It probably increases Elon's share of the combined entity.
It delivers on a promise to investors that he will make money for them, even as the underlying businesses are lousy.
A Starlink satellite uses about 5K Watts of solar power. It needs to dissipate around that amount (+ the sun power on it) just to operate. There are around 10K starlink satellites already in orbit, which means that the Starlink constellation is already effectively equivalent to a 50 Mega-watt (in a rough, back of the envelope feasibility way).
Isn't 50MW already by itself equivalent to the energy consumption of a typical hyperscaler cloud?
Why is starlink possible and other computations are not? Starlink is also already financially viable. Wouldn't it also become significantly cheaper as we improve our orbital launch vehicles?
A single AI rack consumes 60kW, and there is apparently a single DC that alone consumes 650MW.
When Microsoft puts in a DC, the machines are done in units of a "stamp", ie a couple racks together. These aren't scaled by dollar or sqft, but by the MW.
And on top of that... That's a bunch of satellites not even trying to crunch data at top speed. No where near the right order of magnitude.
The energy demand of these DCs is monstrous, I seriously can't imagine something similar being deployed in orbit...
The Outer Space Treaty (1967) has a loophole. If you launch from international waters (planned by SpaceX) and the equipment is not owned by a US-company or other legal entity there is significant legal ambiguity. This is Dogecoin with AI. Exploiting this accountability gap and creating a Grok AI plus free-speech platform in space sounds like a typical Elon endeavour.
That doesn't mean you need a gigawatt of power before achieving anything useful. For training, maybe, but not for inference which scales horizontally.
With satellites you need an orbital slot and launch time, and I honestly don't know how hard it is to get those, but space is pretty big and the only reasons for denying them would be safety. Once those are obtained done you can make satellite inferencing cubes in a factory and just keep launching them on a cadence.
I also strongly suspect, given some background reading, that radiator tech is very far from optimized. Most stuff we put into space so far just doesn't have big cooling needs, so there wasn't a market for advanced space radiator tech. If now there is, there's probably a lot of low hanging fruit (droplet radiators maybe).
Space has some huge downsides:
* Everything is being irradiated all the time. Things need to be radiation hardened or shielded.
* Putting even 1kg into space takes vast amounts of energy. A Falcon 9 burns 260 MJ of fuel per kg into LEO. I imagine the embodied energy in the disposable rocket and liquid oxygen make the total number 2-3x that at least.
* Cooling is a nightmare. The side of the satellite in the sun is very hot, while the side facing space is incredibly cold. No fans or heat sinks - all the heat has to be conducted from the electronics and radiated into space.
* Orbit keeping requires continuous effort. You need some sort of hypergolic rocket, which has the nasty effect of coating all your stuff in horrible corrosive chemicals
* You can't fix anything. Even a tiny failure means writing off the entire system.
* Everything has to be able to operate in a vacuum. No electrolytic capacitors for you!
So I guess the question is - why bother? The only benefit I can think of is very short "days" and "nights" - so you don't need as much solar or as big a battery to power the thing. But that benefit is surely outweighed by the fact you have to blast it all into space? Why not just overbuild the solar and batteries on earth?
You'd be wrong. There's a huge incentive to optimized radiator tech because of things like the international space station and MIR. It's a huge part of the deployment due to life having pretty narrow thermal bands. The added cost to deploy that tech also incentivizes hyper optimization.
Making bigger structures doesn't make that problem easier.
Fun fact, heat pipes were invented by NASA in the 60s to help address this very problem.
Are Earth-based datacenters actually bound by some bottleneck that space-based datacenters would not be? Grid connections or on-site power plants take time to build, yes. How long does it take to build the rocket fleet required to launch a space “datacenter” in a reasonable time window?
This is not a problem that needs to be solved. Certainly not worth investing billions in, and definitely not when run by the biggest scam artist of the 21st century.
We still don’t have any plan I’ve heard of for avoiding a cascade of space debris when satellites collide and turn into lots of fast moving shrapnel. Yes, space is big, but low Earth orbit is a very tiny subset of all space.
The amount of propulsion satellites have before they become unable to maneuver is relatively small and the more satellite traffic there is, the faster each satellite will exhaust their propulsion gasses.
Distributing useful work over so many small objects is a very hard problem, and not even shown to be possible at useful scales for many of the things AI datacenters are doing today. And that's with direct cables - using wireless communication means even less bandwidth between nodes, more noise as the number of nodes grows, and significantly higher power use and complexity for the communication in the first place.
Building data centres in the middle of the sahara desert is still much better in pretty much every metric than in space, be it price, performance, maintainance, efficiency, ease of cooling, pollution/"trash" disposal etc. Even things like communication network connectivity would be easier, as at the amounts of money this constellation mesh would cost you could lay new fibre optic cables to build an entire new global network to anywhere on earth and have new trunk connections to every major hub.
There are advantages to being in space - normally around increased visibility for wireless signals, allowing great distances to be covered at (relatively) low bandwidth. But that comes at an extreme cost. Paying that cost for a use case that simply doesn't get much advantages from those benefits is nonsense.
This is a pump-and-dump bid for investor money. They will line up to give it to him.
It's like his "Mars Colony" junk - and people lap it up, keeping him in the news (in a not explicitly negative light - unlike some recent stories....)
That is clearly not true. How do you power the data center on antarctica? May i remind you it will be in the shadow of earth for half a year.
Of course this doesn't solve the myriad problems, but it does put dissipation squarely in the category of "we've solved similar problems". I agree there's still no good reason to actually do this unless there's a use for all that compute out there in orbit, but that too is happening with immense growth and demand expected for increased pharmaceutical research and various manufacturing capabilities that require low/no gravity.
Space changes this. Laser based optical links offer bandwidth of 100 - 1000 Gbps with much lower power consumption than radio based links. They are more feasible in orbit due to the lack of interference and fogging.
> Building data centres in the middle of the sahara desert is still much better in pretty much every metric
This is not true for the power generation aspect (which is the main motivation for orbital TPUs). Desert solar is a hard problem due to the need for a water supply to keep the panels clear of dust. Also the cooling problem is greatly exacerbated.
The lack of launch costs more than offset the need for extra panels and batteries.
1. The capital costs are higher, you have to expend tons of energy to put it into orbit
2. The maintenance costs are higher because the lifetime of satellites is pretty low
3. Refurbishment is next to impossible
4. Networking is harder, either you are ok with a relatively small datacenter or you have to deal with radio or laser links between satellites
For starlink this isn't as important. Starlink provides something that can't really be provided any other way, but even so just the US uses 176 terawatt-hours of power for data centers so starlink is 1/400th of that assuming your estimate is accurate (and I'm not sure it is, does it account for the night cycle?)
Current satellites get around 150W/kg from solar panels. Cost of launching 1kg to space is ~$2000. So we're at $13.3(3)/Watt. We need to double it because same amount need to be dissipated so let's round it to $27
One NVidia GB200 rack is ~120kW. To just power it, you need to send $3 240 000 worth of payload into space. Then you need to send additional $3 106 000 (rack of them is 1553kg) worth of servers. Plus some extra for piping
At the end of the day I don't really care either way. It ain't my money, and their money isn't going to get back into the economy by sitting in a brokerage portfolio. To get them to spend money this is as good a way as any other, I guess. At least it helps fund a little spaceflight and satellite R&D on the way.
Presumably they're planning on doing in-orbit propellant transfer to reboost the satellites so that they don't have to let their GPUs crash into the ocean...
Ionizing radiation disrupts the crystalline structure of the semiconductor and makes performance worse over time.
High energy protons randomly flip bits, can cause latchup, single event gate rupture, destroy hardware immediately, etc.
These are all things which add weight, complexity and cost.
Propellant transfer to an orbital Starship hasn't even been done yet and that's completely vital to it's intended missions.
Hell, you're going to lose some fraction of chips to entropy every year. What if you could process those into reaction mass?
putting 1KW of solar on land - $2K, putting it into orbit on Starship (current ground-based heavy solar panels, 40kg for 4m2 of 1KW in space) - anywhere between $400 and $4K. Add to that that the costs on Earth will only be growing, while costs in space will be falling.
Ultimately Starship's costs will come down to the bare cost of fuel + oxidizer, 20kg per 1kg in LEO, i.e. less than $10. And if they manage streamlined operations and high reuse. Yet even with $100/kg, it is still better in space than on the ground.
And for cooling that people so complain about without running it in calculator - https://news.ycombinator.com/item?id=46878961
>2. The maintenance costs are higher because the lifetime of satellites is pretty low
it will live those 3-5 years of the GPU lifecycle.
That would make your solar panel (40kg) around $60K to put into space.
Even being generous and assuming you could get it to $100 per kg that's still $4000
There's a lot of land in the middle of nowhere that is going to be cheaper than sending shit to space.
What starship? The fantasy rocket Musk has been promising for 10 years or the real one that has thus far delivered only one banana worth of payload into orbit?
And maintenance and replacing parts and managing flights and ... You're trying to yadda-yadda so much opex here!
(I'm ignoring installation costs etc. because actually creating the satellites is ignored here, too)
Minus one big one: permitting. Every datacentre I know going up right now is spending 90% of their bullshit budget on battlig state and local governments.
All satellites launched into orbit these days are required to have de-orbiting capabilities to "clean up" after EOL.
I dunno, two years ago I would have said municipal zoning probably ain't as hard to ignore as international treaties, but who the hell knows these days.
Source? I can't immediately find anything like that.
xAI’s first data center buildout was in the 300MW range and their second is in the Gigawatt range. There are planned buildouts from other companies even bigger than that.
So data center buildouts in the AI era need 1-2 orders of magnitude more power and cooling than your 50MW estimate.
Even a single NVL72 rack, just one rack, needs 120kW.
Datacenters already exist. Putting datacenters in space does not offer any new capabilities.
The “+ solar power” part is the majority of the energy. Solar panel efficiency is only about 25-30% at beginning-of-life whereas typical albedos are effectively 100%. So your estimate is off by at least a factor of three.
Also, I’m not sure where you got 5 kw from. The area of the satellite is ~100 m2, which means they are intercepting over 100 kw of bolometric solar power.
0. https://www.arccompute.io/solutions/hardware/gpu-servers/sup...
The short answer is that ~100m2 of steel plate at 1400C (just below its melting point) will shed 50MW of power in black body radiation.
So your huge metal plate would radiate (1673/374)^4 = 400 times less heat, i.e. only 125 kW.
In reality, it would radiate much less than that, even if made of copper or silver covered with Vantablack, because the limited thermal conductivity will reduce the temperature for the parts distant from the body.
I would assume such a setup involves multiple stages of heat pumps to from GPU to 1400C radiatoe. Obviously that's going to impact efficiency.
Also I'm not seriously suggesting that 1400C radiators is a reasonable approach to cooling a space data centre. It's just intended to demonstrate how infeasible the idea is.
if the current satellite model dissipates 5kW, you can't just add a GPU (+1kW). maybe removing most of the downlink stuff lets you put in 2 GPUs? so if you had 10k of these, you'd have a pretty high-latency cluster of 20k GPUs.
I'm not saying I'd turn down free access to it, but it's also very cracked. you know, sort of Howard Hughesy.
Is that 5kW of electrical power input at the terminals, or 5kW irradiation onto the panels?
Because that sounds like kind of a lot, for something the size of a fridge.
Aside from the point others have made that 50 MW is small in the context of hyperscalers, if you want to do things like SOTA LLM training, you can't feasibly do it with large numbers of small devices.
Density is key because of latency - you need the nodes to be in close physical proximity to communicate with each other at very high speeds.
For training an LLM, you're ideally going to want individual satellites with power delivery on the order of at least about 20 MW, and that's just for training previous-generation SOTA models. That's nearly 5,000 times more power than a single current Starlink satellite, and nearly 300 times that of the ISS.
You'd need radiator areas in the range of tens of thousands of square meters to handle that. Is it theoretically technically possible? Sure. But it's a long-term project, the kind of thing that Musk will say takes "5 years" that will actually take many decades. And making it economically viable is another story - the OP article points out other issues with that, such as handling hardware upgrades. Starlink's current model relies on many cheap satellites - the equation changes when each one is going to be very, very expensive, large, and difficult to deploy.
A data center is nowhere near that and requires constant physical interventions. How do they suggest to address this?
This isn't quite true. It's very possible that the majority of that power is going into the antennas/lasers which technically means that the energy is being dissipated, but it never became heat in the first place. Also, 5KW solar power likely only means ~3kw of actual electrical consumption (you will over-provision a bit both for when you're behind the earth and also just for safety margin).
A single server in a data center will consume 5-10 kW.
I'm not a space engineer but I'd imagine that smaller satellites can make due with a lot of passive cooling on the exterior of the housing, whereas a shopping-mall sized computer in space would will require a lot of extra plumbing.
You might only care about coding models, but text is dominating the market share right now and Grok is the #2 model for that in arena rankings.
Openrouter is a decent proxy for real world use and Grok is currently 8% of the market: https://openrouter.ai/rankings (and is less than 7% of TypeScript programming)
They have no path to paying for their existence unless they drastically increase usage. There aren't going to be very many big winners in this segment and xAI's expenses are really really big.
Is the plan to have everyone so hopelessly dependent on their product that they grit their teeth and keep on paying?
Think about the stock return over a period - its composed of capital gains and dividends.
Now what happens capital gains disappears and perhaps turns into capital losses? Dividends have to go higher.
What does this mean? Less retained earnings / cashflows that can be re-invested.
Apple is the only one that will come out of this OK. The others will be destroyed for if they dont return cash, the cash balance will be discounted leading to a further reduction in the value of equity. The same thing that happened to Zuckerberg and Meta with the Metaverse fiasco.
Firms in the private sphere will go bust/acquired.
Gemini is practically guaranteed. With the ad model already primed, their financial resources, their traffic to endlessly promote Gemini (ala Chrome), their R&D capabilities around AI, their own chips, crazy access to training data, and so on - they'd have to pull the ultimate goof to mess up here.
Microsoft is toast, short of a miracle. I'd bet against Office and Windows here. As Office goes down, it's going to take Windows down with it. The great Office moat is about to end. The company struggles, the stock struggles, Azure gets spun off (unlock value, institutional pressure), Office + Windows get spun off - the company splits into pieces. The LLMs are an inflection point for Office and Microsoft is super at risk, backwards regarding AI and they're slow. The OpenAI pursuit as it was done, was a gigantic mistake for Microsoft - one of the dumbest strategies in the history of tech, it left them with their pants down. Altman may have killed a king by getting him to be complacent.
Grok is very unlikely to make it (as is). The merger with SpaceX guarantees its death as a competitor to GPT/Gemini/Claude, it's over. Maybe they'll turn Grok into something useful to SpaceX. More likely they'll slip behind and it'll die rapidly like Llama. The merger is because they see the writing on the wall, this is a bailout to the investors (not named Elon) of xAI, as the forced Twitter rollup was a bailout for the investors of Twitter.
Claude is in a weird spot. What they have is not worth $300-$500 billion. Can they figure out how to build a lot more value out of what they have today (and get their finances sustainable), before the clock runs out? Or do they get purchased by Meta, Microsoft, etc.
OpenAI has to rapidly roll out the advertising model and get the burn rate down to meaningless levels, so they're no longer dependent on capital markets for financing (that party is going to end suddenly).
Meta is permanently on the outside looking in. They will never field an in-house competitor to GPT or Gemini that can persistently keep up. Meta doesn't know what it is or why it should be trying to compete with GPT/Gemini/Claude. Their failure (at this) is already guaranteed. They should just acquire GPT 4o and let their aging userbase on FB endlessly talk itself into the grave for the next 30 years while clicking ads.
If Amazon knew what they were doing (they don't right now), they would: immediately split retail + ads and AWS. The ad business ensures that the retail business will continue to thrive and would be highly lucrative. Then have AWS purchase Anthropic when valuations drop, bolt it on to AWS everything. Far less of an anti-trust issue than if what is presently known as Amazon attempted it here and now. Anthropic needs to build a lot on to itself to sustain itself and justify its valuation, AWS already has the answer to that.
If valuations plunge, and OpenAI is not yet sustainable, Microsoft should split itself into pieces and have the Windows-Office division purchase OpenAI as their AI option. It'd be their only path to avoiding anti-trust blocking that acquisition. As is Microsoft would not be allowed to buy OpenAI. Alternatively Microsoft can take a shot at acquiring Anthropic at some point - this seems likely given the internal usage going on at Redmond, the primary question is anti-trust (but in this case, Anthropic is viewed as the #3, so Microsoft would argue it bolsters competition with GPT & Gemini).
