On top of the other very reasonable concerns people have mentioned, I haven't seen anything convincing yet about how anyone plans to cool this kind of high power density installation in a vacuum....
Exemplary
On top of the other very reasonable concerns people have mentioned, I haven't seen anything convincing yet about how anyone plans to cool this kind of high power density installation in a vacuum. You're essentially running the whole thing inside an incredibly effective thermos, while simultaneously pumping huge amounts of additional solar energy into the system to actually power the hardware.
If they're aiming for 5GW with 4km2 of cooling panels (it says "solar and cooling panels that measure roughly 4 kilometers in both width and height", so let's give them the benefit of the doubt and say they mean a full, separate 4km2 each for solar collection and for cooling), that's 312W/m2 to dissipate. That's better than double the efficiency of the ISS cooling system, and more than 35,000x the size.
I'm not saying it's impossible, but extraordinary claims require extraordinary evidence. Don't get me wrong, I absolutely love space tech and if billionaire investors want to pour money into that I'm a lot happier about it than a lot of other things they could be doing! I just don't remotely believe they're going to manage scaling up by more than four orders of magnitude, while achieving more than twice the efficiency of existing systems, all while keeping both the system cost and the launch cost low enough that a one order of magnitude saving on power makes the whole thing financially worthwhile.
But hey, maybe we get some cool and applicable research out of the project before it bankrupts itself! It's like the old system of scientific patronage, except via hubris rather than altruism.
[Edit] I found a whitepaper of theirs from last year: https://starcloudinc.github.io/wp.pdf - they're actually basing their estimates around radiating 633 W/m2, so around 17,000x the size of the ISS system at more than four times the efficiency. They do acknowledge that "To achieve this, Starcloud is developing a lightweight deployable radiator design with a very large area - by far the largest radiators deployed in space". They don't touch on how.
[Edit 2] I was trying to put it in perspective for myself so I looked at a few more numbers: the comparison I landed on is that an oil tanker is about 35,000x the size of a Toyota Prius. That's the amount they're going to need to scale up existing tech to make this work at all [give or take a factor of two], and the amount they're going to need to simultaneously scale down weight and/or launch costs to make it financially viable. All while running at 400 - 500% the efficiency of current systems.
[Edit 3] Fixed a missing factor of two in my scaling for the system size. It doesn't change my skepticism!
Thank you for this. When I first heard of this idea, it seemed absolutely baffling for anyone to want to put a data center in orbit, when so much of data center optimization comes from heat...
Thank you for this. When I first heard of this idea, it seemed absolutely baffling for anyone to want to put a data center in orbit, when so much of data center optimization comes from heat exchanger systems!
It always makes me a little mentally upset when the biggest limiting factor to something is that you can't cool it effectively. Just knowing all that energy is not only going to waste, but...
It always makes me a little mentally upset when the biggest limiting factor to something is that you can't cool it effectively. Just knowing all that energy is not only going to waste, but actively requires even more resources in order to deal with it.
My educated brain understands why, but my lizard brain keeps thinking there has to be a solution that doesn't produce "waste" heat
I very much know what you mean, although I will say I'm a lot more comfortable when it's a (hypothetical) solar installation like this would be - the kids' picture book part of my brain thinks of...
I very much know what you mean, although I will say I'm a lot more comfortable when it's a (hypothetical) solar installation like this would be - the kids' picture book part of my brain thinks of it like sending the "used up" sunlight out the other side again, once you've had it do a bit of work for you, like a stream flowing on past a waterwheel.
Can't say the same for the rocket fuel it'd take to get the thing up there, though... Maybe we'll see some nice self contained installations out in the desert if they sustain funding long enough to build a proof of concept that gets half way to what the orbital one would need to do?
From the article: … … … This is just a tech demo (or a stunt if you’re uncharitable) and I’m skeptical that it will compete with regular data centers on costs any time soon. But they do seem...
From the article:
Last month, the Washington-based company launched a satellite with an Nvidia H100 graphics processing unit, sending a chip into outer space that’s 100 times more powerful than any GPU compute that has been in space before. Starcloud was able to train and run NanoGPT, a large language model created by OpenAI founding member Andrej Karpathy, on the H100 chip in orbit using the complete works of Shakespeare.
…
The company’s Starcloud-1 satellite is also now running and querying responses from Gemma, an open large language model from Google based on the company’s Gemini models, in orbit, marking the first time in history that an LLM has been run on a high-powered Nvidia GPU in outer space, CNBC has learned.
…
Starcloud — a member of the Nvidia Inception program and graduate from Y Combinator and the Google for Startups Cloud AI Accelerator — plans to build a 5-gigawatt orbital data center with solar and cooling panels that measure roughly 4 kilometers in both width and height. A compute cluster of that gigawatt size would produce more power than the largest power plant in the U.S. and would be substantially smaller and cheaper than a terrestrial solar farm of the same capacity, according to Starcloud’s white paper.
…
Along with Starcloud and Nvidia’s efforts, several companies have announced space-based data center missions. On Nov. 4, Google unveiled a “moonshot” initiative titled Project Suncatcher, which aims to put solar-powered satellites into space with Google’s tensor processing units. Privately owned Lonestar Data Holdings is working to put the first-ever commercial lunar data center on the moon’s surface. Aetherflux, founded by former Robinhood co-founder and chief executive Baiju Bhatt, on Tuesday announced a target to deploy an orbital data center satellite in the first quarter of 2027.
This is just a tech demo (or a stunt if you’re uncharitable) and I’m skeptical that it will compete with regular data centers on costs any time soon. But they do seem rather committed to the bit.
I can't claim to have thought of this myself but one possible reason people are really interested in having data centers in space could be the lack of government jurisdiction. I struggle to think...
I can't claim to have thought of this myself but one possible reason people are really interested in having data centers in space could be the lack of government jurisdiction.
I struggle to think that any of the other benefits outweigh the risk and cost of launching all that expensive, fragile, and quickly outdated tech into space and dealing with the challenges of cooling in a vacuum.
I don't think that's the most logical answer. Why not just put your data center in Switzerland or one of the other infamous tax/money havens, you know? There's just zero chance that spending...
I don't think that's the most logical answer. Why not just put your data center in Switzerland or one of the other infamous tax/money havens, you know? There's just zero chance that spending billions building a moonshot project in space to hide from the government is more cost effective than spending a million on lawyers and bribes. (Also, the company is still governed by the rules of wherever they're incorporated, so this wouldn't even work. It's not like the US government is going to be like "rats, I guess we can't apply court orders to Google anymore," you know?)
The only potential reason I could think of, and this is a reaaaaaal stretch, would be for image processing. If you put the AI compute into space, you could mount hundreds of extremely high...
The only potential reason I could think of, and this is a reaaaaaal stretch, would be for image processing. If you put the AI compute into space, you could mount hundreds of extremely high resolution cameras on the satellite and run the image processing in the satellite itself, where streaming that data back to earth with processing would take up a ridiculous amount of bandwidth.
It's probably still way more cost effective to use a powerful directional microwave array or laser transmission systems and brute force the bandwidth with WDM, but maybe there's a tipping point where that becomes unfeasible. The two big applications I could think of would be surveillance (think, running an AI models to scan terabytes of images taken over an entire country for an instance of a license plate using temporal stitching methods to create even higher resolution images), or weather and climate science (scanning extremely high resolution images to track climate change and early indicators of weather).
In the former instance, if you want to use that data for targeting, latency matters a lot. When you're guiding a missile moving at 4 times the speed of sound, 400 MS of latency because of data transfer is the difference between hitting a target and leveling a random building.
It's probably cheaper to do this stuff on the ground, but I don't have any insight into the real costs and technical limitations here.
I like the line of thinking, but yeah, I think your first assumption about it just not being worthwhile was probably still correct! Latency's definitely an interesting one, and something I hadn't...
I like the line of thinking, but yeah, I think your first assumption about it just not being worthwhile was probably still correct!
Latency's definitely an interesting one, and something I hadn't thought of as being potentially important for this kind of work; I can absolutely see it as a priority for military applications, but you also aren't going to need the full model training hardware up there for anything latency sensitive. I could imagine there being a smallish niche for a lightweight, radiation hardened server you can build into a satellite for onboard inference work (maybe something else viable that could actually come out of this project if they get that far!), but it's an edge case, and I can't think of a way it'd require much more than a few GPUs built into a few specific satellites.
For scientific work, you can comfortably do tens of Gbps to a single satellite if you're using all of its bandwidth - they normally serve a few thousand users each at any given time, but the current gen Starlink nodes are 96Gbps, and they're aiming for a terabit on the next generation (obviously apply a Musk-company bullshit filter to that second number, but I don't think it's wildly unrealistic given the advancements we're seeing in ground based networking at the moment).
The highest resolution image sensors I'm aware of - ultra-specialist 200-400 megapixel ones that'll dump multiple gigabytes for a single image - still have less than 50Gbps bandwidth from the sensor itself, so you'd need multiple instruments sending back continuous raw format video without compression to saturate the downlink. Allow for realistic lossless compression and you can likely support 15+ simultaneous streams from those kind of absurd sensors now, going up to 100+ if the next generation of comm systems really do manage a significant improvement.
There's definitely part of me that wants this to work purely because space datacenters in space are straight out of sci-fi, but as with the cooling systems I just can't make the numbers add up. And I will at least take the consolation prize that blasting terabits per second of realtime data down from space is still pretty damn cool!
Ok, but instead you have to [waves arms skyward] pay to launch them into freaking orbit. And it's not like you have long to recoup your costs before your data center needs upgraded hardware...
Starcloud CEO Philip Johnston told CNBC that the company’s orbital data centers will have 10 times lower energy costs than terrestrial data centers.
Ok, but instead you have to [waves arms skyward] pay to launch them into freaking orbit. And it's not like you have long to recoup your costs before your data center needs upgraded hardware launched into orbit, GPU tech iterations can be measured in months, not years.
I wonder if they will take a page out of military hardware sales and allow out-dated assets to be sold to allies and partners who otherwise can't buy or built the technology themselves. 4th...
And it's not like you have long to recoup your costs before your data center needs upgraded hardware ...
I wonder if they will take a page out of military hardware sales and allow out-dated assets to be sold to allies and partners who otherwise can't buy or built the technology themselves. 4th generation aircraft can be sold to pretty much anyone, depending on the avionics, while 5th generation (and 4.5-gen) is reserved for partner countries with a special status so that the manufacturing countries are in control of the other country's capabilities.
For the demonstration unit, possibly, but I don't know how big that unit is. The article is saying the theoretical datacenter-replacement satellite they want to build would sit out from Earth far...
For the demonstration unit, possibly, but I don't know how big that unit is. The article is saying the theoretical datacenter-replacement satellite they want to build would sit out from Earth far enough that there won't be much if any of a shadow cast on its solar panels -- so probably either at a distant polar orbit or reeeeally far out.
Starlink, Iridium, and other "communications" satellites sit at very, very low orbits. Like low enough that fighter aircraft can shoot them down with reasonably conventional rockets (i.e. not a special platform with a custom-built rocket). In fact, Starfighter Space, Inc. just had their IPO -- their plan is to use Vietnam-era fighter jets as the launch platform to put small satellites into space with under-wing rockets. Anyhow, most of these low Earth orbit satellites sit at 300-550 km (up to around 333 mi) or a little higher. Hubble Space Telescope and the ISS only orbit at the upper end of that because we need to be able to reach them routinely via human-manned spacecraft at minimal cost, but in a practical sense 550 km from the surface of the Earth is walking distance compared to, say, reaching the Moon (hiking distance) or the Kuiper Belt (you'll want to make reservations in advance). There is a really cool illustration of man-made satellite orbits on Wikipedia. The ISS is hardly even in "space" as us laypeople think of it (and Katy Perry is still not an astronaut no matter how politicians change the definition).
Putting up the multi-ton satellite that Nvidia and their partners are imagining as a final product (something capable of pulling in 5-gigawatts via solar panel arrays) is going to be the size of a bus, at least, and once the solar array is spread will be absolutely massive. Deorbiting a monstrosity like this, full of heavy heatsinks (because even in "space" they need to dissipate that heat effectively), will mean pulling it down into the ocean in a safe way, the way they typically deorbit large satellites. But maybe by the time we're training models in space (this sentence would've looked weird in the 90s!) we'll actually employ the more environmentally-friendly methods of deorbiting and getting rid of space debris that people have already thought of.
It seems like they might use a Molniya orbit. Edit because I accidentally hit save: This might be good for training data, I suppose, as you could uplink at perihelion and then let it plug and chug...
Edit because I accidentally hit save: This might be good for training data, I suppose, as you could uplink at perihelion and then let it plug and chug for 12 hours.
Also, does Starfighter Space plan to compete with Stratolaunch, because it seems like Stratolaunch would have a much easier time of putting things in orbit, as that's what it was designed to do.
This is so so stupid. There is 0 way this could ever be economically feasible. It'd be easier to put these in the middle of the Sahara or Antarctica vs putting them in orbit. Various problems that...
This is so so stupid. There is 0 way this could ever be economically feasible. It'd be easier to put these in the middle of the Sahara or Antarctica vs putting them in orbit.
Various problems that need to be solved include, but are not limited to:
Getting it up there safely
Shielding it from radiation
Communicating with the ground on a delay.
Handling orbit decay
Never being able to repair/replace anything
Cooling it.
Powering it.
Each one of those problems ranges from "solved at orders of magnitude more expense to a earth based solution" to "unsolved, still going to be expensive, and requires leaps forward in technology that would probably change the way we handle everything in life anyways".
1 and 2 are solved by throwing startup money and government grants into it. 3 might not be as important depending on the application. 4 and 7 are not that much of a problem if it sits in a...
1 and 2 are solved by throwing startup money and government grants into it.
3 might not be as important depending on the application.
4 and 7 are not that much of a problem if it sits in a Lagrange point closer to the sun.
5 They don't care, making a new one might even be more profitable.
6 Just blow into it like soup, bro, just one trillion more and we'll get AGI, bro, we'll science our way out of all problems with it, bro, what's a few extra warming between bros, bro
It's the same kind of thing as asteroid mining. Neat PR sales pitches to clients drunk on scifi. The logistics and costs to even get to one are astronomical let alone setting up a supply chain.
It's the same kind of thing as asteroid mining. Neat PR sales pitches to clients drunk on scifi.
The logistics and costs to even get to one are astronomical let alone setting up a supply chain.
For anyone else who was (like me) wondering "Why...?":
For anyone else who was (like me) wondering "Why...?":
Starcloud wants to show outer space can be a hospitable environment for data centers, particularly as Earth-based facilities strain power grids, consume billions of gallons of water annually and produce hefty greenhouse gas emissions. The electricity consumption of data centers is projected to more than double by 2030, according to data from the International Energy Agency.
Starcloud CEO Philip Johnston told CNBC that the company’s orbital data centers will have 10 times lower energy costs than terrestrial data centers.
“Anything you can do in a terrestrial data center, I’m expecting to be able to be done in space. And the reason we would do it is purely because of the constraints we’re facing on energy terrestrially,” Johnston said in an interview.
Only problem is you can’t fix anything that goes wrong. So you will increase costs quite a bit for redundancy and extra engineering (not to mention launch costs which aren’t cheap). What would a...
Only problem is you can’t fix anything that goes wrong. So you will increase costs quite a bit for redundancy and extra engineering (not to mention launch costs which aren’t cheap). What would a data center look like if every failure that requires hands on the hardware could never be addressed? Every year more and more goes offline forever.
I do think it’s a cool idea. But we’re missing the infrastructure necessary to support it.
I think the next test is probably 'how long can it stay up there and functional' and there will be another article like that, and we'll all discuss the problems they still face, and, eventually,...
I think the next test is probably 'how long can it stay up there and functional' and there will be another article like that, and we'll all discuss the problems they still face, and, eventually, there won't be any more problems and data centers in space will just be normal.
I feel like I just went through this with a lot of things that SpaceX did.
Difference is, it's easily foreseeable that you can refine manufacturing processes so that electric cars would eventually compete with petrol on price point. But blasting infrastructure into...
Difference is, it's easily foreseeable that you can refine manufacturing processes so that electric cars would eventually compete with petrol on price point. But blasting infrastructure into space, along with two or three backups of everything, is never going to be cheaper than shipping one of everything in a truck. Not every new moonshot technology is the next Galileo.
Follow-up question: I understand many of the benefits of running computationally intensive (and heat-intensive) tasks in space (and demo'ing the tech now). But what kind of bandwidth is the final...
Follow-up question:
I understand many of the benefits of running computationally intensive (and heat-intensive) tasks in space (and demo'ing the tech now). But what kind of bandwidth is the final product going to need and do we currently have the price-performance to make that amount of bandwidth even happen in the next couple of years? I know that satellite bandwidth has gotten a lot cheaper and even cellular phone carriers are offering services (along with things like Garmin's inReach satellite texting communication that's been in consumers' hands foe 10+ years), but how much cheaper? Datacenters have staggeringly huge pipes these days so I'm wondering how much bandwidth the data from an AI question-response session would take and how much of it they plan to have completed in orbit.
The article says they're planning not to have a geostationary orbit for the 5-gigawatt satellite so that it can gather solar energy without day/night restrictions, but they would have to get other satellites up just to carry the communication signals to whatever side(s) of the planet the data is being sent. So bandwidth really becomes a huge factor as more hops are added to the transmission. Are there companies looking at developing this tech with geosynchronous orbits with a dedicated ground receiving station? And how expensive is that real estate about to get? Would those companies be looking at more equatorial regions for the ground stations (or perhaps just Texas and Florida)?
I imagine they’re talking up training and not inference because the networking requirements are lower. They still need to upload the training data, though. Perhaps some of it would be preinstalled...
I imagine they’re talking up training and not inference because the networking requirements are lower. They still need to upload the training data, though. Perhaps some of it would be preinstalled before launch?
If they launched a GPU (and hardware to run it) into orbit then it's trivial to include a micro SD card with the works of Shakespeare that they used as training data.
If they launched a GPU (and hardware to run it) into orbit then it's trivial to include a micro SD card with the works of Shakespeare that they used as training data.
That gives a whole new meaning to "Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway. –Andrew Tanenbaum, 1981"!
Perhaps some of it would be preinstalled before launch?
That gives a whole new meaning to "Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway. –Andrew Tanenbaum, 1981"!
We're probably only a few years from 5G speeds from orbit. Consider that Starlink can reliably deliver ~150mbps from LEO, which is staggeringly fast, and companies like ASTS (and spaceX, again)...
We're probably only a few years from 5G speeds from orbit. Consider that Starlink can reliably deliver ~150mbps from LEO, which is staggeringly fast, and companies like ASTS (and spaceX, again) are working to stretch that to mobile devices.
I'm assuming you mean 5G speeds rather than literal 5G. It's physically impossible for standard 5G wavelengths to travel that far. It might be possible to improve some kind of signal beaming...
I'm assuming you mean 5G speeds rather than literal 5G. It's physically impossible for standard 5G wavelengths to travel that far. It might be possible to improve some kind of signal beaming technology, but that's different.
On top of the other very reasonable concerns people have mentioned, I haven't seen anything convincing yet about how anyone plans to cool this kind of high power density installation in a vacuum. You're essentially running the whole thing inside an incredibly effective thermos, while simultaneously pumping huge amounts of additional solar energy into the system to actually power the hardware.
If they're aiming for 5GW with 4km2 of cooling panels (it says "solar and cooling panels that measure roughly 4 kilometers in both width and height", so let's give them the benefit of the doubt and say they mean a full, separate 4km2 each for solar collection and for cooling), that's 312W/m2 to dissipate. That's better than double the efficiency of the ISS cooling system, and more than 35,000x the size.
I'm not saying it's impossible, but extraordinary claims require extraordinary evidence. Don't get me wrong, I absolutely love space tech and if billionaire investors want to pour money into that I'm a lot happier about it than a lot of other things they could be doing! I just don't remotely believe they're going to manage scaling up by more than four orders of magnitude, while achieving more than twice the efficiency of existing systems, all while keeping both the system cost and the launch cost low enough that a one order of magnitude saving on power makes the whole thing financially worthwhile.
But hey, maybe we get some cool and applicable research out of the project before it bankrupts itself! It's like the old system of scientific patronage, except via hubris rather than altruism.
[Edit] I found a whitepaper of theirs from last year: https://starcloudinc.github.io/wp.pdf - they're actually basing their estimates around radiating 633 W/m2, so around 17,000x the size of the ISS system at more than four times the efficiency. They do acknowledge that "To achieve this, Starcloud is developing a lightweight deployable radiator design with a very large area - by far the largest radiators deployed in space". They don't touch on how.
[Edit 2] I was trying to put it in perspective for myself so I looked at a few more numbers: the comparison I landed on is that an oil tanker is about 35,000x the size of a Toyota Prius. That's the amount they're going to need to scale up existing tech to make this work at all [give or take a factor of two], and the amount they're going to need to simultaneously scale down weight and/or launch costs to make it financially viable. All while running at 400 - 500% the efficiency of current systems.
[Edit 3] Fixed a missing factor of two in my scaling for the system size. It doesn't change my skepticism!
Thank you for this. When I first heard of this idea, it seemed absolutely baffling for anyone to want to put a data center in orbit, when so much of data center optimization comes from heat exchanger systems!
It always makes me a little mentally upset when the biggest limiting factor to something is that you can't cool it effectively. Just knowing all that energy is not only going to waste, but actively requires even more resources in order to deal with it.
My educated brain understands why, but my lizard brain keeps thinking there has to be a solution that doesn't produce "waste" heat
I very much know what you mean, although I will say I'm a lot more comfortable when it's a (hypothetical) solar installation like this would be - the kids' picture book part of my brain thinks of it like sending the "used up" sunlight out the other side again, once you've had it do a bit of work for you, like a stream flowing on past a waterwheel.
Can't say the same for the rocket fuel it'd take to get the thing up there, though... Maybe we'll see some nice self contained installations out in the desert if they sustain funding long enough to build a proof of concept that gets half way to what the orbital one would need to do?
From the article:
…
…
…
This is just a tech demo (or a stunt if you’re uncharitable) and I’m skeptical that it will compete with regular data centers on costs any time soon. But they do seem rather committed to the bit.
I can't claim to have thought of this myself but one possible reason people are really interested in having data centers in space could be the lack of government jurisdiction.
I struggle to think that any of the other benefits outweigh the risk and cost of launching all that expensive, fragile, and quickly outdated tech into space and dealing with the challenges of cooling in a vacuum.
I don't think that's the most logical answer. Why not just put your data center in Switzerland or one of the other infamous tax/money havens, you know? There's just zero chance that spending billions building a moonshot project in space to hide from the government is more cost effective than spending a million on lawyers and bribes. (Also, the company is still governed by the rules of wherever they're incorporated, so this wouldn't even work. It's not like the US government is going to be like "rats, I guess we can't apply court orders to Google anymore," you know?)
The only potential reason I could think of, and this is a reaaaaaal stretch, would be for image processing. If you put the AI compute into space, you could mount hundreds of extremely high resolution cameras on the satellite and run the image processing in the satellite itself, where streaming that data back to earth with processing would take up a ridiculous amount of bandwidth.
It's probably still way more cost effective to use a powerful directional microwave array or laser transmission systems and brute force the bandwidth with WDM, but maybe there's a tipping point where that becomes unfeasible. The two big applications I could think of would be surveillance (think, running an AI models to scan terabytes of images taken over an entire country for an instance of a license plate using temporal stitching methods to create even higher resolution images), or weather and climate science (scanning extremely high resolution images to track climate change and early indicators of weather).
In the former instance, if you want to use that data for targeting, latency matters a lot. When you're guiding a missile moving at 4 times the speed of sound, 400 MS of latency because of data transfer is the difference between hitting a target and leveling a random building.
It's probably cheaper to do this stuff on the ground, but I don't have any insight into the real costs and technical limitations here.
I like the line of thinking, but yeah, I think your first assumption about it just not being worthwhile was probably still correct!
Latency's definitely an interesting one, and something I hadn't thought of as being potentially important for this kind of work; I can absolutely see it as a priority for military applications, but you also aren't going to need the full model training hardware up there for anything latency sensitive. I could imagine there being a smallish niche for a lightweight, radiation hardened server you can build into a satellite for onboard inference work (maybe something else viable that could actually come out of this project if they get that far!), but it's an edge case, and I can't think of a way it'd require much more than a few GPUs built into a few specific satellites.
For scientific work, you can comfortably do tens of Gbps to a single satellite if you're using all of its bandwidth - they normally serve a few thousand users each at any given time, but the current gen Starlink nodes are 96Gbps, and they're aiming for a terabit on the next generation (obviously apply a Musk-company bullshit filter to that second number, but I don't think it's wildly unrealistic given the advancements we're seeing in ground based networking at the moment).
The highest resolution image sensors I'm aware of - ultra-specialist 200-400 megapixel ones that'll dump multiple gigabytes for a single image - still have less than 50Gbps bandwidth from the sensor itself, so you'd need multiple instruments sending back continuous raw format video without compression to saturate the downlink. Allow for realistic lossless compression and you can likely support 15+ simultaneous streams from those kind of absurd sensors now, going up to 100+ if the next generation of comm systems really do manage a significant improvement.
There's definitely part of me that wants this to work purely because space datacenters in space are straight out of sci-fi, but as with the cooling systems I just can't make the numbers add up. And I will at least take the consolation prize that blasting terabits per second of realtime data down from space is still pretty damn cool!
Ok, but instead you have to [waves arms skyward] pay to launch them into freaking orbit. And it's not like you have long to recoup your costs before your data center needs upgraded hardware launched into orbit, GPU tech iterations can be measured in months, not years.
I wonder if they will take a page out of military hardware sales and allow out-dated assets to be sold to allies and partners who otherwise can't buy or built the technology themselves. 4th generation aircraft can be sold to pretty much anyone, depending on the avionics, while 5th generation (and 4.5-gen) is reserved for partner countries with a special status so that the manufacturing countries are in control of the other country's capabilities.
Wouldn't these deorbit and burn up long before they could be excessed to anyone? Apparently, Starlink satellites are coming down daily.
For the demonstration unit, possibly, but I don't know how big that unit is. The article is saying the theoretical datacenter-replacement satellite they want to build would sit out from Earth far enough that there won't be much if any of a shadow cast on its solar panels -- so probably either at a distant polar orbit or reeeeally far out.
Starlink, Iridium, and other "communications" satellites sit at very, very low orbits. Like low enough that fighter aircraft can shoot them down with reasonably conventional rockets (i.e. not a special platform with a custom-built rocket). In fact, Starfighter Space, Inc. just had their IPO -- their plan is to use Vietnam-era fighter jets as the launch platform to put small satellites into space with under-wing rockets. Anyhow, most of these low Earth orbit satellites sit at 300-550 km (up to around 333 mi) or a little higher. Hubble Space Telescope and the ISS only orbit at the upper end of that because we need to be able to reach them routinely via human-manned spacecraft at minimal cost, but in a practical sense 550 km from the surface of the Earth is walking distance compared to, say, reaching the Moon (hiking distance) or the Kuiper Belt (you'll want to make reservations in advance). There is a really cool illustration of man-made satellite orbits on Wikipedia. The ISS is hardly even in "space" as us laypeople think of it (and Katy Perry is still not an astronaut no matter how politicians change the definition).
Putting up the multi-ton satellite that Nvidia and their partners are imagining as a final product (something capable of pulling in 5-gigawatts via solar panel arrays) is going to be the size of a bus, at least, and once the solar array is spread will be absolutely massive. Deorbiting a monstrosity like this, full of heavy heatsinks (because even in "space" they need to dissipate that heat effectively), will mean pulling it down into the ocean in a safe way, the way they typically deorbit large satellites. But maybe by the time we're training models in space (this sentence would've looked weird in the 90s!) we'll actually employ the more environmentally-friendly methods of deorbiting and getting rid of space debris that people have already thought of.
I think “hardly even in space” is underselling the difficulty of getting to orbit. Most of the energy usage comes from getting to orbital velocity.
It seems like they might use a Molniya orbit.
Edit because I accidentally hit save: This might be good for training data, I suppose, as you could uplink at perihelion and then let it plug and chug for 12 hours.
Also, does Starfighter Space plan to compete with Stratolaunch, because it seems like Stratolaunch would have a much easier time of putting things in orbit, as that's what it was designed to do.
This is so so stupid. There is 0 way this could ever be economically feasible. It'd be easier to put these in the middle of the Sahara or Antarctica vs putting them in orbit.
Various problems that need to be solved include, but are not limited to:
Each one of those problems ranges from "solved at orders of magnitude more expense to a earth based solution" to "unsolved, still going to be expensive, and requires leaps forward in technology that would probably change the way we handle everything in life anyways".
So yeah...just invent fusion first and then.....
1 and 2 are solved by throwing startup money and government grants into it.
3 might not be as important depending on the application.
4 and 7 are not that much of a problem if it sits in a Lagrange point closer to the sun.
5 They don't care, making a new one might even be more profitable.
6 Just blow into it like soup, bro, just one trillion more and we'll get AGI, bro, we'll science our way out of all problems with it, bro, what's a few extra warming between bros, bro
It's the same kind of thing as asteroid mining. Neat PR sales pitches to clients drunk on scifi.
The logistics and costs to even get to one are astronomical let alone setting up a supply chain.
For anyone else who was (like me) wondering "Why...?":
Only problem is you can’t fix anything that goes wrong. So you will increase costs quite a bit for redundancy and extra engineering (not to mention launch costs which aren’t cheap). What would a data center look like if every failure that requires hands on the hardware could never be addressed? Every year more and more goes offline forever.
I do think it’s a cool idea. But we’re missing the infrastructure necessary to support it.
I think the next test is probably 'how long can it stay up there and functional' and there will be another article like that, and we'll all discuss the problems they still face, and, eventually, there won't be any more problems and data centers in space will just be normal.
I feel like I just went through this with a lot of things that SpaceX did.
Even with zero problems this is still more expensive and less adaptable than the ground-based data centres they'll be competing with.
To be fair, people said that about most technologies attempting to unseat an established tech. Electric cars come to mind, for instance.
Difference is, it's easily foreseeable that you can refine manufacturing processes so that electric cars would eventually compete with petrol on price point. But blasting infrastructure into space, along with two or three backups of everything, is never going to be cheaper than shipping one of everything in a truck. Not every new moonshot technology is the next Galileo.
Follow-up question:
I understand many of the benefits of running computationally intensive (and heat-intensive) tasks in space (and demo'ing the tech now). But what kind of bandwidth is the final product going to need and do we currently have the price-performance to make that amount of bandwidth even happen in the next couple of years? I know that satellite bandwidth has gotten a lot cheaper and even cellular phone carriers are offering services (along with things like Garmin's inReach satellite texting communication that's been in consumers' hands foe 10+ years), but how much cheaper? Datacenters have staggeringly huge pipes these days so I'm wondering how much bandwidth the data from an AI question-response session would take and how much of it they plan to have completed in orbit.
The article says they're planning not to have a geostationary orbit for the 5-gigawatt satellite so that it can gather solar energy without day/night restrictions, but they would have to get other satellites up just to carry the communication signals to whatever side(s) of the planet the data is being sent. So bandwidth really becomes a huge factor as more hops are added to the transmission. Are there companies looking at developing this tech with geosynchronous orbits with a dedicated ground receiving station? And how expensive is that real estate about to get? Would those companies be looking at more equatorial regions for the ground stations (or perhaps just Texas and Florida)?
I imagine they’re talking up training and not inference because the networking requirements are lower. They still need to upload the training data, though. Perhaps some of it would be preinstalled before launch?
If they launched a GPU (and hardware to run it) into orbit then it's trivial to include a micro SD card with the works of Shakespeare that they used as training data.
That gives a whole new meaning to "Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway. –Andrew Tanenbaum, 1981"!
We're probably only a few years from 5G speeds from orbit. Consider that Starlink can reliably deliver ~150mbps from LEO, which is staggeringly fast, and companies like ASTS (and spaceX, again) are working to stretch that to mobile devices.
I'm assuming you mean 5G speeds rather than literal 5G. It's physically impossible for standard 5G wavelengths to travel that far. It might be possible to improve some kind of signal beaming technology, but that's different.
I did mean it colloquially, but you're right that that was sloppy of me. I've edited my post. Thanks!