If space-based data centers are too impractical, and floating Earth-bound data centers are still too impractical … yet data centers near population centers are ALSO problematic … what’s left?
There are places on earth where there’s plenty of moving water (e.g. major waterfalls, rapids, large rivers, and such) but fairly far away from where people live. While of course still far from perfect, and - yes - generally having serious local environmental effects**, would building data centers at/near sites like Angel Falls, Venezuela (assuming political cooperation) or the year-round white-water rivers in the US Pacific Northwest be the best of many sub-optimal options? At least more practical than floatind data centers, yes?
** Though one hopes this could be at least minimized through careful site selection and bleeding-edge near-term tech.
How is the presence of this water going to help? Data centers use water that has been treated and processed. It’s not raw water that is passed directly over equipment. Is the suggestion to also build new water treatment and electrical power plants out there as well?
My understanding : they use both. There’s a closed loop of extremely clean and controlled liquid circulating over the equipment to conduct heat away from the equipment, to a heat exchanger. And that exchanger transfers the heat to a second set of pipes, which can use less-clean water, often to evaporate the heat away via towers or pools. (This is for a facility located on Earth, of course.)
Well, yes, it can be less clean but it’s not going to be random water pulled up from a river or lake, either. If it’s in another set of pipes, it’s going to be treated to some degree, or else there’s going to be some nasty fouling of the equipment in short order.
So, again, that’s a treatment plant (plus the power plant or long-distance high voltage transmission lines) that need to be built and staffed.
Well, one disadvantage would be latency - a data center floating in the middle of the ocean would involve 4 trips to/from satellites, instead of just two.
Cooling isn’t that huge a problem. We already have a lot of energy using things in orbit that need to reject heat. The most obvious is the ISS, which has dedicated heat radiators sticking out from the structure.
Even a current Starlink satellite has to radiate waste heat away from its electronics. So long as you have a high emissivity surface facing into space (ie not towards the Sun or Earth) you can reject useful amounts of heat. The radiators are not as large as the solar panels used to feed the satellite. But it adds a pile of extra stuff, and you need to build the system to move heat about - which means pumps and coolant etc one you get big.
The problem with the data centre in space model is the use case and the compute model. This is classic stuff for designing supercomputers. What is the balance of compute, data, data movement and synchronisation needed? Compute is the easy bit. Data size can vary a lot even for the same problem depensding upon the answers to the next two. How much data do you need to shovel around between compute? That comes as bandwidth, but just as importantly, latency. Synchronisation of distributed compute can become the pacing factor. Our local friendly LLM is based on linear algebra models of insane size. It is very unlikely that you can distribute a training system between separate orbiting units. Latency is too high, even if you can get pretty impressive bandwidths with laser communications.
We are seeing that AI is dividing into training systems and delivery of services, which have very different hardware needs. One assumes that the AI tech bros are thinking that they will deploy thousands of identical service provision systems, all running with pre-canned parameters derived from the very very large earthbound training systems. These can be standalone systems, with little need to communicate. Just how large an individual system needs to be might still be an open question - but the equivalent of an earthbound single rack of compute might be reasonable. So a few 10s of kilowatts maybe. That isn’t insane from a technical point of view. (Just insane for all the other reasons.)
Maybe Elon imagines that Starlink will provide a direct to consumer AI service from a fleet of orbiting service nodes. Maybe your next car will link via Starlink to the orbiting AI service that provides full self drive.
The whole problem is as above. For every advantage that space gives you, a few acres of land somewhere sunny will get you the same for cheap. Way cheap. Commodity solar panels are so cheap it is silly. Battery systems are storming into the wholesale electricity supply systems already. Elon even has a profitable side hustle selling them. Nothing about space data centres makes sense. Even if the launch costs were zero.
Yes, as compared to putting these things in space, floating on the ocean surface, or undersea.
Generally, in these discussions, every idea besides “in city centers, or at least in dense suburbia” is cast aside summarily. Might be good reasons to do this … but then I ask: If every answer but “densely populated areas” is too hard or too expensive, just how seriously are we, collectively, taking the apparent (?) hazards of data centers?
It’s a huge problem. The ISS has to reject about .126 MW of heat. That’s very doable with ammonia loops and radiators. A hyperscale Data Center needs to reject about 1,250 MW of heat.
If I’m reading this document correctly (ATCS Team Overview [pdf]), the ISS panels handle 70 kW of cooling. (See System Performance Overview on page 17.) That’s not nothing, but it’s not going to scale well when most of the payload of the spacecraft is computing. It’s not unusual to have a computer with a 1kW power supply (and equivalent cooling need), and a small datacenter has hundreds of computers. That’s just not doable in space for a reasonable cost.
I don’t think there is any intention to build a hyperscale. Rather the opposite. Think more in terms of tens of kilowatts each. Lots and lots of them. Rather like Starlink.
There were some initial insane depictions of hyperscale designs. But no more.
Cooling might very well be the biggest snag for a data center in space. Space is vacuum, a vacuum is a pretty darned good thermal insulator, and data centers need insane amounts of cooling.
Simple. Use clean water in your cooling water loop, dump the heat through a heat exchanger submerged in the nearest river.
Of course, you’d have to disregard what that heat and temperature increase does to the river’s environment, but who’s counting? We’re building a data center and making money from AI-generated SoMe videos, we ain’t hugging trees.
I don’t know where you’re getting your information from. All of the “AI datacenters in space” talk is coming from Musk and the SpaceX IPO filings. Musk has stated numerous times that moving to space is the only way to scale, e.g.
Putting data centers in space “was always going to be” the solution, Musk said Thursday, “as you can scale a trillion times more than you could on Earth.”
Trillion times! And it doesn’t really matter if each individual satellite is only producing 10s of kW each, as others have pointed out the ISS only produces 10s of kW and the cooling apparatus on that thing dwarfs the living compartment. It’s massive and complex, and every single puny 10s of kW little satellite would need to pack something that large.
But you’re right on one thing, there’s probably no intention to build any of this as it’s just been a scam to tie spaceX to AI and pump the IPO valuation. Because the concept makes no sense with some basic napkin math.
And what’s the power usage of the ISS, compared to the power usage of a datacenter?
The reason why new datacenters aren’t being built in places like that is because all of those sites are already taken by old datacenters. That’s the reason why a lot of tech companies are already headquartered in the Pacific Northwest: High-flow, high-drop rivers provide cheap power and cooling.
This is a “how long is a piece of string” question. The data centre is the size you build it.
Perhaps the more interesting point is that there is a relationship between the area of solar panel and area of heat rejection panels. The heat rejection panels are smaller - mostly because the solar panels are not all that efficient.
Whether you have a space station, comms satellite, or a data centre, you need solar panels, and to reject heat. Managing the heat gets worse the larger the system. Electrical power runs down wires, but heat needs pumping about, and that just gets worse the bigger things get. So you don’t go too big.
Google’s idea for orbiting data satellites use 28kW each. This isn’t difficult. (Just stupid.)
These are the reasons they might actually try to do it, which is different from any scam based reasons they are simply claiming they want to do it.
Don’t discount the science fiction sounding aspects, too. They want space based, extraterritorial data centers to be a mature technology waiting for brain uploads. The dream is immortality based outside the reach of anyone to be able to stop it or have any influence over decisions and power.
If all goes according to plan, in 100 years Musk’s (and friends’) brains will be controlling a multiplanetary human empire with their orbiting and deep space brains. Yes, I know exactly which forum I’m posting in.
If you just want extraterritoriality, then the ocean is much easier. And if you’re controlling a multiplanetary empire, then it’s not extraterritorial any more, it’s your territory, and it doesn’t matter where it is.
You neglected my use of the word “effective” extraterritoriality. An ocean-based free enclave is only extraterritorial until you piss off someone with a blue-water navy. For now, medium-height orbit and higher is nearly immune to armed interference. (The main weakness would be ground-based facilities for building, launching, and communication.)
