At first I thought the poles (of the planet) might be good. The cooling is basically free. But the energy and internet connectivity would be a problem. At the poles you can really only get solar about three months a year, and even then you need a lot of panels. Most of Antarctica is powered diesel because of this.
So the next thought was space. At the time, launching to space was way too costly for it to ever make sense. But now, with much cheaper launches, space is accessible.
Power seems easily solved. You can get lots of free energy from the sun with some modest panels. But to do that requires an odd orbit where you wouldn't be over the same spot on earth, which could make internet access difficult. Or you can go geostationary over a powerful ground station, but then you'd need some really big batteries for all the time you aren't in the sun.
But cooling is a huge problem. Space is cold, but there is no medium to transfer the heat away from the hot objects. I think this will be the biggest sticking point, unless they came up with an innovative solution.
> As conduction and convection to the environment are not available in space, this means the data center will require radiators capable of radiatively dissipating gigawatts of thermal load. To achieve this, Starcloud is developing a lightweight deployable radiator design with a very large area - by far the largest radiators deployed in space - radiating primarily towards deep space...
They claim they can radiate "633.08 W / m^2". At that rate, they're looking at square kilometers of radiators to dissipate gigawatts of thermal load, perhaps hectares of radiators.
They also claim that they can "dramatically increase" heat dissipation with heat pumps.
So, there you have it: "all you have to do" is deploy a few hectares of radiators in space, combined with heat pumps that can dissipate gigawatts of thermal load with no maintenance at all over a lifetime of decades.
This seems like the sort of "not technically impossible" problem that can attract a large amount of VC funding, as VCs buy lottery tickets that the problem can be solved.
> They also claim that they can "dramatically increase" heat dissipation with heat pumps.
Right, great idea. Start with the heat where you don't want it -- in the chip -- and pump it out to where it can't go anywhere. Then you can recirculate the medium back and have slightly older heat that you can mix with the new heat! It'll be a heat party!
It's just like a terrestrial heat pump, where you pump the heat out to where you have a huge environmental sink to transfer the heat to. In space, you have something like a hundred thousand hydrogen atoms per cubic meter to take up the heat. A HUNDRED THOUSAND! That's a bigly number, it must work out. We can always make those atoms go really, really fast!
Did an AI invent this whole scheme?
The whitepaper shows a 4km x 4km solar array, which is 1600 hectares (3200 International Space Stations). Would assume the array they're proposing would be cheaper since its structurally more homogenous, but $480 trillion dollars is a whole lot of money.
Starcloud’s whitepaper suggests a 4 km × 4 km radiator. For comparison, the James Web Space Telescope has a sunshield measuring 21 m × 14 m and the International Space Station measures 109 m × 73 m.
Doubling the radiator temperature would give you 16x more radiated power.
> they're looking at square kilometers of radiators to dissipate gigawatts of thermal load
I mean this is a ridiculous concept. We've never put anything remotely that size into space. To argue that this would be cheaper than putting something underwater or in the middle of nowhere is crazy. I'd rather deal with salt than deal with radiation.
FWIW there's a reason that Sweden has a bunch of datacenters in the north that are peanuts compared to hosting in Virginia.
They're "poorly" connected (by virtue of being a bit out of the way), but the free cooling and power from renewables make them extremely attractive. There was a time where they were the favourite of crypto-miners for the same reason as they would be attractive to AI training farms.
Fortlax has some I believe; https://www.fortlax.se
-----
As for the meat of the paper. Anyone with a passing understanding of space will be quick to point out that:
A) Heat is a problem in space, it's either way-way-way to hot (IE; you're in the path of the Sun) or it's way-way-way too cold (IE; you're out of the sun) and the shift between the two means you need to build for both. You also can't dissipate heat as there's no air to take the heat away.
B) Power is not so abundant and solar panels degrade; a huge amount of satellite building is essentially managing a decline in the capability of hardware. That's part of why there are so many up there.
C) Getting reasonably sized hardware up there is beyond improbable, though I'll grant you that most of the weight in a computer is the cooling components and chassis.
D) Cosmic Rays. No electromagnetic barrier from earth and extremely tight lithographies. I mean... there's a reason NASA is still using CPU's measured in the megahertz range.
AFAIK someone (Mars Ingenuity helicopter team) discovered that some chips handle them much better than others, so they just test a bunch and keep resistant ones.
Their main tech breakthrough would have to be in this area otherwise the company is worthless imo.
This is a super cool idea and seems like perfect investor-bait. That's about where it ends.
Geostationary satellites only go into Earth's shadow on around 20 days on each side of an equinox. That leave 280+ days each year where they are in sun all day. Maybe that's enough to be worth it?
Or if you do need to keep the things working even on those ~80 days a year when they do spend part of the day in shadow maybe they could be powered by energy beamed in from those not in shadow? You'd put a bunch in geostationary orbits spread out evenly so that each is close enough to its neighbors for power beaming.
I wonder if something crazy might work? Could you actually connect adjacent satellites by an actual physical power cable, which would also be in geostationary orbit?
I'd guess you'd actually need two conductors in your cable, carrying current in opposite direction to cancel out interactions with Earth's magnetic field so the system doesn't get pushed out of its orbit (which would probably be bad).
There would probably be gravitational interactions like with the Moon that might also make it hard to keep everything in place, but maybe by purposefully sending different currents in opposite directions on some of the links you could purposefully use interactions with the Earth's magnetic field to move the cable back where you wanted?
If the satellites are connected by cables then maybe they could actually be slightly higher than geostationary but moving faster than circular orbital speed at that altitude so there is a net outward force from that, which could be countered by tension in the power cables to force them into a circular path that is still geostationary.
Surely you'd want to use satellite constellations as relays? There's thousands of those satellites in line-of-sight all the time.
It's strictly superior on pure geometry anyway (I think). You have a finite channel capacity between your satellite and your ground station; but different satellites, in non-overlapping microwave spots, are in separate spatial channels.
Yes, let's go ahead and finish melting the ice caps, great idea
Routing through starlink should have direct LoS at all times.
(probably not)
There is no easy passive cooling in space, getting rid of heat is a major problem. And you need more redundancy because the radiation will crash your computers. And launch is very expensive of course.
And the whole presentation is completely ludicrous. Look at table 1 in the linked PDF and tell me you’re serious. There is no additional cost when sending a datacenter to space except launch cost and shielding? Building a server farm on earth is the same price as building a satellite you can launch on a rocket as long as you use the same computers?
Can I take the other side of this investment? Like an angel funding round, but selling short?
For all the financial inefficiencies of that - objectively - we get to benefit as humanity from the (expensive) mistakes of others.
So here's to whatever this is leading to much better insights about computing in space at best!
I really wish there was a secondary for private equity.
Venture Capitalists are already like reverse insurance companies. They cover lots of people in the hopes that one of them will hit a rare event and it'll pay for the others.
Buying shares of a single startup is sort of the equivalent of betting on a specific horse to win a race. But what's the equivalent of lay betting (betting that a specific horse will NOT win)? Shorting? But you can't short a private company.
But wait, venture capitalists are already betting that their startups will make money. What if they were willing to double down a bit and accept lay bets? Say there was a kind of specialized short agreement that let you say "here is $x, if in N years company Y has less than $z profit/revenue, I get K*$x. Otherwise, VC gets to keep my $x." You could sell it to VCs as a way to do options trading on their own startup investments, plus it'd be a good way to get the wisdom of the crowds or whatever.
Nonetheless getting rid of heat (by radiation) is possible, otherwise people would be roasted inside the ISS.
I'm sure all of these companies are advertising "ChatGPT in Space!" because that's what will generate hype and money, but what they'll actually be planning is very small edge data centers whose job is to reduce latency.
Whether that makes financial sense, I have no idea. But I am sure it's at least physically possible for a small enough data center.
That’s like saying „if you’re thirsty on a ship, getting thrown into the sea is actually really nice because you will be around a lot of water.“.
Physically, you could do it, but it won’t be simpler or cheaper than on earth. Except for constant solar availability, there are only downsides with this.
Reduce latency to where?
Given Y Combinator's vetting process, I'm sure they would have tackled this problem somehow - maybe by feeding the heat into another process? It will be interesting to see how they've solved this.
The vetting process of the fund that quite famously invests in the founders over the idea?
But I agree with your general point. At 100°C, you can radiate about 1kW/m^2. That’s 1000m^2 of radiator per MW of datacenter, assuming you can operate with the radiator at 100°C. You can fudge this a bit with a heat pump (to run the radiator hotter, paying a linear-ish power penalty and gaining a fourth-power radiation benefit), but that’s expensive and that power isn’t free.
Here on Earth, you can cool by conduction or evaporation, which isn’t an option in space.
> Space (with a sunshade) is a nearly perfect medium into which to radiate heat, in the sense that there’s nothing better.
- You can't build 40MW of solar panels for $2M, even with theoretical maximum efficiency. You can't even build the cabling and regulators at that price.
- You need battery storage -- not as your backup -- but as primary source. It is going to cost more than $2M. Batteries are heavy. They are going to cost a lot to launch. This is not even solved on the ground yet.
- You need a heat transport medium to move heat into your massive radiator. Either you use water or you use air or you use heatpipes (metal). You have to pay for the cost and weight and launch expense. This is probably half the weight of the rack and I haven't bothered to do the math about how you transport heat into a 500 foot solar sail.
- Let's not even talk about how you need to colocate multiple other racks for compute and storage. There aren't any 1TBps orbital link technologies.
- Rad shielding? It doesn't work, but I'll let this slide; it seems like the least problematic part of the proposal.
- 15 year lifetime? GPUs are obsolete after 12 months.
I don't want to be the guy who shoots stuff down just for fun, but this doesn't even pass the sniff test. Maybe you can get 10x cheaper power and cooling in space. Still doesn't work.
I guess you need connections too, and maybe a previous exit.
This idea in particular doesn't make any sense... Currently. Maybe in a decade or so with better technology.
Although the prospect of polluting the stars itself with a bunch of computers generating AI slips... We paved Paradise to put up a parking lot
The biggest problem is software. The CUDA stack is not maintained forever and certainly less than 15 years.
Not arguing with your overall point - this company looks ridiculous.
So it skews the economics pretty harshly. I think OP is right - you need good batteries somehow.
How so? Is it not possible to position the satellite in an orbit that keeps it in perpetual sunlight?
Just that it tends to involve heavy AF materials like water
... couldn't you just merge both problems into a solution - your radiators ARE you power source
We should train AI in space [pdf] - https://news.ycombinator.com/item?id=41478241 - Sept 2024 (93 comments)
A bit more here:
Lumen Orbit - https://news.ycombinator.com/item?id=42790424 - Jan 2025 (2 comments)
VCs wanted to get into Lumen Orbit's $11M seed round - https://news.ycombinator.com/item?id=42518284 - Dec 2024 (2 comments)
It's a bit of testing the waters toward what proper URL aggregation might look like. By "URL aggregation" I mean grouping together a bunch of URLs for a story, rather than just one link per submission.
Also, so is heating.
When you're on the sun side, everything is too hot and it is hard to cool. You can do direct cooling, such as water cooling, but you have no radiator to dump the heat to...
When you're in the shadow of the sun you have the opposite problem. Things are way too cold. Cold enough normal electronics can fail.
For reference, the ISS can fluctuate between -250F and 250F.
I'm willing to bet that it is easier to deal with the issues of salt water than it is to deal with the heating and cooling issues combined with difficulty to manually access problems presented by space. Price per pound into orbit is still quite expensive...
It's fairly close, about 1.3 light seconds away. You wouldn't use it for anything realtime, but it would be fine for long AI training jobs.
You could bury the servers underground to shield them from cosmic rays. That would also be good for any people living there.
You could get power from solar panels on peaks near the poles that get light almost all the time. For example, some ridges around Shackleton Crater are sunlit up to ~90% of the time, with short periods of darkness. Use batteries to smooth out the power supply.
For heating and cooling, just use the standard techniques. It's not easy, but it's a solved problem. As a bonus, near the poles, the temperature extremes aren't as bad as at the equator.
You could also sell tickets to tourists. People will pay to see the darndest things.
Radiative cooling works by exploiting the fact that hot objects emit electromagnetic radiation (glow), and hot means everything above absolute zero. The glow carries away energy which cools down the object. One complication is that each glowy object is also going to be absorbing glow from other objects. While the sun, earth, and moon all emit large amounts of glow (again, heat radiation), empty space is around 2.7 Kelvin, which is very cold and has little glow. So the radiative coolers typically need to have line of sight to empty space, which allows them to emit more energy than they absorb.
> Additionally, deep space is cold, which is accurate in that the "effective" ambient temperature is around -270°C, corresponding to the temperature of the cosmic microwave background.
> he mass of radiation shielding scales linearly with the container surface area, whereas the compute per container scales with the volume
> Therefore the mass of shielding needed per compute unit decreases linearly with container size.
m = ρV.
Let's simplify and assume we're using a sphere since this is the most efficient, giving V = 4/3r^3. Your shield is going to be approximately constant density since you need to shield from all directions (can optimize by using other things in your system).
m ∝ ρr^3
I'm not sure what here is decreasing nor what is a linear relationship. To adjust this to a shell you just need to consider the thickness so you can do Δr = r_outer - r_inner and that doesn't take away the cubic relationship.
https://en.wikipedia.org/wiki/Thermal_radiation#Characterist...
https://en.wikipedia.org/wiki/Black-body_radiation
https://www.nasa.gov/smallsat-institute/sst-soa/thermal-cont...
https://ocw.mit.edu/courses/16-851-satellite-engineering-fal...
I wonder how much cooling the solar panels alone would need, when operating at that scale.
Thats tricky. I know the heat exchange components are called radiators but most of the heat they give off is by convection not radiation. (At least here on the ground.) I heard 80%-20% rule of thumb.
But you are right in the broad strokes. Cooling is not easier in space. Mostly because you have no convective heat transfer.
I can help explain why. On earth, we are surrounded by stuff. Radiative cooling relies on thermal radiation leaving an object. Crucially, it also requires the object to absorb less thermal radiation than it emits. On earth we are surrounded by stuff, including air, that emits thermal radiation. There is a window of wavelengths, called the atmospheric window[0], that will allow parts of the thermal radiation out into space, rather than returned back. Imagine shining a flashlight on tinted glass, the light will get through depending on the color. If the light gets through, it has escaped. If not, the light is returned and heats up your surroundings again.
Also on earth the other methods (conduction, convection, and phase changes) are more effective. The earth can be used as a very big heat sink. On a spaceship or satellite, you don't have the extra mass to store the energy, so radiative is the only option.
https://en.wikipedia.org/wiki/Spacecraft_thermal_control?wpr...
> Most spacecraft radiators reject between 100 and 350 W of internally generated electronics waste heat per square meter.
It’s is on earth as well using solar and batteries. What is likely to get cheaper faster? Solar and batteries? Or lifting datacenters to space? The world is almost at the point of deploying 1TW/year of solar, and batteries are catching up. No space required. There aren't a lot of VC investment opportunities speeding the rate of solar and battery deployments though.
I'm not one of those idiots who would claim that "we should focus on terrestrial problems instead of space," but this idea seems to have only downsides.
It's kind of depressing that the only way to make this tech better is to feed it more energy. (And apparently now to send it to space)
Temporarily putting aside (extremely fair) feasibility questions around those two pre-requisites, data centers are a not-bad choice for things to do with unlimited space energy.
Aluminum smelting or growing food are the two I’d think of otherwise, and neither of those can have inputs/outputs beamed to a global network of high-bandwidth satellites.
To me, the cost estimates seem a bit off and conflate capital with running costs.
The main benefit for space at the moment seems to be sidestepping terrestrial regulations.
I think at the core of this there's a risk analysis. At one point I briefly worked in a team in charge of a company's servers, and there were plenty of stories of things gone wrong enough that someone had to drive or fly to the datacenter. These company's datacenters were named after the closest airport for this reason, iirc. A little optimization in case things went very wrong; you always knew where you'd have to fly in to.
Even if an undersea data center could potentially yield cost benefits, it's also significantly riskier in case something goes wrong. How long would it take to physically access a machine? Do you have to bring down other machines to access it? And at scale, things tend to always go wrong.
To comment on the original post, needless to say this is even more complicated, costly and untimely in space.
Just think about the sheer effort required to dump 1 BILLION watts of waste heat into space - the engineering challenges alone make this completely impractical.
Compared to this, Theranos actually looks like a solid investment. At least Holmes had working demos and big-name backers before it all fell apart. This doesn't even pass the basic smell test.
One NVIDIA DGX SuperPOD consumes 10 kW which would be ~500 square feet of solar panels and ~100 square feet of radiator area.
> Their design calls for a cluster of shipping container-style boxes packed with high-speed AI chips. These would be anchored at the centre of a 16 sq km array of solar panels generating up to five gigawatts of power — about 25 per cent more than Drax, Britain’s biggest power station. The mammoth structure would circle the Earth in “sun synchronous” orbit so that it is never in shade
They are literally planning to feed the radiators using a coolant like water and sensible heat at 35 degC to 5 degC. At 5 GW, you then need to be pumping 60 000 liters of water per second.
That's like a tenth of the Sacramento river, going through a 16 sq km array in space and hoping that nothing leaks.
This plan seems about as realistic as Bluthton though. https://www.reddit.com/r/arresteddevelopment/comments/1gtyvv...
The big radiators on the ISS can only dump a few server racks worth of heat.
A CDN for Starlink customers is probably the first use case for servers in space, not training GPT6, which will be a big enough project on familiar territory.
I can't imagine running bleeding edge GPUs in a particle accelerator and getting reasonable results.
To be fair that's mostly part "if it works don't change it" and part "that's how we've always done it". SpX uses newer hardware w/ traditional OSs (linux) w/ lots of redundancy.
Aside from the technical concerns already raised in other comments, I'm also not sure we really want all this private for-profit usage of earth's orbit. The orbital environment is already somewhat congested and people have already been raising concerns about it. There is the potential for it all to spectacularly blow up in our faces and become so polluted that we won't be able to do many launches at all.
it's going to take the management of our shared resources and spaces (orbit) for instance to leave earth, and this becomes especially important as Kessler syndrome risk rises with increasing debris in orbit
private companies launching without public oversight and controls are a recipe for cluttering earth's orbit and leaving us earth-bound for far into the future (same if the public sector launches without care but that seems less likely imo)
Kessler Syndrome: https://en.wikipedia.org/wiki/Kessler_syndrome
Be the start of a geoengineering project that is actively collecting energy and radiating it out into space while performing useful work along the way. Infrared radiation between 8 and 13 micrometers isn't absorbed by the atmosphere and exits the planet.
So, why not build big infra here and if you can build powerful radiators, aim them up and away from the earth, which needs to be cooled a bit anyways!
Starcloud is developing a lightweight deployable radiator design with a very large area - by far the largest radiators deployed in space - radiating primarily towards deep space, which has an average temperature of about 2.7 Kelvin or -270°C. The radiators can be positioned in-line with the solar arrays as shown in Figure 3, with one side exposed to sunlight.
…
Figure 3. A data center in Sun Synchronous Orbit, showing a 4km x 4km deployed solar array and radiators.
1. provide internet. 2. provide CDN. 3. Edge Compute. 4. Full-on cloud.
These guys see to be focussing on what is basically offline processing (AI training).
Datacenters in space makes no sense at all. Even ignoring the huge cost of sending hardware there in the first place, cooling is a massive issue in space. No medium to sink heat into means the only way to cool anything is by running water through giant infrared radiators. Not ideal when cooling is the largest bottleneck in scaling datacenters. Note that they would also have to dissipate the large amounts heat their datacenter satellite gets from being exposed to the Sun.
Also disregard the cost it takes to send a technician for maintenance, of updating hardware, etc.
This won't happen. If a satellite fails they will just write it off. Maintenance would be more expensive than depreciation
I will call it MartianCloud.
>Presumably the real purpose of this company is to refute people who said "We'll just walk over to the superintelligence and pull the plug out", without MIRI needing to argue with them
https://xcancel.com/ESYudkowsky/status/1831899029690839549#m
- radhard sdks aren't offered at sota node sizes.
- they are much much more expensive
If they can partner with a chipmaker (AMD :^)) and use that as a way to collect some meaningful data, that would be useful at the very least.
huh? I was under the impression that cooling in space is an absolute nightmare since radiating heat into vacuum is super hard?
Even the comparatively small and decidedly H100-free ISS needed giant radiators
https://en.wikipedia.org/wiki/External_Active_Thermal_Contro...
That is pretty amazing to say with a straight face. Whatever GPT6 may or may not ever be…
The work is mysterious and important. Praise Kier.
Space? I really don't get it.
Seriously! There is just so much wrong and some of is trivial.
> radiating primarily towards deep space, which has an average temperature of about 2.7 Kelvin or -270°C.
In these locations you have to deal with cooling AND heating. On the moon you swing from -130C (LRO got down to -250C) on the dark side and 121C on the light side. The ISS swings from -160C to 120C. These are too cold for most electronics. Not to mention that these temperature swings create a lot of physical stress on parts, and we're talking about putting up up some of the smallest objects we commercially make? They will rip right off the circuit-board if you don't get it right.
Not to mention that radiating into space is quite difficult. There's a reason we use convection ovens and why we put fans in our computers. It isn't about the temperature of the atmosphere nor the thermal efficiency, it is because convection is just a hell of a lot more efficient. Thermal radiation is like shedding your heat via a lightbulb.
Their claim here is that they can radiate 633W/m2. For supercomputers we're talking on the order of 10s of MW of waste heat. That's 10^7! These are going to be BY FAR the largest radiators in space and going to cost tons of money for the mass alone.
Not to mention the size of the solar panels they'll need... But at least they mention this one: "A 5 GW data center would require a solar array with dimensions of approximately 4 km by 4 km," These are GIGANTIC structures and far larger than anything we've put into space.
> The mass of radiation shielding scales linearly with the container surface area, whereas the compute per container scales with the volume. Therefore the mass of shielding needed per compute unit decreases linearly with container size.
Density (ρ) is mass (m) divided by volume (V): m = ρV. We'll assume a sphere due to its efficient surface area. You use Δr as the shell's thickness: V = 4/3(Δr)^3
Let: m = ρV
Let: V = 4/3(Δr)^3
∴ m ∝ ρ(Δr)^3
> This effect, combined with the shielding afforded by the cooling blocks, means that radiation shielding is proportionally a much smaller concern compared to electronics on typical satellites today.
But also you have to remember that you can't shield your solar panels. To do so would prevent light from reaching them. That leads to a weird constraint here and I would not expect these machines to be meaningfully long lived. The alternative is you could go repair them, but that's expensive too.
-------
------
https://ocw.mit.edu/courses/16-851-satellite-engineering-fal...
https://www.jpl.nasa.gov/nmp/st8/tech/eaftc_tech1.php
https://www.nature.com/articles/s41597-024-03913-w
[0] I know this because I've research for NASA on radiation shields. I got multiple SBIR and STTR grants for this work. Material choices still do matter but the right material is proportional to the radiation level. But the higher the energy level, the less atomic properties matter and the more density does. You can get benefits from the electromagnetic properties of protons and electrons (beta-), but these don't help you with neutrons. That is, until after you slow these things down, which is why there is typically layering.
"Please don't fulminate."
"Don't be curmudgeonly. Thoughtful criticism is fine, but please don't be rigidly or generically negative."
/s (kinda but maybe not really...)
"There are also potential opportunities to leverage the fact that the speed of light in a vacuum is 35% faster than in a typical glass fiber."
This is just wrong. The speed of light in a fiber is ~2/3 slower than in a vacuum. That means that the speed of light in a vacuum is 50% faster than in a fiber, not 35% (~1/3) faster!
Not even English.
what the actual fuck? my boys joseph and ludwig would like to have a word with y'all
in ideal conditions, your gpu putting out 600W will need about a square meter facing deep space to keep it at 80c, this idea is absurd on first principles alone, maybe if you have heat pumps you can push this but then you're dealing with on orbit fluid loops that you can't maintain, as i said, what the fuck?
https://www.youtube.com/watch?v=EsUBRd1O2dU
TL;DR... cooling in space isn't passive, you're on the "inside" of an enormous vacuum flask. And radiative coupling with space is possible from the ground, if that's what you're interested in:
https://www.skycoolsystems.com
But god bless crazy entrepreneurs. Don't ask whether we can, definitely don't ask if we should, just ask whether it makes for good headlines...
I hate it.
Let’s definitely downplay the physics _and_ pollute orbit in service of what will no doubt be producing yet more AI slop. So good.
Is this not a major problem with YC, specifically? Our beloved orange site funded and accelerated these guys.