Hacker News Reader

> why no one brought this up before or outside the context of AI

AI compute is different (slightly higher latency is fine for inference, and there's no issue for training), and there has never been so much backlash against data centers or other infrastructure buildout. If an increasingly-non-minority of politicians get their way it will in fact be cheaper and faster to get some servers shipped to space than it will be to get the permits and build it out here on earth.

Also, most Datacenter maintenance is just dealing with the problems you see with space- power and cooling. Solar is 5x better in space and a lot more consistent. Here on land they're shutting down nuclear and imposing so many new regulations on gas and coal (and now on solar and wind) that there aren't many grids that can support growth at the scale that's being requested.

Nobody is claiming that all compute is going to space, which is what you seem to think you're arguing against. There's high-dollar demand for it right now, so if we want to be multi-planetary it's the perfect time to start tackling the "compute in space" problem that needs to be solved. Or you need to prove to future people that "All compute can be forever located on Earth" which seems a lot harder sell.

> never been so much backlash against data centers

Backlash against the proliferation of space junk still has plenty of room to grow.

But it'll have to start with Sinophobia so nobody wants to touch it. LEO objects are just such a non-issue that you'd have to really be grasping at straws (though the "Data centers will consumes earth's water" people can probably fall for it)

Assume reusable spaceflight eventually brings launch cost close to the cost of fuel. This is close to happening.

The overhead of building out grid and power infrastructure on land would then exceed the installation speed and cost relative to space based deployments.

Also assume the compute that does make it to space has a short shelf life anyways so lack of ability to repair is a non issue. As we scale manufacturing on land this will increasingly be the case.

China has already run experiments and served models from space, so we know the heat dissipation equation is solvable.

Finally you’d arrive at a similar model that’s already proven successful with Starlink but applied to serving inference.

The key question is speed to scale new deployments to meet demand. If the markets demand is near infinite, they will choose to fund space based deployments over slower land deployments.

That makes no sense to me. You've cited one overhead for building on land but just didn't bring up the billions of overheads for deploying satellites. The associated costs of launching this stuff into space blows the price of boring old power grid infrastructure out of the water, even with launch prices assumed to be the cheapest they can be. For instance, the extreme additional costs of equipping each payload with their individual systems of power generation, flight computers, maneuvering systems, structural components, communications infrastructure, heat dissipation hardware, radiation shielding and so much more. You have to pay the price in weight and money for all of these inefficient small-scale components for each unit launch, compared to plugging in an additional server rack and taking advantage of the economies of scale in a centralized data center that already solve all of these things. And that's before you get into the costs of designing, mass-producing and operating these things, compared to the costs of running a normal data center. And that thing about reusability that I mentioned also cuts into the margins. Data centers can keep working even if their hardware is a generation out of date, they can sell off old equipment, they can repair individual components instead of throwing out a whole rack. This is literally setting your GPUs on fire. It's just one more thing that makes it even more expensive.

Heat dissipation in space is possible, of course, but every kilogram you spend on heat is a kilogram you could've spent on something else. When you're talking about boxes that generate so much heat, you're going to need to spend a lot on that ancillary hardware in each unit that, again, makes it even less rational.

Then the concerns about megastructures or distributed computing go unanswered - to my knowledge, we simply don't have the technology for either of these right now. Starlink isn't close to solving it - the bandwidth of a Starlink satellite is nothing in comparison to the bandwidth of a single current-gen server GPU connection.

Energy is not cheaper on earth. Solar in space gets 24/7 power at a 30%+ rate vs the surface. Radiative cooling is passive and cheaper because there are no HVACs or chillers and high temperature chips reduce the need. The ISS does this already. Radiation really isn't a big issue with ECC and redundancy and optical links are fine for batch training.

The only hard part making the math work is the launch cost. They need reusable and reliable starship economics. If they hit that goal, it will become cheaper for pre-training, which is 70%+ of the budget for the frontier models.

> It offers no real advantages over doing the same thing on Earth

Abundant solar energy, free real estate, less regulation, less backlash from NIMBYs, simpler (yes) cooling.

> Energy is far cheaper and more abundant here than in space.

Huh? The sun is obviously the most abundant source of energy in the solar system.

Satellites in a dusk dawn sun synchronous orbit can be fully illuminated 24/7, so they receive ~6x more solar energy than panels on earth. They also don’t need batteries to operate 24/7, and the panels don’t need glass to deflect hail.

> Heat dissipation

Yes, it will require large radiators. They’re mechanically simpler than terrestrial cooling though.

> Radiation shielding

It turns out generative AI is somewhat uniquely robust to occasional bit flips.

- Either the most complex in-space construction ever undertaken, or the most complex distributed computing problem ever undertaken (no, Starlink satellites aren't good enough, we're orders of magnitude away from replicating the speed and reliability of connections within a single room)

It won’t be a single structure, and will only be used for inference, so latency between satellites/racks doesn’t matter.

- Zero flexibility, zero repairability, zero upgradability. Either it's working, or you make it burn up in the atmosphere with no in-between. Add on that the rationality of sending mountains of precision-manufactured tech containing many uncommon metals only for them to be completely lost. This makes the pricing even worse, in addition to - Already high costs for designing, building and launching all that in addition to all the extra weight overhead you're taking in components that don't do computation, when the alternative is building a glorified warehouse in the middle of nowhere.

I think people vastly underestimate how much a fully reusable Starship will change the economics of space operations.

Not only the initial launch costs, but things like refueling and repairing satellites becomes more economical. I wouldn’t be surprise if SpaceX sends Starships to refuel/maintain satellites to keep them in orbit longer. In fact, SpaceX’s own animation shows modular servers sliding in and out of the satellites.