SpaceX wants to put 1 million solar-powered data centers into orbit

[Daniel Nenni]

The company claims it’s a cheaper and more environmentally-friendly alternative to land-based centers.

SpaceX has filed a request with the Federal Communications Commission to launch a constellation of up to 1 million solar-powered satellites that it said will serve as data centers for artificial intelligence.

The company’s filing lays out a grandiose vision, not just describing these planned satellites as “the most efficient way to meet the accelerating demand for AI computing power” but also framing them as “a first step towards becoming a Kardashev II-level civilization — one that can harness the Sun’s full power” while also “ensuring humanity’s multi-planetary future amongst the stars.”

The Verge argued that the 1 million satellite number is unlikely to be approved outright and is probably meant as a starting point for negotiations. The FCC recently gave SpaceX permission to launch an additional 7,500 Starlink satellites, but said it would “defer authorization on the remaining 14,988” proposed satellites.

There are currently around man-made 15,000 satellites orbiting the Earth, according to the European Space Agency, and they’re already creating issues with pollution and debris.

The filing also comes as Amazon — citing a lack of rockets — is seeking an extension on an FCC deadline to have more than 1,600 satellites in orbit. Meanwhile, SpaceX is reportedly considering a merger with two of Elon Musk’s other companies, Tesla and xAI (which already merged with X), ahead of going public.

SpaceX seeks federal approval to launch 1 million solar-powered satellite data centers | TechCrunch

SpaceX's filing claims these satellites will be “a first step towards becoming a Kardashev II-level civilization — one that can harness the Sun’s full power."

techcrunch.com

 

[C-4]

So, how do you cool your chips in space where you can’t just dump the heat into the atmosphere or the sea? And what to do about the part of low earth orbit where the sun sets behind earth?

 

[Barnsley]

What happen when there is so much junk up there it blocks the sun?

 

[IanD]

So, how do you cool your chips in space where you can’t just dump the heat into the atmosphere or the sea? And what to do about the part of low earth orbit where the sun sets behind earth?

You radiate the heat out into space behind the solar panels -- might take similar area to solar panels for input power (see calculations elsewhere), but cooling rig weight could be "challenging"... ;-)

I assume they'd use orbits in a band around the Earth which are always roughly at right-angles to the sun, so they never go into the shadow.

 

[IanD]

What happen when there is so much junk up there it blocks the sun?

Even a million orbiting data centers would occupy only a *tiny* fraction of the orbital area...

“Space is big. You just won't believe how vastly, hugely, mind-bogglingly big it is. I mean, you may think it's a long way down the road to the chemist's, but that's just peanuts to space.”

 

[hist78]

So, how do you cool your chips in space where you can’t just dump the heat into the atmosphere or the sea? And what to do about the part of low earth orbit where the sun sets behind earth?

In order to reach such one million "data centers", each of them needs to be cheap and almost disposable, because there is no easy or cheap way to send groups of engineers or robots to the outer space to fix anything that breaks down due to technical or environmental issues.

Is it real, or is it an idea created for other reasons?

 

[peter]

You radiate the heat out into space behind the solar panels -- might take similar area to solar panels for input power (see calculations elsewhere), but cooling rig weight could be "challenging"... ;-)

I assume they'd use orbits in a band around the Earth which are always roughly at right-angles to the sun, so they never go into the shadow.

The FCC application states 500km to 2000km altitude for the orbital shells at up to 50km increments at 30 degree inclination and sun synchronous orbits

What happen when there is so much junk up there it blocks the sun?

Funny you mention that the FCC application also states: "This satellite system will represent the “first step towards becoming a Kardashev II-level civilization – one that can harness the Sun’s full power,” according to SpaceX"

In order to reach such one million "data centers", each of them needs to be cheap and almost disposable, because there is no easy or cheap way to send groups of engineers or robots to the outer space to fix anything that breaks down due to technical or environmental issues.

Is it real, or is it an idea created for other reasons?

I suppose it provides a lot of autonomy/less regulation and that alone could be a big driver considering Musk's AI ambitions and the uphill battle these companies are going to face building out AI datacenters in areas that are power constrained. As pro nuclear as I am, I don't envision any new nuclear plants being completed within a decade in the US, far too slow to make a dent on the explosive energy demand that datacenters are going to place on the grid in just the next few years. Starlink V3 sats have >100kw of solar modules, and solar arrays in orbit have 2-3 times higher capacity factor than the best places on Earth with unparalleled consistency.

Is it real? If you run the numbers, it comes down to whether or not they can get the launch cost per kg down sufficiently. Starlink has already been a driver for them to do that and really push Starship, but datacenter satellites in space could dramatically increase the demand for launch volume and drive Starship launch costs down, while also: 1. potentially improve latency, 2. no real estate costs/energy costs to operate, 3. reduce physical staffing costs, 4. potentially reduce network connectivity costs utilizing Starlink (will still need ground stations), 5. not deal with any nimbyism or regulatory restraints, 6. potentially be more reliable. It's interesting to see how when Microsoft tested a containerized datacenter in the ocean with an oxygen purged atmosphere, the reliability of the hardware was 7 times higher than the same hardware exposed to air. Hard to say whether such improvements can be seen in space, you do have to contend with stray cosmic rays and CMEs and the like, and mitigating those hazards by rad-hardening, but perhaps this would be a great opportunity to test out how latest generation semiconductors handle the rigors of space compared to their terrestrial counterparts in parallel.

As for cheap and disposable, that's how Starlink satellites are designed. I'd be concerned about making sure there are no collisions and they are safely deorbited after reading EOL. Heat dissipation is manageable, wouldn't be too concerned as these satellites will spend half the time behind the Earth, radiator sizes will get big but it's not a show stopper.

 

[Xebec]

What happen when there is so much junk up there it blocks the sun?

Space is pretty darn big...

 

[Xebec]

In order to reach such one million "data centers", each of them needs to be cheap and almost disposable, because there is no easy or cheap way to send groups of engineers or robots to the outer space to fix anything that breaks down due to technical or environmental issues.

Is it real, or is it an idea created for other reasons?

There'a absolutely no way this is a stock pump! /sarcasm

Slightly more seriously -- I think SpaceX knows how to handle the cheap and semi-disposable piece - the Starlink satellites are designed that way. They also come with solar power, batteries, and some level of compute. They know how to re-enter them into the atmosphere to burn them up, etc.

But the size and cost are obviously the issues. Guesses online think the Starlink satellites can radiate and power up to 1kW continuously (i.e. ball park of a single "Blackwell Ultra" GPU), and have the battery to support that when they're out of the sun (2+ kWh, maybe more for longevity reasons).

So if they just repurposed Starlink satellites with some small enhancements to get them to dock together and communicate more closely, you're looking at ~ 30 Blackwell Ultras per Falcon 9 launch, and maybe ~ 300 per Starship V3-V4 launch. Rough internet based math; that's $500K per GPU on Falcon 9, and (my total guess) maybe 1/5th that - $100K per GPU on Starship.

That cost is insanely high, but permits, taxes, data center buildouts, and energy aren't trivial either on Earth. I'm too lazy to try to make a comparison in this post .

 

[peter]

There'a absolutely no way this is a stock pump! /sarcasm

Slightly more seriously -- I think SpaceX knows how to handle the cheap and semi-disposable piece - the Starlink satellites are designed that way. They also come with solar power, batteries, and some level of compute. They know how to re-enter them into the atmosphere to burn them up, etc.

But the size and cost are obviously the issues. Guesses online think the Starlink satellites can radiate and power up to 1kW continuously (i.e. ball park of a single "Blackwell Ultra" GPU), and have the battery to support that when they're out of the sun (2+ kWh, maybe more for longevity reasons).

So if they just repurposed Starlink satellites with some small enhancements to get them to dock together and communicate more closely, you're looking at ~ 30 Blackwell Ultras per Falcon 9 launch, and maybe ~ 300 per Starship V3-V4 launch. Rough internet based math; that's $500K per GPU on Falcon 9, and (my total guess) maybe 1/5th that - $100K per GPU on Starship.

That cost is insanely high, but permits, taxes, data center buildouts, and energy aren't trivial either on Earth. I'm too lazy to try to make a comparison in this post .

Did some back of the napkin math, it's anywhere between $10 to $30+M per MW of compute for terrestrial datacenters, Jensen Huang said some approach $50M per MW using the latest and greatest, not a trivial cost. Factor in OpEx costs and right now, one of the main weakness is securing energy to get larger datacenters connected to the grid, these costs will rise in the near future as we don't have any solutions with quick turnaround time. I don't believe these CapEx costs are even being factored in at the moment unless the datacenter generates power on-site (not considering backup generators). We're not rolling out nuclear quickly enough to power these datacenters, CCGT and other options are booked for the next few years solid, fuel cells from Bloom isn't exactly cheap but it is viable, solar and batteries can probably be deployed the quickest but will require substantial storage to provide stable uptime around the clock. Build out of power plants in the US on average for CTs is 4 years, even for solar, it's 2-3 years. There's 233GW of requests for power connections in the pipeline for ERCOT alone, just shy of 20% of the total generation capacity of the entire United States. CapEx costs for power generation is anywhere from about $1M per MW to $15+M per MW ranging from CT to nuclear, with solar closer to the lower end of the spectrum.

On the other hand, to build out datacenter in space, looking at some +25 tons per MW for hardware, with $15-25M per MW for the satellite hardware and compute. SpaceX has been using COTS for solar, not bothering with GaA or more exotic multijunction panels, and as mentioned above, solar capacity factor rises tremendously in space. It'll come down to how much Starship can really bring down launch costs. It doesn't have to beat terrestrial costs, even if the upfront cost is a bit higher, the time savings and OpEx costs could swing in favor of LEO datacenters. If they can bring launch costs down to even around $250/kg, I think there's a good shot of it being viable. That's already a magnitude lower than commercial Falcon 9 launch costs (though internal costs are probably already around $1000/kg). They only have to last 5 years tops, so there's room for cost cutting that isn't practical on earth, like using cheaper solar panels, no concerns about encapsulation because of operating in the vacuum of space, short life materials like perovskites could be worth deploying in space whereas currently, their few years of operating life makes them impractical to deploy terrestrially.

By the way, each Starlink v2 mini satellite's solar array is over 100m², they produce over 25kw in sunlight. Starlink v3 for Starship is supposed to be over 120kw. To date, there is nearly 0.4km² of solar panels in space from SpaceX alone, over 100MW of solar in space. This below picture is from 2023, they have over 9300 satellites in space now!

 

[peter]

So, how do you cool your chips in space where you can’t just dump the heat into the atmosphere or the sea? And what to do about the part of low earth orbit where the sun sets behind earth?

Radiative cooling. Use a black body radiation calculator. The amount of radiative surface area needed to cool is less than the size of the solar array, it's entirely doable. Sure, they will need to completely redesign the compute hardware to optimize for satellite packaging, but those issues aren't impossible to resolve.

At 500km atitude, 60% capacity factor, solar panels put out far more energy than their equivalent on Earth. No losses from the atmosphere, no weather to content with. The amount of solar and batteries needed would be quite a bit less for the same amount of energy vs a terrestrial setup. I'd be more concerned about shielding for semis using the most advanced processes but considering how well Starlink has been working and still rapidly iterating, seems like those are concerns their engineers are successfully addressing. The amount of lithium used for batteries, or other materials used in space now is trivial but I'm not sure what the long term ramifications of deploying all this hardware only to have them disintegrate in the atmosphere will be. Not to mention the studies showing aluminum components being aerosolized in the atmosphere during reentry causing ozone destruction, something we're actually successfully reversing with the implementation of the Montreal Protocol. I hate a love-hate relationship with SpaceX, rely entirely on it for high speed communication living off grid, but destroys all my astrophotography imaging which was one of the reasons for moving off grid in the first place.

 

[swka]

Elon has been saying such space based DC will be available in 36 months, I think. Most discussion here are about cost. Like solar efficiency increase in space, I assume the radiation impact will increase multi-fold as well, without earth atmosphere to absorb most of the radiation. With that, it seemed we need to redesign most of our electronics for DC to much higher RAD standard, including possible process technology improvement. 36 months do not seem likely for such HVM production of new development.

Thoughts?

 

[peter]

Elon has been saying such space based DC will be available in 36 months, I think. Most discussion here are about cost. Like solar efficiency increase in space, I assume the radiation impact will increase multi-fold as well, without earth atmosphere to absorb most of the radiation. With that, it seemed we need to redesign most of our electronics for DC to much higher RAD standard, including possible process technology improvement. 36 months do not seem likely for such HVM production of new development.

Thoughts?

Maybe rudimentary PoC systems in that time frame, I don't see HVM for latest and greatest happening within 36 months. Radiation impact does increase, but it's still LEO, and considering the short expected lifespan of the satellites, it seems a few mm of aluminum is enough for shielding and COTS can be utilized without specialized radiation hardening. Google has done some tests using their TPUs and AMD CPUs exposing the system to a proton beam with 5 years of expected exposure and reported the chips survived the test but the HBM did start to suffer some uncorrectable errors, but stated they were acceptable levels for inferencing work. But yeah, 36 months and doing a complete redesign of, say an Nvidia Rubin or Feynman system with cooling and power designed for a satellite bus...does seem a wee bit optimistic. I think they should just focus on getting Starship working and working consistently; once they do that, space datacenters may be inevitable.