ~$4k in mint condition, less than 10% of the spot price of a top-tier trip from today (GB200). Good luck finding another capital asset that depreciates so quickly. There are few data center workloads, except maybe crypto mining for some reason, where it is useful to run a fleet of these.
2. Yes, but after 5 years on the ground you are also replacing batteries, which are very expensive and prone to failure. "The rest of the complex" is stuff (batteries, buildings, high power cabling, plumbing, etc etc) is stuff that just straight up doesn't exist in the space based deployment. Your coat of re-using it in space is 0, because it does not exist. Think of it this way: both modes of deployment have an AI chip and solar array that are near worthless in 5 years. The ground setup has all the stuff I mentioned above; the satellite setup has a chassis and radiator instead. It's not immediately obvious that at mass produced scales the former is costlier than the latter. It may be! But I feel like it will be pretty close.
I am not saying it's a slam dunk. I get that the benefits are marginal or nonexistent at the moment. I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space.
1) I agree that they degrade in cost quickly, but that doesn't mean they don't work. The chip industry is full of examples where less powerful chips are still useful in some capacity. But even if its the reverse, that just underlines my point which is that if chips are depreciating faster, than having to relaunch a whole new data center every 3 years instead of upgrading a groundbound center is worse!
2) Industrial batters don't fail in 5 years...most are targeting 20+. Beyond that, the reason that all that stuff exists on earth and not in space is because its solving for problems that a space center hasn't figured out yet! You have confirmed energy delivery, cooling understood, data transmission addressed, replacement/upgrade cycles understood and addressed, transport understood and addressed. Meanwhile in space, none of these are even understood much less addressed!
3) Let's talk about just that last one, transport. A server rack weighs in at 2-4 thousand lbs. Data center has 1000+ server racks. That's 2-4 million lbs, or 1000-2000 tons. The Falcon heavy can lift 64 tons and the Starship is posited to do 150 tons. So that works out to at best 7 and probably closer to 30 launches to put one data center in space. At 31 launches it would currently costs about $70+ million to get a Falcon launched. So current price just to transport a data center into space based on weight of server racks would be $2 billion dollars. Even if the price for weight falls dramatically, the transport costs would huge. And if AI really takes off, and we are needing to build thousands of data centers, it would take 10s of thousands of launches when SpaceX is doing a total of 150 a year. Not only do you need to figure out the engineering in space for a data center, you need to scale up launching so that you are doing hundreds a day!
"I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space."
I think that if this is true, its because AI has reached AGI and we live in a very different place than we do now.
1. A used A100 (top-tier chip from 5 years ago) is ~$4k in mint condition, less than 10% of the spot price of a top-tier trip from today (GB200). Good luck finding another capital asset that depreciates so quickly. There are few data center workloads, except maybe crypto mining for some reason, where it is useful to run a fleet of these.
2. Yes, but after 5 years on the ground you are also replacing batteries, which are very expensive and prone to failure. "The rest of the complex" is stuff (batteries, buildings, high power cabling, plumbing, etc etc) is stuff that just straight up doesn't exist in the space based deployment. Your coat of re-using it in space is 0, because it does not exist. Think of it this way: both modes of deployment have an AI chip and solar array that are near worthless in 5 years. The ground setup has all the stuff I mentioned above; the satellite setup has a chassis and radiator instead. It's not immediately obvious that at mass produced scales the former is costlier than the latter. It may be! But I feel like it will be pretty close.
I am not saying it's a slam dunk. I get that the benefits are marginal or nonexistent at the moment. I am just saying that thinking a decade out, it feels very plausible that the marginal watt of AI is built out in space.
" just build it on the ground" isn't so easy when there's such massive NIMBY backlash to datacenters these days. No matter how expensive it is to launch stuff into space, it's even more expensive to fight a never-ending legal battle against every single local government in the world.
> AI inference is extremely robust to occasional bitflips
This is goofy. I question whether you understand what a bitflip is. A single bitflip in firmware can cause corruption that can brick the hardware. In software it can crash the program. It doesn't just make a parameter wrong. You're thinking at the absolutely wrong abstraction level here.
> If a single cell in a matrix goes even from 0.0 to 1.0 (the worst case scenario for a single flip)
This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0, it makes the value completely unreadable. What are you even talking about? Datacenter servers don't have ECC memory just for fun.
I understand perfectly well. By mass, 99+% of the silicon being sent up is GPUs and ECC'ed memory. Only a tiny , isolated portion is dedicated to firmware, microcontroller, etc, which can be radiation shielded at relatively low (in terms of weight) costs.
> This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0
I did not say that it would. I was specifically enumerating the worst thing that could happen from a single bit flip of working memory neural net weights, not claiming that this specific switch would necessarily happen. Last I checked (a while ago, I don't know what the state of the art for vector dbs is), neural net weights are generally stored as matrices of small width floats (fp16?), clamped at 0.0 <= x <= 1.0. The worst case scenario, in terms of weighting, is that a single bit flip maximally reverses the weight: makes a near 0 weight go to near 1. Obviously a bit flip could also turn the float into an out of bounds value or NaN, but this would be an easily detectable error, and hence not the worst case scenario.
> What are you even talking about?
Why be mean and condescending? Why not be charitable and curious instead? I'd offer you the same courtesy.
I understand perfectly well. By mass, 99+% of the silicon being sent up is GPUs and ECC'ed memory. Only a tiny , isolated portion is dedicated to firmware, microcontroller, etc, which can be radiation shielded at relatively low (in terms of weight) costs.
> This doesn't even make any sense. A flipped bit somewhere randomly in a data structure doesn't turn a 0.0 into 1.0
I did not say that it would. I was specifically enumerating the worst thing that could happen from a single bit flip of working memory neural net weights, not claiming that this specific switch would necessarily happen. Last I checked (a while ago, I don't know what the state of the art for vector dbs is), neural net weights are generally stored as matrices of small width floats (fp16?), clamped at 0.0 <= x <= 1.0. The worst case scenario, in terms of weighting, is that a single bit flip maximally reverses the weight: makes a near 0 weight go to near 1. Obviously a bit flip could also turn the float into an out of bounds value or NaN, but this would be an easily detectable error, and hence not the worst case scenario.
> What are you even talking about?
Why be mean and condescending? Why not be charitable and curious instead? I'd offer you the same courtesy.
Yeah, I'm involved in amateur radio satellites (https://www.amsat.org/) and it has taught me that space is a very harsh environment for electronics, even relatively simple stuff like UHF/VHF analog radio repeaters.
Whatever NIMBY obstacles might exist, it just seems far easier on basically every front to just build data centers on Earth. For the cost of getting a bunch of electronics to orbit just buy land and pay off the local government to let you build a gigantic solar array.
Yes, there's a big problem with energy generation when your solar screens go into Earth's shade.
Because in space, no one can heat your screen.
[Ducks]
The thing is, earth based solar panels are in the Earth’s shade exactly half the time, and are often in the shade of other things, like big clouds of water vapor. If you have the right sort of orbit, space based solar panels will never be in the shade of anything.
Point taken.
No more bad, stupid, dumb puns from me.
(Spoiler alert: this promise was soon forgotten.)
In that case what is the harm in common carrier requirements?
Yeah this article was absolutely stupid.
I am against hobbling our only successful space company- and a world leading one- because the owner has bad politics and is intemperate on Twitter. Who among us, etc.
The extent to which "maybe Elon will do it" is still the official industrial policy of the United States is ridiculous. He's going to drag us back into PV manufacturing as a part of this, which should be a giant emergency that Congress wants to solve.
I don’t think it’s hobbling the company to say that it has to offer the same deals to all customers, including subsidiaries. We already do that for other important infrastructure companies, like Warren Buffet’s BNSF railroad.
I was under the impression that, as a result of measures taken under Obama and Trump I, we are already self-sufficient in PV cell output?
Anyway, if Musk's engineers can make it dirt-cheap via process improvement and automation, so as to compete with China's "subsidies, no environmental controls, and scale" model, all to the good.
“Musk is trying to do something here that is bad for the country and the world”
Hold on there. Musk is trying to do something to solve one of the world’s most pressing problems - AGAIN! Many of these problems, that Musk attacks, are left-wing priorities (municipal traffic, practical EVs).
He is now focusing his remarkable talents on using space, which he has made accessible, to mitigate our power and climate concerns. This is what you characterize as “bad for humanity”. It’s GREAT for humanity. It’s also true that we should not let Musk exploit his space leadership to block other AI companies from innovating space AI solutions. Your argument for Common Carrier regulations is sound, and maybe what that cures is all you meant with your “bad for humanity” comment. If Musk voted for Kamala I believe you’d have made that clear.
“As a result, this merger poses the very real risk that the combined company will be able to leverage a dominant market position in the space launch industry into a dominant market position in A.I. If nothing else, that is clearly what Musk wants to achieve with this merger.”
I’m pretty sure what Musk wants to achieve is a lower cost of capital for xAI. Turns out there are plenty of people that still take his word at face value though.
I think he's looking for something to do with all the excess launch capacity he'll have once Starship is flying regularly.
https://www.bloomberg.com/opinion/newsletters/2026-02-03/musk-s-moonshot-merger?embedded-checkout=true
That’s interesting. What makes you think that’s the main reason?
Full Self Driving is right around the corner folks!
Applying common carrier rules to space launch makes sense, but the rest of this seems a bit tenuous. Why does the merger matter if Musk already directly controls both companies? What about the fact that SpaceX currently launches Project Kuiper satellites that directly compete with Starlink? Is there any reason to believe they wouldn’t launch rival data centers?
If you agree on the policy upshot, I’m happy to agree to agree!
I’m with you on the common carrier part, but blocking the merger would be actively anti-competitive
I read the piece as an argument for either stopping the merger or establishing common carrier laws, not both.
Agreed, but common carrier regs for commercial space seem like a good idea in general (not just for space data centers), and blocking the merger seems bad unconditionally
I've been in the datacenter and telecom space for 20 years and the general experience of the industry is that anti-trust/common carrier legislation is good in theory but bad in practice. It is *theoretically* good to enable competition with a low capital barrier to entry. In reality, this has disincentivized significant capital investment in digital infrastructure outside of the two boom/bubble cycles unless directly subsidized by the government (which in itself is a whole other saga) - the competitive phone carriers that emerged in the 80s solely as a service on top of Bell infrastructure all failed relatively quickly, did nothing to improve quality of service, and pushed consumer prices so low (good) that there was no incentive for operators to reinvest (bad). There are other drivers at play here, of course, but the right way to deal with the AT&T monopoly in the late 70s would have been to make local exclusivity agreements illegal (something that is less relevant today but still needs to happen) and force consolidation of authority around permitting (something that is more relevant today and still needs to happen). Relative to what actually happened, this would have dramatically increased the competitive barrier of entry, which is bad (in theory), but winnowing the competitors only to well-capitalized players would have pushed competitive infrastructure buildouts earlier, increased mean quality of service and (theoretically) improved consumer pricing more quickly. Since SpaceX has neither a regulatory nor a natural monopoly, there is no good reason *in practice* to enforce the same kind of structure just because they are perceived to be clearly ahead (which is true but we are far enough away from datacenters in space being practical and there are enough competitors trying to solve the cooling issues at scale that it is plausible if not even likely that they won't be by the time this is relevant).
“…this has disincentivized significant capital investment…”
Don’t worry: The Chinese will step up to fill the gap when Matt’s plan goes into effect on the US powerhouses in the field.
The telephone story is interesting. But can you explain why the 3 cell carriers and the handful of cable companies offer such low speeds at such high prices compared to other countries?
The premise of this question isn't entirely correct, but nonetheless the regulatory environment around common carriers has nothing to do with the answer:
1. The median broadband/5G speed in the US generally better than in most other countries, and the exceptions to this are almost universally wealthy petrostates with nationalized access providers (i.e. a lot of capital available to solve the problem via enforced monopoly).
1a. Above 100 Mbps, speed isn't the best determinant of service quality for consumer internet anyway, median latency and lack of bufferbloat are, and by these metrics US ISPs are generally light-years ahead of peers elsewhere, although much of this is due to work by backbone and content providers, not the ISPs themselves.
2a. Median broadband pricing is somewhat higher than in many peer countries. Controlling for service quality in a quantitative way is difficult, but this is almost certainly less true or untrue once you control for service quality.
2b. Median wireless data pricing is dramatically higher than in many peer countries.
3. Density (low) and scale (high) are huge problems in the US that don't exist in most other peer nations in a substantial way. The capital required to launch an infrastructure-owning competitor at national scale is dramatically different here than in an EU nation.
3a. A byproduct of this is that the major MNOs in Europe own a majority of their own towers. This is not the case in the US and dramatically increases US MNO OpEx on a unit basis.
3b. The cost to build infrastructure of any kind is dramatically higher in the US for reasons beyond scale (queue Abundance argument, etc.). This dramatically increases US CapEx on a unit basis.
3c. Government subsidies for broadband/cell access in the US are narrowly tailored in a way that they are not in most peer countries.
3d. All of the above lead unit pricing higher but also create other perverse incentives for operators that are bad for consumers in attempts to grow profits.
"2b. Median wireless data pricing is dramatically higher than in many peer countries."
It seems to me that our wireless providers are much better at price discrimination than peers elsewhere. Bundle everything, especially including devices and help getting new ones up and running, into a single contract for well-off but uninquisitive consumers and sweep a bunch of the surplus off the table, while offering cheaper "utility-type" plans for more sophisticated mid-market consumers and cheap MNVOs for poorer or more budget-conscious ones.
I pay a comparable amount in 2026 for four basic unlimited data lines as I did for 2 separate plans with capped data in China a decade ago, and that's without having gone down to an MVNO. Maybe China has fallen a ton in price since then, but I doubt it's gone so far as to put the shoe clearly on the other foot.
I generally agree (although the proper comparable would be Vodafone/DT in the EU, the Chinese market is too dramatically different IMO). What I am trying to get at in 3d above is, in part, this type of behavior. Of course the Big 3 here are allowed to do a lot of things they would not be able to in the EU, but I don't think the degree to which their behavior is anti-consumer would be as significant if they could make decent margin at what we would consider a promotional service base rate.
Yep. I don't even know if the bundles are really anti-consumer, though the case could be made either way.
I was, however, under the impression that US telecoms infrastructure gets basically a blanket exemption from most of the usual processes that get leveraged by NIMBY nutjobs, no?
In some respects, yes. You don't need zoning approval to build most infrastructure except macro towers and small cells (and even then not in all situations), but there is still local permitting required which has its own sets of challenges and restrictions depending on what and where - and really the issue is the lack of uniformity in what you're going to encounter from jurisdiction to jurisdiction, or in many cases within a jurisdiction depending on what individual in a permitting office you may be dealing with! A lot of success comes down to relationship building with those in control of permitting, which is a great business practice but not exactly how government is supposed to work.
Amusingly this discussion reminded me that I have not yelled at Verizon's Fios people lately and, sure enough, my bill had crept up to their listed long-term rate. A brief chat and it's now chopped in half for the next twelve months and I have an annual calendar reminder in to do this in perpetuity.
Thanks for the detailed answer! I know I oversimplified my question a bit, but this is exactly the kind of perspective I was looking for.
Wait, what? My experience of US 5G has been leaps and bounds and climbed-cliffs ahead of my experience in Europe or Japan, let alone China.
The last time I had a fiber connection abroad was in Beijing in 2016, but it was rated at 500 Mbps and it still sucked. My current 300 Mbps connection here is comparable in price and vastly, vastly better, and has been since we bought our house in 2017.
The plural of anecdote is not data, blah blah, but it's enough to induce a deep sense of skepticism any time someone uses rated speeds to tell me that US telecoms suck.
Having a vibrant competition in space launch is a good thing. Therefore, in the present regulatory environment (i.e., based on whatever is bugging Donald Trump), the best way to keep at least two companies in that competition is to cheer on Jeff Bezos's destruction of the Washington Post.
Yuck.
Full disclosure: I work in the industry, but not for SpaceX. Thoughts are my own etc etc.
I think the bull case here is quite strong. If you look at where the puck is going on two fronts (launch cost per kg, and AI uptake and power consumption growth), this seems like the only long term viable solution. Three variables to consider:
1. Do you think AI demand will continue to grow at least linearly (if not exponentially) basically forever, a la railroads, the internet, or energy production? Set aside short term shocks like bubble pops, and think in 10+ year timelines.
2. Do you think the cost of a kg to orbit is going to go down with scale (of both individual vehicles, and in number of vehicles) in the usual "learning curve" that has applied to every industry from air travel to solar panel production?
3. Do you think we can develop slightly more efficient methods of shedding heat in space?
My personal answer to these is "definitely, definitely, and probably". If yours is too, then the long term trend of putting compute in space is obvious; and the only question becomes "how soon do I need to do it?" If you are a country like Saudi Arabia or China and have basically no barriers to "tiling the desert in solar panels", you can wait a decade before this becomes the only viable path. If you are the US or Europe, where the idea of doubling domestic electricity production is laughably off the table, the time to start doing this is "oh shit, right now". To be clear: the China/Saudi terrestrial approach while your rocket industry takes a decade to catch up is clearly the safer play, but that play is closed to the West for sadly self-inflicted reasons, so it seems like the space long shot is our best bet.
What in the name of all that is holy breeds this weird sense that AI computing is "*the* very lifeblood of the future and the sole terrain on which the New Cold War will be fought," as opposed to "just another technology whose continued development and successful integration into knowledge work will raise productivity growth by a percentage point or two a year until the S-curve flattens at some point in the future?"
The argument of "is AI like any other technology (see power loom, or flight, or nuclear power, or whatever) or qualitatively different in ways that make it hard to reason about?" has been utterly beaten to death elsewhere. Either you buy that thesis, or you don't. I personally do, you seem not to, c'est la vie I suppose.
Edit: just a thought exercise, if you look at America's 10 most valuable companies, roughly 100% of their workers' output is knowledge work. Seems like valuable stuff to me.
Great point on the 10 most valuable companies
All S-curves look exponential at some point.
Unless the theory is the immensely half-baked, technobabble-filled "recursive self-improvement" stuff, there's simply no reason to think we aren't headed for a capex bust in a few years followed by a long period in which white-collar work digests what LLMs can do well and people experiment with other architectures or use compute for other machine learning tools with more niche/technical applications.
The amount of economically useful work that LLMs can do at present, even if they were *snaps fingers* perfectly integrated into every white-collar field today, isn't revolutionary enough to justify either planned capex or valuations, and capability growth is already slowing down on the way to plateauing.
And when we wake up in a few years, China will still conduct 30% of global manufacturing value-add, against like 15% for the US, a figure which is in some ways overstated by exchange rates.
EDIT: Having just caught your edit... sure, if LLMs or any on-the-horizon architecture were actually even theoretically capable of doing a large fraction of that knowledge work, it'd be revolutionary. LLMs specifically are not and there are very good reasons to believe that this will always be the case, and there are to my knowledge no successor architectures posed that could overcome such limitations.
LLMs already do a large fraction of what used to be my job. The latest versions of the frontier models were mostly coded by their predecessors. I am not sure your argument about knowledge work holds up to sustained scrutiny today, let alone if we extrapolate linear improvement. And most evidence indicates that improvement is still exponential.
The fraction of my job they can do rounds to 0, which isn't surprising as it's a mainly human-facing role. Wringing value from our corporate LLM as a search assistant is, in the main, harder than simply filing and documenting things well to begin with. Wringing value from it as a copywriter is harder than writing things well in the first place, at least for me. Wringing value from it as a sounding board is functionally impossible, worse in every way than finding 20 minutes to shoot the shit with a trusted coworker. It just spits out nonsense, on par with the worst corporate buzzword presentations by MBAs with no industry knowledge. Our developers have said they find it useful for hacking stuff together but it requires a lot of policing and checking before integrating any coding outputs into work product.
What's worse is that they can do very little better when used by the engineering firms and infrastructure owners my company serves and pointed squarely at technical tasks with short horizons. The most promising applications we've got at this point, which are the same ones we had last year, and mostly the same ones we posed the year prior, are "code research assistant" and "drawing output checking engine." There's some hope for "automatic annotation engine" but at present they underperform older, purely rules-based, tools on this front. For all three applications, we're very, very deep into "trust but verify" territory as the accuracy rate is in no case over 90%.
All this adds up to an impression, not of complete uselessness, but of "another automation tool to be integrated into white collar work for some marginal productivity gains spaced over a couple decades." It's Excel-All-Over-Again, not Skynet.
"The latest versions of the frontier models were mostly coded by their predecessors."
This... does not comport with what folks in the industry have told me, at all.
The state of the art has changed dramatically in the last three months (like December 2025 is literally the inflection point). Most of the prominent professionals (like, well known legends with major success under their belt) in my field are producing most of their output via LLMs, and even the ones who are not refrain for mostly aesthetic or ethical reasons, not because they doubt the capability. I would check back in with that developer team in a month or two and see what’s changed.
Coding is a closed environment and thus LLMs show better progress. In lots of other fields, that closedness does not exist. And because it does not exist, the applicability is limited. For instance in law, some tasks like first line discovery and review of contracts were being automated and now LLMs can make that more capable.
So I am not doubting that there are some good advances but to extrapolate from programming to other knowledge work is a category error.
Everything except high level "advice and experience" tasks are immediately in the cross hairs. With software, this may at least keep many engineers employed because demand for software has (so far) been quite inelastic. Unless the amount of legal work explodes at a similar scale (for all of our sanity, I hope not...), you're just going to need many fewer lawyers to do the same amount of high-level legal work.
Now, lawyers as a class in particular may succeed in insulating themselves via regulation, since they write all of the regulation, but ex: accountants may well be cooked. I am sure that doing RL on accounting is, if anything, easier than doing it for coding.
I don’t see either of the ideas you mention discussed here. It seems that what this person is claiming is that AI computing is somewhat more important than railroads or telecom infrastructure, and will likely yield several percentage points of annual growth.
This doesn't make sense to me at all. If AI becomes the thing, I don't understand why we wouldn't simply build a bunch of them in Arizona and Nevada under massive solar panel farms. It would be far cheaper to build, maintain, cool, power, and upgrade than launching things into space.
Suppose you build a few, giant plants in Arizona and Nevada. First, those states aren't just Terra Nulla, they have their own political and environmental groups which will fight you on this. Then you have to find a way to concentrate all that solar power into one plant, which is difficult and eventually becomes impossible. Gonna need massive battery backup for nightfall, too. Then you have to distribute all that data back to the rest of the country (or world) which is another incredibly hard engineering problem.
Putting them in space, you completely sidestep all of those issues. Of course, it comes with new challenges, but perhaps those challenges will be less intractable.
"it comes with new challenges, but perhaps those challenges will be less intractable."
There's a lot of work being done in that "perhaps...less intractable." There were over 500 data centers built in 2025 alone. SpaceX did less than 200 launches in 2025. It would take multiple *Starship* much less Falcon launches to put the equivalent of one data center in the space. That's not including dealing with all the other engineer challenges that would have to be overcome simply to build what can already be done on Earth. I wouldn't expect that to be cost competitive for at least a decade and probably two - if it ever does.
You write "at least a decade or two" as if that's some unimaginably long amount of time. That's roughly how long it took Tesla from its founding to where it is now. Companies make 20 yeah investments all the time, nations even more so. The need for data centers isn't going away, and its only getting harder to find places to put them.
This imagines a world where we need to ex: double energy capacity in the next decade or less to support the AI buildout. This would require tiling nearly the entirety of Nevada in solar. By comparison, going to space on an hourly cadence is merely an engineering and capital allocation challenge. It's not obvious whether the regulator and political obstacles of the former are harder to surmount that the lure-space option.
Yeah, I think all of that will be easier to overcome than building AI data centers in space - at least for the next decade or two. Especially at that large of a scale, just how many rockets do you think SpaceX can launch that it could outpace building on the ground?
I guess we’re sort of agreeing then? I am saying this will probably become viable in about a decade, you’re saying it will be a decade or two… seems like a minute difference in timelines, mostly.
Its very hard to think about where we are in 20 years with advancements to AI. Perhaps we've achieved AGI by then! More likely IMO, we get some significant advancements but its not nearly the breakthrough that people imagine now, but more like computer chips have been. Massive changes in many ways, but not transforming everything as many claim. Or something in the middle. AI could help us figure out fusion or help build a space elevator. In which case all these discussion are missing key parts. Or perhaps it leads to nuclear war and we're building stuff in space because half the world in is ruins.
The "slightly" in point 3 is doing a lot of work. Currently radiators would be heavier/larger than the panels.
And in point 2, you mention "normal learning curve" cost reduction for space launches. But that's not what we need. We would need exponential reduction across a few orders of magnitude. That's only ever happened with compute (Moore's law) and perhaps solar.
Space launch is currently a mechanical/chemical pursuit, not an electrical/digital one.
Long story short, you can start studying it "right now", but it won't be a viable industry for a few decades at least.
Would you be able to point me towards anyone who has done an analysis of what the cost breakdown would need to be for things like kg/orbit for data centers in space to be feasible? My prior is that it's 10x the price it would need to be for it to be reasonable, so even if it drops by half that wouldn't be enough. But I'm a lay person in this field and would be super interested to understand how the optimists do their math.
If the value of data centers is also growing (and their other costs are also falling) then the lines cross sooner. What is the halving time for the cost of space launch? Three of those gets you almost to 10x.
Keep in mind that I work in the industry, so am perhaps over bullish relative to a totally dispassionate analysis. Here's a report that predicts a drop from the current customer-facing frontier on F9 (roughly $1.5k/kg, almost certainly lower if done at-cost) to <$100/kg and as low as $33/kg in a little over a decade (warning: PDF): https://ir.citi.com/gps/kdhSENV4r6W%2BZfP44EmqY4zHu%2BDy0vMIZnLqk4CrvkaSl1RIJ943g%2FrFEnNLiT1jB%2BjLJV4P9JM%3D
I mean... this was published in 2022 and has projections that include a fall in total launch costs to around $400/kg by the end of 2025...
It seems to me that "over bullish" doesn't quite cover it. Reading the assumptions they crank into the analysis for 2040 is... harrowing.
We're pricing in a NIMBY tax that is so high that it's preferable to pack up your entire physical infrastructure and send it to space, adding a host of design challenges around heat dissipation, rad hardening, surviving the launch environment?
This is insanity. A world in which space data centers make economic sense is a world in which property rights have degraded to the point to make any gains from building space centers infeasible.
I didn't buy "serious people are considering this" as an argument for it being feasible. No company wants to be the one who admits that they aren't looking into something new and exciting so there's hype reinforcement.
Elon has confidently written off space based solar power, space data centers are only slightly more reasonable.
It's so bizarre to me. Massive data centers are being built today without significant pushback. Any scaling issue with building data centers on terra firma would be even worse in space.
Some data centers are being built without any pushback, but a lot of data centers are getting pushback now. If this gets worse, and the best sites without pushback get used up, then space starts to look better, because all “sites” in space are equal cost and value (at least, for a long long time, since the surface area of orbit is much higher than the entire surface area of earth, let alone the land area that doesn’t have too many disasters).
I get the argument but the case seems vastly overstated. There are 1.5B acres of undeveloped land in the US alone. There's no political crisis in finding land for applications with much less capital and much worse externalities, like industrial meat processing or hog rendering or whatever. There seems to be a huge jump from "we can't build our data center in our first location, downtown Los Angeles" and "I guess we should send it to space!" You can't put these things in Mexico or wherever?
Yeah, it’s hard to tell how significant the ground based obstacles become. But the big thing for space based systems is that if they work once, then you can really get in on the economies of scale, whereas each ground based site has some site specific differences, and there’s more unpredictable weather, legal regulations, and geology on earth on a weekly basis (though perhaps solar flares become a similar issue in space on a yearly basis?)
Solar flares, rockets exploding, etc. etc. I think my main problem is that the solution fast forwards through technological leaps that would make equivalent construction on land much easier. Like this idea is dependent on new ways of cooling GPUs that don't require water or air. Wouldn't that make data centers much more palatable to build on earth?
It reminds me a lot of the basis for terraforming Mars - if there's some ecological disaster on earth we would need to create a habitable environment on Mars. But any technology you create for that would be able to fix the ecological issues here!
I definitely think that terraforming Mars makes less sense than “terraforming” Antarctica!
But space is different enough that it’s not quite so obvious to me that whatever works in space would work on the ground.
I realized after I wrote those two paragraphs that I should ask my partner, since we were having lunch together, and his research is in the optical properties of nano materials. He said that most purely radiative things that could work in space would work on the ground. But he’s been working on materials that take incoming light and re-emit it at higher energy, so that they use incoming light for cooling. If that sort of system could work, then it would work better in space, with 24/7 sunlight, than on the ground. But he did say that cooling in space is one of the big problems, and that putting something that depends on cooling in space is not likely a good idea any time soon.
As a matter of principle, I absolutely agree that if SpaceX is a de facto launch monopoly imposing common carrier-type requirements and other anti-trust enforcement is the correct policy move.
All that said, I think it’s worth noting that Elon is engaging in pre-IPO hucksterism here and that the economic case for data centers in space requires the exhaustion of current cheap energy sources on Earth, a very dramatic decrease in the costs of orbital launch (probably requiring a different mechanism than rocketry— maybe some kind of space elevator), and likely also significant advances in materials science (no convection in space so cooling requires massive radiators)
Right now, there are places on Earth and in developed countries with abundant cheap energy, many of which are in rule of law friendly jurisdictions. In the US, one of these is the Permian Basin in Texas and New Mexico, which has tons of access to cheap wind and solar power and also so much “waste” natural gas from oil production that local methane production often exceeds takeaway capacity and the price at the local gas hub goes negative.
But despite this availability— and land being extremely cheap— hyperscalers have been very cautious about building new data centers in the Permian. This is largely because of a combination of latency issues (proximity to users is super important) and maintenance challenges (Permian is sparsely populated so getting people out to fix problems or update hardware is costly).
Space is uh, considerably worse than the Permian basin on both of these axes. It really, really does not make sense except for say, defense applications where compute collocated with satellites is unusually useful. (A real market but much more limited TAM)
If the data centers are used for training AI models, rather than for inference, latency doesn’t really matter. The other points are important though.
Again, we’re not actually seeing this at scale in terrestrial cheap energy locations like the Permian, Scandinavia (cheap hydro), interior British Columbia (cheap hydro plus cheap gas from the Montney Basin), the Arabian peninsula (abundant solar + gas), etc
I understand that part is the issue is that training is still pretty high-touch and requires close human monitoring (very plausible that this becomes automatable but we’re not there yet), and another part is that any model that’s economically worth its training costs will ultimately generate orders of magnitude more compute in inference use than it did in training (so training-only clusters would only be a small share of the full data center stack)
Yeah, that all seems right.
I think this significantly misses the point Hayes is making. Even though all of those billionaires have divergent views and interests, they are all united (along with other billionaires like Ellison and Benioff and lots of non-tech investors) begin the idea that we're close to building a machine that will replace all white collar workers, and we should built it as fast as possible.
My issue with Hayes statement isn't just that it flattens the divergent views of billionaires, I think it wrongly assumes that every billionaire on earth wants to immiserate white collar workers for their own bottom lines. Billionaires have a lot of different political beliefs and some might actually want a world with mass unemployment because they're bad people, but my suspicion is that many are smart enough to see that that is a bad future for society.
I think the following two statements are true:
- basically all billionaires are all in on developing AI and a positive about doing it
- many people involved in AI, including some of those same billionaires, think that AI will soon be able to replace a huge swath of the workforce
There are lots of differences of opinion about what to do about that and what the implications are, but I take the conjunction of those two claims to be Hayes' point.
"basically all billionaires are all in on developing AI and a positive about doing it"
I don't think this is true at all. Is the Arnault family heavily investing in AI, or Amancio Ortega? I haven't heard Warren Buffet is deeply invested in it.
I think there are many who are, but many others who are focused on the business that got them there and it has nothing to do with AI.
Oprah and Taylor Swift too!
I can't read Hayes mind obviously but I don't think this is what he means at all; it seems much more like a general position-taking statement in which enemies are identified (in the left's current idiom, they are all "billionaires"). The tweet doesn't even make sense as a causal statement: "billionaires" are a group of people and "globalization" and "neoliberalism" are abstract concepts. This strikes me as a thrown-off tweet, a vent, rather than some through-through claim.
You're taking two steps here, interpreting the tweet as "LLM-based AI will do to white collar workers what globalization and neoliberalism did to (American) blue collar workers" and then imputing the spread of LLM-based AI to "billionaires." Fine (and we can argue about that), but highly doubt that's what Hayes means.
I don't think the claim is that billionaires are the causal agents of AI, but that billionaires are all in on building AI.
"but that billionaires are all in on building AI"
Which is clearly wrong. It's nonsensical.
It doesn't take very many if them with their hands in the right pies to make things very bad indeed for the rest of us. The ones helming the big AI companies are not particularly pro-human people.
Dude it's like three real billionaires and a couple paper billionaires that we're actually taking about. Buffett has been anti-tech for 50 years now. The collective Walmart heirs don't give a fuck. Gates is buying farmland. Bloomberg's wealth, outside of the Co., is mostly in real estate. The Koch's are still OG tied to energy infrastructure (a friend ran their venture arm). I can't think of a more diverse "class" than billionaires. Doesn't matter to Hayes though - he'll just keep making shit up.
Bur Hayes has a show on M. Snow. He must know stuff.
I’m not sure that it makes a practical difference, but some people think it makes a moral difference that the billionaires aren’t particularly interested in the immiseration of the white collar class - it’s just a potential foreseeable side effect of the thing they are in fact interested in (along with the side effect of cheap white collar work now available to the working class).
I agree that that's true, but I don't think it makes an important difference. This was true for international trade as well -- the people who advocated for it (including me, still!) didn't want to immiserate anyone.
Salesforce is pitching Agent Force now. The logical end point of that is robot salesmen.
Google, I'm convinced, is going to wind up Larry, Sergey, and a dog that bites them if they interfere with mecha Pichai.
The proprietary self-improving AI stack they own, from data collection to training to silicon design.
I don't think that's the barrier for Oracle in particular -- they sell specific, hard to replace, vital, and extremely expensive software.
It is just not remotely true that Claude Code could easily replicate the Oracle database, or PeopleSoft, or any of the other important Oracle products.
Technical question: how do we know it built the compiler "from scratch" and not based on some other compiler it saw?
It is somewhat laughable that anyone believes they can shape someone else's AI. The government failed to shape social media, which has helped to crater birth rates and raise a generation of nihilists.
Of course, the flip side is that we have a known history of men building the railroads. Without Vanderbilt and others, we wouldn’t have had trains, which opened up interstate trade along with the entire country. While that time was the movement of goods and people, today we are moving information.
The same millionaires of that day, private wealth and industry built it, and the country benefited. Elon Musk is a jerk with an imagination. I worked 35 years in the automotive industry. I can give a number of reasons why Musk's little electric car business was going to fail. Nobody believed he would deliver.
So, no, I would do much to harm Altman, Bezos, or Musk. Bezos and Boeing, and likely some other global rocket companies, can compete against Musk. People can pay to have a data center built in space. It has been called the commercialization of space. That is exactly what is happening.
I would suggest Matt and the Progressives let it happen, just like the railroads happened.
While I find the article overwrought, citing social media, of all things, as a reason not to regulate industries also seems underbaked.
If I had my way we'd regulate social media right the fuck out of existence.
This is what I wrote. I don’t understand your comment
The government failed to shape social media
Matt is not trying to stop it; he specifically said to treat it the way we treated the railroads and impose common carrier rules. Matt has a touch of Elon Derangement Syndrome, which is probably why he is jumping the gun on this issue, but his point seems valid.
Common Carrier mostly means that the federal regulators have wide latitude to set arbitrary rules. If you're a business this is bad for the same reason Trump is bad- you are at risk of getting jerked around by capricious officials.
Given Elon's status, it also seems like a guaranteed outcome, and is a backdoor to let populist dems cripple the company.
I think for this, you need a tort first, you can't just do it preemptively.
Exactly. This article is too Dune-brained.
Bezos will offer rides, or Boeing will if Elon won’t. I would let Elon build the data center for proof of concept. Hamstinging him now is foolhardy.
Will Amazon and Boeing offer significant competition to SpaceX within the next decade?
I would like to think so, hard to answer but I believe Boeing is still in the astronaut program with its capsule
They are, but that's mostly due to government reluctance to be reliant on a sole provider. Legacy space is woefully uncompetitive these days.
If SpaceX has a real competitor in the next decade it's not going to be the ULA.
The relevant standard under Section 7 (the typical statute for challenging a merger) is "substantially to lessen competition, or to tend to create a monopoly" in a relevant antitrust market.
I don't think there's a serious argument presented here that that standard is met.
First, which market? The market for LLMs? The market for space datacenters? The latter doesn't even exist, it can hardly be said to have a substantial reduction in competition.
I've worked on a several of anti-trust investigations and litigations. This post has the same problem you critique anti-trust maximalists for: it's all vibes. If you want to win anti-trust cases in court, you need sound theories, backed by specific evidence, and economics that supports the case.
Frankly, this comes off as "you don't like Musk, therefore we must find a reason to object to the deal".
Wow -- preemptively constrain a risky venture to create a new industry, just in case it works, for nakedly partisan reasons, on transparently spurious grounds of anti-trust? Not what I would expect from you.
It’s not for partisan reasons - it’s for the clear threat to competitiveness of vertically integrating a non-competitive utility with a single participant in the downstream market.