I've been saying for a long time that we should consider remote areas for building datacenters for batch processing.

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.

Scott Manley has published a video a few months ago explaining why putting data centers in space is an absolutely terrible idea. Lumen Orbit, the company mentioned, is a former name of Starcloud.

https://www.youtube.com/watch?v=d-YcVLq98Ew

I’m still not sure how people can believe this, this makes zero sense to me.

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?

Unless they've figured out some impressive cooling tech, which I would expect would be worth more than the rest of their company combined, then this is pretty much DoA. "More efficient cooling architecture taking advantage of higher ΔT in space" would indeed be useful if you had a nice medium to radiate into. It turns out that thermal radiation is incredibly poor into the vacuum of space lol.

I had a good laugh.

- 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.

Their whitepaper (https://starcloudinc.github.io/wp.pdf) had a thread last fall:

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)

I'm not an engineer, so maybe I'm wrong, but isn't cooling famously difficult in space?

It might make more sense to put data centers on the Moon.

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.

The use-case for this is not anything technical, it's putting your data center outside of any jurisdiction.

How does “passive cooling” work in space?

What does space give us that Earth does not in this scenario? Free real estate? They only mention falling costs for deployment.

This is the best piece of satirical writing that I've read in quite a long time. It's genuinely amazing, every sentence is a delight. And the Musk/Altman/Zuck quotes to start, the stupid little graphs and the "equation" at the end of page 2. I'm in stitches!

It sounds like their vision for space-based data centers presupposes nearly-free energy costs, delivered via a colossal solar farm made possible by falling launch costs.

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.

Microsoft had/has the Natick project which was an undersea data center testbed which allegedly had a bunch of benefits. That doesn't seem to have gone anywhere - or at least isn't really scaling up. I'd imagine the ongoing operational costs of space are worse than the ocean?

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.

This has to be one of the dumbest ideas I've seen posted here.

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.

Most expensive IT call ever when you need to go fix that one fried power convertor that everyone said wouldn't fail.

Can this possibly make financial sense even if launch costs were zero?

One NVIDIA DGX SuperPOD consumes 10 kW which would be ~500 square feet of solar panels and ~100 square feet of radiator area.

Isn’t cooling already a major issue for spacecraft?

The big radiators on the ISS can only dump a few server racks worth of heat.

I feel like YC's $500K investment would last 0.25 days at most for this type of business.

Definitely an out of this world idea. I wonder if their micro datacenter is going to be self-sufficient power wise using only solar energy? And how would they address the hardware failures that are likely when you train large language models at scale?

How do they plan on addressing the solar radiation issue? What is the solar flare risk?

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.

Aside from the obvious cooling issues people have already mentioned, isn't cosmic radiation also very unkind to modern ultra dense silicon? AIUI they tend to use really old silicon processes in space stuff for that reason, and even then they have to build in redundant compute to mitigate logic errors that probably wouldn't happen on Earth.

Reading the paper this sounds like space Theranos. If they start producing results then I'd double check to make sure it's not just calculated on regular data centres and that they're just pretending its from their space stations.

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.

Why couldn't you deploy all those solar panels on earth. Then you could supply the grid when you weren't training models, you could recycle your components when they got old, you wouldn't need launches, you could maintain things if they break, and sure, spend a bunch of money to figure out large radiators that could radiate heat out of the atmosphere.

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!

I was wondering if these server racks in space would need to be specifically designed for enough radiative cooling. Apparently the answer is yes: the radiators would be expansive and placed on the reverse side of the solar panels.

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.

https://www.starcloud.com/wp

wow, one year back, I had made a prediction to a friend that this is the direction that Starlink will head in. I was thinking it would proceed like this:

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).

Solar Radiation and bitrot/damage, how you solving it? Whats your shielding stack?

Ycombinator has one legitimate function: dissipating excess looted wealth.

Can't wait to experience a Gigawatt DC re-entering a la Cosmos 482.

Make something hard harder, just because

space debris, radiation and no maintenance. The buzzwords sure sound cool, but make absolutely no sense.

Eliezer Yudkowsky's take:

>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

I miss when tech was exciting

They closed a very large seed round late last year (FKA "Lumen Orbit"): https://techcrunch.com/2024/10/24/lumen-orbit-closed-one-of-...

It makes so much sense that there are many comments saying why this is a bad idea. In fact I agree with a lot if not all of these comments. However, I still have to commend how big of a dream this is and that they got some sort funding behind them. I say let big dreams possibly fail although I admit it’s not my money being spent.

- radhard electronics require a different sdk to maintain good seu/tid, and also require a higher layout overhead to maintain these metrics.

- 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.

>passive cooling

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...

> able to train large models like GPT6

That is pretty amazing to say with a straight face. Whatever GPT6 may or may not ever be…

Formerly known as "Lumen Orbit", or for short, "Lumon".

The work is mysterious and important. Praise Kier.

Canadian north makes sense (very cheap electricity, ridiculously easy heat management).

Space? I really don't get it.

And what happens to these datacenters when the underlying GPU tech becomes obsolete within 2-3 years?

This doesn't pass the sniff test. Please, show me the napkin math where this remotely adds up.

Does abundant cooling in space mean there's a better way to radiate away heat than on Earth or just that the heat doesn't contribute to heating up Earth or something more complicated? (Asking because I thought cooling was a big problem in space.)

If this and/or the policies, leadership, and mindset that enable these kinds of things interest you, I highly recommend Abundance by Ezra Klein and Derek Thompson. IIRC, they outline data centers in space in his utopian introduction.

Simply tow the data center out of the environment![1]

[1] https://www.youtube.com/watch?v=3m5qxZm_JqM

Very ambitious but it seems futile if you’re not building the rockets yourself. Personally I’m more bullish on figuring out how to use analog chips to train models.

Neat, I wonder if https://star-catcher.com/ could provide power.

Suggest changing the elevator pitch text - it’s never going to make sense to train a frontier model in space. In our lifetime at least.

How I sniffed and stole training data and model data over the air to "Starcloud" posts gonna be crazy amount in abundance

How did this get funding?

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.

[flagged]

I guess since we decided Idiocracy was aspirational, why not Wall-E as well.

/s (kinda but maybe not really...)

There is a lot of handwaving going on here. The below, however, is an egregious misunderstanding of how percentages work. That, or it betrays very sloppy work. Doing it right would actually make the specific case they make stronger.

"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!

When I read something like this I fell that I am wasting my time working on a B2B SaaS.

Is this an elaborate joke? I believe so given all the very real physical problems with this, but “GPT-6” training was what pushed it undoubtedly over the edge to pure joke.

> passive cooling

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?

These venture capital scams are getting more and more creative...

YC jumped the shark

Maybe we can put the quantum computers in space too.

Cooling in space was covered by XKCD's Randall Munroe in pretty entertaining detail here:

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...

Wait this is not comedy? Is PG on holiday?

Space is really good low temperature compute. It’s perfect for hosting quantum computers.

Dang that's very cool. As long as up and down bandwidth stay strong and reliable.

Thanks.

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.

It feels defeating to see so much money being thrown at what is either an obvious grift or someone hoarding world changing technology. There’s a reason data centers aren’t in space… If someone has the technology to make it feasible, they are doing the world a massive disservice by not sharing it.

is this a joke? isn't cooling in space a really big problem? how does it make sense to run a data centre with huge cooling requirements in a place where cooling is very difficult to accomplish?

I think what I like the most about this whole thing is that the term "cloud computing" (which I always thought idiotic) will now finally become somewhat more meaningful.

Not to be confused with StarTree Cloud:

https://startree.ai/products/startree-cloud

Latency?

All these comments are acting like this is a major problem with VCs in general.

Is this not a major problem with YC, specifically? Our beloved orange site funded and accelerated these guys.