Item — HN Reader

Also, Pixxel, Skyroot, and Agnikul are targeting orbital data center launches this year for defense applications (which is what the entire ODC story is about). The US NRO already uses India's Pixxel [0] along with Finland's ICEYE (which is now co-manufacturing synthetic aperture sats in India with Agnikul [1]) for hyperspectral scanning.

Edit: can't reply

> But.. why?

Missile Defense and C4ISR [2]. Seconds matter, so most of the newer generation of missile defense systems are experimenting with how to offload compute at the edge to reduce C2 latency.

Most orbital sat startups in the US, China, and India are partially backed by military and intel oriented funds (eg. Starcloud and IQT/In-Q-Tel).

-----

[0] - https://www.nro.gov/news-media-featured-stories/news-media-p...

[1] - https://www.livemint.com/companies/news/agnikul-cosmos-iceye...

[2] - https://idsa.in/wp-content/uploads/2026/01/book-MISSION-SUDA...

3 comments

This borders on science fiction. 88,000-and 1-million satellite clusters (as claimed in your link) are hard to take seriously, especially with the possibility of Kessler Syndrome. Also:

1. How do you cool your chips? Vacuum is a thermal insulator, so radiators are required to remove heat. nVidia (or even ASICs) require much and specialised cooling.

2. How does one radiation harden a H100?

3. I'm also seeing where TCO for these are 78x their terrestrial equivalents [0]. Is that financial sustainable?

And there still remain issues with power supply, regulations, and bandwidth. This feels more like a thought experiment rather than an actual serious engineering or business case.

[0] https://www.abiresearch.com/blog/data-centers-in-space

The projections are questionable, but this is something the US and China are experimenting with as well.

1 and 2 are still open questions, but these are not aimed to be commercial grade DCs - this is basically edge compute (think a handful of racks). 3 is not a problem for defense usecases. (EDIT: Discussion here seems to point out that data OP is using might be flawed [1]).

Ignoring the fact that just about every orbital data center startup in the US is funded by IQT and China's CMF has been doing something similar is bad from a defense perspective.

Based on dealflow, these aren't being targeted for consumer usecases in the short-term and whoever has been saying that is misleading.

> especially with the possibility of Kessler Syndrome...

India, Russia, China, and the US don't care about Kessler Syndrome - they have already launched and deployed ASATs. This also comes after India and China had a near collision in 2024 that was treated as an offensive action [0].

[0] - https://www.bloomberg.com/news/articles/2025-09-22/india-pla...

[1] - https://news.ycombinator.com/item?id=49039873

    3 is not a problem for defense usecases.

Cost is the critical constraint for defense programs. Yes, militaries spend a shit-ton of money, but that's usually in a penny wise and pound foolish way. A dollar spent on your project is a dollar that's not spent on the innumerable things generals think are critical to their future wars like boondoggles, missiles, planes, ships, missile defense systems, etc.

The exception is if you're one of those rare projects. A starlink type constellation or launch capabilities certainly could be, but it's hard to imagine generals getting excited about compute in space just for the sake of it.

It is critical to avoid a boondoggle gap.

This reminds me of the railgun. Basic math and physics tells us that not only would the (very expensive) barrels wear out very quickly, but that it would have had to be fitted on on a nuclear-powered pocket battlecruiser.

Worse, the technology for firing any meaningful payloads from an electric gun (8" Small Diameter Bomb equivalents, guided, airburst, incendiary) simply does not exist.

Same as with Musk's California Vacuum Tunnel (which diverted attention from passenger rail). And his Neuralink. In the 1980s, it was nuclear pumped space-based lasers and Soviet particle beam weapons.

All of the above can be debunked with 2 years undergrad physics and a Casio calculator. Yet they were still taken seriously by high-level politicians and business, some of whom were deeply connected with the military-industrial complex.

People have shrugging of questions like 1 and 2, only for 3 to hit them very hard. But we will see if they can launch 600 of these satellites as they claim.

As raynier and pfdietz pointed out [0], the assumptions you are using aren't necessarily correct.

[0] - https://news.ycombinator.com/item?id=49039873

Soviets definitely experimented with airborne CO2 lasers.

> 1. How do you cool your chips? Vacuum is a thermal insulator, so radiators are required to remove heat.

This is space 101 and honestly, at this point, I consider bringing this up as disqualifying from giving criticism on the topic.

> nVidia (or even ASICs) require much and specialised cooling.

Whatever. They generate N kW of heat, you need to shed N kW of heat, and you need to fit that within your budget. End of worry.

Pro tip: if you can slice your problem so that individual satellite needs to shed less than M kW of heat, where M corresponds to how much heat some existing, deployed satellite platform handles, you can just do your initial design around that satellite platform, replacing the "business payload" with yours.

This is e.g. how the unfairly criticized recent "satellite swarm TPUs" paper from Google handled it. Everyone who brought up cooling revealed themselves as not having read the first page and not having thought about it seriously for more than 30 seconds.

> 2. How does one radiation harden a H100?

See that paper for some ideas and considerations, as this is part of what they focused on, after solving cooling by sizing the per-satellite payload power needs to Starlink.

Yeah, there's no solution for heat dissolution at this time.

You can run a few cpus in space, you are a few seconds closer to the data to figure out something basic, but you cant do that much calculation. If you do much cpu intensive stuff, you just get hot quickly and you can't dissipate the heat. I'm sure new ways to cool down will be developed, but there's not even any experimental techniques, right?

ISS collects ~250 kW solar power. All of that needs eventually to be dissipated - the energy removed by communication radiowaves is small. ISS routinely handles this thermal question for decades.

There is nothing magical with cooling in space. Just a different environment - a rather well studied already, we send satellites to space for some 70 years almost. Saying there is no solution is incorrect - we have options, we have numbers, we can point to concrete questions and answers.

1. By using some form of liquid droplet radiator, as described in these:

https://en.wikipedia.org/wiki/Liquid_droplet_radiator

https://ntrs.nasa.gov/citations/19850005591

https://ntrs.nasa.gov/citations/19810062992

https://spectrum.ieee.org/orbital-data-centers-heat

2. By producing them on intel 18A, like Starfire:

https://duckduckgo.com/?q=intel+foundry+18A+starfire

( Or going for different substrates, like Gallium-arsenides, Gallium-nitrides, Silicon on Sapphire, by using https://en.wikipedia.org/wiki/Heterojunction_bipolar_transis... on https://en.wikipedia.org/wiki/Indium_phosphide 's , enabling insane speeds & photonics.

To be fabbed here, possibly:

https://www.nist.gov/news-events/news/2026/06/department-com...

Or whichever other new metamaterial may come along. )

Now imagine the medium of these liquid droplets to directly flow through micro-capillaries in on-die cooling channels.

Good Morning! Ring, Ding, Ding, Rrrrriiiinnnnng!

(1) seems to be the big question to me. Surely it would be much cheaper to simply stick a data centre up in the mountains somewhere, low enough to be fairly easy to reach but high enough to be nice and chilly?

Why do you need radiation hardening? We don't need hyper deterministic systems in those setups.

You need your chip to not completely fail. Latch-up can destroy things. A complete failure of a SERDES may dramatically reduce the utility of the whole system.

Explain to the class why Kessler Syndrome isn’t possible with large LEO constellations.

It is in fact possible to have the Kessler syndrome in LEO, the debris doesn't just immediately fall out of the sky so in practice you just need to put more satellites into a specific orbit to reach the critical density where one satellite breaking leads to a runaway chain reaction.

In actual reality it's estimated that several segments of LEO have already reached criticality, especially around 600 km: https://conference.sdo.esoc.esa.int/proceedings/sdc9/paper/3...

In a form of a question: does low earth orbit has big enough drag to be effectively do automatic cleanup every several years?

The answer is positive. LEO satellites need orbital corrections every few weeks, and some, several times a week.

In case of collision in LEO, wouldn't the debris have chance to go high altitude and cause further collisions?

No. It'll stay in essentially the same orbit.

During collision, some of the energy will be lost, so the fragments will necessarily have a lower orbit.

The only thing that can accelerate fragments into a substantially higher orbit is the energy of elastic deformation of material during the explosion.

Mein Gott, you're not just making this up ( https://www.datacenterdynamics.com/en/news/neevcloud-and-agn... ).

But.. why?

Musk bandwagon-hopping? Investor bamboozlery? (Same diff?)

Is my borderline childish understanding of basic physics THAT off because I cannot see the value or utility (beyond a small smear of niche edge cases) of space-based compute?

- ed, disclosure: You seem to have edited your response whilst I was typing mine, adding in valuable links. Thanks!

> Is my borderline childish understanding of basic physics THAT off because I cannot see the value or utility (beyond a small smear of niche edge cases) of space-based compute?

In the real world, "physics" is not necessarily the gating factor. There is a major concern about the environmental footprint of terrestrial data centers, to the point where major U.S. states are enacting moratoriums: https://www.governor.ny.gov/news/first-statewide-moratorium-.... These legal and social roadblocks must be accounted for in analyzing the viability of orbital data centers.

If the "physics" tells you that your satellite cannot radiate heat away from your nVidia GPU cluster because each H100 needs 1.1 meter square of radiator, then opinions do not matter. The same applies to power supply and bandwidth.

The radiator is much more efficient than the solar panels (for obvious reasons), so that’s not the limiting factor.

As to power, Caltech is already at 2-3 pounds per square meter: https://magazine.caltech.edu/post/sspp-space-solar-power-pro...

> Another way to think about it: An SSPP spacecraft with a 60-meter-by-60-meter surface area made using today’s space PV-cell technology would cost $36 million and weigh nearly 9,000 pounds, or almost as much as a Ford F-450 truck. With the ultra-lightweight PV-cell technology Atwater envisions, it would cost just $450,000 and weigh about 300 pounds, or about as much as an IKEA three-seat sofa

That’s megawatt-level solar power under 5 tons using today’s leading edge technology. Starship super heavy can launch 100 tons into LEO.

As to bandwidth, Starlink V3 backhaul capacity is 1 terabit. Microwave radio frequencies have an insane amount of bandwidth.

ISS today generates and radiates away about 120 KW of energy with its old tech 3250 sq m of panels panels and it's current radiators. That's what 3 H100 racks need

There may be an economic challenge - which seems to be the sort of problem mass manufacturing can solve very well.

There may be a compute model & latency problem, how do you organize model training when racks are much further apart than in traditional data centres (although speed of light is 50% faster in vacuum than glass fibre). But that's algorithms.

Relative to everything else in orbit, powering a rack of compute and some comms per satellite seems not really to be a physics problem.

What are the assumptions behind that 1.1 m^2 figure?

If I understand correctly, NY is concerned about the power and water consumption. 1. How do they get power in space? (solar) Do the same on earth at 1/100 the cost (or build 50x the number of solar to account for atmospheric loss and you will still come out ahead) 2. How do they get water in space (they don't). Whatever they do in space to not need water, do the same on earth! 3. People don't like living near datacenters - put them in remote areas hundreds of miles away from people, after all space datacenters won't have employees . Again, at 1/100 or 1/1000 the cost.

The point is that while power and water is scarce in space, the NY state has no jurisdiction there. It might literally be easier to build data center in space than it is to overcome social and political pressure in NY.

> Is my borderline childish understanding of basic physics THAT off because I cannot see the value or utility (beyond a small smear of niche edge cases) of space-based compute?

Don't worry, most of the people online didn't get the memo on this.

One of the major selling points of orbital compute is power supply - more specifically, it turns out that, compared to beaming power from space, it's projected to be cheaper to move compute upwell instead - atmosphere losses for beamed power are just too big. Of course this doesn't matter if you can get cheap, clean power from elsewhere (e.g. nuclear).

This is in general data center case. Here, GP says the motivation is reducing C2 RTT, which makes sense for military applications.

The unit economic tradeoffs aren't quite that straightforward (beamed power is lossy enough to require more panels[1], but solar cells and satellite structures have longer useful lives than inference chips so over longer time periods you likely end up launching more stuff into space with disposable datacentres anyway[2], particularly given the datacentres also need bigger radiators. Other issues which favour power being beamed to the ground are not exposing those expensive chips to radiation, and being able to replace them on cycles dictated by inference chip innovation or end-of-life rather than fixed cycles depending on a satellite propellant budget...

The other comparison point is of course plain old terrestrial power sources, including energy storage in the mix if it relies heavily on solar power without a sufficiently global grid. Not having to launch into space buys a lot of power.

C2 RTT times and edge computation of large datasets collected in space make much more sense since unit economics aren't the driving factor, but are unlikely to need datacentres on the same scale as inference compute for the general population (another reason why this use case makes more sense)

[1]but I think the underrated problem with beamed power isn't just the low maturity of the technology, but that as soon as you start talking about sending power to earth via RF or laser you're going to encounter political opposition that makes objections to terrestrial datacentres seem tame.... [2]chips can in theory be replaced on orbit, but the "million satellite" filings are disposable. Replenishing propellant for a few large power stations, potentially on longer cycles, is a simpler task.

> but solar cells and satellite structures have longer useful lives than inference chips

The limiting factor on the lifetime of any satellite is fuel. Especially for large structures that are influenced a lot by atmospheric drag and solar wind.

> beamed power is lossy enough to require more panels[1]

What kills that is simply costs of land. You have to put these larger rectennas somewhere, and land ain't getting cheaper.

Could we have giant nuclear reactors in space? What's the power loss from beaming down power?

> Could we have giant nuclear reactors in space?

We could. The usual stumbling block is how to ship fissile material upwell without the risk of a launch failure spreading highly radioactive material over several countries.

> What's the power loss from beaming down power?

I think it's about 10% from atmosphere alone, but you have to add losses from other components in the system, including light -> current -> RF and RF -> current legs, and I've seen estimates ranging from 15% to 40% efficiency end to end; this random article includes breakdown with estimates, that multiplies down to 37.5% efficiency end-to-end.

https://www.sciencedirect.com/topics/earth-and-planetary-sci...

EDIT:

This system design gives 7-14% end-to-end efficiency: https://arxiv.org/pdf/2206.08373

EDIT2:

Also to spell out another non-obvious aspect of beamed power, it turns out that it's not the efficiency that's the limiting factor per se, but land - you can improve efficiency by building larger rectennas, but it gets very expensive very quickly once you consider paying for land under them.

We already have a giant fusion reactor in space

See:

SNAP-10A

BES-5 (oops, sorry 'bout that Canada!)

TOPAZ-I

Kiwi, Phoebus, and NRX (Mars here we come!)

RD-0410 (Dossvidanya Solar System!)

SP-100

TOPAZ-II

Kilopower

Everybody keeps talking about cheap, clean power. But where I live the price of power has gone up dramatically in recent years. Free power (well, marginally free, anyway) is a major plus.

In France, where we have relatively cheap and clean energy, they plan to build GW DCs with GWs setups of diesel backup generators. When these will run, the pollution will be staggering.

My concern is more in the realms of cooling. I know there's the potential for lots of 'free' energy up there, but how do you then ensure your space-based array of GPU farms bleed all of the resultant heat?

This is like "how do you ensure you can bring back home all the groceries you bought" kind of problem. Space 101. We know how much heat we can bleed off, how, and how fast, and this gives us bounds on how much power we can use, and that is the starting point - you design everything around that.

From a Defense perspective, it's acceptable if multiple ODC fails and you have to re-launch another one. This is why these are being treated as part of a mesh. These aren't supposed to be a commercial DC and are intended to be a mesh of multiple racks in orbit.

The fact that the US, China, Russia, and India have already deployed ASATs means a Kessler effect if a question of when and not if.

> it turns out that, compared to beaming power from space, it's projected to be cheaper to move compute upwell instead

This is damning with the faintest of praise. Beaming power from space is a terrible idea.

> targeting orbital data center launches

I am regularly assured by HN denizens that orbiting data centers will never work.

Because they won't for commercial usecases (which was never the point of an ODC).

For C4ISR usecases they solve the latency problem and are worth the spend needed.

Everyone on HN is talking about a commercial usecase that isn't under serious consideration.

It's funny, people here said the same thing about Starlink. "Sure, it's technically possible, but you could only lose money by doing something like that. There's just no market for it."

What I repeatedly read here was it is physically not possible to build an ODC because of heat problems and alpha particle problems.

At 1GW scale, I'd agree. But if you need around 40-50 racks to do image processing and alerting, that's largely a solved problem.

The whole thesis around ODCs is to basically mass deploy a bunch of replaceable racks en masse and consistently, such that any geospatial intel can be processed at the edge.