Moving The Cloud to orbit

[edzieba]

This is less convincing. Fiber-connected users at the edge often still have a tough time getting more than ~10mbps, yet they can easily consume many multiples of that between their cloud compute instance and their data storage service. Even with Ka band RF Starlink is getting something like 20 Gbps to each of their groundstation gateways? That's not exactly chump-change.

It's also not easily scaleable. If you need more bandwidth at a groundside datacentre than your existing fibre link can provide, you can lay additional links and aggregrate them trivially.
For a satellite ground station, you can only get as much bandwidth as the visible constellation overhead has available. If an existing groundstation can achieve and aggregate 20Gbps from talking to all the satellites visible overhead, then putting down a groundstation next to it would add zero bandwidth. Worse, if the datacentre next door also puts in a 20Gbps groundstation, you now only have 10Gbps each available at any one time. And if you need to geographically separate your ground stations so they can see 'fresh' satellites, then you may as well take the fibre links you now need to lay to those ground stations and run them to an exchange instead, skipping the satellites altogether.

Megaconstallations are fantastic for spreading a little bit of bandwidth ("little" in gigabit terms) consistently over a massive area. They're absolutely abysmal at achieving very high data rates to a single point.

[VSECOTSPE]

The report finds that data centers consumed about 4.4% of total U.S. electricity in 2023 and are expected to consume approximately 6.7 to 12% of total U.S. electricity by 2028. The report indicates that total data center electricity usage climbed from 58 TWh in 2014 to 176 TWh in 2023 and estimates an increase between 325 to 580 TWh by 2028.

https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers#:~:text=The%20report%20finds%20that%20data,total%20U.S.%20electricity%20by%202028.

The earth intercepts 174,000 TW (terawatts) of energy from the sun. 70,000 TW makes it to the surface. About 20,000 TW of that energy is captured by land plants, from rain forests to the soil microbes of the Mojave. Most of that captured energy drives chemical processes in the cells of the plants. About 200 TW remains to become food and fiber, which feeds animals and microbes and fungi. While the fraction used by animals and fungi is small, these organisms are the recyclers of nature, converting dead plants into the materials to make more plants. Interrupting those flows damages the biosphere.

A healthy adult consumes 2000 Calories a day, which is about 100 watts. That is about half a terawatt for all the humans alive today. We consume about 14 TW of manufactured energy globally. A sustainable world of 10 billion humans, living at first-world standards, might consume 100 TW of manufactured energy by the end of the century. Obviously we are not going to get that by burning 200 TW of plant product, and robbing the other animals and fungi. We are not going to get 100 TW by wind and solar, because without gigantic year-capacity energy storage devices, we must displace huge amounts of habitat to have enough production capacity for winter. Our bite of the biosphere is already too large. What to do?

In space, energy is not interrupted by clouds, or winter, or night. One square kilometer of sunlight is 1.36 billion watts. Solar cells in space, with 15% efficency and 70% availability, can deliver 140 million watts per square kilometer. A 200nm thick, graded junction InP solar cell is highly radiation resistant, and weighs 1 gram per square meter, ten tons per square kilometer with a 10x thicker substrate. At $10,000/kg launch cost, that is 70 cents per watt - and launch can be orders of magnitude cheaper.

One pixel in the picture above, 75 by 75 kilometers, is 800 billion watts, 0.8 TW. Florida is 170,000 square kilometers, and that area could produce 24 TW in space. The highlight in the middle of the marble represents more than 100 TW. Because of clouds, and night, and winter, the amount of area needed for ground solar would be far larger.

http://server-sky.com/212acres

[RedLineTrain]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  Data centers devoted to inference tend to not need as much buffering because there many jobs going at once that can smooth each other out.

Because AI training data centers need so much special care and feeding, I suspect that they would not be suitable for orbit quite yet.

[sdsds]

Megaconstallations are [...] absolutely abysmal at achieving very high data rates to a single point.

Ah, thanks for your explanation of that. It certainly limits the use cases where "the cloud in orbit" might be competitive since there's no practical way to run fiber through the atmosphere. Is there any hope for 6G-style millimeter wave photonics?

Reportedly in 2023 NASA established a laser link from its TBIRD optical payload to Earth with a throughput of 200Gbps.
https://www.cailabs.com/blog/aerospace-and-defense/exceeding-throughput-limits-with-laser-communications/

[DanClemmensen]

i don't see how "cloud in orbit" can work. The latency from a user on the ground to any given orbital server wll vary continuously from about 10 ms to about 70 ms as the server orbits. I suppose you could artificially fix the latency at 70 ms, but that's not very good service. Sure, the cloud could make all servers virtual and hop them to keep each virtual server near a fixed location, the the bandwidth to support this would massively exceed the user bandwidth.

[Twark_Main]

i don't see how "cloud in orbit" can work. The latency from a user on the ground to any given orbital server wll vary continuously from about 10 ms to about 70 ms as the server orbits.

That latency and variation is fine for the vast majority of workloads.

There are certainly issues putting servers in space, but this isn't a major show-stopper.

[Robotbeat]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

[edzieba]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

1) That means you would need to size the solar array to peak power demand and not to average power demand, and deal with dumping that power somewhere when actual demand dips below that peak output (a PV cell is active as long as it is illuminated).
2) Training workloads are 'bursty', even during training operations, with spikes on the order of the actual peak draw occurring on the sub-second scale. This is even more of a problem for a solar-powered datacentre (in space or on the surface) because solar power is DC, not AC. This means there are no physically spinning generators to provide literal (mechanical) and electrical inertia to the grid to buffer load spikes. Worse still, for an orbiting datacentre the spacecraft must handle 100% of the fluctuations internally, as there is no grid connection.

[Lee Jay]

We consume about 14 TW of manufactured energy globally. A sustainable world of 10 billion humans, living at first-world standards, might consume 100 TW of manufactured energy by the end of the century.

I know it's not a quote from you, but this particular claim now has a name - "the Primary Energy Fallacy".  It's the idea that all energy we consume in fossil fuels has to be replaced by energy consumed in electricity in the future.

To give you an idea how utterly false that is, I was "consuming" 77MWH per year in total energy for home and transportation.  When I electrified my fossil energy consumption went to zero and I now consume 10MWH per year in total energy consumption, with 12.5MWH per year produced by the solar on my roof.

In other words, there's no good reason that getting 10 billion humans to first-world standards couldn't be done while consuming *less* energy than we consume now, not 7 times more.

[Twark_Main]

We consume about 14 TW of manufactured energy globally. A sustainable world of 10 billion humans, living at first-world standards, might consume 100 TW of manufactured energy by the end of the century.

I know it's not a quote from you, but this particular claim now has a name - "the Primary Energy Fallacy".  It's the idea that all energy we consume in fossil fuels has to be replaced by energy consumed in electricity in the future.

To give you an idea how utterly false that is, I was "consuming" 77MWH per year in total energy for home and transportation.  When I electrified my fossil energy consumption went to zero and I now consume 10MWH per year in total energy consumption, with 12.5MWH per year produced by the solar on my roof.

In other words, there's no good reason that getting 10 billion humans to first-world standards couldn't be done while consuming *less* energy than we consume now, not 7 times more.

See also:

https://en.wikipedia.org/wiki/2000-watt_society

https://www.2000-watt-society.org/what

[sdsds]

i don't see how "cloud in orbit" can work. The latency from a user on the ground to any given orbital server wll vary continuously from about 10 ms to about 70 ms as the server orbits. I suppose you could artificially fix the latency at 70 ms, but that's not very good service. Sure, the cloud could make all servers virtual and hop them to keep each virtual server near a fixed location, the the bandwidth to support this would massively exceed the user bandwidth.

As others indicate this amount of latency variation is likely not critical. If it were, could the terrestrial gateway buffer the flow to make the realized latency between each packet match the average latency of of the flow (40 ms in the given case)?

[Lee Jay]

Putting a server farm underwater makes it hard to maintain.
Putting a server farm in space makes it hard to power, hard to cool and VERY hard to maintain.

[Robotbeat]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

1) That means you would need to size the solar array to peak power demand and not to average power demand, and deal with dumping that power somewhere when actual demand dips below that peak output (a PV cell is active as long as it is illuminated).
2) Training workloads are 'bursty', even during training operations, with spikes on the order of the actual peak draw occurring on the sub-second scale. This is even more of a problem for a solar-powered datacentre (in space or on the surface) because solar power is DC, not AC. This means there are no physically spinning generators to provide literal (mechanical) and electrical inertia to the grid to buffer load spikes. Worse still, for an orbiting datacentre the spacecraft must handle 100% of the fluctuations internally, as there is no grid connection.

Oh please. Sub-second fluctuations? They’re called capacitors.

I used to build large computer systems for datacenters and even ran a small datacenter for a while. The server power supplies already have to deal with all of the constraints you’re talking about. This isn’t exotic.

[Robotbeat]

Putting a server farm underwater makes it hard to maintain.
Putting a server farm in space makes it hard to power, hard to cool and VERY hard to maintain.

It’s not harder to cool or power unless you already think large reusable launch vehicles won’t happen, in which case the whole thing won’t happen anyway.

[Lee Jay]

Putting a server farm underwater makes it hard to maintain.
Putting a server farm in space makes it hard to power, hard to cool and VERY hard to maintain.

It’s not harder to cool or power unless you already think large reusable launch vehicles won’t happen, in which case the whole thing won’t happen anyway.

You think generating 150MW of electricity and dissipating 150MW of heat is just as easy on orbit as it is on the ground?  ISS generates and dissipates about 75kW.  So you need 2000 times that amount to match a modern good-sized data center (we were recently studying a data center at work that has three 50MW halls).

[Lee Jay]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

1) That means you would need to size the solar array to peak power demand and not to average power demand, and deal with dumping that power somewhere when actual demand dips below that peak output (a PV cell is active as long as it is illuminated).
2) Training workloads are 'bursty', even during training operations, with spikes on the order of the actual peak draw occurring on the sub-second scale. This is even more of a problem for a solar-powered datacentre (in space or on the surface) because solar power is DC, not AC. This means there are no physically spinning generators to provide literal (mechanical) and electrical inertia to the grid to buffer load spikes. Worse still, for an orbiting datacentre the spacecraft must handle 100% of the fluctuations internally, as there is no grid connection.

Oh please. Sub-second fluctuations? They’re called capacitors.

I used to build large computer systems for datacenters and even ran a small datacenter for a while. The server power supplies already have to deal with all of the constraints you’re talking about. This isn’t exotic.

We've been studying the electrical side of data centers at work.  In many cases, we see multi-kHz fluctuations in power draw in the MW range.  It's kind of a huge issue for the power grid.  The causes are in some cases complex.

[sdsds]

I'm one of those pathetic amateur electronics designers who don't really distinguish between capacitors, super-capacitors and rechargeable batteries. From a systems engineering perspective isn't the functionality essentially the same? They all create a buffer between something that's variable and something the designer wants to keep reasonably constant. On the scale of a datacenter are batteries the obvious choice?

[edzieba]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

1) That means you would need to size the solar array to peak power demand and not to average power demand, and deal with dumping that power somewhere when actual demand dips below that peak output (a PV cell is active as long as it is illuminated).
2) Training workloads are 'bursty', even during training operations, with spikes on the order of the actual peak draw occurring on the sub-second scale. This is even more of a problem for a solar-powered datacentre (in space or on the surface) because solar power is DC, not AC. This means there are no physically spinning generators to provide literal (mechanical) and electrical inertia to the grid to buffer load spikes. Worse still, for an orbiting datacentre the spacecraft must handle 100% of the fluctuations internally, as there is no grid connection.

Oh please. Sub-second fluctuations? They’re called capacitors.

I used to build large computer systems for datacenters and even ran a small datacenter for a while. The server power supplies already have to deal with all of the constraints you’re talking about. This isn’t exotic.

Check the link. Training datacentres have substantially different load characteristics than regular HPC or other sites.

[Lee Jay]

I'm one of those pathetic amateur electronics designers who don't really distinguish between capacitors, super-capacitors and rechargeable batteries. From a systems engineering perspective isn't the functionality essentially the same?

No.

They all create a buffer between something that's variable and something the designer wants to keep reasonably constant. On the scale of a datacenter are batteries the obvious choice?

They all serve different functions.

Capacitors store 6 orders of magnitude less energy than ultra-capacitors but have very low inductance.  That means you use them to filter switching noise, not store energy for more than microseconds.  Ultracapacitors can store a good bit of energy but they aren't constant voltage sources like batteries and they still store about 1/10th as much as batteries for a given weight.  Batteries are constant voltage sources and store a lot of energy (relative to the other two) but have high inductance and thus relatively slow response time, so you don't use them for filtering microsecond-level switching.  It's common to put capacitors in parallel with batteries for this reason.

[RedLineTrain]

There is turning out to be a few types of data centers.  For the AI data centers, there are two types:  training and inference.  Data centers devoted to training need buffered power because there are dramatic usage spikes.  Buffering requires batteries, which are very heavy.  …

This isn’t true. In orbit, you either have short, 40 minute shadow periods (way shorter than 12-18 hour times of low solar in Earth) which means very small batteries (as batteries are INCREDIBLY good at high power but are heavy for long duration storage… although better than people think)or are in a sun synchronous or high orbit and don’t need batteries at all(other than very short periods of eclipse perhaps).

Training workloads last for days, weeks, or months at a time. So in fact they’re MORE consistent than inference, which goes based on demand from humans.

That's not the way it works, unfortunately.

https://semianalysis.com/2025/06/25/ai-training-load-fluctuations-at-gigawatt-scale-risk-of-power-grid-blackout/

Not only is the scale massive, but AI training workloads have a very unique load profile, unexpectedly rising and falling from full load to nearly idle in fractions of a second. . .  The issue caught leading AI labs by surprise. Meta’s LLaMa 3 paper mentions challenges with power fluctuations, and that is “only” a 24,000 H100 Cluster (30MW of IT capacity).